Datasets:

memray

/

semeval

like 0

train_C-41

Evaluating Adaptive Resource Management for Distributed Real-Time Embedded Systems

A challenging problem faced by researchers and developers of distributed real-time and embedded (DRE) systems is devising and implementing effective adaptive resource management strategies that can meet end-to-end quality of service (QoS) requirements in varying operational conditions. This paper presents two contributions to research in adaptive resource management for DRE systems. First, we describe the structure and functionality of the Hybrid Adaptive Resourcemanagement Middleware (HyARM), which provides adaptive resource management using hybrid control techniques for adapting to workload fluctuations and resource availability. Second, we evaluate the adaptive behavior of HyARM via experiments on a DRE multimedia system that distributes video in real-time. Our results indicate that HyARM yields predictable, stable, and high system performance, even in the face of fluctuating workload and resource availability.

1. INTRODUCTION Achieving end-to-end real-time quality of service (QoS) is particularly important for open distributed real-time and embedded (DRE) systems that face resource constraints, such as limited computing power and network bandwidth. Overutilization of these system resources can yield unpredictable and unstable behavior, whereas under-utilization can yield excessive system cost. A promising approach to meeting these end-to-end QoS requirements effectively, therefore, is to develop and apply adaptive middleware [10, 15], which is software whose functional and QoS-related properties can be modified either statically or dynamically. Static modifications are carried out to reduce footprint, leverage capabilities that exist in specific platforms, enable functional subsetting, and/or minimize hardware/software infrastructure dependencies. Objectives of dynamic modifications include optimizing system responses to changing environments or requirements, such as changing component interconnections, power-levels, CPU and network bandwidth availability, latency/jitter, and workload. In open DRE systems, adaptive middleware must make such modifications dependably, i.e., while meeting stringent end-to-end QoS requirements, which requires the specification and enforcement of upper and lower bounds on system resource utilization to ensure effective use of system resources. To meet these requirements, we have developed the Hybrid Adaptive Resource-management Middleware (HyARM), which is an open-source1 distributed resource management middleware. HyARM is based on hybrid control theoretic techniques [8], which provide a theoretical framework for designing control of complex system with both continuous and discrete dynamics. In our case study, which involves a distributed real-time video distribution system, the task of adaptive resource management is to control the utilization of the different resources, whose utilizations are described by continuous variables. We achieve this by adapting the resolution of the transmitted video, which is modeled as a continuous variable, and by changing the frame-rate and the compression, which are modeled by discrete actions. We have implemented HyARM atop The ACE ORB (TAO) [13], which is an implementation of the Real-time CORBA specification [12]. Our results show that (1) HyARM ensures effective system resource utilization and (2) end-to-end QoS requirements of higher priority applications are met, even in the face of fluctuations in workload. The remainder of the paper is organized as follows: Section 2 describes the architecture, functionality, and resource utilization model of our DRE multimedia system case study; Section 3 explains the structure and functionality of HyARM; Section 4 evaluates the adaptive behavior of HyARM via experiments on our multimedia system case study; Section 5 compares our research on HyARM with related work; and Section 6 presents concluding remarks. 1 The code and examples for HyARM are available at www. dre.vanderbilt.edu/∼nshankar/HyARM/. Article 7 2. CASE STUDY: DRE MULTIMEDIA SYSTEM This section describes the architecture and QoS requirements of our DRE multimedia system. 2.1 Multimedia System Architecture Wireless Link Wireless Link Wireless Link ` ` ` Physical Link Physical Link Physical Link Base Station End Receiver End Receiver End Receiver` Physical Link End Receiver UAV Camera Video Encoder Camera Video Encoder Camera Video Encoder UAV Camera Video Encoder Camera Video Encoder Camera Video Encoder UAV Camera Video Encoder Camera Video Encoder Camera Video Encoder Figure 1: DRE Multimedia System Architecture The architecture for our DRE multimedia system is shown in Figure 1 and consists of the following entities: (1)Data source (video capture by UAV), where video is captured (related to subject of interest) by camera(s) on each UAV, followed by encoding of raw video using a specific encoding scheme and transmitting the video to the next stage in the pipeline. (2)Data distributor (base station), where the video is processed to remove noise, followed by retransmission of the processed video to the next stage in the pipeline. (3) Sinks (command and control center), where the received video is again processed to remove noise, then decoded and finally rendered to end user via graphical displays. Significant improvements in video encoding/decoding and (de)compression techniques have been made as a result of recent advances in video encoding and compression techniques [14]. Common video compression schemes are MPEG1, MPEG-2, Real Video, and MPEG-4. Each compression scheme is characterized by its resource requirement, e.g., the computational power to (de)compress the video signal and the network bandwidth required to transmit the compressed video signal. Properties of the compressed video, such as resolution and frame-rate determine both the quality and the resource requirements of the video. Our multimedia system case study has the following endto-end real-time QoS requirements: (1) latency, (2) interframe delay (also know as jitter), (3) frame rate, and (4) picture resolution. These QoS requirements can be classified as being either hard or soft. Hard QoS requirements should be met by the underlying system at all times, whereas soft QoS requirements can be missed occasionally.2 For our case study, we treat QoS requirements such as latency and jitter as harder QoS requirements and strive to meet these requirements at all times. In contrast, we treat QoS requirements such as video frame rate and picture resolution as softer QoS requirements and modify these video properties adaptively to handle dynamic changes in resource availabil2 Although hard and soft are often portrayed as two discrete requirement sets, in practice they are usually two ends of a continuum ranging from softer to harder rather than two disjoint points. ity effectively. 2.2 DRE Multimedia System Rresources There are two primary types of resources in our DRE multimedia system: (1) processors that provide computational power available at the UAVs, base stations, and end receivers and (2) network links that provide communication bandwidth between UAVs, base stations, and end receivers. The computing power required by the video capture and encoding tasks depends on dynamic factors, such as speed of the UAV, speed of the subject (if the subject is mobile), and distance between UAV and the subject. The wireless network bandwidth available to transmit video captured by UAVs to base stations also depends on the wireless connectivity between the UAVs and the base station, which in-turn depend on dynamic factors such as the speed of the UAVs and the relative distance between UAVs and base stations. The bandwidth of the link between the base station and the end receiver is limited, but more stable than the bandwidth of the wireless network. Resource requirements and availability of resources are subjected to dynamic changes. Two classes of applications - QoS-enabled and best-effort - use the multimedia system infrastructure described above to transmit video to their respective receivers. QoS-enabled class of applications have higher priority over best-effort class of application. In our study, emergency response applications belong to QoS-enabled and surveillance applications belong to best-effort class. For example, since a stream from an emergency response application is of higher importance than a video stream from a surveillance application, it receives more resources end-to-end. Since resource availability significantly affects QoS, we use current resource utilization as the primary indicator of system performance. We refer to the current level of system resource utilization as the system condition. Based on this definition, we can classify system conditions as being either under, over, or effectively utilized. Under-utilization of system resources occurs when the current resource utilization is lower than the desired lower bound on resource utilization. In this system condition, residual system resources (i.e., network bandwidth and computational power) are available in large amounts after meeting end-to-end QoS requirements of applications. These residual resources can be used to increase the QoS of the applications. For example, residual CPU and network bandwidth can be used to deliver better quality video (e.g., with greater resolution and higher frame rate) to end receivers. Over-utilization of system resources occurs when the current resource utilization is higher than the desired upper bound on resource utilization. This condition can arise from loss of resources - network bandwidth and/or computing power at base station, end receiver or at UAV - or may be due to an increase in resource demands by applications. Over-utilization is generally undesirable since the quality of the received video (such as resolution and frame rate) and timeliness properties (such as latency and jitter) are degraded and may result in an unstable (and thus ineffective) system. Effective resource utilization is the desired system condition since it ensures that end-to-end QoS requirements of the UAV-based multimedia system are met and utilization of both system resources, i.e., network bandwidth and computational power, are within their desired utilization bounds. Article 7 Section 3 describes techniques we applied to achieve effective utilization, even in the face of fluctuating resource availability and/or demand. 3. OVERVIEW OF HYARM This section describes the architecture of the Hybrid Adaptive Resource-management Middleware (HyARM). HyARM ensures efficient and predictable system performance by providing adaptive resource management, including monitoring of system resources and enforcing bounds on application resource utilization. 3.1 HyARM Structure and Functionality Resource Utilization Legend Resource Allocation Application Parameters Figure 2: HyARM Architecture HyARM is composed of three types of entities shown in Figure 2 and described below: Resource monitors observe the overall resource utilization for each type of resource and resource utilization per application. In our multimedia system, there are resource monitors for CPU utilization and network bandwidth. CPU monitors observe the CPU resource utilization of UAVs, base station, and end receivers. Network bandwidth monitors observe the network resource utilization of (1) wireless network link between UAVs and the base station and (2) wired network link between the base station and end receivers. The central controller maintains the system resource utilization below a desired bound by (1) processing periodic updates it receives from resource monitors and (2) modifying the execution of applications accordingly, e.g., by using different execution algorithms or operating the application with increased/decreased QoS. This adaptation process ensures that system resources are utilized efficiently and end-to-end application QoS requirements are met. In our multimedia system, the HyARM controller determines the value of application parameters such as (1) video compression schemes, such as Real Video and MPEG-4, and/or (2) frame rate, and (3) picture resolution. From the perspective of hybrid control theoretic techniques [8], the different video compression schemes and frame rate form the discrete variables of application execution and picture resolution forms the continuous variables. Application adapters modify application execution according to parameters recommended by the controller and ensures that the operation of the application is in accordance with the recommended parameters. In the current mplementation of HyARM, the application adapter modifies the input parameters to the application that affect application QoS and resource utilization - compression scheme, frame rate, and picture resolution. In our future implementations, we plan to use resource reservation mechanisms such as Differentiated Service [7, 3] and Class-based Kernel Resource Management [4] to provision/reserve network and CPU resources. In our multimedia system, the application adapter ensures that the video is encoded at the recommended frame rate and resolution using the specified compression scheme. 3.2 Applying HyARM to the Multimedia System Case Study HyARM is built atop TAO [13], a widely used open-source implementation of Real-time CORBA [12]. HyARM can be applied to ensure efficient, predictable and adaptive resource management of any DRE system where resource availability and requirements are subject to dynamic change. Figure 3 shows the interaction of various parts of the DRE multimedia system developed with HyARM, TAO, and TAO"s A/V Streaming Service. TAO"s A/V Streaming service is an implementation of the CORBA A/V Streaming Service specification. TAO"s A/V Streaming Service is a QoS-enabled video distribution service that can transfer video in real-time to one or more receivers. We use the A/V Streaming Service to transmit the video from the UAVs to the end receivers via the base station. Three entities of Receiver UAV TAO Resource Utilization HyARM Central Controller A/V Streaming Service : Sender MPEG1 MPEG4 Real Video HyARM Resource Monitor A/V Streaming Service : Receiver Compressed Video Compressed Video Application HyARM Application Adapter Remote Object Call Control Inputs Resource Utilization Resource Utilization / Control Inputs Control Inputs Legend Figure 3: Developing the DRE Multimedia System with HyARM HyARM, namely the resource monitors, central controller, and application adapters are built as CORBA servants, so they can be distributed throughout a DRE system. Resource monitors are remote CORBA objects that update the central controller periodically with the current resource utilization. Application adapters are collocated with applications since the two interact closely. As shown in Figure 3, UAVs compress the data using various compression schemes, such as MPEG1, MPEG4, and Real Video, and uses TAO"s A/V streaming service to transmit the video to end receivers. HyARM"s resource monitors continuously observe the system resource utilization and notify the central controller with the current utilization. 3 The interaction between the controller and the resource monitors uses the Observer pattern [5]. When the controller receives resource utilization updates from monitors, it computes the necessary modifications to application(s) parameters and notifies application adapter(s) via a remote operation call. Application adapter(s), that are collocated with the application, modify the input parameters to the application - in our case video encoder - to modify the application resource utilization and QoS. 3 The base station is not included in the figure since it only retransmits the video received from UAVs to end receivers. Article 7 4. PERFORMANCE RESULTS AND ANALYSIS This section first describes the testbed that provides the infrastructure for our DRE multimedia system, which was used to evaluate the performance of HyARM. We then describe our experiments and analyze the results obtained to empirically evaluate how HyARM behaves during underand over-utilization of system resources. 4.1 Overview of the Hardware and Software Testbed Our experiments were performed on the Emulab testbed at University of Utah. The hardware configuration consists of two nodes acting as UAVs, one acting as base station, and one as end receiver. Video from the two UAVs were transmitted to a base station via a LAN configured with the following properties: average packet loss ratio of 0.3 and bandwidth 1 Mbps. The network bandwidth was chosen to be 1 Mbps since each UAV in the DRE multimedia system is allocated 250 Kbps. These parameters were chosen to emulate an unreliable wireless network with limited bandwidth between the UAVs and the base station. From the base station, the video was retransmitted to the end receiver via a reliable wireline link of 10 Mbps bandwidth with no packet loss. The hardware configuration of all the nodes was chosen as follows: 600 MHz Intel Pentium III processor, 256 MB physical memory, 4 Intel EtherExpress Pro 10/100 Mbps Ethernet ports, and 13 GB hard drive. A real-time version of Linux - TimeSys Linux/NET 3.1.214 based on RedHat Linux 9was used as the operating system for all nodes. The following software packages were also used for our experiments: (1) Ffmpeg 0.4.9-pre1, which is an open-source library (http: //www.ffmpeg.sourceforge.net/download.php) that compresses video into MPEG-2, MPEG-4, Real Video, and many other video formats. (2) Iftop 0.16, which is an opensource library (http://www.ex-parrot.com/∼pdw/iftop/) we used for monitoring network activity and bandwidth utilization. (3) ACE 5.4.3 + TAO 1.4.3, which is an opensource (http://www.dre.vanderbilt.edu/TAO) implementation of the Real-time CORBA [12] specification upon which HyARM is built. TAO provides the CORBA Audio/Video (A/V) Streaming Service that we use to transmit the video from the UAVs to end receivers via the base station. 4.2 Experiment Configuration Our experiment consisted of two (emulated) UAVs that simultaneously send video to the base station using the experimentation setup described in Section 4.1. At the base station, video was retransmitted to the end receivers (without any modifications), where it was stored to a file. Each UAV hosted two applications, one QoS-enabled application (emergency response), and one best-effort application (surveillance). Within each UAV, computational power is shared between the applications, while the network bandwidth is shared among all applications. To evaluate the QoS provided by HyARM, we monitored CPU utilization at the two UAVs, and network bandwidth utilization between the UAV and the base station. CPU resource utilization was not monitored at the base station and the end receiver since they performed no computationallyintensive operations. The resource utilization of the 10 Mpbs physical link between the base station and the end receiver does not affect QoS of applications and is not monitored by HyARM since it is nearly 10 times the 1 MB bandwidth of the LAN between the UAVs and the base station. The experiment also monitors properties of the video that affect the QoS of the applications, such as latency, jitter, frame rate, and resolution. The set point on resource utilization for each resource was specified at 0.69, which is the upper bound typically recommended by scheduling techniques, such as rate monotonic algorithm [9]. Since studies [6] have shown that human eyes can perceive delays more than 200ms, we use this as the upper bound on jitter of the received video. QoS requirements for each class of application is specified during system initialization and is shown in Table 1. 4.3 Empirical Results and Analysis This section presents the results obtained from running the experiment described in Section 4.2 on our DRE multimedia system testbed. We used system resource utilization as a metric to evaluate the adaptive resource management capabilities of HyARM under varying input work loads. We also used application QoS as a metric to evaluate HyARM"s capabilities to support end-to-end QoS requirements of the various classes of applications in the DRE multimedia system. We analyze these results to explain the significant differences in system performance and application QoS. Comparison of system performance is decomposed into comparison of resource utilization and application QoS. For system resource utilization, we compare (1) network bandwidth utilization of the local area network and (2) CPU utilization at the two UAV nodes. For application QoS, we compare mean values of video parameters, including (1) picture resolution, (2) frame rate, (3) latency, and (4) jitter. Comparison of resource utilization. Over-utilization of system resources in DRE systems can yield an unstable system. In contrast, under-utilization of system resources increases system cost. Figure 4 and Figure 5 compare the system resource utilization with and without HyARM. Figure 4 shows that HyARM maintains system utilization close to the desired utilization set point during fluctuation in input work load by transmitting video of higher (or lower) QoS for QoS-enabled (or best-effort) class of applications during over (or under) utilization of system resources. Figure 5 shows that without HyARM, network utilization was as high as 0.9 during increase in workload conditions, which is greater than the utilization set point of 0.7 by 0.2. As a result of over-utilization of resources, QoS of the received video, such as average latency and jitter, was affected significantly. Without HyARM, system resources were either under-utilized or over-utilized, both of which are undesirable. In contrast, with HyARM, system resource utilization is always close to the desired set point, even during fluctuations in application workload. During sudden fluctuation in application workload, system conditions may be temporarily undesirable, but are restored to the desired condition within several sampling periods. Temporary over-utilization of resources is permissible in our multimedia system since the quality of the video may be degraded for a short period of time, though application QoS will be degraded significantly if poor quality video is transmitted for a longer period of time. Comparison of application QoS. Figures 6, Figure 7, and Table 2 compare latency, jitter, resolution, and frameArticle 7 Class Resolution Frame Rate Latency (msec ) Jitter (msec) QoS Enabled 1024 x 768 25 200 200 Best-effort 320 x 240 15 300 250 Table 1: Application QoS Requirements Figure 4: Resource utilization with HyARM Figure 5: Resource utilization without HyARM rate of the received video, respectively. Table 2 shows that HyARM increases the resolution and frame video of QoSenabled applications, but decreases the resolution and frame rate of best effort applications. During over utilization of system resources, resolution and frame rate of lower priority applications are reduced to adapt to fluctuations in application workload and to maintain the utilization of resources at the specified set point. It can be seen from Figure 6 and Figure 7 that HyARM reduces the latency and jitter of the received video significantly. These figures show that the QoS of QoS-enabled applications is greatly improved by HyARM. Although application parameters, such as frame rate and resolutions, which affect the soft QoS requirements of best-effort applications may be compromised, the hard QoS requirements, such as latency and jitter, of all applications are met. HyARM responds to fluctuation in resource availability and/or demand by constant monitoring of resource utilization. As shown in Figure 4, when resources utilization increases above the desired set point, HyARM lowers the utilization by reducing the QoS of best-effort applications. This adaptation ensures that enough resources are available for QoS-enabled applications to meet their QoS needs. Figures 6 and 7 show that the values of latency and jitter of the received video of the system with HyARM are nearly half of the corresponding value of the system without HyARM. With HyARM, values of these parameters are well below the specified bounds, whereas without HyARM, these value are significantly above the specified bounds due to overutilization of the network bandwidth, which leads to network congestion and results in packet loss. HyARM avoids this by reducing video parameters such as resolution, frame-rate, and/or modifying the compression scheme used to compress the video. Our conclusions from analyzing the results described above are that applying adaptive middleware via hybrid control to DRE system helps to (1) improve application QoS, (2) increase system resource utilization, and (3) provide better predictability (lower latency and inter-frame delay) to QoSenabled applications. These improvements are achieved largely due to monitoring of system resource utilization, efficient system workload management, and adaptive resource provisioning by means of HyARM"s network/CPU resource monitors, application adapter, and central controller, respectively. 5. RELATED WORK A number of control theoretic approaches have been applied to DRE systems recently. These techniques aid in overcoming limitations with traditional scheduling approaches that handle dynamic changes in resource availability poorly and result in a rigidly scheduled system that adapts poorly to change. A survey of these techniques is presented in [1]. One such approach is feedback control scheduling (FCS) [2, 11]. FCS algorithms dynamically adjust resource allocation by means of software feedback control loops. FCS algorithms are modeled and designed using rigorous controltheoretic methodologies. These algorithms provide robust and analytical performance assurances despite uncertainties in resource availability and/or demand. Although existing FCS algorithms have shown promise, these algorithms often assume that the system has continuous control variable(s) that can continuously be adjusted. While this assumption holds for certain classes of systems, there are many classes of DRE systems, such as avionics and total-ship computing environments that only support a finite a priori set of discrete configurations. The control variables in such systems are therefore intrinsically discrete. HyARM handles both continuous control variables, such as picture resolution, and discrete control variable, such as discrete set of frame rates. HyARM can therefore be applied to system that support continuous and/or discrete set of control variables. The DRE multimedia system as described in Section 2 is an example DRE system that offers both continuous (picture resolution) and discrete set (frame-rate) of control variables. These variables are modified by HyARM to achieve efficient resource utilization and improved application QoS. 6. CONCLUDING REMARKS Article 7 Figure 6: Comparison of Video Latency Figure 7: Comparison of Video Jitter Source Picture Size / Frame Rate With HyARM Without HyARM UAV1 QoS Enabled Application 1122 X 1496 / 25 960 X 720 / 20 UAV1 Best-effort Application 288 X 384 / 15 640 X 480 / 20 UAV2 QoS Enabled Application 1126 X 1496 / 25 960 X 720 / 20 UAV2 Best-effort Application 288 X 384 / 15 640 X 480 / 20 Table 2: Comparison of Video Quality Many distributed real-time and embedded (DRE) systems demand end-to-end quality of service (QoS) enforcement from their underlying platforms to operate correctly. These systems increasingly run in open environments, where resource availability is subject to dynamic change. To meet end-to-end QoS in dynamic environments, DRE systems can benefit from an adaptive middleware that monitors system resources, performs efficient application workload management, and enables efficient resource provisioning for executing applications. This paper described HyARM, an adaptive middleware, that provides effective resource management to DRE systems. HyARM employs hybrid control techniques to provide the adaptive middleware capabilities, such as resource monitoring and application adaptation that are key to providing the dynamic resource management capabilities for open DRE systems. We employed HyARM to a representative DRE multimedia system that is implemented using Real-time CORBA and CORBA A/V Streaming Service. We evaluated the performance of HyARM in a system composed of three distributed resources and two classes of applications with two applications each. Our empirical results indicate that HyARM ensures (1) efficient resource utilization by maintaining the resource utilization of system resources within the specified utilization bounds, (2) QoS requirements of QoS-enabled applications are met at all times. Overall, HyARM ensures efficient, predictable, and adaptive resource management for DRE systems. 7. REFERENCES [1] T. F. Abdelzaher, J. Stankovic, C. Lu, R. Zhang, and Y. Lu. Feddback Performance Control in Software Services. IEEE: Control Systems, 23(3), June 2003. [2] L. Abeni, L. Palopoli, G. Lipari, and J. Walpole. Analysis of a reservation-based feedback scheduler. In IEEE Real-Time Systems Symposium, Dec. 2002. [3] S. Blake, D. Black, M. Carlson, E. Davies, Z. Wang, and W. Weiss. An architecture for differentiated services. Network Information Center RFC 2475, Dec. 1998. [4] H. Franke, S. Nagar, C. Seetharaman, and V. Kashyap. Enabling Autonomic Workload Management in Linux. In Proceedings of the International Conference on Autonomic Computing (ICAC), New York, New York, May 2004. IEEE. [5] E. Gamma, R. Helm, R. Johnson, and J. Vlissides. Design Patterns: Elements of Reusable Object-Oriented Software. Addison-Wesley, Reading, MA, 1995. [6] G. Ghinea and J. P. Thomas. Qos impact on user perception and understanding of multimedia video clips. In MULTIMEDIA "98: Proceedings of the sixth ACM international conference on Multimedia, pages 49-54, Bristol, United Kingdom, 1998. ACM Press. [7] Internet Engineering Task Force. Differentiated Services Working Group (diffserv) Charter. www.ietf.org/html.charters/diffserv-charter.html, 2000. [8] X. Koutsoukos, R. Tekumalla, B. Natarajan, and C. Lu. Hybrid Supervisory Control of Real-Time Systems. In 11th IEEE Real-Time and Embedded Technology and Applications Symposium, San Francisco, California, Mar. 2005. [9] J. Lehoczky, L. Sha, and Y. Ding. The Rate Monotonic Scheduling Algorithm: Exact Characterization and Average Case Behavior. In Proceedings of the 10th IEEE Real-Time Systems Symposium (RTSS 1989), pages 166-171. IEEE Computer Society Press, 1989. [10] J. Loyall, J. Gossett, C. Gill, R. Schantz, J. Zinky, P. Pal, R. Shapiro, C. Rodrigues, M. Atighetchi, and D. Karr. Comparing and Contrasting Adaptive Middleware Support in Wide-Area and Embedded Distributed Object Applications. In Proceedings of the 21st International Conference on Distributed Computing Systems (ICDCS-21), pages 625-634. IEEE, Apr. 2001. [11] C. Lu, J. A. Stankovic, G. Tao, and S. H. Son. Feedback Control Real-Time Scheduling: Framework, Modeling, and Algorithms. Real-Time Systems Journal, 23(1/2):85-126, July 2002. [12] Object Management Group. Real-time CORBA Specification, OMG Document formal/02-08-02 edition, Aug. 2002. [13] D. C. Schmidt, D. L. Levine, and S. Mungee. The Design and Performance of Real-Time Object Request Brokers. Computer Communications, 21(4):294-324, Apr. 1998. [14] Thomas Sikora. Trends and Perspectives in Image and Video Coding. In Proceedings of the IEEE, Jan. 2005. [15] X. Wang, H.-M. Huang, V. Subramonian, C. Lu, and C. Gill. CAMRIT: Control-based Adaptive Middleware for Real-time Image Transmission. In Proc. of the 10th IEEE Real-Time and Embedded Tech. and Applications Symp. (RTAS), Toronto, Canada, May 2004. Article 7

real-time video distribution system;dynamic environment;hybrid system;video encoding/decoding;quality of service;streaming service;hybrid adaptive resourcemanagement middleware;distributed real-time embedded system;distribute real-time embed system;adaptive resource management;service end-to-end quality;hybrid control technique;service quality;real-time corba specification;end-to-end quality of service;resource reservation mechanism

train_C-42

Demonstration of Grid-Enabled Ensemble Kalman Filter Data Assimilation Methodology for Reservoir Characterization

Ensemble Kalman filter data assimilation methodology is a popular approach for hydrocarbon reservoir simulations in energy exploration. In this approach, an ensemble of geological models and production data of oil fields is used to forecast the dynamic response of oil wells. The Schlumberger ECLIPSE software is used for these simulations. Since models in the ensemble do not communicate, message-passing implementation is a good choice. Each model checks out an ECLIPSE license and therefore, parallelizability of reservoir simulations depends on the number licenses available. We have Grid-enabled the ensemble Kalman filter data assimilation methodology for the TIGRE Grid computing environment. By pooling the licenses and computing resources across the collaborating institutions using GridWay metascheduler and TIGRE environment, the computational accuracy can be increased while reducing the simulation runtime. In this paper, we provide an account of our efforts in Gridenabling the ensemble Kalman Filter data assimilation methodology. Potential benefits of this approach, observations and lessons learned will be discussed.

1. INTRODUCTION Grid computing [1] is an emerging collaborative computing paradigm to extend institution/organization specific high performance computing (HPC) capabilities greatly beyond local resources. Its importance stems from the fact that ground breaking research in strategic application areas such as bioscience and medicine, energy exploration and environmental modeling involve strong interdisciplinary components and often require intercampus collaborations and computational capabilities beyond institutional limitations. The Texas Internet Grid for Research and Education (TIGRE) [2,3] is a state funded cyberinfrastructure development project carried out by five (Rice, A&M, TTU, UH and UT Austin) major university systems - collectively called TIGRE Institutions. The purpose of TIGRE is to create a higher education Grid to sustain and extend research and educational opportunities across Texas. TIGRE is a project of the High Performance Computing across Texas (HiPCAT) [4] consortium. The goal of HiPCAT is to support advanced computational technologies to enhance research, development, and educational activities. The primary goal of TIGRE is to design and deploy state-of-the-art Grid middleware that enables integration of computing systems, storage systems and databases, visualization laboratories and displays, and even instruments and sensors across Texas. The secondary goal is to demonstrate the TIGRE capabilities to enhance research and educational opportunities in strategic application areas of interest to the State of Texas. These are bioscience and medicine, energy exploration and air quality modeling. Vision of the TIGRE project is to foster interdisciplinary and intercampus collaborations, identify novel approaches to extend academic-government-private partnerships, and become a competitive model for external funding opportunities. The overall goal of TIGRE is to support local, campus and regional user interests and offer avenues to connect with national Grid projects such as Open Science Grid [5], and TeraGrid [6]. Within the energy exploration strategic application area, we have Grid-enabled the ensemble Kalman Filter (EnKF) [7] approach for data assimilation in reservoir modeling and demonstrated the extensibility of the application using the TIGRE environment and the GridWay [8] metascheduler. Section 2 provides an overview of the TIGRE environment and capabilities. Application description and the need for Grid-enabling EnKF methodology is provided in Section 3. The implementation details and merits of our approach are discussed in Section 4. Conclusions are provided in Section 5. Finally, observations and lessons learned are documented in Section 6. 2. TIGRE ENVIRONMENT The TIGRE Grid middleware consists of minimal set of components derived from a subset of the Virtual Data Toolkit (VDT) [9] which supports a variety of operating systems. The purpose of choosing a minimal software stack is to support applications at hand, and to simplify installation and distribution of client/server stacks across TIGRE sites. Additional components will be added as they become necessary. The PacMan [10] packaging and distribution mechanism is employed for TIGRE client/server installation and management. The PacMan distribution mechanism involves retrieval, installation, and often configuration of the packaged software. This approach allows the clients to keep current, consistent versions of TIGRE software. It also helps TIGRE sites to install the needed components on resources distributed throughout the participating sites. The TIGRE client/server stack consists of an authentication and authorization layer, Globus GRAM4-based job submission via web services (pre-web services installations are available up on request). The tools for handling Grid proxy generation, Grid-enabled file transfer and Grid-enabled remote login are supported. The pertinent details of TIGRE services and tools for job scheduling and management are provided below. 2.1. Certificate Authority The TIGRE security infrastructure includes a certificate authority (CA) accredited by the International Grid Trust Federation (IGTF) for issuing X. 509 user and resource Grid certificates [11]. The Texas Advanced Computing Center (TACC), University of Texas at Austin is the TIGRE"s shared CA. The TIGRE Institutions serve as Registration Authorities (RA) for their respective local user base. For up-to-date information on securing user and resource certificates and their installation instructions see ref [2]. The users and hosts on TIGRE are identified by their distinguished name (DN) in their X.509 certificate provided by the CA. A native Grid-mapfile that contains a list of authorized DNs is used to authenticate and authorize user job scheduling and management on TIGRE site resources. At Texas Tech University, the users are dynamically allocated one of the many generic pool accounts. This is accomplished through the Grid User Management System (GUMS) [12]. 2.2. Job Scheduling and Management The TIGRE environment supports GRAM4-based job submission via web services. The job submission scripts are generated using XML. The web services GRAM translates the XML scripts into target cluster specific batch schedulers such as LSF, PBS, or SGE. The high bandwidth file transfer protocols such as GridFTP are utilized for staging files in and out of the target machine. The login to remote hosts for compilation and debugging is only through GSISSH service which requires resource authentication through X.509 certificates. The authentication and authorization of Grid jobs are managed by issuing Grid certificates to both users and hosts. The certificate revocation lists (CRL) are updated on a daily basis to maintain high security standards of the TIGRE Grid services. The TIGRE portal [2] documentation area provides a quick start tutorial on running jobs on TIGRE. 2.3. Metascheduler The metascheduler interoperates with the cluster level batch schedulers (such as LSF, PBS) in the overall Grid workflow management. In the present work, we have employed GridWay [8] metascheduler - a Globus incubator project - to schedule and manage jobs across TIGRE. The GridWay is a light-weight metascheduler that fully utilizes Globus functionalities. It is designed to provide efficient use of dynamic Grid resources by multiple users for Grid infrastructures built on top of Globus services. The TIGRE site administrator can control the resource sharing through a powerful built-in scheduler provided by GridWay or by extending GridWay"s external scheduling module to provide their own scheduling policies. Application users can write job descriptions using GridWay"s simple and direct job template format (see Section 4 for details) or standard Job Submission Description Language (JSDL). See section 4 for implementation details. 2.4. Customer Service Management System A TIGRE portal [2] was designed and deployed to interface users and resource providers. It was designed using GridPort [13] and is maintained by TACC. The TIGRE environment is supported by open source tools such as the Open Ticket Request System (OTRS) [14] for servicing trouble tickets, and MoinMoin [15] Wiki for TIGRE content and knowledge management for education, outreach and training. The links for OTRS and Wiki are consumed by the TIGRE portal [2] - the gateway for users and resource providers. The TIGRE resource status and loads are monitored by the Grid Port Information Repository (GPIR) service of the GridPort toolkit [13] which interfaces with local cluster load monitoring service such as Ganglia. The GPIR utilizes cron jobs on each resource to gather site specific resource characteristics such as jobs that are running, queued and waiting for resource allocation. 3. ENSEMBLE KALMAN FILTER APPLICATION The main goal of hydrocarbon reservoir simulations is to forecast the production behavior of oil and gas field (denoted as field hereafter) for its development and optimal management. In reservoir modeling, the field is divided into several geological models as shown in Figure 1. For accurate performance forecasting of the field, it is necessary to reconcile several geological models to the dynamic response of the field through history matching [16-20]. Figure 1. Cross-sectional view of the Field. Vertical layers correspond to different geological models and the nails are oil wells whose historical information will be used for forecasting the production behavior. (Figure Ref:http://faculty.smu.edu/zchen/research.html). The EnKF is a Monte Carlo method that works with an ensemble of reservoir models. This method utilizes crosscovariances [21] between the field measurements and the reservoir model parameters (derived from several models) to estimate prediction uncertainties. The geological model parameters in the ensemble are sequentially updated with a goal to minimize the prediction uncertainties. Historical production response of the field for over 50 years is used in these simulations. The main advantage of EnKF is that it can be readily linked to any reservoir simulator, and can assimilate latest production data without the need to re-run the simulator from initial conditions. Researchers in Texas are large subscribers of the Schlumberger ECLIPSE [22] package for reservoir simulations. In the reservoir modeling, each geological model checks out an ECLIPSE license. The simulation runtime of the EnKF methodology depends on the number of geological models used, number of ECLIPSE licenses available, production history of the field, and propagated uncertainties in history matching. The overall EnKF workflow is shown Figure 2. Figure 2. Ensemble Kaman Filter Data Assimilation Workflow. Each site has L licenses. At START, the master/control process (EnKF main program) reads the simulation configuration file for number (N) of models, and model-specific input files. Then, N working directories are created to store the output files. At the end of iteration, the master/control process collects the output files from N models and post processes crosscovariances [21] to estimate the prediction uncertainties. This information will be used to update models (or input files) for the next iteration. The simulation continues until the production histories are exhausted. Typical EnKF simulation with N=50 and field histories of 50-60 years, in time steps ranging from three months to a year, takes about three weeks on a serial computing environment. In parallel computing environment, there is no interprocess communication between the geological models in the ensemble. However, at the end of each simulation time-step, model-specific output files are to be collected for analyzing cross covariances [21] and to prepare next set of input files. Therefore, master-slave model in messagepassing (MPI) environment is a suitable paradigm. In this approach, the geological models are treated as slaves and are distributed across the available processors. The master Cluster or (TIGRE/GridWay) START Read Configuration File Create N Working Directories Create N Input files Model l Model 2 Model N. . . ECLIPSE on site A ECLIPSE on Site B ECLIPSE on Site Z Collect N Model Outputs, Post-process Output files END . . . process collects model-specific output files, analyzes and prepares next set of input files for the simulation. Since each geological model checks out an ECLIPSE license, parallelizability of the simulation depends on the number of licenses available. When the available number of licenses is less than the number of models in the ensemble, one or more of the nodes in the MPI group have to handle more than one model in a serial fashion and therefore, it takes longer to complete the simulation. A Petroleum Engineering Department usually procures 10-15 ECLIPSE licenses while at least ten-fold increase in the number of licenses would be necessary for industry standard simulations. The number of licenses can be increased by involving several Petroleum Engineering Departments that support ECLIPSE package. Since MPI does not scale very well for applications that involve remote compute clusters, and to get around the firewall issues with license servers across administrative domains, Grid-enabling the EnKF workflow seems to be necessary. With this motivation, we have implemented Grid-enabled EnKF workflow for the TIGRE environment and demonstrated parallelizability of the application across TIGRE using GridWay metascheduler. Further details are provided in the next section. 4. IMPLEMENTATION DETAILS To Grid-enable the EnKF approach, we have eliminated the MPI code for parallel processing and replaced with N single processor jobs (or sub-jobs) where, N is the number of geological models in the ensemble. These model-specific sub-jobs were distributed across TIGRE sites that support ECLIPSE package using the GridWay [8] metascheduler. For each sub-job, we have constructed a GridWay job template that specifies the executable, input and output files, and resource requirements. Since the TIGRE compute resources are not expected to change frequently, we have used static resource discovery policy for GridWay and the sub-jobs were scheduled dynamically across the TIGRE resources using GridWay. Figure 3 represents the sub-job template file for the GridWay metascheduler. Figure 3. GridWay Sub-Job Template In Figure 3, REQUIREMENTS flag is set to choose the resources that satisfy the application requirements. In the case of EnKF application, for example, we need resources that support ECLIPSE package. ARGUMENTS flag specifies the model in the ensemble that will invoke ECLIPSE at a remote site. INPUT_FILES is prepared by the EnKF main program (or master/control process) and is transferred by GridWay to the remote site where it is untared and is prepared for execution. Finally, OUTPUT_FILES specifies the name and location where the output files are to be written. The command-line features of GridWay were used to collect and process the model-specific outputs to prepare new set of input files. This step mimics MPI process synchronization in master-slave model. At the end of each iteration, the compute resources and licenses are committed back to the pool. Table 1 shows the sub-jobs in TIGRE Grid via GridWay using gwps command and for clarity, only selected columns were shown . USER JID DM EM NAME HOST pingluo 88 wrap pend enkf.jt antaeus.hpcc.ttu.edu/LSF pingluo 89 wrap pend enkf.jt antaeus.hpcc.ttu.edu/LSF pingluo 90 wrap actv enkf.jt minigar.hpcc.ttu.edu/LSF pingluo 91 wrap pend enkf.jt minigar.hpcc.ttu.edu/LSF pingluo 92 wrap done enkf.jt cosmos.tamu.edu/PBS pingluo 93 wrap epil enkf.jt cosmos.tamu.edu/PBS Table 1. Job scheduling across TIGRE using GridWay Metascheduler. DM: Dispatch state, EM: Execution state, JID is the job id and HOST corresponds to site specific cluster and its local batch scheduler. When a job is submitted to GridWay, it will go through a series of dispatch (DM) and execution (EM) states. For DM, the states include pend(ing), prol(og), wrap(per), epil(og), and done. DM=prol means the job has been scheduled to a resource and the remote working directory is in preparation. DM=warp implies that GridWay is executing the wrapper which in turn executes the application. DM=epil implies the job has finished running at the remote site and results are being transferred back to the GridWay server. Similarly, when EM=pend implies the job is waiting in the queue for resource and the job is running when EM=actv. For complete list of message flags and their descriptions, see the documentation in ref [8]. We have demonstrated the Grid-enabled EnKF runs using GridWay for TIGRE environment. The jobs are so chosen that the runtime doesn"t exceed more than a half hour. The simulation runs involved up to 20 jobs between A&M and TTU sites with TTU serving 10 licenses. For resource information, see Table I. One of the main advantages of Grid-enabled EnKF simulation is that both the resources and licenses are released back to the pool at the end of each simulation time step unlike in the case of MPI implementation where licenses and nodes are locked until the completion of entire simulation. However, the fact that each sub-job gets scheduled independently via GridWay could possibly incur another time delay caused by waiting in queue for execution in each simulation time step. Such delays are not expected EXECUTABLE=runFORWARD REQUIREMENTS=HOSTNAME=cosmos.tamu.edu | HOSTNAME=antaeus.hpcc.ttu.edu | HOSTNAME=minigar.hpcc.ttu.edu | ARGUMENTS=001 INPUT_FILES=001.in.tar OUTPUT_FILES=001.out.tar in MPI implementation where the node is blocked for processing sub-jobs (model-specific calculation) until the end of the simulation. There are two main scenarios for comparing Grid and cluster computing approaches. Scenario I: The cluster is heavily loaded. The conceived average waiting time of job requesting large number of CPUs is usually longer than waiting time of jobs requesting single CPU. Therefore, overall waiting time could be shorter in Grid approach which requests single CPU for each sub-job many times compared to MPI implementation that requests large number of CPUs at a single time. It is apparent that Grid scheduling is beneficial especially when cluster is heavily loaded and requested number of CPUs for the MPI job is not readily available. Scenario II: The cluster is relatively less loaded or largely available. It appears the MPI implementation is favorable compared to the Grid scheduling. However, parallelizability of the EnKF application depends on the number of ECLIPSE licenses and ideally, the number of licenses should be equal to the number of models in the ensemble. Therefore, if a single institution does not have sufficient number of licenses, the cluster availability doesn"t help as much as it is expected. Since the collaborative environment such as TIGRE can address both compute and software resource requirements for the EnKF application, Grid-enabled approach is still advantageous over the conventional MPI implementation in any of the above scenarios. 5. CONCLUSIONS AND FUTURE WORK TIGRE is a higher education Grid development project and its purpose is to sustain and extend research and educational opportunities across Texas. Within the energy exploration application area, we have Grid-enabled the MPI implementation of the ensemble Kalman filter data assimilation methodology for reservoir characterization. This task was accomplished by removing MPI code for parallel processing and replacing with single processor jobs one for each geological model in the ensemble. These single processor jobs were scheduled across TIGRE via GridWay metascheduler. We have demonstrated that by pooling licenses across TIGRE sites, more geological models can be handled in parallel and therefore conceivably better simulation accuracy. This approach has several advantages over MPI implementation especially when a site specific cluster is heavily loaded and/or the number licenses required for the simulation is more than those available at a single site. Towards the future work, it would be interesting to compare the runtime between MPI, and Grid implementations for the EnKF application. This effort could shed light on quality of service (QoS) of Grid environments in comparison with cluster computing. Another aspect of interest in the near future would be managing both compute and license resources to address the job (or processor)-to-license ratio management. 6. OBSERVATIONS AND LESSIONS LEARNED The Grid-enabling efforts for EnKF application have provided ample opportunities to gather insights on the visibility and promise of Grid computing environments for application development and support. The main issues are industry standard data security and QoS comparable to cluster computing. Since the reservoir modeling research involves proprietary data of the field, we had to invest substantial efforts initially in educating the application researchers on the ability of Grid services in supporting the industry standard data security through role- and privilege-based access using X.509 standard. With respect to QoS, application researchers expect cluster level QoS with Grid environments. Also, there is a steep learning curve in Grid computing compared to the conventional cluster computing. Since Grid computing is still an emerging technology, and it spans over several administrative domains, Grid computing is still premature especially in terms of the level of QoS although, it offers better data security standards compared to commodity clusters. It is our observation that training and outreach programs that compare and contrast the Grid and cluster computing environments would be a suitable approach for enhancing user participation in Grid computing. This approach also helps users to match their applications and abilities Grids can offer. In summary, our efforts through TIGRE in Grid-enabling the EnKF data assimilation methodology showed substantial promise in engaging Petroleum Engineering researchers through intercampus collaborations. Efforts are under way to involve more schools in this effort. These efforts may result in increased collaborative research, educational opportunities, and workforce development through graduate/faculty research programs across TIGRE Institutions. 7. ACKNOWLEDGMENTS The authors acknowledge the State of Texas for supporting the TIGRE project through the Texas Enterprise Fund, and TIGRE Institutions for providing the mechanism, in which the authors (Ravi Vadapalli, Taesung Kim, and Ping Luo) are also participating. The authors thank the application researchers Prof. Akhil Datta-Gupta of Texas A&M University and Prof. Lloyd Heinze of Texas Tech University for their discussions and interest to exploit the TIGRE environment to extend opportunities in research and development. 8. REFERENCES [1] Foster, I. and Kesselman, C. (eds.) 2004. The Grid: Blueprint for a new computing infrastructure (The Elsevier series in Grid computing) [2] TIGRE Portal: http://tigreportal.hipcat.net [3] Vadapalli, R. Sill, A., Dooley, R., Murray, M., Luo, P., Kim, T., Huang, M., Thyagaraja, K., and Chaffin, D. 2007. Demonstration of TIGRE environment for Grid enabled/suitable applications. 8th IEEE/ACM Int. Conf. on Grid Computing, Sept 19-21, Austin [4] The High Performance Computing across Texas Consortium http://www.hipcat.net [5] Pordes, R. Petravick, D. Kramer, B. Olson, D. Livny, M. Roy, A. Avery, P. Blackburn, K. Wenaus, T. Würthwein, F. Foster, I. Gardner, R. Wilde, M. Blatecky, A. McGee, J. and Quick, R. 2007. The Open Science Grid, J. Phys Conf Series http://www.iop.org/EJ/abstract/1742-6596/78/1/012057 and http://www.opensciencegrid.org [6] Reed, D.A. 2003. Grids, the TeraGrid and Beyond, Computer, vol 30, no. 1 and http://www.teragrid.org [7] Evensen, G. 2006. Data Assimilation: The Ensemble Kalman Filter, Springer [8] Herrera, J. Huedo, E. Montero, R. S. and Llorente, I. M. 2005. Scientific Programming, vol 12, No. 4. pp 317-331 [9] Avery, P. and Foster, I. 2001. The GriPhyN project: Towards petascale virtual data grids, technical report GriPhyN-200115 and http://vdt.cs.wisc.edu [10] The PacMan documentation and installation guide http://physics.bu.edu/pacman/htmls [11] Caskey, P. Murray, M. Perez, J. and Sill, A. 2007. Case studies in identify management for virtual organizations, EDUCAUSE Southwest Reg. Conf., Feb 21-23, Austin, TX. http://www.educause.edu/ir/library/pdf/SWR07058.pdf [12] The Grid User Management System (GUMS) https://www.racf.bnl.gov/Facility/GUMS/index.html [13] Thomas, M. and Boisseau, J. 2003. Building grid computing portals: The NPACI grid portal toolkit, Grid computing: making the global infrastructure a reality, Chapter 28, Berman, F. Fox, G. Thomas, M. Boisseau, J. and Hey, T. (eds), John Wiley and Sons, Ltd, Chichester [14] Open Ticket Request System http://otrs.org [15] The MoinMoin Wiki Engine http://moinmoin.wikiwikiweb.de [16] Vasco, D.W. Yoon, S. and Datta-Gupta, A. 1999. Integrating dynamic data into high resolution reservoir models using streamline-based analytic sensitivity coefficients, Society of Petroleum Engineers (SPE) Journal, 4 (4). [17] Emanuel, A. S. and Milliken, W. J. 1998. History matching finite difference models with 3D streamlines, SPE 49000, Proc of the Annual Technical Conf and Exhibition, Sept 2730, New Orleans, LA. [18] Nævdal, G. Johnsen, L.M. Aanonsen, S.I. and Vefring, E.H. 2003. Reservoir monitoring and Continuous Model Updating using Ensemble Kalman Filter, SPE 84372, Proc of the Annual Technical Conf and Exhibition, Oct 5-8, Denver, CO. [19] Jafarpour B. and McLaughlin, D.B. 2007. History matching with an ensemble Kalman filter and discrete cosine parameterization, SPE 108761, Proc of the Annual Technical Conf and Exhibition, Nov 11-14, Anaheim, CA [20] Li, G. and Reynolds, A. C. 2007. An iterative ensemble Kalman filter for data assimilation, SPE 109808, Proc of the SPE Annual Technical Conf and Exhibition, Nov 11-14, Anaheim, CA [21] Arroyo-Negrete, E. Devagowda, D. Datta-Gupta, A. 2006. Streamline assisted ensemble Kalman filter for rapid and continuous reservoir model updating. Proc of the Int. Oil & Gas Conf and Exhibition, SPE 104255, Dec 5-7, China [22] ECLIPSE Reservoir Engineering Software http://www.slb.com/content/services/software/reseng/index.a sp

pooling license;grid-enabling;ensemble kalman filter;and gridway;cyberinfrastructure development project;tigre grid computing environment;grid computing;hydrocarbon reservoir simulation;gridway metascheduler;enkf;datum assimilation methodology;high performance computing;tigre;energy exploration;tigre grid middleware;strategic application area;reservoir model

train_C-44

MSP: Multi-Sequence Positioning of Wireless Sensor Nodes∗

Wireless Sensor Networks have been proposed for use in many location-dependent applications. Most of these need to identify the locations of wireless sensor nodes, a challenging task because of the severe constraints on cost, energy and effective range of sensor devices. To overcome limitations in existing solutions, we present a Multi-Sequence Positioning (MSP) method for large-scale stationary sensor node localization in outdoor environments. The novel idea behind MSP is to reconstruct and estimate two-dimensional location information for each sensor node by processing multiple one-dimensional node sequences, easily obtained through loosely guided event distribution. Starting from a basic MSP design, we propose four optimizations, which work together to increase the localization accuracy. We address several interesting issues, such as incomplete (partial) node sequences and sequence flip, found in the Mirage test-bed we built. We have evaluated the MSP system through theoretical analysis, extensive simulation as well as two physical systems (an indoor version with 46 MICAz motes and an outdoor version with 20 MICAz motes). This evaluation demonstrates that MSP can achieve an accuracy within one foot, requiring neither additional costly hardware on sensor nodes nor precise event distribution. It also provides a nice tradeoff between physical cost (anchors) and soft cost (events), while maintaining localization accuracy.

1 Introduction Although Wireless Sensor Networks (WSN) have shown promising prospects in various applications [5], researchers still face several challenges for massive deployment of such networks. One of these is to identify the location of individual sensor nodes in outdoor environments. Because of unpredictable flow dynamics in airborne scenarios, it is not currently feasible to localize sensor nodes during massive UVA-based deployment. On the other hand, geometric information is indispensable in these networks, since users need to know where events of interest occur (e.g., the location of intruders or of a bomb explosion). Previous research on node localization falls into two categories: range-based approaches and range-free approaches. Range-based approaches [13, 17, 19, 24] compute per-node location information iteratively or recursively based on measured distances among target nodes and a few anchors which precisely know their locations. These approaches generally require costly hardware (e.g., GPS) and have limited effective range due to energy constraints (e.g., ultrasound-based TDOA [3, 17]). Although range-based solutions can be suitably used in small-scale indoor environments, they are considered less cost-effective for large-scale deployments. On the other hand, range-free approaches [4, 8, 10, 13, 14, 15] do not require accurate distance measurements, but localize the node based on network connectivity (proximity) information. Unfortunately, since wireless connectivity is highly influenced by the environment and hardware calibration, existing solutions fail to deliver encouraging empirical results, or require substantial survey [2] and calibration [24] on a case-by-case basis. Realizing the impracticality of existing solutions for the large-scale outdoor environment, researchers have recently proposed solutions (e.g., Spotlight [20] and Lighthouse [18]) for sensor node localization using the spatiotemporal correlation of controlled events (i.e., inferring nodes" locations based on the detection time of controlled events). These solutions demonstrate that long range and high accuracy localization can be achieved simultaneously with little additional cost at sensor nodes. These benefits, however, come along with an implicit assumption that the controlled events can be precisely distributed to a specified location at a specified time. We argue that precise event distribution is difficult to achieve, especially at large scale when terrain is uneven, the event distribution device is not well calibrated and its position is difficult to maintain (e.g., the helicopter-mounted scenario in [20]). To address these limitations in current approaches, in this paper we present a multi-sequence positioning (MSP) method 15 for large-scale stationary sensor node localization, in deployments where an event source has line-of-sight to all sensors. The novel idea behind MSP is to estimate each sensor node"s two-dimensional location by processing multiple easy-to-get one-dimensional node sequences (e.g., event detection order) obtained through loosely-guided event distribution. This design offers several benefits. First, compared to a range-based approach, MSP does not require additional costly hardware. It works using sensors typically used by sensor network applications, such as light and acoustic sensors, both of which we specifically consider in this work. Second, compared to a range-free approach, MSP needs only a small number of anchors (theoretically, as few as two), so high accuracy can be achieved economically by introducing more events instead of more anchors. And third, compared to Spotlight, MSP does not require precise and sophisticated event distribution, an advantage that significantly simplifies the system design and reduces calibration cost. This paper offers the following additional intellectual contributions: • We are the first to localize sensor nodes using the concept of node sequence, an ordered list of sensor nodes, sorted by the detection time of a disseminated event. We demonstrate that making full use of the information embedded in one-dimensional node sequences can significantly improve localization accuracy. Interestingly, we discover that repeated reprocessing of one-dimensional node sequences can further increase localization accuracy. • We propose a distribution-based location estimation strategy that obtains the final location of sensor nodes using the marginal probability of joint distribution among adjacent nodes within the sequence. This new algorithm outperforms the widely adopted Centroid estimation [4, 8]. • To the best of our knowledge, this is the first work to improve the localization accuracy of nodes by adaptive events. The generation of later events is guided by localization results from previous events. • We evaluate line-based MSP on our new Mirage test-bed, and wave-based MSP in outdoor environments. Through system implementation, we discover and address several interesting issues such as partial sequence and sequence flips. To reveal MSP performance at scale, we provide analytic results as well as a complete simulation study. All the simulation and implementation code is available online at http://www.cs.umn.edu/∼zhong/MSP. The rest of the paper is organized as follows. Section 2 briefly surveys the related work. Section 3 presents an overview of the MSP localization system. In sections 4 and 5, basic MSP and four advanced processing methods are introduced. Section 6 describes how MSP can be applied in a wave propagation scenario. Section 7 discusses several implementation issues. Section 8 presents simulation results, and Section 9 reports an evaluation of MSP on the Mirage test-bed and an outdoor test-bed. Section 10 concludes the paper. 2 Related Work Many methods have been proposed to localize wireless sensor devices in the open air. Most of these can be classified into two categories: range-based and range-free localization. Range-based localization systems, such as GPS [23], Cricket [17], AHLoS [19], AOA [16], Robust Quadrilaterals [13] and Sweeps [7], are based on fine-grained point-topoint distance estimation or angle estimation to identify pernode location. Constraints on the cost, energy and hardware footprint of each sensor node make these range-based methods undesirable for massive outdoor deployment. In addition, ranging signals generated by sensor nodes have a very limited effective range because of energy and form factor concerns. For example, ultrasound signals usually effectively propagate 20-30 feet using an on-board transmitter [17]. Consequently, these range-based solutions require an undesirably high deployment density. Although the received signal strength indicator (RSSI) related [2, 24] methods were once considered an ideal low-cost solution, the irregularity of radio propagation [26] seriously limits the accuracy of such systems. The recently proposed RIPS localization system [11] superimposes two RF waves together, creating a low-frequency envelope that can be accurately measured. This ranging technique performs very well as long as antennas are well oriented and environmental factors such as multi-path effects and background noise are sufficiently addressed. Range-free methods don"t need to estimate or measure accurate distances or angles. Instead, anchors or controlled-event distributions are used for node localization. Range-free methods can be generally classified into two types: anchor-based and anchor-free solutions. • For anchor-based solutions such as Centroid [4], APIT [8], SeRLoc [10], Gradient [13] , and APS [15], the main idea is that the location of each node is estimated based on the known locations of the anchor nodes. Different anchor combinations narrow the areas in which the target nodes can possibly be located. Anchor-based solutions normally require a high density of anchor nodes so as to achieve good accuracy. In practice, it is desirable to have as few anchor nodes as possible so as to lower the system cost. • Anchor-free solutions require no anchor nodes. Instead, external event generators and data processing platforms are used. The main idea is to correlate the event detection time at a sensor node with the known space-time relationship of controlled events at the generator so that detection time-stamps can be mapped into the locations of sensors. Spotlight [20] and Lighthouse [18] work in this fashion. In Spotlight [20], the event distribution needs to be precise in both time and space. Precise event distribution is difficult to achieve without careful calibration, especially when the event-generating devices require certain mechanical maneuvers (e.g., the telescope mount used in Spotlight). All these increase system cost and reduce localization speed. StarDust [21], which works much faster, uses label relaxation algorithms to match light spots reflected by corner-cube retro-reflectors (CCR) with sensor nodes using various constraints. Label relaxation algorithms converge only when a sufficient number of robust constraints are obtained. Due to the environmental impact on RF connectivity constraints, however, StarDust is less accurate than Spotlight. In this paper, we propose a balanced solution that avoids the limitations of both anchor-based and anchor-free solutions. Unlike anchor-based solutions [4, 8], MSP allows a flexible tradeoff between the physical cost (anchor nodes) with the soft 16 1 A B 2 3 4 5 Target nodeAnchor node 1A 5 3 B2 4 1 B2 5A 43 1A25B4 3 1 52 AB 4 3 1 2 3 5 4 (b) (c)(d) (a) Event 1 Node Sequence generated by event 1 Event 3 Node Sequence generated by event 2 Node Sequence generated by event 3 Node Sequence generated by event 4 Event 2 Event 4 Figure 1. The MSP System Overview cost (localization events). MSP uses only a small number of anchors (theoretically, as few as two). Unlike anchor-free solutions, MSP doesn"t need to maintain rigid time-space relationships while distributing events, which makes system design simpler, more flexible and more robust to calibration errors. 3 System Overview MSP works by extracting relative location information from multiple simple one-dimensional orderings of nodes. Figure 1(a) shows a layout of a sensor network with anchor nodes and target nodes. Target nodes are defined as the nodes to be localized. Briefly, the MSP system works as follows. First, events are generated one at a time in the network area (e.g., ultrasound propagations from different locations, laser scans with diverse angles). As each event propagates, as shown in Figure 1(a), each node detects it at some particular time instance. For a single event, we call the ordering of nodes, which is based on the sequential detection of the event, a node sequence. Each node sequence includes both the targets and the anchors as shown in Figure 1(b). Second, a multi-sequence processing algorithm helps to narrow the possible location of each node to a small area (Figure 1(c)). Finally, a distributionbased estimation method estimates the exact location of each sensor node, as shown in Figure 1(d). Figure 1 shows that the node sequences can be obtained much more economically than accurate pair-wise distance measurements between target nodes and anchor nodes via ranging methods. In addition, this system does not require a rigid time-space relationship for the localization events, which is critical but hard to achieve in controlled event distribution scenarios (e.g., Spotlight [20]). For the sake of clarity in presentation, we present our system in two cases: • Ideal Case, in which all the node sequences obtained from the network are complete and correct, and nodes are time-synchronized [12, 9]. • Realistic Deployment, in which (i) node sequences can be partial (incomplete), (ii) elements in sequences could flip (i.e., the order obtained is reversed from reality), and (iii) nodes are not time-synchronized. To introduce the MSP algorithm, we first consider a simple straight-line scan scenario. Then, we describe how to implement straight-line scans as well as other event types, such as sound wave propagation. 1 A 2 3 4 5 B C 6 7 8 9 Straight-line Scan 1 Straight-lineScan2 8 1 5 A 6 C 4 3 7 2 B 9 3 1 C 5 9 2 A 4 6 B 7 8 Target node Anchor node Figure 2. Obtaining Multiple Node Sequences 4 Basic MSP Let us consider a sensor network with N target nodes and M anchor nodes randomly deployed in an area of size S. The top-level idea for basic MSP is to split the whole sensor network area into small pieces by processing node sequences. Because the exact locations of all the anchors in a node sequence are known, all the nodes in this sequence can be divided into O(M +1) parts in the area. In Figure 2, we use numbered circles to denote target nodes and numbered hexagons to denote anchor nodes. Basic MSP uses two straight lines to scan the area from different directions, treating each scan as an event. All the nodes react to the event sequentially generating two node sequences. For vertical scan 1, the node sequence is (8,1,5,A,6,C,4,3,7,2,B,9), as shown outside the right boundary of the area in Figure 2; for horizontal scan 2, the node sequence is (3,1,C,5,9,2,A,4,6,B,7,8), as shown under the bottom boundary of the area in Figure 2. Since the locations of the anchor nodes are available, the anchor nodes in the two node sequences actually split the area vertically and horizontally into 16 parts, as shown in Figure 2. To extend this process, suppose we have M anchor nodes and perform d scans from different angles, obtaining d node sequences and dividing the area into many small parts. Obviously, the number of parts is a function of the number of anchors M, the number of scans d, the anchors" location as well as the slop k for each scan line. According to the pie-cutting theorem [22], the area can be divided into O(M2d2) parts. When M and d are appropriately large, the polygon for each target node may become sufficiently small so that accurate estimation can be achieved. We emphasize that accuracy is affected not only by the number of anchors M, but also by the number of events d. In other words, MSP provides a tradeoff between the physical cost of anchors and the soft cost of events. Algorithm 1 depicts the computing architecture of basic MSP. Each node sequence is processed within line 1 to 8. For each node, GetBoundaries() in line 5 searches for the predecessor and successor anchors in the sequence so as to determine the boundaries of this node. Then in line 6 UpdateMap() shrinks the location area of this node according to the newly obtained boundaries. After processing all sequences, Centroid Estimation (line 11) set the center of gravity of the final polygon as the estimated location of the target node. Basic MSP only makes use of the order information between a target node and the anchor nodes in each sequence. Actually, we can extract much more location information from 17 Algorithm 1 Basic MSP Process Output: The estimated location of each node. 1: repeat 2: GetOneUnprocessedSeqence(); 3: repeat 4: GetOneNodeFromSequenceInOrder(); 5: GetBoundaries(); 6: UpdateMap(); 7: until All the target nodes are updated; 8: until All the node sequences are processed; 9: repeat 10: GetOneUnestimatedNode(); 11: CentroidEstimation(); 12: until All the target nodes are estimated; each sequence. Section 5 will introduce advanced MSP, in which four novel optimizations are proposed to improve the performance of MSP significantly. 5 Advanced MSP Four improvements to basic MSP are proposed in this section. The first three improvements do not need additional sensing and communication in the networks but require only slightly more off-line computation. The objective of all these improvements is to make full use of the information embedded in the node sequences. The results we have obtained empirically indicate that the implementation of the first two methods can dramatically reduce the localization error, and that the third and fourth methods are helpful for some system deployments. 5.1 Sequence-Based MSP As shown in Figure 2, each scan line and M anchors, splits the whole area into M + 1 parts. Each target node falls into one polygon shaped by scan lines. We noted that in basic MSP, only the anchors are used to narrow down the polygon of each target node, but actually there is more information in the node sequence that we can made use of. Let"s first look at a simple example shown in Figure 3. The previous scans narrow the locations of target node 1 and node 2 into two dashed rectangles shown in the left part of Figure 3. Then a new scan generates a new sequence (1, 2). With knowledge of the scan"s direction, it is easy to tell that node 1 is located to the left of node 2. Thus, we can further narrow the location area of node 2 by eliminating the shaded part of node 2"s rectangle. This is because node 2 is located on the right of node 1 while the shaded area is outside the lower boundary of node 1. Similarly, the location area of node 1 can be narrowed by eliminating the shaded part out of node 2"s right boundary. We call this procedure sequence-based MSP which means that the whole node sequence needs to be processed node by node in order. Specifically, sequence-based MSP follows this exact processing rule: 1 2 1 2 1 2 Lower boundary of 1 Upper boundary of 1 Lower boundary of 2 Upper boundary of 2 New sequence New upper boundary of 1 New Lower boundary of 2 EventPropagation Figure 3. Rule Illustration in Sequence Based MSP Algorithm 2 Sequence-Based MSP Process Output: The estimated location of each node. 1: repeat 2: GetOneUnprocessedSeqence(); 3: repeat 4: GetOneNodeByIncreasingOrder(); 5: ComputeLowbound(); 6: UpdateMap(); 7: until The last target node in the sequence; 8: repeat 9: GetOneNodeByDecreasingOrder(); 10: ComputeUpbound(); 11: UpdateMap(); 12: until The last target node in the sequence; 13: until All the node sequences are processed; 14: repeat 15: GetOneUnestimatedNode(); 16: CentroidEstimation(); 17: until All the target nodes are estimated; Elimination Rule: Along a scanning direction, the lower boundary of the successor"s area must be equal to or larger than the lower boundary of the predecessor"s area, and the upper boundary of the predecessor"s area must be equal to or smaller than the upper boundary of the successor"s area. In the case of Figure 3, node 2 is the successor of node 1, and node 1 is the predecessor of node 2. According to the elimination rule, node 2"s lower boundary cannot be smaller than that of node 1 and node 1"s upper boundary cannot exceed node 2"s upper boundary. Algorithm 2 illustrates the pseudo code of sequence-based MSP. Each node sequence is processed within line 3 to 13. The sequence processing contains two steps: Step 1 (line 3 to 7): Compute and modify the lower boundary for each target node by increasing order in the node sequence. Each node"s lower boundary is determined by the lower boundary of its predecessor node in the sequence, thus the processing must start from the first node in the sequence and by increasing order. Then update the map according to the new lower boundary. Step 2 (line 8 to 12): Compute and modify the upper boundary for each node by decreasing order in the node sequence. Each node"s upper boundary is determined by the upper boundary of its successor node in the sequence, thus the processing must start from the last node in the sequence and by decreasing order. Then update the map according to the new upper boundary. After processing all the sequences, for each node, a polygon bounding its possible location has been found. Then, center-ofgravity-based estimation is applied to compute the exact location of each node (line 14 to 17). An example of this process is shown in Figure 4. The third scan generates the node sequence (B,9,2,7,4,6,3,8,C,A,5,1). In addition to the anchor split lines, because nodes 4 and 7 come after node 2 in the sequence, node 4 and 7"s polygons could be narrowed according to node 2"s lower boundary (the lower right-shaded area); similarly, the shaded area in node 2"s rectangle could be eliminated since this part is beyond node 7"s upper boundary indicated by the dotted line. Similar eliminating can be performed for node 3 as shown in the figure. 18 1 A 2 3 4 5 B C 6 7 8 9 Straight-line Scan 1 Straight-lineScan2 Straight-line Scan 3 Target node Anchor node Figure 4. Sequence-Based MSP Example 1 A 2 3 4 5 B C 6 7 8 9 Straight-line Scan 1 Straight-lineScan2 Straight-line Scan 3 Reprocessing Scan 1 Target node Anchor node Figure 5. Iterative MSP: Reprocessing Scan 1 From above, we can see that the sequence-based MSP makes use of the information embedded in every sequential node pair in the node sequence. The polygon boundaries of the target nodes obtained in prior could be used to further split other target nodes" areas. Our evaluation in Sections 8 and 9 shows that sequence-based MSP considerably enhances system accuracy. 5.2 Iterative MSP Sequence-based MSP is preferable to basic MSP because it extracts more information from the node sequence. In fact, further useful information still remains! In sequence-based MSP, a sequence processed later benefits from information produced by previously processed sequences (e.g., the third scan in Figure 5). However, the first several sequences can hardly benefit from other scans in this way. Inspired by this phenomenon, we propose iterative MSP. The basic idea of iterative MSP is to process all the sequences iteratively several times so that the processing of each single sequence can benefit from the results of other sequences. To illustrate the idea more clearly, Figure 4 shows the results of three scans that have provided three sequences. Now if we process the sequence (8,1,5,A,6,C,4,3,7,2,B,9) obtained from scan 1 again, we can make progress, as shown in Figure 5. The reprocessing of the node sequence 1 provides information in the way an additional vertical scan would. From sequencebased MSP, we know that the upper boundaries of nodes 3 and 4 along the scan direction must not extend beyond the upper boundary of node 7, therefore the grid parts can be eliminated (a) Central of Gravity (b) Joint Distribution 1 2 2 1 1 2 1 2 2 1 1 2 2 1 1 2 Figure 6. Example of Joint Distribution Estimation …... vm ap[0] vm ap[1] vm ap[2] vm ap[3] Combine m ap Figure 7. Idea of DBE MSP for Each Node for the nodes 3 and node 4, respectively, as shown in Figure 5. From this example, we can see that iterative processing of the sequence could help further shrink the polygon of each target node, and thus enhance the accuracy of the system. The implementation of iterative MSP is straightforward: process all the sequences multiple times using sequence-based MSP. Like sequence-based MSP, iterative MSP introduces no additional event cost. In other words, reprocessing does not actually repeat the scan physically. Evaluation results in Section 8 will show that iterative MSP contributes noticeably to a lower localization error. Empirical results show that after 5 iterations, improvements become less significant. In summary, iterative processing can achieve better performance with only a small computation overhead. 5.3 Distribution-Based Estimation After determining the location area polygon for each node, estimation is needed for a final decision. Previous research mostly applied the Center of Gravity (COG) method [4] [8] [10] which minimizes average error. If every node is independent of all others, COG is the statistically best solution. In MSP, however, each node may not be independent. For example, two neighboring nodes in a certain sequence could have overlapping polygon areas. In this case, if the marginal probability of joint distribution is used for estimation, better statistical results are achieved. Figure 6 shows an example in which node 1 and node 2 are located in the same polygon. If COG is used, both nodes are localized at the same position (Figure 6(a)). However, the node sequences obtained from two scans indicate that node 1 should be to the left of and above node 2, as shown in Figure 6(b). The high-level idea of distribution-based estimation proposed for MSP, which we call DBE MSP, is illustrated in Figure 7. The distributions of each node under the ith scan (for the ith node sequence) are estimated in node.vmap[i], which is a data structure for remembering the marginal distribution over scan i. Then all the vmaps are combined to get a single map and weighted estimation is used to obtain the final location. For each scan, all the nodes are sorted according to the gap, which is the diameter of the polygon along the direction of the scan, to produce a second, gap-based node sequence. Then, the estimation starts from the node with the smallest gap. This is because it is statistically more accurate to assume a uniform distribution of the node with smaller gap. For each node processed in order from the gap-based node sequence, either if 19 Pred. node"s area Predecessor node exists: conditional distribution based on pred. node"s area Alone: Uniformly Distributed Succ. node"s area Successor node exists: conditional distribution based on succ. node"s area Succ. node"s area Both predecessor and successor nodes exist: conditional distribution based on both of them Pred. node"s area Figure 8. Four Cases in DBE Process no neighbor node in the original event-based node sequence shares an overlapping area, or if the neighbors have not been processed due to bigger gaps, a uniform distribution Uniform() is applied to this isolated node (the Alone case in Figure 8). If the distribution of its neighbors sharing overlapped areas has been processed, we calculate the joint distribution for the node. As shown in Figure 8, there are three possible cases depending on whether the distribution of the overlapping predecessor and/or successor nodes have/has already been estimated. The estimation"s strategy of starting from the most accurate node (smallest gap node) reduces the problem of estimation error propagation. The results in the evaluation section indicate that applying distribution-based estimation could give statistically better results. 5.4 Adaptive MSP So far, all the enhancements to basic MSP focus on improving the multi-sequence processing algorithm given a fixed set of scan directions. All these enhancements require only more computing time without any overhead to the sensor nodes. Obviously, it is possible to have some choice and optimization on how events are generated. For example, in military situations, artillery or rocket-launched mini-ultrasound bombs can be used for event generation at some selected locations. In adaptive MSP, we carefully generate each new localization event so as to maximize the contribution of the new event to the refinement of localization, based on feedback from previous events. Figure 9 depicts the basic architecture of adaptive MSP. Through previous localization events, the whole map has been partitioned into many small location areas. The idea of adaptive MSP is to generate the next localization event to achieve best-effort elimination, which ideally could shrink the location area of individual node as much as possible. We use a weighted voting mechanism to evaluate candidate localization events. Every node wants the next event to split its area evenly, which would shrink the area fast. Therefore, every node votes for the parameters of the next event (e.g., the scan angle k of the straight-line scan). Since the area map is maintained centrally, the vote is virtually done and there is no need for the real sensor nodes to participate in it. After gathering all the voting results, the event parameters with the most votes win the election. There are two factors that determine the weight of each vote: • The vote for each candidate event is weighted according to the diameter D of the node"s location area. Nodes with bigger location areas speak louder in the voting, because Map Partitioned by the Localization Events Diameter of Each Area Candidate Localization Events Evaluation Trigger Next Localization Evet Figure 9. Basic Architecture of Adaptive MSP 2 3 Diameter D3 1 1 3k 2 3k 3 3k 4 3k 5 3k 6 3k 1 3k 2 3k 3 3k 6 3k4 3k 5 3k Weight el small i opt i j ii j i S S DkkDfkWeight arg ),(,()( ⋅=∆= 1 3 opt k Target node Anchor node Center of Gravity Node 3's area Figure 10. Candidate Slops for Node 3 at Anchor 1 overall system error is reduced mostly by splitting the larger areas. • The vote for each candidate event is also weighted according to its elimination efficiency for a location area, which is defined as how equally in size (or in diameter) an event can cut an area. In other words, an optimal scan event cuts an area in the middle, since this cut shrinks the area quickly and thus reduces localization uncertainty quickly. Combining the above two aspects, the weight for each vote is computed according to the following equation (1): Weight(k j i ) = f(Di,△(k j i ,k opt i )) (1) k j i is node i"s jth supporting parameter for next event generation; Di is diameter of node i"s location area; △(k j i ,k opt i ) is the distance between k j i and the optimal parameter k opt i for node i, which should be defined to fit the specific application. Figure 10 presents an example for node 1"s voting for the slopes of the next straight-line scan. In the system, there are a fixed number of candidate slopes for each scan (e.g., k1,k2,k3,k4...). The location area of target node 3 is shown in the figure. The candidate events k1 3,k2 3,k3 3,k4 3,k5 3,k6 3 are evaluated according to their effectiveness compared to the optimal ideal event which is shown as a dotted line with appropriate weights computed according to equation (1). For this specific example, as is illustrated in the right part of Figure 10, f(Di,△(k j i ,kopt i )) is defined as the following equation (2): Weight(kj i ) = f(Di,△(kj i ,kopt i )) = Di · Ssmall Slarge (2) Ssmall and Slarge are the sizes of the smaller part and larger part of the area cut by the candidate line respectively. In this case, node 3 votes 0 for the candidate lines that do not cross its area since Ssmall = 0. We show later that adaptive MSP improves localization accuracy in WSNs with irregularly shaped deployment areas. 20 5.5 Overhead and MSP Complexity Analysis This section provides a complexity analysis of the MSP design. We emphasize that MSP adopts an asymmetric design in which sensor nodes need only to detect and report the events. They are blissfully oblivious to the processing methods proposed in previous sections. In this section, we analyze the computational cost on the node sequence processing side, where resources are plentiful. According to Algorithm 1, the computational complexity of Basic MSP is O(d · N · S), and the storage space required is O(N · S), where d is the number of events, N is the number of target nodes, and S is the area size. According to Algorithm 2, the computational complexity of both sequence-based MSP and iterative MSP is O(c·d ·N ·S), where c is the number of iterations and c = 1 for sequencebased MSP, and the storage space required is O(N ·S). Both the computational complexity and storage space are equal within a constant factor to those of basic MSP. The computational complexity of the distribution-based estimation (DBE MSP) is greater. The major overhead comes from the computation of joint distributions when both predecessor and successor nodes exit. In order to compute the marginal probability, MSP needs to enumerate the locations of the predecessor node and the successor node. For example, if node A has predecessor node B and successor node C, then the marginal probability PA(x,y) of node A"s being at location (x,y) is: PA(x,y) = ∑ i ∑ j ∑ m ∑ n 1 NB,A,C ·PB(i, j)·PC(m,n) (3) NB,A,C is the number of valid locations for A satisfying the sequence (B, A, C) when B is at (i, j) and C is at (m,n); PB(i, j) is the available probability of node B"s being located at (i, j); PC(m,n) is the available probability of node C"s being located at (m,n). A naive algorithm to compute equation (3) has complexity O(d · N · S3). However, since the marginal probability indeed comes from only one dimension along the scanning direction (e.g., a line), the complexity can be reduced to O(d · N · S1.5) after algorithm optimization. In addition, the final location areas for every node are much smaller than the original field S; therefore, in practice, DBE MSP can be computed much faster than O(d ·N ·S1.5). 6 Wave Propagation Example So far, the description of MSP has been solely in the context of straight-line scan. However, we note that MSP is conceptually independent of how the event is propagated as long as node sequences can be obtained. Clearly, we can also support wave-propagation-based events (e.g., ultrasound propagation, air blast propagation), which are polar coordinate equivalences of the line scans in the Cartesian coordinate system. This section illustrates the effects of MSP"s implementation in the wave propagation-based situation. For easy modelling, we have made the following assumptions: • The wave propagates uniformly in all directions, therefore the propagation has a circle frontier surface. Since MSP does not rely on an accurate space-time relationship, a certain distortion in wave propagation is tolerable. If any directional wave is used, the propagation frontier surface can be modified accordingly. 1 3 5 9 Target node Anchor node Previous Event location A 2 Center of Gravity 4 8 7 B 6 C A line of preferred locations for next event Figure 11. Example of Wave Propagation Situation • Under the situation of line-of-sight, we allow obstacles to reflect or deflect the wave. Reflection and deflection are not problems because each node reacts only to the first detected event. Those reflected or deflected waves come later than the line-of-sight waves. The only thing the system needs to maintain is an appropriate time interval between two successive localization events. • We assume that background noise exists, and therefore we run a band-pass filter to listen to a particular wave frequency. This reduces the chances of false detection. The parameter that affects the localization event generation here is the source location of the event. The different distances between each node and the event source determine the rank of each node in the node sequence. Using the node sequences, the MSP algorithm divides the whole area into many non-rectangular areas as shown in Figure 11. In this figure, the stars represent two previous event sources. The previous two propagations split the whole map into many areas by those dashed circles that pass one of the anchors. Each node is located in one of the small areas. Since sequence-based MSP, iterative MSP and DBE MSP make no assumptions about the type of localization events and the shape of the area, all three optimization algorithms can be applied for the wave propagation scenario. However, adaptive MSP needs more explanation. Figure 11 illustrates an example of nodes" voting for next event source locations. Unlike the straight-line scan, the critical parameter now is the location of the event source, because the distance between each node and the event source determines the rank of the node in the sequence. In Figure 11, if the next event breaks out along/near the solid thick gray line, which perpendicularly bisects the solid dark line between anchor C and the center of gravity of node 9"s area (the gray area), the wave would reach anchor C and the center of gravity of node 9"s area at roughly the same time, which would relatively equally divide node 9"s area. Therefore, node 9 prefers to vote for the positions around the thick gray line. 7 Practical Deployment Issues For the sake of presentation, until now we have described MSP in an ideal case where a complete node sequence can be obtained with accurate time synchronization. In this section we describe how to make MSP work well under more realistic conditions. 21 7.1 Incomplete Node Sequence For diverse reasons, such as sensor malfunction or natural obstacles, the nodes in the network could fail to detect localization events. In such cases, the node sequence will not be complete. This problem has two versions: • Anchor nodes are missing in the node sequence If some anchor nodes fail to respond to the localization events, then the system has fewer anchors. In this case, the solution is to generate more events to compensate for the loss of anchors so as to achieve the desired accuracy requirements. • Target nodes are missing in the node sequence There are two consequences when target nodes are missing. First, if these nodes are still be useful to sensing applications, they need to use other backup localization approaches (e.g., Centroid) to localize themselves with help from their neighbors who have already learned their own locations from MSP. Secondly, since in advanced MSP each node in the sequence may contribute to the overall system accuracy, dropping of target nodes from sequences could also reduce the accuracy of the localization. Thus, proper compensation procedures such as adding more localization events need to be launched. 7.2 Localization without Time Synchronization In a sensor network without time synchronization support, nodes cannot be ordered into a sequence using timestamps. For such cases, we propose a listen-detect-assemble-report protocol, which is able to function independently without time synchronization. listen-detect-assemble-report requires that every node listens to the channel for the node sequence transmitted from its neighbors. Then, when the node detects the localization event, it assembles itself into the newest node sequence it has heard and reports the updated sequence to other nodes. Figure 12 (a) illustrates an example for the listen-detect-assemble-report protocol. For simplicity, in this figure we did not differentiate the target nodes from anchor nodes. A solid line between two nodes stands for a communication link. Suppose a straight line scans from left to right. Node 1 detects the event, and then it broadcasts the sequence (1) into the network. Node 2 and node 3 receive this sequence. When node 2 detects the event, node 2 adds itself into the sequence and broadcasts (1, 2). The sequence propagates in the same direction with the scan as shown in Figure 12 (a). Finally, node 6 obtains a complete sequence (1,2,3,5,7,4,6). In the case of ultrasound propagation, because the event propagation speed is much slower than that of radio, the listendetect-assemble-report protocol can work well in a situation where the node density is not very high. For instance, if the distance between two nodes along one direction is 10 meters, the 340m/s sound needs 29.4ms to propagate from one node to the other. While normally the communication data rate is 250Kbps in the WSN (e.g., CC2420 [1]), it takes only about 2 ∼ 3 ms to transmit an assembled packet for one hop. One problem that may occur using the listen-detectassemble-report protocol is multiple partial sequences as shown in Figure 12 (b). Two separate paths in the network may result in two sequences that could not be further combined. In this case, since the two sequences can only be processed as separate sequences, some order information is lost. Therefore the 1,2,5,4 1,3,7,4 1,2,3,5 1,2,3,5,7,4 1,2,3,5,7 1,2,3,5 1,3 1,2 1 2 3 5 7 4 6 1 1 1,3 1,2,3,5,7,4,6 1,2,5 1,3,7 1,3 1,2 1 2 3 5 7 4 6 1 1 1,3,7,4,6 1,2,5,4,6 (a) (b) (c) 1,3,2,5 1,3,2,5,7,4 1,3,2,5,7 1,3,2,5 1,3 1,2 1 2 3 5 7 4 6 1 1 1,3 1,3,2,5,7,4,6 Event Propagation Event Propagation Event Propagation Figure 12. Node Sequence without Time Synchronization accuracy of the system would decrease. The other problem is the sequence flip problem. As shown in Figure 12 (c), because node 2 and node 3 are too close to each other along the scan direction, they detect the scan almost simultaneously. Due to the uncertainty such as media access delay, two messages could be transmitted out of order. For example, if node 3 sends out its report first, then the order of node 2 and node 3 gets flipped in the final node sequence. The sequence flip problem would appear even in an accurately synchronized system due to random jitter in node detection if an event arrives at multiple nodes almost simultaneously. A method addressing the sequence flip is presented in the next section. 7.3 Sequence Flip and Protection Band Sequence flip problems can be solved with and without time synchronization. We firstly start with a scenario applying time synchronization. Existing solutions for time synchronization [12, 6] can easily achieve sub-millisecond-level accuracy. For example, FTSP [12] achieves 16.9µs (microsecond) average error for a two-node single-hop case. Therefore, we can comfortably assume that the network is synchronized with maximum error of 1000µs. However, when multiple nodes are located very near to each other along the event propagation direction, even when time synchronization with less than 1ms error is achieved in the network, sequence flip may still occur. For example, in the sound wave propagation case, if two nodes are less than 0.34 meters apart, the difference between their detection timestamp would be smaller than 1 millisecond. We find that sequence flip could not only damage system accuracy, but also might cause a fatal error in the MSP algorithm. Figure 13 illustrates both detrimental results. In the left side of Figure 13(a), suppose node 1 and node 2 are so close to each other that it takes less than 0.5ms for the localization event to propagate from node 1 to node 2. Now unfortunately, the node sequence is mistaken to be (2,1). So node 1 is expected to be located to the right of node 2, such as at the position of the dashed node 1. According to the elimination rule in sequencebased MSP, the left part of node 1"s area is cut off as shown in the right part of Figure 13(a). This is a potentially fatal error, because node 1 is actually located in the dashed area which has been eliminated by mistake. During the subsequent eliminations introduced by other events, node 1"s area might be cut off completely, thus node 1 could consequently be erased from the map! Even in cases where node 1 still survives, its area actually does not cover its real location. 22 1 2 12 2 Lower boundary of 1 Upper boundary of 1 Flipped Sequence Fatal Elimination Error EventPropagation 1 1 Fatal Error 1 1 2 12 2 Lower boundary of 1 Upper boundary of 1 Flipped Sequence Safe Elimination EventPropagation 1 1 New lower boundary of 1 1 B (a) (b) B: Protection band Figure 13. Sequence Flip and Protection Band Another problem is not fatal but lowers the localization accuracy. If we get the right node sequence (1,2), node 1 has a new upper boundary which can narrow the area of node 1 as in Figure 3. Due to the sequence flip, node 1 loses this new upper boundary. In order to address the sequence flip problem, especially to prevent nodes from being erased from the map, we propose a protection band compensation approach. The basic idea of protection band is to extend the boundary of the location area a little bit so as to make sure that the node will never be erased from the map. This solution is based on the fact that nodes have a high probability of flipping in the sequence if they are near to each other along the event propagation direction. If two nodes are apart from each other more than some distance, say, B, they rarely flip unless the nodes are faulty. The width of a protection band B, is largely determined by the maximum error in system time synchronization and the localization event propagation speed. Figure 13(b) presents the application of the protection band. Instead of eliminating the dashed part in Figure 13(a) for node 1, the new lower boundary of node 1 is set by shifting the original lower boundary of node 2 to the left by distance B. In this case, the location area still covers node 1 and protects it from being erased. In a practical implementation, supposing that the ultrasound event is used, if the maximum error of system time synchronization is 1ms, two nodes might flip with high probability if the timestamp difference between the two nodes is smaller than or equal to 1ms. Accordingly, we set the protection band B as 0.34m (the distance sound can propagate within 1 millisecond). By adding the protection band, we reduce the chances of fatal errors, although at the cost of localization accuracy. Empirical results obtained from our physical test-bed verified this conclusion. In the case of using the listen-detect-assemble-report protocol, the only change we need to make is to select the protection band according to the maximum delay uncertainty introduced by the MAC operation and the event propagation speed. To bound MAC delay at the node side, a node can drop its report message if it experiences excessive MAC delay. This converts the sequence flip problem to the incomplete sequence problem, which can be more easily addressed by the method proposed in Section 7.1. 8 Simulation Evaluation Our evaluation of MSP was conducted on three platforms: (i) an indoor system with 46 MICAz motes using straight-line scan, (ii) an outdoor system with 20 MICAz motes using sound wave propagation, and (iii) an extensive simulation under various kinds of physical settings. In order to understand the behavior of MSP under numerous settings, we start our evaluation with simulations. Then, we implemented basic MSP and all the advanced MSP methods for the case where time synchronization is available in the network. The simulation and implementation details are omitted in this paper due to space constraints, but related documents [25] are provided online at http://www.cs.umn.edu/∼zhong/MSP. Full implementation and evaluation of system without time synchronization are yet to be completed in the near future. In simulation, we assume all the node sequences are perfect so as to reveal the performance of MSP achievable in the absence of incomplete node sequences or sequence flips. In our simulations, all the anchor nodes and target nodes are assumed to be deployed uniformly. The mean and maximum errors are averaged over 50 runs to obtain high confidence. For legibility reasons, we do not plot the confidence intervals in this paper. All the simulations are based on the straight-line scan example. We implement three scan strategies: • Random Scan: The slope of the scan line is randomly chosen at each time. • Regular Scan: The slope is predetermined to rotate uniformly from 0 degree to 180 degrees. For example, if the system scans 6 times, then the scan angles would be: 0, 30, 60, 90, 120, and 150. • Adaptive Scan: The slope of each scan is determined based on the localization results from previous scans. We start with basic MSP and then demonstrate the performance improvements one step at a time by adding (i) sequencebased MSP, (ii) iterative MSP, (iii) DBE MSP and (iv) adaptive MSP. 8.1 Performance of Basic MSP The evaluation starts with basic MSP, where we compare the performance of random scan and regular scan under different configurations. We intend to illustrate the impact of the number of anchors M, the number of scans d, and target node density (number of target nodes N in a fixed-size region) on the localization error. Table 1 shows the default simulation parameters. The error of each node is defined as the distance between the estimated location and the real position. We note that by default we only use three anchors, which is considerably fewer than existing range-free solutions [8, 4]. Impact of the Number of Scans: In this experiment, we compare regular scan with random scan under a different number of scans from 3 to 30 in steps of 3. The number of anchors Table 1. Default Configuration Parameters Parameter Description Field Area 200×200 (Grid Unit) Scan Type Regular (Default)/Random Scan Anchor Number 3 (Default) Scan Times 6 (Default) Target Node Number 100 (Default) Statistics Error Mean/Max Random Seeds 50 runs 23 0 5 10 15 20 25 30 0 10 20 30 40 50 60 70 80 90 Mean Error and Max Error VS Scan Time Scan Time Error Max Error of Random Scan Max Error of Regular Scan Mean Error of Random Scan Mean Error of Regular Scan (a) Error vs. Number of Scans 0 5 10 15 20 25 30 0 10 20 30 40 50 60 Mean Error and Max Error VS Anchor Number Anchor Number Error Max Error of Random Scan Max Error of Regular Scan Mean Error of Random Scan Mean Error of Regular Scan (b) Error vs. Anchor Number 0 50 100 150 200 10 20 30 40 50 60 70 Mean Error and Max Error VS Target Node Number Target Node Number Error Max Error of Random Scan Max Error of Regular Scan Mean Error of Random Scan Mean Error of Regular Scan (c) Error vs. Number of Target Nodes Figure 14. Evaluation of Basic MSP under Random and Regular Scans 0 5 10 15 20 25 30 0 10 20 30 40 50 60 70 Basic MSP VS Sequence Based MSP II Scan Time Error Max Error of Basic MSP Max Error of Seq MSP Mean Error of Basic MSP Mean Error of Seq MSP (a) Error vs. Number of Scans 0 5 10 15 20 25 30 0 5 10 15 20 25 30 35 40 45 50 Basic MSP VS Sequence Based MSP I Anchor Number Error Max Error of Basic MSP Max Error of Seq MSP Mean Error of Basic MSP Mean Error of Seq MSP (b) Error vs. Anchor Number 0 50 100 150 200 5 10 15 20 25 30 35 40 45 50 55 Basic MSP VS Sequence Based MSP III Target Node Number Error Max Error of Basic MSP Max Error of Seq MSP Mean Error of Basic MSP Mean Error of Seq MSP (c) Error vs. Number of Target Nodes Figure 15. Improvements of Sequence-Based MSP over Basic MSP is 3 by default. Figure 14(a) indicates the following: (i) as the number of scans increases, the localization error decreases significantly; for example, localization errors drop more than 60% from 3 scans to 30 scans; (ii) statistically, regular scan achieves better performance than random scan under identical number of scans. However, the performance gap reduces as the number of scans increases. This is expected since a large number of random numbers converges to a uniform distribution. Figure 14(a) also demonstrates that MSP requires only a small number of anchors to perform very well, compared to existing range-free solutions [8, 4]. Impact of the Number of Anchors: In this experiment, we compare regular scan with random scan under different number of anchors from 3 to 30 in steps of 3. The results shown in Figure 14(b) indicate that (i) as the number of anchor nodes increases, the localization error decreases, and (ii) statistically, regular scan obtains better results than random scan with identical number of anchors. By combining Figures 14(a) and 14(b), we can conclude that MSP allows a flexible tradeoff between physical cost (anchor nodes) and soft cost (localization events). Impact of the Target Node Density: In this experiment, we confirm that the density of target nodes has no impact on the accuracy, which motivated the design of sequence-based MSP. In this experiment, we compare regular scan with random scan under different number of target nodes from 10 to 190 in steps of 20. Results in Figure 14(c) show that mean localization errors remain constant across different node densities. However, when the number of target nodes increases, the average maximum error increases. Summary: From the above experiments, we can conclude that in basic MSP, regular scan are better than random scan under different numbers of anchors and scan events. This is because regular scans uniformly eliminate the map from different directions, while random scans would obtain sequences with redundant overlapping information, if two scans choose two similar scanning slopes. 8.2 Improvements of Sequence-Based MSP This section evaluates the benefits of exploiting the order information among target nodes by comparing sequence-based MSP with basic MSP. In this and the following sections, regular scan is used for straight-line scan event generation. The purpose of using regular scan is to keep the scan events and the node sequences identical for both sequence-based MSP and basic MSP, so that the only difference between them is the sequence processing procedure. Impact of the Number of Scans: In this experiment, we compare sequence-based MSP with basic MSP under different number of scans from 3 to 30 in steps of 3. Figure 15(a) indicates significant performance improvement in sequence-based MSP over basic MSP across all scan settings, especially when the number of scans is large. For example, when the number of scans is 30, errors in sequence-based MSP are about 20% of that of basic MSP. We conclude that sequence-based MSP performs extremely well when there are many scan events. Impact of the Number of Anchors: In this experiment, we use different number of anchors from 3 to 30 in steps of 3. As seen in Figure 15(b), the mean error and maximum error of sequence-based MSP is much smaller than that of basic MSP. Especially when there is limited number of anchors in the system, e.g., 3 anchors, the error rate was almost halved by using sequence-based MSP. This phenomenon has an interesting explanation: the cutting lines created by anchor nodes are exploited by both basic MSP and sequence-based MSP, so as the 24 0 2 4 6 8 10 0 5 10 15 20 25 30 35 40 45 50 Basic MSP VS Iterative MSP Iterative Times Error Max Error of Iterative Seq MSP Mean Error of Iterative Seq MSP Max Error of Basic MSP Mean Error of Basic MSP Figure 16. Improvements of Iterative MSP 0 2 4 6 8 10 12 14 16 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 DBE VS Non−DBE Error CumulativeDistrubutioinFunctions(CDF) Mean Error CDF of DBE MSP Mean Error CDF of Non−DBE MSP Max Error CDF of DBE MSP Max Error CDF of Non−DBE MSP Figure 17. Improvements of DBE MSP 0 20 40 60 80 100 0 10 20 30 40 50 60 70 Adaptive MSP for 500by80 Target Node Number Error Max Error of Regualr Scan Max Error of Adaptive Scan Mean Error of Regualr Scan Mean Error of Adaptive Scan (a) Adaptive MSP for 500 by 80 field 0 10 20 30 40 50 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 Mean Error CDF at Different Angle Steps in Adaptive Scan Mean Error CumulativeDistrubutioinFunctions(CDF) 5 Degree Angle Step Adaptive 10 Degree Angle Step Adaptive 20 Degree Angle Step Adaptive 30 Degree Step Regular Scan (b) Impact of the Number of Candidate Events Figure 18. The Improvements of Adaptive MSP number of anchor nodes increases, anchors tend to dominate the contribution. Therefore the performance gaps lessens. Impact of the Target Node Density: Figure 15(c) demonstrates the benefits of exploiting order information among target nodes. Since sequence-based MSP makes use of the information among the target nodes, having more target nodes contributes to the overall system accuracy. As the number of target nodes increases, the mean error and maximum error of sequence-based MSP decreases. Clearly the mean error in basic MSP is not affected by the number of target nodes, as shown in Figure 15(c). Summary: From the above experiments, we can conclude that exploiting order information among target nodes can improve accuracy significantly, especially when the number of events is large but with few anchors. 8.3 Iterative MSP over Sequence-Based MSP In this experiment, the same node sequences were processed iteratively multiple times. In Figure 16, the two single marks are results from basic MSP, since basic MSP doesn"t perform iterations. The two curves present the performance of iterative MSP under different numbers of iterations c. We note that when only a single iteration is used, this method degrades to sequence-based MSP. Therefore, Figure 16 compares the three methods to one another. Figure 16 shows that the second iteration can reduce the mean error and maximum error dramatically. After that, the performance gain gradually reduces, especially when c > 5. This is because the second iteration allows earlier scans to exploit the new boundaries created by later scans in the first iteration. Such exploitation decays quickly over iterations. 8.4 DBE MSP over Iterative MSP Figure 17, in which we augment iterative MSP with distribution-based estimation (DBE MSP), shows that DBE MSP could bring about statistically better performance. Figure 17 presents cumulative distribution localization errors. In general, the two curves of the DBE MSP lay slightly to the left of that of non-DBE MSP, which indicates that DBE MSP has a smaller statistical mean error and averaged maximum error than non-DBE MSP. We note that because DBE is augmented on top of the best solution so far, the performance improvement is not significant. When we apply DBE on basic MSP methods, the improvement is much more significant. We omit these results because of space constraints. 8.5 Improvements of Adaptive MSP This section illustrates the performance of adaptive MSP over non-adaptive MSP. We note that feedback-based adaptation can be applied to all MSP methods, since it affects only the scanning angles but not the sequence processing. In this experiment, we evaluated how adaptive MSP can improve the best solution so far. The default angle granularity (step) for adaptive searching is 5 degrees. Impact of Area Shape: First, if system settings are regular, the adaptive method hardly contributes to the results. For a square area (regular), the performance of adaptive MSP and regular scans are very close. However, if the shape of the area is not regular, adaptive MSP helps to choose the appropriate localization events to compensate. Therefore, adaptive MSP can achieve a better mean error and maximum error as shown in Figure 18(a). For example, adaptive MSP improves localization accuracy by 30% when the number of target nodes is 10. Impact of the Target Node Density: Figure 18(a) shows that when the node density is low, adaptive MSP brings more benefit than when node density is high. This phenomenon makes statistical sense, because the law of large numbers tells us that node placement approaches a truly uniform distribution when the number of nodes is increased. Adaptive MSP has an edge 25 Figure 19. The Mirage Test-bed (Line Scan) Figure 20. The 20-node Outdoor Experiments (Wave) when layout is not uniform. Impact of Candidate Angle Density: Figure 18(b) shows that the smaller the candidate scan angle step, the better the statistical performance in terms of mean error. The rationale is clear, as wider candidate scan angles provide adaptive MSP more opportunity to choose the one approaching the optimal angle. 8.6 Simulation Summary Starting from basic MSP, we have demonstrated step-bystep how four optimizations can be applied on top of each other to improve localization performance. In other words, these optimizations are compatible with each other and can jointly improve the overall performance. We note that our simulations were done under assumption that the complete node sequence can be obtained without sequence flips. In the next section, we present two real-system implementations that reveal and address these practical issues. 9 System Evaluation In this section, we present a system implementation of MSP on two physical test-beds. The first one is called Mirage, a large indoor test-bed composed of six 4-foot by 8-foot boards, illustrated in Figure 19. Each board in the system can be used as an individual sub-system, which is powered, controlled and metered separately. Three Hitachi CP-X1250 projectors, connected through a Matorx Triplehead2go graphics expansion box, are used to create an ultra-wide integrated display on six boards. Figure 19 shows that a long tilted line is generated by the projectors. We have implemented all five versions of MSP on the Mirage test-bed, running 46 MICAz motes. Unless mentioned otherwise, the default setting is 3 anchors and 6 scans at the scanning line speed of 8.6 feet/s. In all of our graphs, each data point represents the average value of 50 trials. In the outdoor system, a Dell A525 speaker is used to generate 4.7KHz sound as shown in Figure 20. We place 20 MICAz motes in the backyard of a house. Since the location is not completely open, sound waves are reflected, scattered and absorbed by various objects in the vicinity, causing a multi-path effect. In the system evaluation, simple time synchronization mechanisms are applied on each node. 9.1 Indoor System Evaluation During indoor experiments, we encountered several realworld problems that are not revealed in the simulation. First, sequences obtained were partial due to misdetection and message losses. Second, elements in the sequences could flip due to detection delay, uncertainty in media access, or error in time synchronization. We show that these issues can be addressed by using the protection band method described in Section 7.3. 9.1.1 On Scanning Speed and Protection Band In this experiment, we studied the impact of the scanning speed and the length of protection band on the performance of the system. In general, with increasing scanning speed, nodes have less time to respond to the event and the time gap between two adjacent nodes shrinks, leading to an increasing number of partial sequences and sequence flips. Figure 21 shows the node flip situations for six scans with distinct angles under different scan speeds. The x-axis shows the distance between the flipped nodes in the correct node sequence. y-axis shows the total number of flips in the six scans. This figure tells us that faster scan brings in not only increasing number of flips, but also longer-distance flips that require wider protection band to prevent from fatal errors. Figure 22(a) shows the effectiveness of the protection band in terms of reducing the number of unlocalized nodes. When we use a moderate scan speed (4.3feet/s), the chance of flipping is rare, therefore we can achieve 0.45 feet mean accuracy (Figure 22(b)) with 1.6 feet maximum error (Figure 22(c)). With increasing speeds, the protection band needs to be set to a larger value to deal with flipping. Interesting phenomena can be observed in Figures 22: on one hand, the protection band can sharply reduce the number of unlocalized nodes; on the other hand, protection bands enlarge the area in which a target would potentially reside, introducing more uncertainty. Thus there is a concave curve for both mean and maximum error when the scan speed is at 8.6 feet/s. 9.1.2 On MSP Methods and Protection Band In this experiment, we show the improvements resulting from three different methods. Figure 23(a) shows that a protection band of 0.35 feet is sufficient for the scan speed of 8.57feet/s. Figures 23(b) and 23(c) show clearly that iterative MSP (with adaptation) achieves best performance. For example, Figures 23(b) shows that when we set the protection band at 0.05 feet, iterative MSP achieves 0.7 feet accuracy, which is 42% more accurate than the basic design. Similarly, Figures 23(b) and 23(c) show the double-edged effects of protection band on the localization accuracy. 0 5 10 15 20 0 20 40 (3) Flip Distribution for 6 Scans at Line Speed of 14.6feet/s Flips Node Distance in the Ideal Node Sequence 0 5 10 15 20 0 20 40 (2) Flip Distribution for 6 Scans at Line Speed of 8.6feet/s Flips 0 5 10 15 20 0 20 40 (1) Flip Distribution for 6 Scans at Line Speed of 4.3feet/s Flips Figure 21. Number of Flips for Different Scan Speeds 26 0 0.2 0.4 0.6 0.8 1 0 2 4 6 8 10 12 14 16 18 20 Unlocalized Node Number(Line Scan at Different Speed) Protection Band (in feet) UnlocalizedNodeNumber Scan Line Speed: 14.6feet/s Scan Line Speed: 8.6feet/s Scan Line Speed: 4.3feet/s (a) Number of Unlocalized Nodes 0 0.2 0.4 0.6 0.8 1 0.4 0.5 0.6 0.7 0.8 0.9 1 1.1 Mean Error(Line Scan at Different Speed) Protection Band (in feet) Error(infeet) Scan Line Speed:14.6feet/s Scan Line Speed: 8.6feet/s Scan Line Speed: 4.3feet/s (b) Mean Localization Error 0 0.2 0.4 0.6 0.8 1 1.5 2 2.5 3 3.5 4 Max Error(Line Scan at Different Speed) Protection Band (in feet) Error(infeet) Scan Line Speed: 14.6feet/s Scan Line Speed: 8.6feet/s Scan Line Speed: 4.3feet/s (c) Max Localization Error Figure 22. Impact of Protection Band and Scanning Speed 0 0.2 0.4 0.6 0.8 1 0 2 4 6 8 10 12 14 16 18 20 Unlocalized Node Number(Scan Line Speed 8.57feet/s) Protection Band (in feet) Numberofunlocalizednodeoutof46 Unlocalized node of Basic MSP Unlocalized node of Sequence Based MSP Unlocalized node of Iterative MSP (a) Number of Unlocalized Nodes 0 0.2 0.4 0.6 0.8 1 0.7 0.8 0.9 1 1.1 1.2 1.3 1.4 Mean Error(Scan Line Speed 8.57feet/s) Protection Band (in feet) Error(infeet) Mean Error of Basic MSP Mean Error of Sequence Based MSP Mean Error of Iterative MSP (b) Mean Localization Error 0 0.2 0.4 0.6 0.8 1 1.5 2 2.5 3 3.5 4 Max Error(Scan Line Speed 8.57feet/s) Protection Band (in feet) Error(infeet) Max Error of Basic MSP Max Error of Sequence Based MSP Max Error of Iterative MSP (c) Max Localization Error Figure 23. Impact of Protection Band under Different MSP Methods 3 4 5 6 7 8 9 10 11 0 0.5 1 1.5 2 2.5 Unlocalized Node Number(Protection Band: 0.35 feet) Anchor Number UnlocalizedNodeNumber 4 Scan Events at Speed 8.75feet/s 6 Scan Events at Speed 8.75feet/s 8 Scan Events at Speed 8.75feet/s (a) Number of Unlocalized Nodes 3 4 5 6 7 8 9 10 11 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 Mean Error(Protection Band: 0.35 feet) Anchor Number Error(infeet) Mean Error of 4 Scan Events at Speed 8.75feet/s Mean Error of 6 Scan Events at Speed 8.75feet/s Mean Error of 8 Scan Events at Speed 8.75feet/s (b) Mean Localization Error 3 4 5 6 7 8 9 10 11 0.8 1 1.2 1.4 1.6 1.8 2 2.2 2.4 2.6 2.8 Max Error(Protection Band: 0.35 feet) Anchor Number Error(infeet) Max Error of 4 Scan Events at Speed 8.75feet/s Max Error of 6 Scan Events at Speed 8.75feet/s Max Error of 8 Scan Events at Speed 8.75feet/s (c) Max Localization Error Figure 24. Impact of the Number of Anchors and Scans 9.1.3 On Number of Anchors and Scans In this experiment, we show a tradeoff between hardware cost (anchors) with soft cost (events). Figure 24(a) shows that with more cutting lines created by anchors, the chance of unlocalized nodes increases slightly. We note that with a 0.35 feet protection band, the percentage of unlocalized nodes is very small, e.g., in the worst-case with 11 anchors, only 2 out of 46 nodes are not localized due to flipping. Figures 24(b) and 24(c) show the tradeoff between number of anchors and the number of scans. Obviously, with the number of anchors increases, the error drops significantly. With 11 anchors we can achieve a localization accuracy as low as 0.25 ∼ 0.35 feet, which is nearly a 60% improvement. Similarly, with increasing number of scans, the accuracy drops significantly as well. We can observe about 30% across all anchor settings when we increase the number of scans from 4 to 8. For example, with only 3 anchors, we can achieve 0.6-foot accuracy with 8 scans. 9.2 Outdoor System Evaluation The outdoor system evaluation contains two parts: (i) effective detection distance evaluation, which shows that the node sequence can be readily obtained, and (ii) sound propagation based localization, which shows the results of wavepropagation-based localization. 9.2.1 Effective Detection Distance Evaluation We firstly evaluate the sequence flip phenomenon in wave propagation. As shown in Figure 25, 20 motes were placed as five groups in front of the speaker, four nodes in each group at roughly the same distances to the speaker. The gap between each group is set to be 2, 3, 4 and 5 feet respectively in four experiments. Figure 26 shows the results. The x-axis in each subgraph indicates the group index. There are four nodes in each group (4 bars). The y-axis shows the detection rank (order) of each node in the node sequence. As distance between each group increases, number of flips in the resulting node sequence 27 Figure 25. Wave Detection 1 2 3 4 5 0 5 10 15 20 2 feet group distance Rank Group Index 1 2 3 4 5 0 5 10 15 20 3 feet group distance Rank Group Index 1 2 3 4 5 0 5 10 15 20 4 feet group distance Rank Group Index 1 2 3 4 5 0 5 10 15 20 5 feet group distance Rank Group Index Figure 26. Ranks vs. Distances 0 2 4 6 8 10 12 14 16 18 20 22 24 0 2 4 6 8 10 12 14 Y-Dimension(feet) X-Dimension (feet) Node 0 2 4 6 8 10 12 14 16 18 20 22 24 0 2 4 6 8 10 12 14 Y-Dimension(feet) X-Dimension (feet) Anchor Figure 27. Localization Error (Sound) decreases. For example, in the 2-foot distance subgraph, there are quite a few flips between nodes in adjacent and even nonadjacent groups, while in the 5-foot subgraph, flips between different groups disappeared in the test. 9.2.2 Sound Propagation Based Localization As shown in Figure 20, 20 motes are placed as a grid including 5 rows with 5 feet between each row and 4 columns with 4 feet between each column. Six 4KHz acoustic wave propagation events are generated around the mote grid by a speaker. Figure 27 shows the localization results using iterative MSP (3 times iterative processing) with a protection band of 3 feet. The average error of the localization results is 3 feet and the maximum error is 5 feet with one un-localized node. We found that sequence flip in wave propagation is more severe than that in the indoor, line-based test. This is expected due to the high propagation speed of sound. Currently we use MICAz mote, which is equipped with a low quality microphone. We believe that using a better speaker and more events, the system can yield better accuracy. Despite the hardware constrains, the MSP algorithm still successfully localized most of the nodes with good accuracy. 10 Conclusions In this paper, we present the first work that exploits the concept of node sequence processing to localize sensor nodes. We demonstrated that we could significantly improve localization accuracy by making full use of the information embedded in multiple easy-to-get one-dimensional node sequences. We proposed four novel optimization methods, exploiting order and marginal distribution among non-anchor nodes as well as the feedback information from early localization results. Importantly, these optimization methods can be used together, and improve accuracy additively. The practical issues of partial node sequence and sequence flip were identified and addressed in two physical system test-beds. We also evaluated performance at scale through analysis as well as extensive simulations. Results demonstrate that requiring neither costly hardware on sensor nodes nor precise event distribution, MSP can achieve a sub-foot accuracy with very few anchor nodes provided sufficient events. 11 References [1] CC2420 Data Sheet. Avaiable at http://www.chipcon.com/. [2] P. Bahl and V. N. Padmanabhan. Radar: An In-Building RF-Based User Location and Tracking System. In IEEE Infocom "00. [3] M. Broxton, J. Lifton, and J. Paradiso. Localizing A Sensor Network via Collaborative Processing of Global Stimuli. In EWSN "05. [4] N. Bulusu, J. Heidemann, and D. Estrin. GPS-Less Low Cost Outdoor Localization for Very Small Devices. IEEE Personal Communications Magazine, 7(4), 2000. [5] D. Culler, D. Estrin, and M. Srivastava. Overview of Sensor Networks. IEEE Computer Magazine, 2004. [6] J. Elson, L. Girod, and D. Estrin. Fine-Grained Network Time Synchronization Using Reference Broadcasts. In OSDI "02. [7] D. K. Goldenberg, P. Bihler, M. Gao, J. Fang, B. D. Anderson, A. Morse, and Y. Yang. Localization in Sparse Networks Using Sweeps. In MobiCom "06. [8] T. He, C. Huang, B. M. Blum, J. A. Stankovic, and T. Abdelzaher. RangeFree Localization Schemes in Large-Scale Sensor Networks. In MobiCom "03. [9] B. Kusy, P. Dutta, P. Levis, M. Mar, A. Ledeczi, and D. Culler. Elapsed Time on Arrival: A Simple and Versatile Primitive for Canonical Time Synchronization Services. International Journal of ad-hoc and Ubiquitous Computing, 2(1), 2006. [10] L. Lazos and R. Poovendran. SeRLoc: Secure Range-Independent Localization for Wireless Sensor Networks. In WiSe "04. [11] M. Maroti, B. Kusy, G. Balogh, P. Volgyesi, A. Nadas, K. Molnar, S. Dora, and A. Ledeczi. Radio Interferometric Geolocation. In SenSys "05. [12] M. Maroti, B. Kusy, G. Simon, and A. Ledeczi. The Flooding Time Synchronization Protocol. In SenSys "04. [13] D. Moore, J. Leonard, D. Rus, and S. Teller. Robust Distributed Network Localization with Noise Range Measurements. In SenSys "04. [14] R. Nagpal and D. Coore. An Algorithm for Group Formation in an Amorphous Computer. In PDCS "98. [15] D. Niculescu and B. Nath. ad-hoc Positioning System. In GlobeCom "01. [16] D. Niculescu and B. Nath. ad-hoc Positioning System (APS) Using AOA. In InfoCom "03. [17] N. B. Priyantha, A. Chakraborty, and H. Balakrishnan. The Cricket Location-Support System. In MobiCom "00. [18] K. R¨omer. The Lighthouse Location System for Smart Dust. In MobiSys "03. [19] A. Savvides, C. C. Han, and M. B. Srivastava. Dynamic Fine-Grained Localization in ad-hoc Networks of Sensors. In MobiCom "01. [20] R. Stoleru, T. He, J. A. Stankovic, and D. Luebke. A High-Accuracy, Low-Cost Localization System for Wireless Sensor Networks. In SenSys "05. [21] R. Stoleru, P. Vicaire, T. He, and J. A. Stankovic. StarDust: a Flexible Architecture for Passive Localization in Wireless Sensor Networks. In SenSys "06. [22] E. W. Weisstein. Plane Division by Lines. mathworld.wolfram.com. [23] B. H. Wellenhoff, H. Lichtenegger, and J. Collins. Global Positions System: Theory and Practice,Fourth Edition. Springer Verlag, 1997. [24] K. Whitehouse. The Design of Calamari: an ad-hoc Localization System for Sensor Networks. In University of California at Berkeley, 2002. [25] Z. Zhong. MSP Evaluation and Implementation Report. Avaiable at http://www.cs.umn.edu/∼zhong/MSP. [26] G. Zhou, T. He, and J. A. Stankovic. Impact of Radio Irregularity on Wireless Sensor Networks. In MobiSys "04. 28

marginal distribution;node localization;multi-sequence positioning;listen-detect-assemble-report protocol;event distribution;range-based approach;spatiotemporal correlation;localization;node sequence process;distribution-based location estimation;massive uva-based deploment;wireless sensor network

train_C-45

StarDust: A Flexible Architecture for Passive Localization in Wireless Sensor Networks

The problem of localization in wireless sensor networks where nodes do not use ranging hardware, remains a challenging problem, when considering the required location accuracy, energy expenditure and the duration of the localization phase. In this paper we propose a framework, called StarDust, for wireless sensor network localization based on passive optical components. In the StarDust framework, sensor nodes are equipped with optical retro-reflectors. An aerial device projects light towards the deployed sensor network, and records an image of the reflected light. An image processing algorithm is developed for obtaining the locations of sensor nodes. For matching a node ID to a location we propose a constraint-based label relaxation algorithm. We propose and develop localization techniques based on four types of constraints: node color, neighbor information, deployment time for a node and deployment location for a node. We evaluate the performance of a localization system based on our framework by localizing a network of 26 sensor nodes deployed in a 120 × 60 ft2 area. The localization accuracy ranges from 2 ft to 5 ft while the localization time ranges from 10 milliseconds to 2 minutes.

1 Introduction Wireless Sensor Networks (WSN) have been envisioned to revolutionize the way humans perceive and interact with the surrounding environment. One vision is to embed tiny sensor devices in outdoor environments, by aerial deployments from unmanned air vehicles. The sensor nodes form a network and collaborate (to compensate for the extremely scarce resources available to each of them: computational power, memory size, communication capabilities) to accomplish the mission. Through collaboration, redundancy and fault tolerance, the WSN is then able to achieve unprecedented sensing capabilities. A major step forward has been accomplished by developing systems for several domains: military surveillance [1] [2] [3], habitat monitoring [4] and structural monitoring [5]. Even after these successes, several research problems remain open. Among these open problems is sensor node localization, i.e., how to find the physical position of each sensor node. Despite the attention the localization problem in WSN has received, no universally acceptable solution has been developed. There are several reasons for this. On one hand, localization schemes that use ranging are typically high end solutions. GPS ranging hardware consumes energy, it is relatively expensive (if high accuracy is required) and poses form factor challenges that move us away from the vision of dust size sensor nodes. Ultrasound has a short range and is highly directional. Solutions that use the radio transceiver for ranging either have not produced encouraging results (if the received signal strength indicator is used) or are sensitive to environment (e.g., multipath). On the other hand, localization schemes that only use the connectivity information for inferring location information are characterized by low accuracies: ≈ 10 ft in controlled environments, 40−50 ft in realistic ones. To address these challenges, we propose a framework for WSN localization, called StarDust, in which the complexity associated with the node localization is completely removed from the sensor node. The basic principle of the framework is localization through passivity: each sensor node is equipped with a corner-cube retro-reflector and possibly an optical filter (a coloring device). An aerial vehicle projects light onto the deployment area and records images containing retro-reflected light beams (they appear as luminous spots). Through image processing techniques, the locations of the retro-reflectors (i.e., sensor nodes) is deter57 mined. For inferring the identity of the sensor node present at a particular location, the StarDust framework develops a constraint-based node ID relaxation algorithm. The main contributions of our work are the following. We propose a novel framework for node localization in WSNs that is very promising and allows for many future extensions and more accurate results. We propose a constraint-based label relaxation algorithm for mapping node IDs to the locations, and four constraints (node, connectivity, time and space), which are building blocks for very accurate and very fast localization systems. We develop a sensor node hardware prototype, called a SensorBall. We evaluate the performance of a localization system for which we obtain location accuracies of 2 − 5 ft with a localization duration ranging from 10 milliseconds to 2 minutes. We investigate the range of a system built on our framework by considering realities of physical phenomena that occurs during light propagation through the atmosphere. The rest of the paper is structured as follows. Section 2 is an overview of the state of art. The design of the StarDust framework is presented in Section 3. One implementation and its performance evaluation are in Sections 4 and 5, followed by a suite of system optimization techniques, in Section 6. In Section 7 we present our conclusions. 2 Related Work We present the prior work in localization in two major categories: the range-based, and the range-free schemes. The range-based localization techniques have been designed to use either more expensive hardware (and hence higher accuracy) or just the radio transceiver. Ranging techniques dependent on hardware are the time-of-flight (ToF) and the time-difference-of-arrival(TDoA). Solutions that use the radio are based on the received signal strength indicator (RSSI) and more recently on radio interferometry. The ToF localization technique that is most widely used is the GPS. GPS is a costly solution for a high accuracy localization of a large scale sensor network. AHLoS [6] employs a TDoA ranging technique that requires extensive hardware and solves relatively large nonlinear systems of equations. The Cricket location-support system (TDoA) [7] can achieve a location granularity of tens of inches with highly directional and short range ultrasound transceivers. In [2] the location of a sniper is determined in an urban terrain, by using the TDoA between an acoustic wave and a radio beacon. The PushPin project [8] uses the TDoA between ultrasound pulses and light flashes for node localization. The RADAR system [9] uses the RSSI to build a map of signal strengths as emitted by a set of beacon nodes. A mobile node is located by the best match, in the signal strength space, with a previously acquired signature. In MAL [10], a mobile node assists in measuring the distances (acting as constraints) between nodes until a rigid graph is generated. The localization problem is formulated as an on-line state estimation in a nonlinear dynamic system [11]. A cooperative ranging that attempts to achieve a global positioning from distributed local optimizations is proposed in [12]. A very recent, remarkable, localization technique is based on radio interferometry, RIPS [13], which utilizes two transmitters to create an interfering signal. The frequencies of the emitters are very close to each other, thus the interfering signal will have a low frequency envelope that can be easily measured. The ranging technique performs very well. The long time required for localization and multi-path environments pose significant challenges. Real environments create additional challenges for the range based localization schemes. These have been emphasized by several studies [14] [15] [16]. To address these challenges, and others (hardware cost, the energy expenditure, the form factor, the small range, localization time), several range-free localization schemes have been proposed. Sensor nodes use primarily connectivity information for inferring proximity to a set of anchors. In the Centroid localization scheme [17], a sensor node localizes to the centroid of its proximate beacon nodes. In APIT [18] each node decides its position based on the possibility of being inside or outside of a triangle formed by any three beacons within node"s communication range. The Gradient algorithm [19], leverages the knowledge about the network density to infer the average one hop length. This, in turn, can be transformed into distances to nodes with known locations. DV-Hop [20] uses the hop by hop propagation capability of the network to forward distances to landmarks. More recently, several localization schemes that exploit the sensing capabilities of sensor nodes, have been proposed. Spotlight [21] creates well controlled (in time and space) events in the network while the sensor nodes detect and timestamp this events. From the spatiotemporal knowledge for the created events and the temporal information provided by sensor nodes, nodes" spatial information can be obtained. In a similar manner, the Lighthouse system [22] uses a parallel light beam, that is emitted by an anchor which rotates with a certain period. A sensor node detects the light beam for a period of time, which is dependent on the distance between it and the light emitting device. Many of the above localization solutions target specific sets of requirements and are useful for specific applications. StarDust differs in that it addresses a particular demanding set of requirements that are not yet solved well. StarDust is meant for localizing air dropped nodes where node passiveness, high accuracy, low cost, small form factor and rapid localization are all required. Many military applications have such requirements. 3 StarDust System Design The design of the StarDust system (and its name) was inspired by the similarity between a deployed sensor network, in which sensor nodes indicate their presence by emitting light, and the Universe consisting of luminous and illuminated objects: stars, galaxies, planets, etc. The main difficulty when applying the above ideas to the real world is the complexity of the hardware that needs to be put on a sensor node so that the emitted light can be detected from thousands of feet. The energy expenditure for producing an intense enough light beam is also prohibitive. Instead, what we propose to use for sensor node localization is a passive optical element called a retro-reflector. The most common retro-reflective optical component is a Corner-Cube Retroreflector (CCR), shown in Figure 1(a). It consists of three mutually perpendicular mirrors. The inter58 (a) (b) Figure 1. Corner-Cube Retroreflector (a) and an array of CCRs molded in plastic (b) esting property of this optical component is that an incoming beam of light is reflected back, towards the source of the light, irrespective of the angle of incidence. This is in contrast with a mirror, which needs to be precisely positioned to be perpendicular to the incident light. A very common and inexpensive implementation of an array of CCRs is the retroreflective plastic material used on cars and bicycles for night time detection, shown in Figure 1(b). In the StarDust system, each node is equipped with a small (e.g. 0.5in2) array of CCRs and the enclosure has self-righting capabilities that orient the array of CCRs predominantly upwards. It is critical to understand that the upward orientation does not need to be exact. Even when large angular variations from a perfectly upward orientation are present, a CCR will return the light in the exact same direction from which it came. In the remaining part of the section, we present the architecture of the StarDust system and the design of its main components. 3.1 System Architecture The envisioned sensor network localization scenario is as follows: • The sensor nodes are released, possibly in a controlled manner, from an aerial vehicle during the night. • The aerial vehicle hovers over the deployment area and uses a strobe light to illuminate it. The sensor nodes, equipped with CCRs and optical filters (acting as coloring devices) have self-righting capabilities and retroreflect the incoming strobe light. The retro-reflected light is either white, as the originating source light, or colored, due to optical filters. • The aerial vehicle records a sequence of two images very close in time (msec level). One image is taken when the strobe light is on, the other when the strobe light is off. The acquired images are used for obtaining the locations of sensor nodes (which appear as luminous spots in the image). • The aerial vehicle executes the mapping of node IDs to the identified locations in one of the following ways: a) by using the color of a retro-reflected light, if a sensor node has a unique color; b) by requiring sensor nodes to establish neighborhood information and report it to a base station; c) by controlling the time sequence of sensor nodes deployment and recording additional imLight Emitter Sensor Node i Transfer Function Φi(λ) Ψ(λ) Φ(Ψ(λ)) Image Processing Node ID Matching Radio Model R G(Λ,E) Central Device V" V" Figure 2. The StarDust system architecture ages; d) by controlling the location where a sensor node is deployed. • The computed locations are disseminated to the sensor network. The architecture of the StarDust system is shown in Figure 2. The architecture consists of two main components: the first is centralized and it is located on a more powerful device. The second is distributed and it resides on all sensor nodes. The Central Device consists of the following: the Light Emitter, the Image Processing module, the Node ID Mapping module and the Radio Model. The distributed component of the architecture is the Transfer Function, which acts as a filter for the incoming light. The aforementioned modules are briefly described below: • Light Emitter - It is a strobe light, capable of producing very intense, collimated light pulses. The emitted light is non-monochromatic (unlike a laser) and it is characterized by a spectral density Ψ(λ), a function of the wavelength. The emitted light is incident on the CCRs present on sensor nodes. • Transfer Function Φ(Ψ(λ)) - This is a bandpass filter for the incident light on the CCR. The filter allows a portion of the original spectrum, to be retro-reflected. From here on, we will refer to the transfer function as the color of a sensor node. • Image Processing - The Image Processing module acquires high resolution images. From these images the locations and the colors of sensor nodes are obtained. If only one set of pictures can be taken (i.e., one location of the light emitter/image analysis device), then the map of the field is assumed to be known as well as the distance between the imaging device and the field. The aforementioned assumptions (field map and distance to it) are not necessary if the images can be simultaneously taken from different locations. It is important to remark here that the identity of a node can not be directly obtained through Image Processing alone, unless a specific characteristic of a sensor node can be identified in the image. • Node ID Matching - This module uses the detected locations and through additional techniques (e.g., sensor node coloring and connectivity information (G(Λ,E)) from the deployed network) to uniquely identify the sensor nodes observed in the image. The connectivity information is represented by neighbor tables sent from 59 Algorithm 1 Image Processing 1: Background filtering 2: Retro-reflected light recognition through intensity filtering 3: Edge detection to obtain the location of sensor nodes 4: Color identification for each detected sensor node each sensor node to the Central Device. • Radio Model - This component provides an estimate of the radio range to the Node ID Matching module. It is only used by node ID matching techniques that are based on the radio connectivity in the network. The estimate of the radio range R is based on the sensor node density (obtained through the Image Processing module) and the connectivity information (i.e., G(Λ,E)). The two main components of the StarDust architecture are the Image Processing and the Node ID Mapping. Their design and analysis is presented in the sections that follow. 3.2 Image Processing The goal of the Image Processing Algorithm (IPA) is to identify the location of the nodes and their color. Note that IPA does not identify which node fell where, but only what is the set of locations where the nodes fell. IPA is executed after an aerial vehicle records two pictures: one in which the field of deployment is illuminated and one when no illuminations is present. Let Pdark be the picture of the deployment area, taken when no light was emitted and Plight be the picture of the same deployment area when a strong light beam was directed towards the sensor nodes. The proposed IPA has several steps, as shown in Algorithm 1. The first step is to obtain a third picture Pfilter where only the differences between Pdark and Plight remain. Let us assume that Pdark has a resolution of n × m, where n is the number of pixels in a row of the picture, while m is the number of pixels in a column of the picture. Then Pdark is composed of n × m pixels noted Pdark(i, j), i ∈ 1 ≤ i ≤ n,1 ≤ j ≤ m. Similarly Plight is composed of n × m pixels noted Plight(i, j), 1 ≤ i ≤ n,1 ≤ j ≤ m. Each pixel P is described by an RGB value where the R value is denoted by PR, the G value is denoted by PG, and the B value is denoted by PB. IPA then generates the third picture, Pfilter, through the following transformations: PR filter(i, j) = PR light(i, j)−PR dark(i, j) PG filter(i, j) = PG light(i, j)−PG dark(i, j) PB filter(i, j) = PB light(i, j)−PB dark(i, j) (1) After this transformation, all the features that appeared in both Pdark and Plight are removed from Pfilter. This simplifies the recognition of light retro-reflected by sensor nodes. The second step consists of identifying the elements contained in Pfilter that retro-reflect light. For this, an intensity filter is applied to Pfilter. First IPA converts Pfilter into a grayscale picture. Then the brightest pixels are identified and used to create Preflect. This step is eased by the fact that the reflecting nodes should appear much brighter than any other illuminated object in the picture. Support: Q(λk) ni P1 ... P2 ... PN λ1 ... λk ... λN Figure 3. Probabilistic label relaxation The third step runs an edge detection algorithm on Preflect to identify the boundary of the nodes present. A tool such as Matlab provides a number of edge detection techniques. We used the bwboundaries function. For the obtained edges, the location (x,y) (in the image) of each node is determined by computing the centroid of the points constituting its edges. Standard computer graphics techniques [23] are then used to transform the 2D locations of sensor nodes detected in multiple images into 3D sensor node locations. The color of the node is obtained as the color of the pixel located at (x,y) in Plight. 3.3 Node ID Matching The goal of the Node ID Matching module is to obtain the identity (node ID) of a luminous spot in the image, detected to be a sensor node. For this, we define V = {(x1,y1),(x2,y2),...,(xm,ym)} to be the set of locations of the sensor nodes, as detected by the Image Processing module and Λ = {λ1,λ2,...,λm} to be the set of unique node IDs assigned to the m sensor nodes, before deployment. From here on, we refer to node IDs as labels. We model the problem of finding the label λj of a node ni as a probabilistic label relaxation problem, frequently used in image processing/understanding. In the image processing domain, scene labeling (i.e., identifying objects in an image) plays a major role. The goal of scene labeling is to assign a label to each object detected in an image, such that an appropriate image interpretation is achieved. It is prohibitively expensive to consider the interactions among all the objects in an image. Instead, constraints placed among nearby objects generate local consistencies and through iteration, global consistencies can be obtained. The main idea of the sensor node localization through probabilistic label relaxation is to iteratively compute the probability of each label being the correct label for a sensor node, by taking into account, at each iteration, the support for a label. The support for a label can be understood as a hint or proof, that a particular label is more likely to be the correct one, when compared with the other potential labels for a sensor node. We pictorially depict this main idea in Figure 3. As shown, node ni has a set of candidate labels {λ1,...,λk}. Each of the labels has a different value for the Support function Q(λk). We defer the explanation of how the Support function is implemented until the subsections that follow, where we provide four concrete techniques. Formally, the algorithm is outlined in Algorithm 2, where the equations necessary for computing the new probability Pni(λk) for a label λk of a node ni, are expressed by the 60 Algorithm 2 Label Relaxation 1: for each sensor node ni do 2: assign equal prob. to all possible labels 3: end for 4: repeat 5: converged ← true 6: for each sensor node ni do 7: for each each label λj of ni do 8: compute the Support label λj: Equation 4 9: end for 10: compute K for the node ni: Equation 3 11: for each each label λj do 12: update probability of label λj: Equation 2 13: if |new prob.−old prob.| ≥ ε then 14: converged ← false 15: end if 16: end for 17: end for 18: until converged = true following equations: Ps+1 ni (λk) = 1 Kni Ps ni (λk)Qs ni (λk) (2) where Kni is a normalizing constant, given by: Kni = N ∑ k=1 Ps ni (λk)Qs ni (λk) (3) and Qs ni (λk) is: Qs ni (λk) = support for label λk of node ni (4) The label relaxation algorithm is iterative and it is polynomial in the size of the network(number of nodes). The pseudo-code is shown in Algorithm 2. It initializes the probabilities associated with each possible label, for a node ni, through a uniform distribution. At each iteration s, the algorithm updates the probability associated with each label, by considering the Support Qs ni (λk) for each candidate label of a sensor node. In the sections that follow, we describe four different techniques for implementing the Support function: based on node coloring, radio connectivity, the time of deployment (time) and the location of deployment (space). While some of these techniques are simplistic, they are primitives which, when combined, can create powerful localization systems. These design techniques have different trade-offs, which we will present in Section 3.3.6. 3.3.1 Relaxation with Color Constraints The unique mapping between a sensor node"s position (identified by the image processing) and a label can be obtained by assigning a unique color to each sensor node. For this we define C = {c1,c2,...,cn} to be the set of unique colors available and M : Λ → C to be a one-to-one mapping of labels to colors. This mapping is known prior to the sensor node deployment (from node manufacturing). In the case of color constrained label relaxation, the support for label λk is expressed as follows: Qs ni (λk) = 1 (5) As a result, the label relaxation algorithm (Algorithm 2) consists of the following steps: one label is assigned to each sensor node (lines 1-3 of the algorithm), implicitly having a probability Pni(λk) = 1 ; the algorithm executes a single iteration, when the support function, simply, reiterates the confidence in the unique labeling. However, it is often the case that unique colors for each node will not be available. It is interesting to discuss here the influence that the size of the coloring space (i.e., |C|) has on the accuracy of the localization algorithm. Several cases are discussed below: • If |C| = 0, no colors are used and the sensor nodes are equipped with simple CCRs that reflect back all the incoming light (i.e., no filtering, and no coloring of the incoming light). From the image processing system, the position of sensor nodes can still be obtained. Since all nodes appear white, no single sensor node can be uniquely identified. • If |C| = m − 1 then there are enough unique colors for all nodes (one node remains white, i.e. no coloring), the problem is trivially solved. Each node can be identified, based on its unique color. This is the scenario for the relaxation with color constraints. • If |C| ≥ 1, there are several options for how to partition the coloring space. If C = {c1} one possibility is to assign the color c1 to a single node, and leave the remaining m−1 sensor nodes white, or to assign the color c1 to more than one sensor node. One can observe that once a color is assigned uniquely to a sensor node, in effect, that sensor node is given the status of anchor, or node with known location. It is interesting to observe that there is an entire spectrum of possibilities for how to partition the set of sensor nodes in equivalence classes (where an equivalence class is represented by one color), in order to maximize the success of the localization algorithm. One of the goals of this paper is to understand how the size of the coloring space and its partitioning affect localization accuracy. Despite the simplicity of this method of constraining the set of labels that can be assigned to a node, we will show that this technique is very powerful, when combined with other relaxation techniques. 3.3.2 Relaxation with Connectivity Constraints Connectivity information, obtained from the sensor network through beaconing, can provide additional information for locating sensor nodes. In order to gather connectivity information, the following need to occur: 1) after deployment, through beaconing of HELLO messages, sensor nodes build their neighborhood tables; 2) each node sends its neighbor table information to the Central device via a base station. First, let us define G = (Λ,E) to be the weighted connectivity graph built by the Central device from the received neighbor table information. In G the edge (λi,λj) has a 61 λ1 λ2 ... λN ni nj gi2,j2 λ1 λ2 ... λN Pj,λ1 Pj,λ2 ... Pj,λN Pi,λ1 Pi,λ1 ... Pi,λN gi2,jm Figure 4. Label relaxation with connectivity constraints weight gij represented by the number of beacons sent by λj and received by λi. In addition, let R be the radio range of the sensor nodes. The main idea of the connectivity constrained label relaxation is depicted in Figure 4 in which two nodes ni and nj have been assigned all possible labels. The confidence in each of the candidate labels for a sensor node, is represented by a probability, shown in a dotted rectangle. It is important to remark that through beaconing and the reporting of neighbor tables to the Central Device, a global view of all constraints in the network can be obtained. It is critical to observe that these constraints are among labels. As shown in Figure 4 two constraints exist between nodes ni and nj. The constraints are depicted by gi2,j2 and gi2,jM, the number of beacons sent the labels λj2 and λjM and received by the label λi2. The support for the label λk of sensor node ni, resulting from the interaction (i.e., within radio range) with sensor node nj is given by: Qs ni (λk) = M ∑ m=1 gλkλm Ps nj (λm) (6) As a result, the localization algorithm (Algorithm 3 consists of the following steps: all labels are assigned to each sensor node (lines 1-3 of the algorithm), and implicitly each label has a probability initialized to Pni(λk) = 1/|Λ|; in each iteration, the probabilities for the labels of a sensor node are updated, when considering the interaction with the labels of sensor nodes within R. It is important to remark that the identity of the nodes within R is not known, only the candidate labels and their probabilities. The relaxation algorithm converges when, during an iteration, the probability of no label is updated by more than ε. The label relaxation algorithm based on connectivity constraints, enforces such constraints between pairs of sensor nodes. For a large scale sensor network deployment, it is not feasible to consider all pairs of sensor nodes in the network. Hence, the algorithm should only consider pairs of sensor nodes that are within a reasonable communication range (R). We assume a circular radio range and a symmetric connectivity. In the remaining part of the section we propose a simple analytical model that estimates the radio range R for medium-connected networks (less than 20 neighbors per R). We consider the following to be known: the size of the deployment field (L), the number of sensor nodes deployed (N) Algorithm 3 Localization 1: Estimate the radio range R 2: Execute the Label Relaxation Algorithm with Support Function given by Equation 6 for neighbors less than R apart 3: for each sensor node ni do 4: node identity is λk with max. prob. 5: end for and the total number of unidirectional (i.e., not symmetric) one-hop radio connections in the network (k). For our analysis, we uniformly distribute the sensor nodes in a square area of length L, by using a grid of unit length L/ √ N. We use the substitution u = L/ √ N to simplify the notation, in order to distinguish the following cases: if u ≤ R ≤ √ 2u each node has four neighbors (the expected k = 4N); if √ 2u ≤ R ≤ 2u each node has eight neighbors (the expected k = 8N); if 2u ≤ R ≤ √ 5u each node has twelve neighbors ( the expected k = 12N); if √ 5u ≤ R ≤ 3u each node has twenty neighbors (the expected k = 20N) For a given t = k/4N we take R to be the middle of the interval. As an example, if t = 5 then R = (3 + √ 5)u/2. A quadratic fitting for R over the possible values of t, produces the following closed-form solution for the communication range R, as a function of network connectivity k, assuming L and N constant: R(k) = L √ N −0.051 k 4N 2 +0.66 k 4N +0.6 (7) We investigate the accuracy of our model in Section 5.2.1. 3.3.3 Relaxation with Time Constraints Time constraints can be treated similarly with color constraints. The unique identification of a sensor node can be obtained by deploying sensor nodes individually, one by one, and recording a sequence of images. The sensor node that is identified as new in the last picture (it was not identified in the picture before last) must be the last sensor node dropped. In a similar manner with color constrained label relaxation, the time constrained approach is very simple, but may take too long, especially for large scale systems. While it can be used in practice, it is unlikely that only a time constrained label relaxation is used. As we will see, by combining constrained-based primitives, realistic localization systems can be implemented. The support function for the label relaxation with time constraints is defined identically with the color constrained relaxation: Qs ni (λk) = 1 (8) The localization algorithm (Algorithm 2 consists of the following steps: one label is assigned to each sensor node (lines 1-3 of the algorithm), and implicitly having a probability Pni(λk) = 1 ; the algorithm executes a single iteration, 62 D1 D2 D4 D 3Node Label-1 Label-2 Label-3 Label-4 0.2 0.1 0.5 0.2 Figure 5. Relaxation with space constraints 0 0.2 0.4 0.6 0.8 1 1.2 1.4 0 1 2 3 4 5 6 7 8 PDF Distance D σ = 0.5 σ = 1 σ = 2 Figure 6. Probability distribution of distances -4 -3 -2 -1 0 1 2 3 4 X -4 -3 -2 -1 0 1 2 3 4 Y 0 0.2 0.4 0.6 0.8 1 Node Density Figure 7. Distribution of nodes when the support function, simply, reiterates the confidence in the unique labeling. 3.3.4 Relaxation with Space Constraints Spatial information related to sensor deployment can also be employed as another input to the label relaxation algorithm. To do that, we use two types of locations: the node location pn and the label location pl. The former pn is defined as the position of nodes (xn,yn,zn) after deployment, which can be obtained through Image Processing as mentioned in Section 3.3. The latter pl is defined as the location (xl,yl,zl) where a node is dropped. We use Dni λm to denote the horizontal distance between the location of the label λm and the location of the node ni. Clearly, Dni λm = (xn −xl)2 +(yn −yl)2. At the time of a sensor node release, the one-to-one mapping between the node and its label is known. In other words, the label location is the same as the node location at the release time. After release, the label location information is partially lost due to the random factors such as wind and surface impact. However, statistically, the node locations are correlated with label locations. Such correlation depends on the airdrop methods employed and environments. For the sake of simplicity, let"s assume nodes are dropped from the air through a helicopter hovering in the air. Wind can be decomposed into three components X,Y and Z. Only X and Y affect the horizontal distance a node can travel. According to [24], we can assume that X and Y follow an independent normal distribution. Therefore, the absolute value of the wind speed follows a Rayleigh distribution. Obviously the higher the wind speed is, the further a node would land away horizontally from the label location. If we assume that the distance D is a function of the wind speed V [25] [26], we can obtain the probability distribution of D under a given wind speed distribution. Without loss of generality, we assume that D is proportional to the wind speed. Therefore, D follows the Rayleigh distribution as well. As shown in Figure 5, the spatial-based relaxation is a recursive process to assign the probability that a nodes has a certain label by using the distances between the location of a node with multiple label locations. We note that the distribution of distance D affects the probability with which a label is assigned. It is not necessarily true that the nearest label is always chosen. For example, if D follows the Rayleigh(σ2) distribution, we can obtain the Probability Density Function (PDF) of distances as shown in Figure 6. This figure indicates that the possibility of a node to fall vertically is very small under windy conditions (σ > 0), and that the distance D is affected by the σ. The spatial distribution of nodes for σ = 1 is shown in Figure 7. Strong wind with a high σ value leads to a larger node dispersion. More formally, given a probability density function PDF(D), the support for label λk of sensor node ni can be formulated as: Qs ni (λk) = PDF(Dni λk ) (9) It is interesting to point out two special cases. First, if all nodes are released at once (i.e., only one label location for all released nodes), the distance D from a node to all labels is the same. In this case, Ps+1 ni (λk) = Ps ni (λk), which indicates that we can not use the spatial-based relaxation to recursively narrow down the potential labels for a node. Second, if nodes are released at different locations that are far away from each other, we have: (i) If node ni has label λk, Ps ni (λk) → 1 when s → ∞, (ii) If node ni does not have label λk, Ps ni (λk) → 0 when s → ∞. In this second scenario, there are multiple labels (one label per release), hence it is possible to correlate release times (labels) with positions on the ground. These results indicate that spatial-based relaxation can label the node with a very high probability if the physical separation among nodes is large. 3.3.5 Relaxation with Color and Connectivity Constraints One of the most interesting features of the StarDust architecture is that it allows for hybrid localization solutions to be built, depending on the system requirements. One example is a localization system that uses the color and connectivity constraints. In this scheme, the color constraints are used for reducing the number of candidate labels for sensor nodes, to a more manageable value. As a reminder, in the connectivity constrained relaxation, all labels are candidate labels for each sensor node. The color constraints are used in the initialization phase of Algorithm 3 (lines 1-3). After the initialization, the standard connectivity constrained relaxation algorithm is used. For a better understanding of how the label relaxation algorithm works, we give a concrete example, exemplified in Figure 8. In part (a) of the figure we depict the data structures 63 11 8 4 1 12 9 7 5 3 ni nj 12 8 10 11 10 0.2 0.2 0.2 0.2 0.2 0.25 0.25 0.25 0.25 (a) 11 8 4 1 12 9 7 5 3 ni nj 12 8 10 11 10 0.2 0.2 0.2 0.2 0.2 0.32 0 0.68 0 (b) Figure 8. A step through the algorithm. After initialization (a) and after the 1st iteration for node ni (b) associated with nodes ni and nj after the initialization steps of the algorithm (lines 1-6), as well as the number of beacons between different labels (as reported by the network, through G(Λ,E)). As seen, the potential labels (shown inside the vertical rectangles) are assigned to each node. Node ni can be any of the following: 11,8,4,1. Also depicted in the figure are the probabilities associated with each of the labels. After initialization, all probabilities are equal. Part (b) of Figure 8 shows the result of the first iteration of the localization algorithm for node ni, assuming that node nj is the first wi chosen in line 7 of Algorithm 3. By using Equation 6, the algorithm computes the support Q(λi) for each of the possible labels for node ni. Once the Q(λi)"s are computed, the normalizing constant, given by Equation 3 can be obtained. The last step of the iteration is to update the probabilities associated with all potential labels of node ni, as given by Equation 2. One interesting problem, which we explore in the performance evaluation section, is to assess the impact the partitioning of the color set C has on the accuracy of localization. When the size of the coloring set is smaller than the number of sensor nodes (as it is the case for our hybrid connectivity/color constrained relaxation), the system designer has the option of allowing one node to uniquely have a color (acting as an anchor), or multiple nodes. Intuitively, by assigning one color to more than one node, more constraints (distributed) can be enforced. 3.3.6 Relaxation Techniques Analysis The proposed label relaxation techniques have different trade-offs. For our analysis of the trade-offs, we consider the following metrics of interest: the localization time (duration), the energy consumed (overhead), the network size (scale) that can be handled by the technique and the localization accuracy. The parameters of interest are the following: the number of sensor nodes (N), the energy spent for one aerial drop (εd), the energy spent in the network for collecting and reporting neighbor information εb and the time Td taken by a sensor node to reach the ground after being aerially deployed. The cost comparison of the different label relaxation techniques is shown in Table 1. As shown, the relaxation techniques based on color and space constraints have the lowest localization duration, zero, for all practical purposes. The scalability of the color based relaxation technique is, however, limited to the number of (a) (b) Figure 9. SensorBall with self-righting capabilities (a) and colored CCRs (b) unique color filters that can be built. The narrower the Transfer Function Ψ(λ), the larger the number of unique colors that can be created. The manufacturing costs, however, are increasing as well. The scalability issue is addressed by all other label relaxation techniques. Most notably, the time constrained relaxation, which is very similar to the colorconstrained relaxation, addresses the scale issue, at a higher deployment cost. Criteria Color Connectivity Time Space Duration 0 NTb NTd 0 Overhead εd εd +Nεb Nεd εd Scale |C| |N| |N| |N| Accuracy High Low High Medium Table 1. Comparison of label relaxation techniques 4 System Implementation The StarDust localization framework, depicted in Figure 2, is flexible in that it enables the development of new localization systems, based on the four proposed label relaxation schemes, or the inclusion of other, yet to be invented, schemes. For our performance evaluation we implemented a version of the StarDust framework, namely the one proposed in Section 3.3.5, where the constraints are based on color and connectivity. The Central device of the StarDust system consists of the following: the Light Emitter - we used a common-off-theshelf flash light (QBeam, 3 million candlepower); the image acquisition was done with a 3 megapixel digital camera (Sony DSC-S50) which provided the input to the Image Processing algorithm, implemented in Matlab. For sensor nodes we built a custom sensor node, called SensorBall, with self-righting capabilities, shown in Figure 9(a). The self-righting capabilities are necessary in order to orient the CCR predominantly upwards. The CCRs that we used were inexpensive, plastic molded, night time warning signs commonly available on bicycles, as shown in Figure 9(b). We remark here the low quality of the CCRs we used. The reflectivity of each CCR (there are tens molded in the plastic container) is extremely low, and each CCR is not built with mirrors. A reflective effect is achieved by employing finely polished plastic surfaces. We had 5 colors available, in addition to the standard CCR, which reflects all the incoming light (white CCR). For a slightly higher price (ours were 20cents/piece), better quality CCRs can be employed. 64 Figure 10. The field in the dark Figure 11. The illuminated field Figure 12. The difference: Figure 10Figure 11 Higher quality (better mirrors) would translate in more accurate image processing (better sensor node detection) and smaller form factor for the optical component (an array of CCRs with a smaller area can be used). The sensor node platform we used was the micaZ mote. The code that runs on each node is a simple application which broadcasts 100 beacons, and maintains a neighbor table containing the percentage of successfully received beacons, for each neighbor. On demand, the neighbor table is reported to a base station, where the node ID mapping is performed. 5 System Evaluation In this section we present the performance evaluation of a system implementation of the StarDust localization framework. The three major research questions that our evaluation tries to answer are: the feasibility of the proposed framework (can sensor nodes be optically detected at large distances), the localization accuracy of one actual implementation of the StarDust framework, and whether or not atmospheric conditions can affect the recognition of sensor nodes in an image. The first two questions are investigated by evaluating the two main components of the StarDust framework: the Image Processing and the Node ID Matching. These components have been evaluated separately mainly because of lack of adequate facilities. We wanted to evaluate the performance of the Image Processing Algorithm in a long range, realistic, experimental set-up, while the Node ID Matching required a relatively large area, available for long periods of time (for connectivity data gathering). The third research question is investigated through a computer modeling of atmospheric phenomena. For the evaluation of the Image Processing module, we performed experiments in a football stadium where we deploy 6 sensor nodes in a 3×2 grid. The distance between the Central device and the sensor nodes is approximately 500 ft. The metrics of interest are the number of false positives and false negatives in the Image Processing Algorithm. For the evaluation of the Node ID Mapping component, we deploy 26 sensor nodes in an 120 × 60 ft2 flat area of a stadium. In order to investigate the influence the radio connectivity has on localization accuracy, we vary the height above ground of the deployed sensor nodes. Two set-ups are used: one in which the sensor nodes are on the ground, and the second one, in which the sensor nodes are raised 3 inches above ground. From here on, we will refer to these two experimental set-ups as the low connectivity and the high connectivity networks, respectively because when nodes are on the ground the communication range is low resulting in less neighbors than when the nodes are elevated and have a greater communication range. The metrics of interest are: the localization error (defined as the distance between the computed location and the true location - known from the manual placement), the percentage of nodes correctly localized, the convergence of the label relaxation algorithm, the time to localize and the robustness of the node ID mapping to errors in the Image Processing module. The parameters that we vary experimentally are: the angle under which images are taken, the focus of the camera, and the degree of connectivity. The parameters that we vary in simulations (subsequent to image acquisition and connectivity collection) the number of colors, the number of anchors, the number of false positives or negatives as input to the Node ID Matching component, the distance between the imaging device and sensor network (i.e., range), atmospheric conditions (light attenuation coefficient) and CCR reflectance (indicative of its quality). 5.1 Image Processing For the IPA evaluation, we deploy 6 sensor nodes in a 3 × 2 grid. We take 13 sets of pictures using different orientations of the camera and different zooming factors. All pictures were taken from the same location. Each set is composed of a picture taken in the dark and of a picture taken with a light beam pointed at the nodes. We process the pictures offline using a Matlab implementation of IPA. Since we are interested in the feasibility of identifying colored sensor nodes at large distance, the end result of our IPA is the 2D location of sensor nodes (position in the image). The transformation to 3D coordinates can be done through standard computer graphics techniques [23]. One set of pictures obtained as part of our experiment is shown in Figures 10 and 11. The execution of our IPA algorithm results in Figure 12 which filters out the background, and Figure 13 which shows the output of the edge detection step of IPA. The experimental results are depicted in Figure 14. For each set of pictures the graph shows the number of false positives (the IPA determines that there is a node 65 Figure 13. Retroreflectors detected in Figure 12 0 1 2 3 1 2 3 4 5 6 7 8 9 10 11 Experiment Number Count False Positives False Negatives Figure 14. False Positives and Negatives for the 6 nodes while there is none), and the number of false negatives (the IPA determines that there is no node while there is one). In about 45% of the cases, we obtained perfect results, i.e., no false positives and no false negatives. In the remaining cases, we obtained a number of false positives of at most one, and a number of false negatives of at most two. We exclude two pairs of pictures from Figure 14. In the first excluded pair, we obtain 42 false positives and in the second pair 10 false positives and 7 false negatives. By carefully examining the pictures, we realized that the first pair was taken out of focus and that a car temporarily appeared in one of the pictures of the second pair. The anomaly in the second set was due to the fact that we waited too long to take the second picture. If the pictures had been taken a few milliseconds apart, the car would have been represented on either both or none of the pictures and the IPA would have filtered it out. 5.2 Node ID Matching We evaluate the Node ID Matching component of our system by collecting empirical data (connectivity information) from the outdoor deployment of 26 nodes in the 120×60 ft2 area. We collect 20 sets of data for the high connectivity and low connectivity network deployments. Off-line we investigate the influence of coloring on the metrics of interest, by randomly assigning colors to the sensor nodes. For one experimental data set we generate 50 random assignments of colors to sensor nodes. It is important to observe that, for the evaluation of the Node ID Matching algorithm (color and connectivity constrained), we simulate the color assignment to sensor nodes. As mentioned in Section 4 the size of the coloring space available to us was 5 (5 colors). Through simulations of color assignment (not connectivity) we are able to investigate the influence that the size of the coloring space has on the accuracy of localization. The value of the param0 10 20 30 40 50 60 0 10 20 30 40 50 60 70 80 90 Distance [feet] Count Connected Not Connected Figure 15. The number of existing and missing radio connections in the sparse connectivity experiment 0 10 20 30 40 50 60 70 0 10 20 30 40 50 60 70 80 90 10 11 12 Distance [feet] Count Connected Not Connected Figure 16. The number of existing and missing radio connections in the high connectivity experiment eter ε used in Algorithm 2 was 0.001. The results presented here represent averages over the randomly generated colorings and over all experimental data sets. We first investigate the accuracy of our proposed Radio Model, and subsequently use the derived values for the radio range in the evaluation of the Node ID matching component. 5.2.1 Radio Model From experiments, we obtain the average number of observed beacons (k, defined in Section 3.3.2) for the low connectivity network of 180 beacons and for the high connectivity network of 420 beacons. From our Radio Model (Equation 7, we obtain a radio range R = 25 ft for the low connectivity network and R = 40 ft for the high connectivity network. To estimate the accuracy of our simple model, we plot the number of radio links that exist in the networks, and the number of links that are missing, as functions of the distance between nodes. The results are shown in Figures 15 and 16. We define the average radio range R to be the distance over which less than 20% of potential radio links, are missing. As shown in Figure 15, the radio range is between 20 ft and 25 ft. For the higher connectivity network, the radio range was between 30 ft and 40 ft. We choose two conservative estimates of the radio range: 20 ft for the low connectivity case and 35 ft for the high connectivity case, which are in good agreement with the values predicted by our Radio Model. 5.2.2 Localization Error vs. Coloring Space Size In this experiment we investigate the effect of the number of colors on the localization accuracy. For this, we randomly assign colors from a pool of a given size, to the sensor nodes. 66 0 5 10 15 20 25 30 35 40 45 0 5 10 15 20 Number of Colors LocalizationError[feet] R=15feet R=20feet R=25feet Figure 17. Localization error 0 10 20 30 40 50 60 70 80 90 100 0 5 10 15 20 Number of Colors %CorrectLocalized[x100] R=15feet R=20feet R=25feet Figure 18. Percentage of nodes correctly localized We then execute the localization algorithm, which uses the empirical data. The algorithm is run for three different radio ranges: 15, 20 and 25 ft, to investigate its influence on the localization error. The results are depicted in Figure 17 (localization error) and Figure 18 (percentage of nodes correctly localized). As shown, for an estimate of 20 ft for the radio range (as predicted by our Radio Model) we obtain the smallest localization errors, as small as 2 ft, when enough colors are used. Both Figures 17 and 18 confirm our intuition that a larger number of colors available significantly decrease the error in localization. The well known fact that relaxation algorithms do not always converge, was observed during our experiments. The percentage of successful runs (when the algorithm converged) is depicted in Figure 19. As shown, in several situations, the algorithm failed to converge (the algorithm execution was stopped after 100 iterations per node). If the algorithm does not converge in a predetermined number of steps, it will terminate and the label with the highest probability will provide the identity of the node. It is very probable that the chosen label is incorrect, since the probabilities of some of labels are constantly changing (with each iteration).The convergence of relaxation based algorithms is a well known issue. 5.2.3 Localization Error vs. Color Uniqueness As mentioned in the Section 3.3.1, a unique color gives a sensor node the statute of an anchor. A sensor node that is an anchor can unequivocally be identified through the Image Processing module. In this section we investigate the effect unique colors have on the localization accuracy. Specifically, we want to experimentally verify our intuition that assigning more nodes to a color can benefit the localization accuracy, by enforcing more constraints, as opposed to uniquely assigning a color to a single node. 90 95 100 105 0 5 10 15 20 Number of Colors ConvergenceRate[x100] R=15feet R=20feet R=25feet Figure 19. Convergence error 0 2 4 6 8 10 12 14 16 4 6 8 Number of Colors LocalizationError[feet] 0 anchors 2 anchors 4 anchors 6 anchors 8 anchors Figure 20. Localization error vs. number of colors For this, we fix the number of available colors to either 4, 6 or 8 and vary the number of nodes that are given unique colors, from 0, up to the maximum number of colors (4, 6 or 8). Naturally, if we have a maximum number of colors of 4, we can assign at most 4 anchors. The experimental results are depicted in Figure 20 (localization error) and Figure 21 (percentage of sensor node correctly localized). As expected, the localization accuracy increases with the increase in the number of colors available (larger coloring space). Also, for a given size of the coloring space (e.g., 6 colors available), if more colors are uniquely assigned to sensor nodes then the localization accuracy decreases. It is interesting to observe that by assigning colors uniquely to nodes, the benefit of having additional colors is diminished. Specifically, if 8 colors are available and all are assigned uniquely, the system would be less accurately localized (error ≈ 7 ft), when compared to the case of 6 colors and no unique assignments of colors (≈ 5 ft localization error). The same trend, of a less accurate localization can be observed in Figure 21, which shows the percentage of nodes correctly localized (i.e., 0 ft localization error). As shown, if we increase the number of colors that are uniquely assigned, the percentage of nodes correctly localized decreases. 5.2.4 Localization Error vs. Connectivity We collected empirical data for two network deployments with different degrees of connectivity (high and low) in order to assess the influence of connectivity on location accuracy. The results obtained from running our localization algorithm are depicted in Figure 22 and Figure 23. We varied the number of colors available and assigned no anchors (i.e., no unique assignments of colors). In both scenarios, as expected, localization error decrease with an increase in the number of colors. It is interesting to observe, however, that the low connectivity scenario im67 0 20 40 60 80 100 120 140 4 6 8 Number of Colors %CorrectLocalized[x100] 0 anchors 2 anchors 4 anchors 6 anchors 8 anchors Figure 21. Percentage of nodes correctly localized vs. number of colors 0 5 10 15 20 25 30 35 40 45 0 2 4 6 8 10 12 Number of Colors LocalizationError[feet] Low Connectivity High Connectivity Figure 22. Localization error vs. number of colors proves the localization accuracy quicker, from the additional number of colors available. When the number of colors becomes relatively large (twelve for our 26 sensor node network), both scenarios (low and high connectivity) have comparable localization errors, of less that 2 ft. The same trend of more accurate location information is evidenced by Figure 23 which shows that the percentage of nodes that are localized correctly grows quicker for the low connectivity deployment. 5.3 Localization Error vs. Image Processing Errors So far we investigated the sources for error in localization that are intrinsic to the Node ID Matching component. As previously presented, luminous objects can be mistakenly detected to be sensor nodes during the location detection phase of the Image Processing module. These false positives can be eliminated by the color recognition procedure of the Image Processing module. More problematic are false negatives (when a sensor node does not reflect back enough light to be detected). They need to be handled by the localization algorithm. In this case, the localization algorithm is presented with two sets of nodes of different sizes, that need to be matched: one coming from the Image Processing (which misses some nodes) and one coming from the network, with the connectivity information (here we assume a fully connected network, so that all sensor nodes report their connectivity information). In this experiment we investigate how Image Processing errors (false negatives) influence the localization accuracy. For this evaluation, we ran our localization algorithm with empirical data, but dropped a percentage of nodes from the list of nodes detected by the Image Processing algorithm (we artificially introduced false negatives in the Image Process0 10 20 30 40 50 60 70 80 90 100 0 2 4 6 8 10 12 Number of Colors %CorrectLocalized[x100] Low Connectivity High Connectivity Figure 23. Percentage of nodes correctly localized 0 2 4 6 8 10 12 14 0 4 8 12 16 % False Negatives [x100] LocalizationError[feet] 4 colors 8 colors 12 colors Figure 24. Impact of false negatives on the localization error ing). The effect of false negatives on localization accuracy is depicted in Figure 24. As seen in the figure if the number of false negatives is 15%, the error in position estimation doubles when 4 colors are available. It is interesting to observe that the scenario when more colors are available (e.g., 12 colors) is being affected more drastically than the scenario with less colors (e.g., 4 colors). The benefit of having more colors available is still being maintained, at least for the range of colors we investigated (4 through 12 colors). 5.4 Localization Time In this section we look more closely at the duration for each of the four proposed relaxation techniques and two combinations of them: color-connectivity and color-time. We assume that 50 unique color filters can be manufactured, that the sensor network is deployed from 2,400 ft (necessary for the time-constrained relaxation) and that the time required for reporting connectivity grows linearly, with an initial reporting period of 160sec, as used in a real world tracking application [1]. The localization duration results, as presented in Table 1, are depicted in Figure 25. As shown, for all practical purposes the time required by the space constrained relaxation techniques is 0sec. The same applies to the color constrained relaxation, for which the localization time is 0sec (if the number of colors is sufficient). Considering our assumptions, only for a network of size 50 the color constrained relaxation works. The localization duration for all other network sizes (100, 150 and 200) is infinite (i.e., unique color assignments to sensor nodes can not be made, since only 50 colors are unique), when only color constrained relaxation is used. Both the connectivity constrained and time constrained techniques increase linearly with the network size (for the time constrained, the Central device deploys sensor nodes one by one, recording an image after the time a sensor node is expected to reach the 68 0 500 1000 1500 2000 2500 3000 Color Connectivity Time Space ColorConenctivity Color-Time Localization technique Localizationtime[sec] 50 nodes 100 nodes 150 nodes 200 nodes Figure 25. Localization time for different label relaxation schemes 0 2000 4000 6000 8000 0 0.5 1 1.5 2 2.5 3 3.5 4 r [feet] C r 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0 Figure 26. Apparent contrast in a clear atmosphere 0 2000 4000 6000 8000 0 0.5 1 1.5 2 2.5 3 3.5 4 r [feet] C r 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0 Figure 27. Apparent contrast in a hazing atmosphere ground). It is interesting to notice in Figure 25 the improvement in the localization time obtained by simply combining the color and the connectivity constrained techniques. The localization duration in this case is identical with the connectivity constrained technique. The combination of color and time constrained relaxations is even more interesting. For a reasonable localization duration of 52seconds a perfect (i.e., 0 ft localization error) localization system can be built. In this scenario, the set of sensor nodes is split in batches, with each batch having a set of unique colors. It would be very interesting to consider other scenarios, where the strength of the space constrained relaxation (0sec for any sensor network size) is used for improving the other proposed relaxation techniques. We leave the investigation and rigorous classification of such technique combination for future work. 5.5 System Range In this section we evaluate the feasibility of the StarDust localization framework when considering the realities of light propagation through the atmosphere. The main factor that determines the range of our system is light scattering, which redirects the luminance of the source into the medium (in essence equally affecting the luminosity of the target and of the background). Scattering limits the visibility range by reducing the apparent contrast between the target and its background (approaches zero, as the distance increases). The apparent contrast Cr is quantitatively expressed by the formula: Cr = (Nt r −Nb r )/Nb r (10) where Nt r and Nb r are the apparent target radiance and apparent background radiance at distance r from the light source, respectively. The apparent radiance Nt r of a target at a distance r from the light source, is given by: Nt r = Na + Iρte−2σr πr2 (11) where I is the intensity of the light source, ρt is the target reflectance, σ is the spectral attenuation coefficient (≈ 0.12km−1 and ≈ 0.60km−1 for a clear and a hazy atmosphere, respectively) and Na is the radiance of the atmospheric backscatter, and it can be expressed as follows: Na = Gσ2I 2π 2σrZ 0.02σr e−x x2 dx (12) where G = 0.24 is a backscatter gain. The apparent background radiance Nb r is given by formulas similar with Equations 11 and 12, where only the target reflectance ρt is substituted with the background reflectance ρb. It is important to remark that when Cr reaches its lower limit, no increase in the source luminance or receiver sensitivity will increase the range of the system. From Equations 11 and 12 it can be observed that the parameter which can be controlled and can influence the range of the system is ρt, the target reflectance. Figures 26 and 27 depict the apparent contrast Cr as a function of the distance r for a clear and for a hazy atmosphere, respectively. The apparent contrast is investigated for reflectance coefficients ρt ranging from 0.3 to 1.0 (perfect reflector). For a contrast C of at least 0.5, as it can be seen in Figure 26 a range of approximately 4,500 ft can be achieved if the atmosphere is clear. The performance dramatically deteriorates, when the atmospheric conditions are problematic. As shown in Figure 27 a range of up to 1,500 ft is achievable, when using highly reflective CCR components. While our light source (3 million candlepower) was sufficient for a range of a few hundred feet, we remark that there exist commercially available light sources (20 million candlepower) or military (150 million candlepower [27]), powerful enough for ranges of a few thousand feet. 6 StarDust System Optimizations In this section we describe extensions of the proposed architecture that can constitute future research directions. 6.1 Chained Constraint Primitives In this paper we proposed four primitives for constraintbased relaxation algorithms: color, connectivity, time and space. To demonstrate the power that can be obtained by combining them, we proposed and evaluated one combination of such primitives: color and connectivity. An interesting research direction to pursue could be to chain more than two of these primitives. An example of such chain is: color, temporal, spatial and connectivity. Other research directions could be to use voting scheme for deciding which primitive to use or assign different weights to different relaxation algorithms. 69 6.2 Location Learning If after several iterations of the algorithm, none of the label probabilities for a node ni converges to a higher value, the confidence in our labeling of that node is relatively low. It would be interesting to associate with a node, more than one label (implicitly more than one location) and defer the label assignment decision until events are detected in the network (if the network was deployed for target tracking). 6.3 Localization in Rugged Environments The initial driving force for the StarDust localization framework was to address the sensor node localization in extremely rugged environments. Canopies, dense vegetation, extremely obstructing environments pose significant challenges for sensor nodes localization. The hope, and our original idea, was to consider the time period between the aerial deployment and the time when the sensor node disappears under the canopy. By recording the last visible position of a sensor node (as seen from the aircraft) a reasonable estimate of the sensor node location can be obtained. This would require that sensor nodes posses self-righting capabilities, while in mid-air. Nevertheless, we remark on the suitability of our localization framework for rugged, non-line-of-sight environments. 7 Conclusions StarDust solves the localization problem for aerial deployments where passiveness, low cost, small form factor and rapid localization are required. Results show that accuracy can be within 2 ft and localization time within milliseconds. StarDust also shows robustness with respect to errors. We predict the influence the atmospheric conditions can have on the range of a system based on the StarDust framework, and show that hazy environments or daylight can pose significant challenges. Most importantly, the properties of StarDust support the potential for even more accurate localization solutions as well as solutions for rugged, non-line-of-sight environments. 8 References [1] T. He, S. Krishnamurthy, J. A. Stankovic, T. Abdelzaher, L. Luo, R. Stoleru, T. Yan, L. Gu, J. Hui, and B. Krogh, An energy-efficient surveillance system using wireless sensor networks, in MobiSys, 2004. [2] G. Simon, M. Maroti, A. Ledeczi, G. Balogh, B. Kusy, A. Nadas, G. Pap, J. Sallai, and K. Frampton, Sensor network-based countersniper system, in SenSys, 2004. [3] A. Arora, P. Dutta, and B. Bapat, A line in the sand: A wireless sensor network for trage detection, classification and tracking, in Computer Networks, 2004. [4] R. Szewczyk, A. Mainwaring, J. Polastre, J. Anderson, and D. Culler, An analysis of a large scale habitat monitoring application, in ACM SenSys, 2004. [5] N. Xu, S. Rangwala, K. K. Chintalapudi, D. Ganesan, A. Broad, R. Govindan, and D. Estrin, A wireless sensor network for structural monitoring, in ACM SenSys, 2004. [6] A. Savvides, C. Han, and M. Srivastava, Dynamic fine-grained localization in ad-hoc networks of sensors, in Mobicom, 2001. [7] N. Priyantha, A. Chakraborty, and H. Balakrishnan, The cricket location-support system, in Mobicom, 2000. [8] M. Broxton, J. Lifton, and J. Paradiso, Localizing a sensor network via collaborative processing of global stimuli, in EWSN, 2005. [9] P. Bahl and V. N. Padmanabhan, Radar: An in-building rf-based user location and tracking system, in IEEE Infocom, 2000. [10] N. Priyantha, H. Balakrishnan, E. Demaine, and S. Teller, Mobileassisted topology generation for auto-localization in sensor networks, in IEEE Infocom, 2005. [11] P. N. Pathirana, A. Savkin, S. Jha, and N. Bulusu, Node localization using mobile robots in delay-tolerant sensor networks, IEEE Transactions on Mobile Computing, 2004. [12] C. Savarese, J. M. Rabaey, and J. Beutel, Locationing in distribued ad-hoc wireless sensor networks, in ICAASSP, 2001. [13] M. Maroti, B. Kusy, G. Balogh, P. Volgyesi, A. Nadas, K. Molnar, S. Dora, and A. Ledeczi, Radio interferometric geolocation, in ACM SenSys, 2005. [14] K. Whitehouse, A. Woo, C. Karlof, F. Jiang, and D. Culler, The effects of ranging noise on multi-hop localization: An empirical study, in IPSN, 2005. [15] Y. Kwon, K. Mechitov, S. Sundresh, W. Kim, and G. Agha, Resilient localization for sensor networks in outdoor environment, UIUC, Tech. Rep., 2004. [16] R. Stoleru and J. A. Stankovic, Probability grid: A location estimation scheme for wireless sensor networks, in SECON, 2004. [17] N. Bulusu, J. Heidemann, and D. Estrin, GPS-less low cost outdoor localization for very small devices, IEEE Personal Communications Magazine, 2000. [18] T. He, C. Huang, B. Blum, J. A. Stankovic, and T. Abdelzaher, Range-Free localization schemes in large scale sensor networks, in ACM Mobicom, 2003. [19] R. Nagpal, H. Shrobe, and J. Bachrach, Organizing a global coordinate system from local information on an ad-hoc sensor network, in IPSN, 2003. [20] D. Niculescu and B. Nath, ad-hoc positioning system, in IEEE GLOBECOM, 2001. [21] R. Stoleru, T. He, J. A. Stankovic, and D. Luebke, A high-accuracy low-cost localization system for wireless sensor networks, in ACM SenSys, 2005. [22] K. R¨omer, The lighthouse location system for smart dust, in ACM/USENIX MobiSys, 2003. [23] R. Y. Tsai, A versatile camera calibration technique for highaccuracy 3d machine vision metrology using off-the-shelf tv cameras and lenses, IEEE JRA, 1987. [24] C. L. Archer and M. Z. Jacobson, Spatial and temporal distributions of U.S. winds and wind power at 80m derived from measurements, Geophysical Research Jrnl., 2003. [25] Team for advanced flow simulation and modeling. [Online]. Available: http://www.mems.rice.edu/TAFSM/RES/ [26] K. Stein, R. Benney, T. Tezduyar, V. Kalro, and J. Leonard, 3-D computation of parachute fluid-structure interactions - performance and control, in Aerodynamic Decelerator Systems Conference, 1999. [27] Headquarters Department of the Army, Technical manual for searchlight infrared AN/GSS-14(V)1, 1982. 70

range;unique mapping;performance;image processing;connectivity;localization;scene labeling;probability;sensor node;aerial vehicle;consistency;wireless sensor network;corner-cube retro-reflector

train_C-46

TSAR: A Two Tier Sensor Storage Architecture Using Interval Skip Graphs

Archival storage of sensor data is necessary for applications that query, mine, and analyze such data for interesting features and trends. We argue that existing storage systems are designed primarily for flat hierarchies of homogeneous sensor nodes and do not fully exploit the multi-tier nature of emerging sensor networks, where an application can comprise tens of tethered proxies, each managing tens to hundreds of untethered sensors. We present TSAR, a fundamentally different storage architecture that envisions separation of data from metadata by employing local archiving at the sensors and distributed indexing at the proxies. At the proxy tier, TSAR employs a novel multi-resolution ordered distributed index structure, the Interval Skip Graph, for efficiently supporting spatio-temporal and value queries. At the sensor tier, TSAR supports energy-aware adaptive summarization that can trade off the cost of transmitting metadata to the proxies against the overhead of false hits resulting from querying a coarse-grain index. We implement TSAR in a two-tier sensor testbed comprising Stargatebased proxies and Mote-based sensors. Our experiments demonstrate the benefits and feasibility of using our energy-efficient storage architecture in multi-tier sensor networks.

1. Introduction 1.1 Motivation Many different kinds of networked data-centric sensor applications have emerged in recent years. Sensors in these applications sense the environment and generate data that must be processed, filtered, interpreted, and archived in order to provide a useful infrastructure to its users. To achieve its goals, a typical sensor application needs access to both live and past sensor data. Whereas access to live data is necessary in monitoring and surveillance applications, access to past data is necessary for applications such as mining of sensor logs to detect unusual patterns, analysis of historical trends, and post-mortem analysis of particular events. Archival storage of past sensor data requires a storage system, the key attributes of which are: where the data is stored, whether it is indexed, and how the application can access this data in an energy-efficient manner with low latency. There have been a spectrum of approaches for constructing sensor storage systems. In the simplest, sensors stream data or events to a server for long-term archival storage [3], where the server often indexes the data to permit efficient access at a later time. Since sensors may be several hops from the nearest base station, network costs are incurred; however, once data is indexed and archived, subsequent data accesses can be handled locally at the server without incurring network overhead. In this approach, the storage is centralized, reads are efficient and cheap, while writes are expensive. Further, all data is propagated to the server, regardless of whether it is ever used by the application. An alternate approach is to have each sensor store data or events locally (e.g., in flash memory), so that all writes are local and incur no communication overheads. A read request, such as whether an event was detected by a particular sensor, requires a message to be sent to the sensor for processing. More complex read requests are handled by flooding. For instance, determining if an intruder was detected over a particular time interval requires the request to be flooded to all sensors in the system. Thus, in this approach, the storage is distributed, writes are local and inexpensive, while reads incur significant network overheads. Requests that require flooding, due to the lack of an index, are expensive and may waste precious sensor resources, even if no matching data is stored at those sensors. Research efforts such as Directed Diffusion [17] have attempted to reduce these read costs, however, by intelligent message routing. Between these two extremes lie a number of other sensor storage systems with different trade-offs, summarized in Table 1. The geographic hash table (GHT) approach [24, 26] advocates the use of an in-network index to augment the fully distributed nature of sensor storage. In this approach, each data item has a key associated with it, and a distributed or geographic hash table is used to map keys to nodes that store the corresponding data items. Thus, writes cause data items to be sent to the hashed nodes and also trigger updates to the in-network hash table. A read request requires a lookup in the in-network hash table to locate the node that stores the data 39 item; observe that the presence of an index eliminates the need for flooding in this approach. Most of these approaches assume a flat, homogeneous architecture in which every sensor node is energy-constrained. In this paper, we propose a novel storage architecture called TSAR1 that reflects and exploits the multi-tier nature of emerging sensor networks, where the application is comprised of tens of tethered sensor proxies (or more), each controlling tens or hundreds of untethered sensors. TSAR is a component of our PRESTO [8] predictive storage architecture, which combines archival storage with caching and prediction. We believe that a fundamentally different storage architecture is necessary to address the multi-tier nature of future sensor networks. Specifically, the storage architecture needs to exploit the resource-rich nature of proxies, while respecting resource constraints at the remote sensors. No existing sensor storage architecture explicitly addresses this dichotomy in the resource capabilities of different tiers. Any sensor storage system should also carefully exploit current technology trends, which indicate that the capacities of flash memories continue to rise as per Moore"s Law, while their costs continue to plummet. Thus it will soon be feasible to equip each sensor with 1 GB of flash storage for a few tens of dollars. An even more compelling argument is the energy cost of flash storage, which can be as much as two orders of magnitude lower than that for communication. Newer NAND flash memories offer very low write and erase energy costs - our comparison of a 1GB Samsung NAND flash storage [16] and the Chipcon CC2420 802.15.4 wireless radio [4] in Section 6.2 indicates a 1:100 ratio in per-byte energy cost between the two devices, even before accounting for network protocol overheads. These trends, together with the energy-constrained nature of untethered sensors, indicate that local storage offers a viable, energy-efficient alternative to communication in sensor networks. TSAR exploits these trends by storing data or events locally on the energy-efficient flash storage at each sensor. Sensors send concise identifying information, which we term metadata, to a nearby proxy; depending on the representation used, this metadata may be an order of magnitude or more smaller than the data itself, imposing much lower communication costs. The resource-rich proxies interact with one another to construct a distributed index of the metadata reported from all sensors, and thus an index of the associated data stored at the sensors. This index provides a unified, logical view of the distributed data, and enables an application to query and read past data efficiently - the index is used to pinpoint all data that match a read request, followed by messages to retrieve that data from the corresponding sensors. In-network index lookups are eliminated, reducing network overheads for read requests. This separation of data, which is stored at the sensors, and the metadata, which is stored at the proxies, enables TSAR to reduce energy overheads at the sensors, by leveraging resources at tethered proxies. 1.2 Contributions This paper presents TSAR, a novel two-tier storage architecture for sensor networks. To the best of our knowledge, this is the first sensor storage system that is explicitly tailored for emerging multitier sensor networks. Our design and implementation of TSAR has resulted in four contributions. At the core of the TSAR architecture is a novel distributed index structure based on interval skip graphs that we introduce in this paper. This index structure can store coarse summaries of sensor data and organize them in an ordered manner to be easily search1 TSAR: Tiered Storage ARchitecture for sensor networks. able. This data structure has O(log n) expected search and update complexity. Further, the index provides a logically unified view of all data in the system. Second, at the sensor level, each sensor maintains a local archive that stores data on flash memory. Our storage architecture is fully stateless at each sensor from the perspective of the metadata index; all index structures are maintained at the resource-rich proxies, and only direct requests or simple queries on explicitly identified storage locations are sent to the sensors. Storage at the remote sensor is in effect treated as appendage of the proxy, resulting in low implementation complexity, which makes it ideal for small, resourceconstrained sensor platforms. Further, the local store is optimized for time-series access to archived data, as is typical in many applications. Each sensor periodically sends a summary of its data to a proxy. TSAR employs a novel adaptive summarization technique that adapts the granularity of the data reported in each summary to the ratio of false hits for application queries. More fine grain summaries are sent whenever more false positives are observed, thereby balancing the energy cost of metadata updates and false positives. Third, we have implemented a prototype of TSAR on a multi-tier testbed comprising Stargate-based proxies and Mote-based sensors. Our implementation supports spatio-temporal, value, and rangebased queries on sensor data. Fourth, we conduct a detailed experimental evaluation of TSAR using a combination of EmStar/EmTOS [10] and our prototype. While our EmStar/EmTOS experiments focus on the scalability of TSAR in larger settings, our prototype evaluation involves latency and energy measurements in a real setting. Our results demonstrate the logarithmic scaling property of the sparse skip graph and the low latency of end-to-end queries in a duty-cycled multi-hop network . The remainder of this paper is structured as follows. Section 2 presents key design issues that guide our work. Section 3 and 4 present the proxy-level index and the local archive and summarization at a sensor, respectively. Section 5 discusses our prototype implementation, and Section 6 presents our experimental results. We present related work in Section 7 and our conclusions in Section 8. 2. Design Considerations In this section, we first describe the various components of a multi-tier sensor network assumed in our work. We then present a description of the expected usage models for this system, followed by several principles addressing these factors which guide the design of our storage system. 2.1 System Model We envision a multi-tier sensor network comprising multiple tiers - a bottom tier of untethered remote sensor nodes, a middle tier of tethered sensor proxies, and an upper tier of applications and user terminals (see Figure 1). The lowest tier is assumed to form a dense deployment of lowpower sensors. A canonical sensor node at this tier is equipped with low-power sensors, a micro-controller, and a radio as well as a significant amount of flash memory (e.g., 1GB). The common constraint for this tier is energy, and the need for a long lifetime in spite of a finite energy constraint. The use of radio, processor, RAM, and the flash memory all consume energy, which needs to be limited. In general, we assume radio communication to be substantially more expensive than accesses to flash memory. The middle tier consists of power-rich sensor proxies that have significant computation, memory and storage resources and can use 40 Table 1: Characteristics of sensor storage systems System Data Index Reads Writes Order preserving Centralized store Centralized Centralized index Handled at store Send to store Yes Local sensor store Fully distributed No index Flooding, diffusion Local No GHT/DCS [24] Fully distributed In-network index Hash to node Send to hashed node No TSAR/PRESTO Fully distributed Distributed index at proxies Proxy lookup + sensor query Local plus index update Yes User Unified Logical Store Queries (time, space, value) Query Response Cache Query forwarding Proxy Remote Sensors Local Data Archive on Flash Memory Interval Skip Graph Query forwarding summaries start index end index linear traversal Query Response Cache-miss triggered query forwarding summaries Figure 1: Architecture of a multi-tier sensor network. these resources continuously. In urban environments, the proxy tier would comprise a tethered base-station class nodes (e.g., Crossbow Stargate), each with with multiple radios-an 802.11 radio that connects it to a wireless mesh network and a low-power radio (e.g. 802.15.4) that connects it to the sensor nodes. In remote sensing applications [10], this tier could comprise a similar Stargate node with a solar power cell. Each proxy is assumed to manage several tens to hundreds of lower-tier sensors in its vicinity. A typical sensor network deployment will contain multiple geographically distributed proxies. For instance, in a building monitoring application, one sensor proxy might be placed per floor or hallway to monitor temperature, heat and light sensors in their vicinity. At the highest tier of our infrastructure are applications that query the sensor network through a query interface[20]. In this work, we focus on applications that require access to past sensor data. To support such queries, the system needs to archive data on a persistent store. Our goal is to design a storage system that exploits the relative abundance of resources at proxies to mask the scarcity of resources at the sensors. 2.2 Usage Models The design of a storage system such as TSAR is affected by the queries that are likely to be posed to it. A large fraction of queries on sensor data can be expected to be spatio-temporal in nature. Sensors provide information about the physical world; two key attributes of this information are when a particular event or activity occurred and where it occurred. Some instances of such queries include the time and location of target or intruder detections (e.g., security and monitoring applications), notifications of specific types of events such as pressure and humidity values exceeding a threshold (e.g., industrial applications), or simple data collection queries which request data from a particular time or location (e.g., weather or environment monitoring). Expected queries of such data include those requesting ranges of one or more attributes; for instance, a query for all image data from cameras within a specified geographic area for a certain period of time. In addition, it is often desirable to support efficient access to data in a way that maintains spatial and temporal ordering. There are several ways of supporting range queries, such as locality-preserving hashes such as are used in DIMS [18]. However, the most straightforward mechanism, and one which naturally provides efficient ordered access, is via the use of order-preserving data structures. Order-preserving structures such as the well-known B-Tree maintain relationships between indexed values and thus allow natural access to ranges, as well as predecessor and successor operations on their key values. Applications may also pose value-based queries that involve determining if a value v was observed at any sensor; the query returns a list of sensors and the times at which they observed this value. Variants of value queries involve restricting the query to a geographical region, or specifying a range (v1, v2) rather than a single value v. Value queries can be handled by indexing on the values reported in the summaries. Specifically, if a sensor reports a numerical value, then the index is constructed on these values. A search involves finding matching values that are either contained in the search range (v1, v2) or match the search value v exactly. Hybrid value and spatio-temporal queries are also possible. Such queries specify a time interval, a value range and a spatial region and request all records that match these attributes - find all instances where the temperature exceeded 100o F at location R during the month of August. These queries require an index on both time and value. In TSAR our focus is on range queries on value or time, with planned extensions to include spatial scoping. 2.3 Design Principles Our design of a sensor storage system for multi-tier networks is based on the following set of principles, which address the issues arising from the system and usage models above. • Principle 1: Store locally, access globally: Current technology allows local storage to be significantly more energyefficient than network communication, while technology trends show no signs of erasing this gap in the near future. For maximum network life a sensor storage system should leverage the flash memory on sensors to archive data locally, substituting cheap memory operations for expensive radio transmission. But without efficient mechanisms for retrieval, the energy gains of local storage may be outweighed by communication costs incurred by the application in searching for data. We believe that if the data storage system provides the abstraction of a single logical store to applications, as 41 does TSAR, then it will have additional flexibility to optimize communication and storage costs. • Principle 2: Distinguish data from metadata: Data must be identified so that it may be retrieved by the application without exhaustive search. To do this, we associate metadata with each data record - data fields of known syntax which serve as identifiers and may be queried by the storage system. Examples of this metadata are data attributes such as location and time, or selected or summarized data values. We leverage the presence of resource-rich proxies to index metadata for resource-constrained sensors. The proxies share this metadata index to provide a unified logical view of all data in the system, thereby enabling efficient, low-latency lookups. Such a tier-specific separation of data storage from metadata indexing enables the system to exploit the idiosyncrasies of multi-tier networks, while improving performance and functionality. • Principle 3: Provide data-centric query support: In a sensor application the specific location (i.e. offset) of a record in a stream is unlikely to be of significance, except if it conveys information concerning the location and/or time at which the information was generated. We thus expect that applications will be best served by a query interface which allows them to locate data by value or attribute (e.g. location and time), rather than a read interface for unstructured data. This in turn implies the need to maintain metadata in the form of an index that provides low cost lookups. 2.4 System Design TSAR embodies these design principles by employing local storage at sensors and a distributed index at the proxies. The key features of the system design are as follows: In TSAR, writes occur at sensor nodes, and are assumed to consist of both opaque data as well as application-specific metadata. This metadata is a tuple of known types, which may be used by the application to locate and identify data records, and which may be searched on and compared by TSAR in the course of locating data for the application. In a camera-based sensing application, for instance, this metadata might include coordinates describing the field of view, average luminance, and motion values, in addition to basic information such as time and sensor location. Depending on the application, this metadata may be two or three orders of magnitude smaller than the data itself, for instance if the metadata consists of features extracted from image or acoustic data. In addition to storing data locally, each sensor periodically sends a summary of reported metadata to a nearby proxy. The summary contains information such as the sensor ID, the interval (t1, t2) over which the summary was generated, a handle identifying the corresponding data record (e.g. its location in flash memory), and a coarse-grain representation of the metadata associated with the record. The precise data representation used in the summary is application-specific; for instance, a temperature sensor might choose to report the maximum and minimum temperature values observed in an interval as a coarse-grain representation of the actual time series. The proxy uses the summary to construct an index; the index is global in that it stores information from all sensors in the system and it is distributed across the various proxies in the system. Thus, applications see a unified view of distributed data, and can query the index at any proxy to get access to data stored at any sensor. Specifically, each query triggers lookups in this distributed index and the list of matches is then used to retrieve the corresponding data from the sensors. There are several distributed index and lookup methods which might be used in this system; however, the index structure described in Section 3 is highly suited for the task. Since the index is constructed using a coarse-grain summary, instead of the actual data, index lookups will yield approximate matches. The TSAR summarization mechanism guarantees that index lookups will never yield false negatives - i.e. it will never miss summaries which include the value being searched for. However, index lookups may yield false positives, where a summary matches the query but when queried the remote sensor finds no matching value, wasting network resources. The more coarse-grained the summary, the lower the update overhead and the greater the fraction of false positives, while finer summaries incur update overhead while reducing query overhead due to false positives. Remote sensors may easily distinguish false positives from queries which result in search hits, and calculate the ratio between the two; based on this ratio, TSAR employs a novel adaptive technique that dynamically varies the granularity of sensor summaries to balance the metadata overhead and the overhead of false positives. 3. Data Structures At the proxy tier, TSAR employs a novel index structure called the Interval Skip Graph, which is an ordered, distributed data structure for finding all intervals that contain a particular point or range of values. Interval skip graphs combine Interval Trees [5], an interval-based binary search tree, with Skip Graphs [1], a ordered, distributed data structure for peer-to-peer systems [13]. The resulting data structure has two properties that make it ideal for sensor networks. First, it has O(log n) search complexity for accessing the first interval that matches a particular value or range, and constant complexity for accessing each successive interval. Second, indexing of intervals rather than individual values makes the data structure ideal for indexing summaries over time or value. Such summary-based indexing is a more natural fit for energyconstrained sensor nodes, since transmitting summaries incurs less energy overhead than transmitting all sensor data. Definitions: We assume that there are Np proxies and Ns sensors in a two-tier sensor network. Each proxy is responsible for multiple sensor nodes, and no assumption is made about the number of sensors per proxy. Each sensor transmits interval summaries of data or events regularly to one or more proxies that it is associated with, where interval i is represented as [lowi, highi]. These intervals can correspond to time or value ranges that are used for indexing sensor data. No assumption is made about the size of an interval or about the amount of overlap between intervals. Range queries on the intervals are posed by users to the network of proxies and sensors; each query q needs to determine all index values that overlap the interval [lowq, highq]. The goal of the interval skip graph is to index all intervals such that the set that overlaps a query interval can be located efficiently. In the rest of this section, we describe the interval skip graph in greater detail. 3.1 Skip Graph Overview In order to inform the description of the Interval Skip Graph, we first provide a brief overview of the Skip Graph data structure; for a more extensive description the reader is referred to [1]. Figure 2 shows a skip graph which indexes 8 keys; the keys may be seen along the bottom, and above each key are the pointers associated with that key. Each data element, consisting of a key and its associated pointers, may reside on a different node in the network, 42 7 9 13 17 21 25 311 level 0 level 1 level 2 key single skip graph element (each may be on different node) find(21) node-to-node messages Figure 2: Skip Graph of 8 Elements [6,14] [9,12] [14,16] [15,23] [18,19] [20,27] [21,30][2,5] 5 14 14 16 23 23 27 30 [low,high] max contains(13) match no match halt Figure 3: Interval Skip Graph [6,14] [9,12] [14,16] [15,23] [18,19] [20,27] [21,30] [2,5] Node 1 Node 2 Node 3 level 2 level 1 level 0 Figure 4: Distributed Interval Skip Graph and pointers therefore identify both a remote node as well as a data element on that node. In this figure we may see the following properties of a skip graph: • Ordered index: The keys are members of an ordered data type, for instance integers. Lookups make use of ordered comparisons between the search key and existing index entries. In addition, the pointers at the lowest level point directly to the successor of each item in the index. • In-place indexing: Data elements remain on the nodes where they were inserted, and messages are sent between nodes to establish links between those elements and others in the index. • Log n height: There are log2 n pointers associated with each element, where n is the number of data elements indexed. Each pointer belongs to a level l in [0... log2 n − 1], and together with some other pointers at that level forms a chain of n/2l elements. • Probabilistic balance: Rather than relying on re-balancing operations which may be triggered at insert or delete, skip graphs implement a simple random balancing mechanism which maintains close to perfect balance on average, with an extremely low probability of significant imbalance. • Redundancy and resiliency: Each data element forms an independent search tree root, so searches may begin at any node in the network, eliminating hot spots at a single search root. In addition the index is resilient against node failure; data on the failed node will not be accessible, but remaining data elements will be accessible through search trees rooted on other nodes. In Figure 2 we see the process of searching for a particular value in a skip graph. The pointers reachable from a single data element form a binary tree: a pointer traversal at the highest level skips over n/2 elements, n/4 at the next level, and so on. Search consists of descending the tree from the highest level to level 0, at each level comparing the target key with the next element at that level and deciding whether or not to traverse. In the perfectly balanced case shown here there are log2 n levels of pointers, and search will traverse 0 or 1 pointers at each level. We assume that each data element resides on a different node, and measure search cost by the number messages sent (i.e. the number of pointers traversed); this will clearly be O(log n). Tree update proceeds from the bottom, as in a B-Tree, with the root(s) being promoted in level as the tree grows. In this way, for instance, the two chains at level 1 always contain n/2 entries each, and there is never a need to split chains as the structure grows. The update process then consists of choosing which of the 2l chains to insert an element into at each level l, and inserting it in the proper place in each chain. Maintaining a perfectly balanced skip graph as shown in Figure 2 would be quite complex; instead, the probabilistic balancing method introduced in Skip Lists [23] is used, which trades off a small amount of overhead in the expected case in return for simple update and deletion. The basis for this method is the observation that any element which belongs to a particular chain at level l can only belong to one of two chains at level l+1. To insert an element we ascend levels starting at 0, randomly choosing one of the two possible chains at each level, an stopping when we reach an empty chain. One means of implementation (e.g. as described in [1]) is to assign each element an arbitrarily long random bit string. Each chain at level l is then constructed from those elements whose bit strings match in the first l bits, thus creating 2l possible chains at each level and ensuring that each chain splits into exactly two chains at the next level. Although the resulting structure is not perfectly balanced, following the analysis in [23] we can show that the probability of it being significantly out of balance is extremely small; in addition, since the structure is determined by the random number stream, input data patterns cannot cause the tree to become imbalanced. 3.2 Interval Skip Graph A skip graph is designed to store single-valued entries. In this section, we introduce a novel data structure that extends skip graphs to store intervals [lowi, highi] and allows efficient searches for all intervals covering a value v, i.e. {i : lowi ≤ v ≤ highi}. Our data structure can be extended to range searches in a straightforward manner. The interval skip graph is constructed by applying the method of augmented search trees, as described by Cormen, Leiserson, and Rivest [5] and applied to binary search trees to create an Interval Tree. The method is based on the observation that a search structure based on comparison of ordered keys, such as a binary tree, may also be used to search on a secondary key which is non-decreasing in the first key. Given a set of intervals sorted by lower bound - lowi ≤ lowi+1 - we define the secondary key as the cumulative maximum, maxi = maxk=0...i (highk). The set of intervals intersecting a value v may then be found by searching for the first interval (and thus the interval with least lowi) such that maxi ≥ v. We then 43 traverse intervals in increasing order lower bound, until we find the first interval with lowi > v, selecting those intervals which intersect v. Using this approach we augment the skip graph data structure, as shown in Figure 3, so that each entry stores a range (lower bound and upper bound) and a secondary key (cumulative maximum of upper bound). To efficiently calculate the secondary key maxi for an entry i, we take the greatest of highi and the maximum values reported by each of i"s left-hand neighbors. To search for those intervals containing the value v, we first search for v on the secondary index, maxi, and locate the first entry with maxi ≥ v. (by the definition of maxi, for this data element maxi = highi.) If lowi > v, then this interval does not contain v, and no other intervals will, either, so we are done. Otherwise we traverse the index in increasing order of mini, returning matching intervals, until we reach an entry with mini > v and we are done. Searches for all intervals which overlap a query range, or which completely contain a query range, are straightforward extensions of this mechanism. Lookup Complexity: Lookup for the first interval that matches a given value is performed in a manner very similar to an interval tree. The complexity of search is O(log n). The number of intervals that match a range query can vary depending on the amount of overlap in the intervals being indexed, as well as the range specified in the query. Insert Complexity: In an interval tree or interval skip list, the maximum value for an entry need only be calculated over the subtree rooted at that entry, as this value will be examined only when searching within the subtree rooted at that entry. For a simple interval skip graph, however, this maximum value for an entry must be computed over all entries preceding it in the index, as searches may begin anywhere in the data structure, rather than at a distinguished root element. It may be easily seen that in the worse case the insertion of a single interval (one that covers all existing intervals in the index) will trigger the update of all entries in the index, for a worst-case insertion cost of O(n). 3.3 Sparse Interval Skip Graph The final extensions we propose take advantage of the difference between the number of items indexed in a skip graph and the number of systems on which these items are distributed. The cost in network messages of an operation may be reduced by arranging the data structure so that most structure traversals occur locally on a single node, and thus incur zero network cost. In addition, since both congestion and failure occur on a per-node basis, we may eliminate links without adverse consequences if those links only contribute to load distribution and/or resiliency within a single node. These two modifications allow us to achieve reductions in asymptotic complexity of both update and search. As may be in Section 3.2, insert and delete cost on an interval skip graph has a worst case complexity of O(n), compared to O(log n) for an interval tree. The main reason for the difference is that skip graphs have a full search structure rooted at each element, in order to distribute load and provide resilience to system failures in a distributed setting. However, in order to provide load distribution and failure resilience it is only necessary to provide a full search structure for each system. If as in TSAR the number of nodes (proxies) is much smaller than the number of data elements (data summaries indexed), then this will result in significant savings. Implementation: To construct a sparse interval skip graph, we ensure that there is a single distinguished element on each system, the root element for that system; all searches will start at one of these root elements. When adding a new element, rather than splitting lists at increasing levels l until the element is in a list with no others, we stop when we find that the element would be in a list containing no root elements, thus ensuring that the element is reachable from all root elements. An example of applying this optimization may be seen in Figure 5. (In practice, rather than designating existing data elements as roots, as shown, it may be preferable to insert null values at startup.) When using the technique of membership vectors as in [1], this may be done by broadcasting the membership vectors of each root element to all other systems, and stopping insertion of an element at level l when it does not share an l-bit prefix with any of the Np root elements. The expected number of roots sharing a log2Np-bit prefix is 1, giving an expected expected height for each element of log2Np +O(1). An alternate implementation, which distributes information concerning root elements at pointer establishment time, is omitted due to space constraints; this method eliminates the need for additional messages. Performance: In a (non-interval) sparse skip graph, since the expected height of an inserted element is now log2 Np + O(1), expected insertion complexity is O(log Np), rather than O(log n), where Np is the number of root elements and thus the number of separate systems in the network. (In the degenerate case of a single system we have a skip list; with splitting probability 0.5 the expected height of an individual element is 1.) Note that since searches are started at root elements of expected height log2 n, search complexity is not improved. For an interval sparse skip graph, update performance is improved considerably compared to the O(n) worst case for the nonsparse case. In an augmented search structure such as this, an element only stores information for nodes which may be reached from that element-e.g. the subtree rooted at that element, in the case of a tree. Thus, when updating the maximum value in an interval tree, the update is only propagated towards the root. In a sparse interval skip graph, updates to a node only propagate towards the Np root elements, for a worst-case cost of Np log2 n. Shortcut search: When beginning a search for a value v, rather than beginning at the root on that proxy, we can find the element that is closest to v (e.g. using a secondary local index), and then begin the search at that element. The expected distance between this element and the search terminus is log2 Np, and the search will now take on average log2 Np + O(1) steps. To illustrate this optimization, in Figure 4 depending on the choice of search root, a search for [21, 30] beginning at node 2 may take 3 network hops, traversing to node 1, then back to node 2, and finally to node 3 where the destination is located, for a cost of 3 messages. The shortcut search, however, locates the intermediate data element on node 2, and then proceeds directly to node 3 for a cost of 1 message. Performance: This technique may be applied to the primary key search which is the first of two insertion steps in an interval skip graph. By combining the short-cut optimization with sparse interval skip graphs, the expected cost of insertion is now O(log Np), independent of the size of the index or the degree of overlap of the inserted intervals. 3.4 Alternative Data Structures Thus far we have only compared the sparse interval skip graph with similar structures from which it is derived. A comparison with several other data structures which meet at least some of the requirements for the TSAR index is shown in Table 2. 44 Table 2: Comparison of Distributed Index Structures Range Query Support Interval Representation Re-balancing Resilience Small Networks Large Networks DHT, GHT no no no yes good good Local index, flood query yes yes no yes good bad P-tree, RP* (distributed B-Trees) yes possible yes no good good DIMS yes no yes yes yes yes Interval Skipgraph yes yes no yes good good [6,14] [9,12] [14,16] [15,23] [18,19] [20,27] [21,30][2,5] Roots Node 1 Node 2 Figure 5: Sparse Interval Skip Graph The hash-based systems, DHT [25] and GHT [26], lack the ability to perform range queries and are thus not well-suited to indexing spatio-temporal data. Indexing locally using an appropriate singlenode structure and then flooding queries to all proxies is a competitive alternative for small networks; for large networks the linear dependence on the number of proxies becomes an issue. Two distributed B-Trees were examined - P-Trees [6] and RP* [19]. Each of these supports range queries, and in theory could be modified to support indexing of intervals; however, they both require complex re-balancing, and do not provide the resilience characteristics of the other structures. DIMS [18] provides the ability to perform spatio-temporal range queries, and has the necessary resilience to failures; however, it cannot be used index intervals, which are used by TSAR"s data summarization algorithm. 4. Data Storage and Summarization Having described the proxy-level index structure, we turn to the mechanisms at the sensor tier. TSAR implements two key mechanisms at the sensor tier. The first is a local archival store at each sensor node that is optimized for resource-constrained devices. The second is an adaptive summarization technique that enables each sensor to adapt to changing data and query characteristics. The rest of this section describes these mechanisms in detail. 4.1 Local Storage at Sensors Interval skip graphs provide an efficient mechanism to lookup sensor nodes containing data relevant to a query. These queries are then routed to the sensors, which locate the relevant data records in the local archive and respond back to the proxy. To enable such lookups, each sensor node in TSAR maintains an archival store of sensor data. While the implementation of such an archival store is straightforward on resource-rich devices that can run a database, sensors are often power and resource-constrained. Consequently, the sensor archiving subsystem in TSAR is explicitly designed to exploit characteristics of sensor data in a resource-constrained setting. Timestamp Calibration Parameters Opaque DataData/Event Attributes size Figure 6: Single storage record Sensor data has very distinct characteristics that inform our design of the TSAR archival store. Sensors produce time-series data streams, and therefore, temporal ordering of data is a natural and simple way of storing archived sensor data. In addition to simplicity, a temporally ordered store is often suitable for many sensor data processing tasks since they involve time-series data processing. Examples include signal processing operations such as FFT, wavelet transforms, clustering, similarity matching, and target detection. Consequently, the local archival store is a collection of records, designed as an append-only circular buffer, where new records are appended to the tail of the buffer. The format of each data record is shown in Figure 6. Each record has a metadata field which includes a timestamp, sensor settings, calibration parameters, etc. Raw sensor data is stored in the data field of the record. The data field is opaque and application-specific-the storage system does not know or care about interpreting this field. A camera-based sensor, for instance, may store binary images in this data field. In order to support a variety of applications, TSAR supports variable-length data fields; as a result, record sizes can vary from one record to another. Our archival store supports three operations on records: create, read, and delete. Due to the append-only nature of the store, creation of records is simple and efficient. The create operation simply creates a new record and appends it to the tail of the store. Since records are always written at the tail, the store need not maintain a free space list. All fields of the record need to be specified at creation time; thus, the size of the record is known a priori and the store simply allocates the the corresponding number of bytes at the tail to store the record. Since writes are immutable, the size of a record does not change once it is created. proxy proxy proxy record 3 record summary local archive in flash memory data summary start,end offset time interval sensor summary sent to proxy Insert summaries into interval skip graph Figure 7: Sensor Summarization 45 The read operation enables stored records to be retrieved in order to answer queries. In a traditional database system, efficient lookups are enabled by maintaining a structure such as a B-tree that indexes certain keys of the records. However, this can be quite complex for a small sensor node with limited resources. Consequently, TSAR sensors do not maintain any index for the data stored in their archive. Instead, they rely on the proxies to maintain this metadata index-sensors periodically send the proxy information summarizing the data contained in a contiguous sequence of records, as well as a handle indicating the location of these records in flash memory. The mechanism works as follows: In addition to the summary of sensor data, each node sends metadata to the proxy containing the time interval corresponding to the summary, as well as the start and end offsets of the flash memory location where the raw data corresponding is stored (as shown in Figure 7). Thus, random access is enabled at granularity of a summary-the start offset of each chunk of records represented by a summary is known to the proxy. Within this collection, records are accessed sequentially. When a query matches a summary in the index, the sensor uses these offsets to access the relevant records on its local flash by sequentially reading data from the start address until the end address. Any queryspecific operation can then be performed on this data. Thus, no index needs to be maintained at the sensor, in line with our goal of simplifying sensor state management. The state of the archive is captured in the metadata associated with the summaries, and is stored and maintained at the proxy. While we anticipate local storage capacity to be large, eventually there might be a need to overwrite older data, especially in high data rate applications. This may be done via techniques such as multi-resolution storage of data [9], or just simply by overwriting older data. When older data is overwritten, a delete operation is performed, where an index entry is deleted from the interval skip graph at the proxy and the corresponding storage space in flash memory at the sensor is freed. 4.2 Adaptive Summarization The data summaries serve as glue between the storage at the remote sensor and the index at the proxy. Each update from a sensor to the proxy includes three pieces of information: the summary, a time period corresponding to the summary, and the start and end offsets for the flash archive. In general, the proxy can index the time interval representing a summary or the value range reported in the summary (or both). The former index enables quick lookups on all records seen during a certain interval, while the latter index enables quick lookups on all records matching a certain value. As described in Section 2.4, there is a trade-off between the energy used in sending summaries (and thus the frequency and resolution of those summaries) and the cost of false hits during queries. The coarser and less frequent the summary information, the less energy required, while false query hits in turn waste energy on requests for non-existent data. TSAR employs an adaptive summarization technique that balances the cost of sending updates against the cost of false positives. The key intuition is that each sensor can independently identify the fraction of false hits and true hits for queries that access its local archive. If most queries result in true hits, then the sensor determines that the summary can be coarsened further to reduce update costs without adversely impacting the hit ratio. If many queries result in false hits, then the sensor makes the granularity of each summary finer to reduce the number and overhead of false hits. The resolution of the summary depends on two parametersthe interval over which summaries of the data are constructed and transmitted to the proxy, as well as the size of the applicationspecific summary. Our focus in this paper is on the interval over which the summary is constructed. Changing the size of the data summary can be performed in an application-specific manner (e.g. using wavelet compression techniques as in [9]) and is beyond the scope of this paper. Currently, TSAR employs a simple summarization scheme that computes the ratio of false and true hits and decreases (increases) the interval between summaries whenever this ratio increases (decreases) beyond a threshold. 5. TSAR Implementation We have implemented a prototype of TSAR on a multi-tier sensor network testbed. Our prototype employs Crossbow Stargate nodes to implement the proxy tier. Each Stargate node employs a 400MHz Intel XScale processor with 64MB RAM and runs the Linux 2.4.19 kernel and EmStar release 2.1. The proxy nodes are equipped with two wireless radios, a Cisco Aironet 340-based 802.11b radio and a hostmote bridge to the Mica2 sensor nodes using the EmStar transceiver. The 802.11b wireless network is used for inter-proxy communication within the proxy tier, while the wireless bridge enables sensor-proxy communication. The sensor tier consists of Crossbow Mica2s and Mica2dots, each consisting of a 915MHz CC1000 radio, a BMAC protocol stack, a 4 Mb on-board flash memory and an ATMega 128L processor. The sensor nodes run TinyOS 1.1.8. In addition to the on-board flash, the sensor nodes can be equipped with external MMC/SD flash cards using a custom connector. The proxy nodes can be equipped with external storage such as high-capacity compact flash (up to 4GB), 6GB micro-drives, or up to 60GB 1.8inch mobile disk drives. Since sensor nodes may be several hops away from the nearest proxy, the sensor tier employs multi-hop routing to communicate with the proxy tier. In addition, to reduce the power consumption of the radio while still making the sensor node available for queries, low power listening is enabled, in which the radio receiver is periodically powered up for a short interval to sense the channel for transmissions, and the packet preamble is extended to account for the latency until the next interval when the receiving radio wakes up. Our prototype employs the MultiHopLEPSM routing protocol with the BMAC layer configured in the low-power mode with a 11% duty cycle (one of the default BMAC [22] parameters) Our TSAR implementation on the Mote involves a data gathering task that periodically obtains sensor readings and logs these reading to flash memory. The flash memory is assumed to be a circular append-only store and the format of the logged data is depicted in Figure 6. The Mote sends a report to the proxy every N readings, summarizing the observed data. The report contains: (i) the address of the Mote, (ii) a handle that contains an offset and the length of the region in flash memory containing data referred to by the summary, (iii) an interval (t1, t2) over which this report is generated, (iv) a tuple (low, high) representing the minimum and the maximum values observed at the sensor in the interval, and (v) a sequence number. The sensor updates are used to construct a sparse interval skip graph that is distributed across proxies, via network messages between proxies over the 802.11b wireless network. Our current implementation supports queries that request records matching a time interval (t1, t2) or a value range (v1, v2). Spatial constraints are specified using sensor IDs. Given a list of matching intervals from the skip graph, TSAR supports two types of messages to query the sensor: lookup and fetch. A lookup message triggers a search within the corresponding region in flash memory and returns the number of matching records in that memory region (but does not retrieve data). In contrast, a fetch message not only 46 0 10 20 30 40 50 60 70 80 128512 1024 2048 4096 NumberofMessages Index size (entries) Insert (skipgraph) Insert (sparse skipgraph) Initial lookup (a) James Reserve Data 0 10 20 30 40 50 60 70 80 512 1024 2048 4096 NumberofMessages Index size (entries) Insert (skipgraph) Insert (sparse skipgraph) Initial lookup (b) Synthetic Data Figure 8: Skip Graph Insert Performance triggers a search but also returns all matching data records to the proxy. Lookup messages are useful for polling a sensor, for instance, to determine if a query matches too many records. 6. Experimental Evaluation In this section, we evaluate the efficacy of TSAR using our prototype and simulations. The testbed for our experiments consists of four Stargate proxies and twelve Mica2 and Mica2dot sensors; three sensors each are assigned to each proxy. Given the limited size of our testbed, we employ simulations to evaluate the behavior of TSAR in larger settings. Our simulation employs the EmTOS emulator [10], which enables us to run the same code in simulation and the hardware platform. Rather than using live data from a real sensor, to ensure repeatable experiments, we seed each sensor node with a dataset (i.e., a trace) that dictates the values reported by that node to the proxy. One section of the flash memory on each sensor node is programmed with data points from the trace; these observations are then replayed during an experiment, logged to the local archive (located in flash memory, as well), and reported to the proxy. The first dataset used to evaluate TSAR is a temperature dataset from James Reserve [27] that includes data from eleven temperature sensor nodes over a period of 34 days. The second dataset is synthetically generated; the trace for each sensor is generated using a uniformly distributed random walk though the value space. Our experimental evaluation has four parts. First, we run EmTOS simulations to evaluate the lookup, update and delete overhead for sparse interval skip graphs using the real and synthetic datasets. Second, we provide summary results from micro-benchmarks of the storage component of TSAR, which include empirical characterization of the energy costs and latency of reads and writes for the flash memory chip as well as the whole mote platform, and comparisons to published numbers for other storage and communication technologies. These micro-benchmarks form the basis for our full-scale evaluation of TSAR on a testbed of four Stargate proxies and twelve Motes. We measure the end-to-end query latency in our multi-hop testbed as well as the query processing overhead at the mote tier. Finally, we demonstrate the adaptive summarization capability at each sensor node. The remainder of this section presents our experimental results. 6.1 Sparse Interval Skip Graph Performance This section evaluates the performance of sparse interval skip graphs by quantifying insert, lookup and delete overheads. We assume a proxy tier with 32 proxies and construct sparse interval skip graphs of various sizes using our datasets. For each skip 0 5 10 15 20 25 30 35 409620481024512 NumberofMessages Index size (entries) Initial lookup Traversal (a) James Reserve Data 0 2 4 6 8 10 12 14 409620481024512 NumberofMessages Index size (entries) Initial lookup Traversal (b) Synthetic Data Figure 9: Skip Graph Lookup Performance 0 10 20 30 40 50 60 70 1 4 8 16 24 32 48 Numberofmessages Number of proxies Skipgraph insert Sparse skipgraph insert Initial lookup (a) Impact of Number of Proxies 0 20 40 60 80 100 120 512 1024 2048 4096 NumberofMessages Index size (entries) Insert (redundant) Insert (non-redundant) Lookup (redundant) Lookup (non-redundant) (b) Impact of Redundant Summaries Figure 10: Skip Graph Overheads graph, we evaluate the cost of inserting a new value into the index. Each entry was deleted after its insertion, enabling us to quantify the delete overhead as well. Figure 8(a) and (b) quantify the insert overhead for our two datasets: each insert entails an initial traversal that incurs log n messages, followed by neighbor pointer update at increasing levels, incurring a cost of 4 log n messages. Our results demonstrate this behavior, and show as well that performance of delete-which also involves an initial traversal followed by pointer updates at each level-incurs a similar cost. Next, we evaluate the lookup performance of the index structure. Again, we construct skip graphs of various sizes using our datasets and evaluate the cost of a lookup on the index structure. Figures 9(a) and (b) depict our results. There are two components for each lookup-the lookup of the first interval that matches the query and, in the case of overlapping intervals, the subsequent linear traversal to identify all matching intervals. The initial lookup can be seen to takes log n messages, as expected. The costs of the subsequent linear traversal, however, are highly data dependent. For instance, temperature values for the James Reserve data exhibit significant spatial correlations, resulting in significant overlap between different intervals and variable, high traversal cost (see Figure 9(a)). The synthetic data, however, has less overlap and incurs lower traversal overhead as shown in Figure 9(b). Since the previous experiments assumed 32 proxies, we evaluate the impact of the number of proxies on skip graph performance. We vary the number of proxies from 10 to 48 and distribute a skip graph with 4096 entries among these proxies. We construct regular interval skip graphs as well as sparse interval skip graphs using these entries and measure the overhead of inserts and lookups. Thus, the experiment also seeks to demonstrate the benefits of sparse skip graphs over regular skip graphs. Figure 10(a) depicts our results. In regular skip graphs, the complexity of insert is O(log2n) in the 47 expected case (and O(n) in the worst case) where n is the number of elements. This complexity is unaffected by changing the number of proxies, as indicated by the flat line in the figure. Sparse skip graphs require fewer pointer updates; however, their overhead is dependent on the number of proxies, and is O(log2Np) in the expected case, independent of n. This can be seen to result in significant reduction in overhead when the number of proxies is small, which decreases as the number of proxies increases. Failure handling is an important issue in a multi-tier sensor architecture since it relies on many components-proxies, sensor nodes and routing nodes can fail, and wireless links can fade. Handling of many of these failure modes is outside the scope of this paper; however, we consider the case of resilience of skip graphs to proxy failures. In this case, skip graph search (and subsequent repair operations) can follow any one of the other links from a root element. Since a sparse skip graph has search trees rooted at each node, searching can then resume once the lookup request has routed around the failure. Together, these two properties ensure that even if a proxy fails, the remaining entries in the skip graph will be reachable with high probability-only the entries on the failed proxy and the corresponding data at the sensors becomes inaccessible. To ensure that all data on sensors remains accessible, even in the event of failure of a proxy holding index entries for that data, we incorporate redundant index entries. TSAR employs a simple redundancy scheme where additional coarse-grain summaries are used to protect regular summaries. Each sensor sends summary data periodically to its local proxy, but less frequently sends a lowerresolution summary to a backup proxy-the backup summary represents all of the data represented by the finer-grained summaries, but in a lossier fashion, thus resulting in higher read overhead (due to false hits) if the backup summary is used. The cost of implementing this in our system is low - Figure 10(b) shows the overhead of such a redundancy scheme, where a single coarse summary is send to a backup for every two summaries sent to the primary proxy. Since a redundant summary is sent for every two summaries, the insert cost is 1.5 times the cost in the normal case. However, these redundant entries result in only a negligible increase in lookup overhead, due the logarithmic dependence of lookup cost on the index size, while providing full resilience to any single proxy failure. 6.2 Storage Microbenchmarks Since sensors are resource-constrained, the energy consumption and the latency at this tier are important measures for evaluating the performance of a storage architecture. Before performing an endto-end evaluation of our system, we provide more detailed information on the energy consumption of the storage component used to implement the TSAR local archive, based on empirical measurements. In addition we compare these figures to those for other local storage technologies, as well as to the energy consumption of wireless communication, using information from the literature. For empirical measurements we measure energy usage for the storage component itself (i.e. current drawn by the flash chip), as well as for the entire Mica2 mote. The power measurements in Table 3 were performed for the AT45DB041 [15] flash memory on a Mica2 mote, which is an older NOR flash device. The most promising technology for low-energy storage on sensing devices is NAND flash, such as the Samsung K9K4G08U0M device [16]; published power numbers for this device are provided in the table. Published energy requirements for wireless transmission using the Chipcon [4] CC2420 radio (used in MicaZ and Telos motes) are provided for comparison, assuming Energy Energy/byte Mote flash Read 256 byte page 58µJ* / 136µJ* total 0.23µJ* Write 256 byte page 926µJ* / 1042µJ* total 3.6µJ* NAND Flash Read 512 byte page 2.7µJ 1.8nJ Write 512 byte page 7.8µJ 15nJ Erase 16K byte sector 60µJ 3.7nJ CC2420 radio Transmit 8 bits (-25dBm) 0.8µJ 0.8µJ Receive 8 bits 1.9µJ 1.9µJ Mote AVR processor In-memory search, 256 bytes 1.8µJ 6.9nJ Table 3: Storage and Communication Energy Costs (*measured values) 0 200 400 600 800 1000 1 2 3 Latency(ms) Number of hops (a) Multi-hop query performance 0 100 200 300 400 500 1 5121024 2048 4096 Latency(ms) Index size (entries) Sensor communication Proxy communication Sensor lookup, processing (b) Query Performance Figure 11: Query Processing Latency zero network and protocol overhead. Comparing the total energy cost for writing flash (erase + write) to the total cost for communication (transmit + receive), we find that the NAND flash is almost 150 times more efficient than radio communication, even assuming perfect network protocols. 6.3 Prototype Evaluation This section reports results from an end-to-end evaluation of the TSAR prototype involving both tiers. In our setup, there are four proxies connected via 802.11 links and three sensors per proxy. The multi-hop topology was preconfigured such that sensor nodes were connected in a line to each proxy, forming a minimal tree of depth 0 400 800 1200 1600 0 20 40 60 80 100 120 140 160 Retrievallatency(ms) Archived data retrieved (bytes) (a) Data Query and Fetch Time 0 2 4 6 8 10 12 4 8 16 32 Latency(ms) Number of 34-byte records searched (b) Sensor query processing delay Figure 12: Query Latency Components 48 3. Due to resource constraints we were unable to perform experiments with dozens of sensor nodes, however this topology ensured that the network diameter was as large as for a typical network of significantly larger size. Our evaluation metric is the end-to-end latency of query processing. A query posed on TSAR first incurs the latency of a sparse skip graph lookup, followed by routing to the appropriate sensor node(s). The sensor node reads the required page(s) from its local archive, processes the query on the page that is read, and transmits the response to the proxy, which then forwards it to the user. We first measure query latency for different sensors in our multi-hop topology. Depending on which of the sensors is queried, the total latency increases almost linearly from about 400ms to 1 second, as the number of hops increases from 1 to 3 (see Figure 11(a)). Figure 11(b) provides a breakdown of the various components of the end-to-end latency. The dominant component of the total latency is the communication over one or more hops. The typical time to communicate over one hop is approximately 300ms. This large latency is primarily due to the use of a duty-cycled MAC layer; the latency will be larger if the duty cycle is reduced (e.g. the 2% setting as opposed to the 11.5% setting used in this experiment), and will conversely decrease if the duty cycle is increased. The figure also shows the latency for varying index sizes; as expected, the latency of inter-proxy communication and skip graph lookups increases logarithmically with index size. Not surprisingly, the overhead seen at the sensor is independent of the index size. The latency also depends on the number of packets transmitted in response to a query-the larger the amount of data retrieved by a query, the greater the latency. This result is shown in Figure 12(a). The step function is due to packetization in TinyOS; TinyOS sends one packet so long as the payload is smaller than 30 bytes and splits the response into multiple packets for larger payloads. As the data retrieved by a query is increased, the latency increases in steps, where each step denotes the overhead of an additional packet. Finally, Figure 12(b) shows the impact of searching and processing flash memory regions of increasing sizes on a sensor. Each summary represents a collection of records in flash memory, and all of these records need to be retrieved and processed if that summary matches a query. The coarser the summary, the larger the memory region that needs to be accessed. For the search sizes examined, amortization of overhead when searching multiple flash pages and archival records, as well as within the flash chip and its associated driver, results in the appearance of sub-linear increase in latency with search size. In addition, the operation can be seen to have very low latency, in part due to the simplicity of our query processing, requiring only a compare operation with each stored element. More complex operations, however, will of course incur greater latency. 6.4 Adaptive Summarization When data is summarized by the sensor before being reported to the proxy, information is lost. With the interval summarization method we are using, this information loss will never cause the proxy to believe that a sensor node does not hold a value which it in fact does, as all archived values will be contained within the interval reported. However, it does cause the proxy to believe that the sensor may hold values which it does not, and forward query messages to the sensor for these values. These false positives constitute the cost of the summarization mechanism, and need to be balanced against the savings achieved by reducing the number of reports. The goal of adaptive summarization is to dynamically vary the summary size so that these two costs are balanced. 0 0.1 0.2 0.3 0.4 0.5 0 5 10 15 20 25 30 Fractionoftruehits Summary size (number of records) (a) Impact of summary size 0 5 10 15 20 25 30 35 0 5000 10000 15000 20000 25000 30000 Summarizationsize(num.records) Normalized time (units) query rate 0.2 query rate 0.03 query rate 0.1 (b) Adaptation to query rate Figure 13: Impact of Summarization Granularity Figure 13(a) demonstrates the impact of summary granularity on false hits. As the number of records included in a summary is increased, the fraction of queries forwarded to the sensor which match data held on that sensor (true positives) decreases. Next, in Figure 13(b) we run the a EmTOS simulation with our adaptive summarization algorithm enabled. The adaptive algorithm increases the summary granularity (defined as the number of records per summary) when Cost(updates) Cost(falsehits) > 1 + and reduces it if Cost(updates) Cost(falsehits) > 1 − , where is a small constant. To demonstrate the adaptive nature of our technique, we plot a time series of the summarization granularity. We begin with a query rate of 1 query per 5 samples, decrease it to 1 every 30 samples, and then increase it again to 1 query every 10 samples. As shown in Figure 13(b), the adaptive technique adjusts accordingly by sending more fine-grain summaries at higher query rates (in response to the higher false hit rate), and fewer, coarse-grain summaries at lower query rates. 7. Related Work In this section, we review prior work on storage and indexing techniques for sensor networks. While our work addresses both problems jointly, much prior work has considered them in isolation. The problem of archival storage of sensor data has received limited attention in sensor network literature. ELF [7] is a logstructured file system for local storage on flash memory that provides load leveling and Matchbox is a simple file system that is packaged with the TinyOS distribution [14]. Both these systems focus on local storage, whereas our focus is both on storage at the remote sensors as well as providing a unified view of distributed data across all such local archives. Multi-resolution storage [9] is intended for in-network storage and search in systems where there is significant data in comparison to storage resources. In contrast, TSAR addresses the problem of archival storage in two-tier systems where sufficient resources can be placed at the edge sensors. The RISE platform [21] being developed as part of the NODE project at UCR addresses the issues of hardware platform support for large amounts of storage in remote sensor nodes, but not the indexing and querying of this data. In order to efficiently access a distributed sensor store, an index needs to be constructed of the data. Early work on sensor networks such as Directed Diffusion [17] assumes a system where all useful sensor data was stored locally at each sensor, and spatially scoped queries are routed using geographic co-ordinates to locations where the data is stored. Sources publish the events that they detect, and sinks with interest in specific events can subscribe to these events. The Directed Diffusion substrate routes queries to specific locations 49 if the query has geographic information embedded in it (e.g.: find temperature in the south-west quadrant), and if not, the query is flooded throughout the network. These schemes had the drawback that for queries that are not geographically scoped, search cost (O(n) for a network of n nodes) may be prohibitive in large networks with frequent queries. Local storage with in-network indexing approaches address this issue by constructing indexes using frameworks such as Geographic Hash Tables [24] and Quad Trees [9]. Recent research has seen a growing body of work on data indexing schemes for sensor networks[26][11][18]. One such scheme is DCS [26], which provides a hash function for mapping from event name to location. DCS constructs a distributed structure that groups events together spatially by their named type. Distributed Index of Features in Sensornets (DIFS [11]) and Multi-dimensional Range Queries in Sensor Networks (DIM [18]) extend the data-centric storage approach to provide spatially distributed hierarchies of indexes to data. While these approaches advocate in-network indexing for sensor networks, we believe that indexing is a task that is far too complicated to be performed at the remote sensor nodes since it involves maintaining significant state and large tables. TSAR provides a better match between resource requirements of storage and indexing and the availability of resources at different tiers. Thus complex operations such as indexing and managing metadata are performed at the proxies, while storage at the sensor remains simple. In addition to storage and indexing techniques specific to sensor networks, many distributed, peer-to-peer and spatio-temporal index structures are relevant to our work. DHTs [25] can be used for indexing events based on their type, quad-tree variants such as Rtrees [12] can be used for optimizing spatial searches, and K-D trees [2] can be used for multi-attribute search. While this paper focuses on building an ordered index structure for range queries, we will explore the use of other index structures for alternate queries over sensor data. 8. Conclusions In this paper, we argued that existing sensor storage systems are designed primarily for flat hierarchies of homogeneous sensor nodes and do not fully exploit the multi-tier nature of emerging sensor networks. We presented the design of TSAR, a fundamentally different storage architecture that envisions separation of data from metadata by employing local storage at the sensors and distributed indexing at the proxies. At the proxy tier, TSAR employs a novel multi-resolution ordered distributed index structure, the Sparse Interval Skip Graph, for efficiently supporting spatio-temporal and range queries. At the sensor tier, TSAR supports energy-aware adaptive summarization that can trade-off the energy cost of transmitting metadata to the proxies against the overhead of false hits resulting from querying a coarser resolution index structure. We implemented TSAR in a two-tier sensor testbed comprising Stargatebased proxies and Mote-based sensors. Our experimental evaluation of TSAR demonstrated the benefits and feasibility of employing our energy-efficient low-latency distributed storage architecture in multi-tier sensor networks. 9. REFERENCES [1] James Aspnes and Gauri Shah. Skip graphs. In Fourteenth Annual ACM-SIAM Symposium on Discrete Algorithms, pages 384-393, Baltimore, MD, USA, 12-14 January 2003. [2] Jon Louis Bentley. Multidimensional binary search trees used for associative searching. Commun. ACM, 18(9):509-517, 1975. [3] Philippe Bonnet, J. E. Gehrke, and Praveen Seshadri. Towards sensor database systems. In Proceedings of the Second International Conference on Mobile Data Management., January 2001. [4] Chipcon. CC2420 2.4 GHz IEEE 802.15.4 / ZigBee-ready RF transceiver, 2004. [5] Thomas H. Cormen, Charles E. Leiserson, Ronald L. Rivest, and Clifford Stein. Introduction to Algorithms. The MIT Press and McGraw-Hill, second edition edition, 2001. [6] Adina Crainiceanu, Prakash Linga, Johannes Gehrke, and Jayavel Shanmugasundaram. Querying Peer-to-Peer Networks Using P-Trees. Technical Report TR2004-1926, Cornell University, 2004. [7] Hui Dai, Michael Neufeld, and Richard Han. ELF: an efficient log-structured flash file system for micro sensor nodes. In SenSys "04: Proceedings of the 2nd international conference on Embedded networked sensor systems, pages 176-187, New York, NY, USA, 2004. ACM Press. [8] Peter Desnoyers, Deepak Ganesan, Huan Li, and Prashant Shenoy. PRESTO: A predictive storage architecture for sensor networks. In Tenth Workshop on Hot Topics in Operating Systems (HotOS X)., June 2005. [9] Deepak Ganesan, Ben Greenstein, Denis Perelyubskiy, Deborah Estrin, and John Heidemann. An evaluation of multi-resolution storage in sensor networks. In Proceedings of the First ACM Conference on Embedded Networked Sensor Systems (SenSys)., 2003. [10] L. Girod, T. Stathopoulos, N. Ramanathan, J. Elson, D. Estrin, E. Osterweil, and T. Schoellhammer. A system for simulation, emulation, and deployment of heterogeneous sensor networks. In Proceedings of the Second ACM Conference on Embedded Networked Sensor Systems, Baltimore, MD, 2004. [11] B. Greenstein, D. Estrin, R. Govindan, S. Ratnasamy, and S. Shenker. DIFS: A distributed index for features in sensor networks. Elsevier Journal of ad-hoc Networks, 2003. [12] Antonin Guttman. R-trees: a dynamic index structure for spatial searching. In SIGMOD "84: Proceedings of the 1984 ACM SIGMOD international conference on Management of data, pages 47-57, New York, NY, USA, 1984. ACM Press. [13] Nicholas Harvey, Michael B. Jones, Stefan Saroiu, Marvin Theimer, and Alec Wolman. Skipnet: A scalable overlay network with practical locality properties. In In proceedings of the 4th USENIX Symposium on Internet Technologies and Systems (USITS "03), Seattle, WA, March 2003. [14] Jason Hill, Robert Szewczyk, Alec Woo, Seth Hollar, David Culler, and Kristofer Pister. System architecture directions for networked sensors. In Proceedings of the Ninth International Conference on Architectural Support for Programming Languages and Operating Systems (ASPLOS-IX), pages 93-104, Cambridge, MA, USA, November 2000. ACM. [15] Atmel Inc. 4-megabit 2.5-volt or 2.7-volt DataFlash AT45DB041B, 2005. [16] Samsung Semiconductor Inc. K9W8G08U1M, K9K4G08U0M: 512M x 8 bit / 1G x 8 bit NAND flash memory, 2003. [17] Chalermek Intanagonwiwat, Ramesh Govindan, and Deborah Estrin. Directed diffusion: A scalable and robust communication paradigm for sensor networks. In Proceedings of the Sixth Annual International Conference on Mobile Computing and Networking, pages 56-67, Boston, MA, August 2000. ACM Press. [18] Xin Li, Young-Jin Kim, Ramesh Govindan, and Wei Hong. Multi-dimensional range queries in sensor networks. In Proceedings of the First ACM Conference on Embedded Networked Sensor Systems (SenSys)., 2003. to appear. [19] Witold Litwin, Marie-Anne Neimat, and Donovan A. Schneider. RP*: A family of order preserving scalable distributed data structures. In VLDB "94: Proceedings of the 20th International Conference on Very Large Data Bases, pages 342-353, San Francisco, CA, USA, 1994. [20] Samuel Madden, Michael Franklin, Joseph Hellerstein, and Wei Hong. TAG: a tiny aggregation service for ad-hoc sensor networks. In OSDI, Boston, MA, 2002. [21] A. Mitra, A. Banerjee, W. Najjar, D. Zeinalipour-Yazti, D.Gunopulos, and V. Kalogeraki. High performance, low power sensor platforms featuring gigabyte scale storage. In SenMetrics 2005: Third International Workshop on Measurement, Modeling, and Performance Analysis of Wireless Sensor Networks, July 2005. [22] J. Polastre, J. Hill, and D. Culler. Versatile low power media access for wireless sensor networks. In Proceedings of the Second ACM Conference on Embedded Networked Sensor Systems (SenSys), November 2004. [23] William Pugh. Skip lists: a probabilistic alternative to balanced trees. Commun. ACM, 33(6):668-676, 1990. [24] S. Ratnasamy, D. Estrin, R. Govindan, B. Karp, L. Yin S. Shenker, and F. Yu. Data-centric storage in sensornets. In ACM First Workshop on Hot Topics in Networks, 2001. [25] S. Ratnasamy, P. Francis, M. Handley, R. Karp, and S. Shenker. A scalable content addressable network. In Proceedings of the 2001 ACM SIGCOMM Conference, 2001. [26] S. Ratnasamy, B. Karp, L. Yin, F. Yu, D. Estrin, R. Govindan, and S. Shenker. GHT - a geographic hash-table for data-centric storage. In First ACM International Workshop on Wireless Sensor Networks and their Applications, 2002. [27] N. Xu, E. Osterweil, M. Hamilton, and D. Estrin. http://www.lecs.cs.ucla.edu/˜nxu/ess/. James Reserve Data. 50

analysis;archival storage;datum separation;sensor datum;index method;homogeneous architecture;flash storage;metada;distributed index structure;interval skip graph;interval tree;spatial scoping;geographic hash table;separation of datum;flooding;multi-tier sensor network;wireless sensor network;archive

train_C-48

Multi-dimensional Range Queries in Sensor Networks∗

In many sensor networks, data or events are named by attributes. Many of these attributes have scalar values, so one natural way to query events of interest is to use a multidimensional range query. An example is: List all events whose temperature lies between 50◦ and 60◦ , and whose light levels lie between 10 and 15. Such queries are useful for correlating events occurring within the network. In this paper, we describe the design of a distributed index that scalably supports multi-dimensional range queries. Our distributed index for multi-dimensional data (or DIM) uses a novel geographic embedding of a classical index data structure, and is built upon the GPSR geographic routing algorithm. Our analysis reveals that, under reasonable assumptions about query distributions, DIMs scale quite well with network size (both insertion and query costs scale as O( √ N)). In detailed simulations, we show that in practice, the insertion and query costs of other alternatives are sometimes an order of magnitude more than the costs of DIMs, even for moderately sized network. Finally, experiments on a small scale testbed validate the feasibility of DIMs.

1. INTRODUCTION In wireless sensor networks, data or events will be named by attributes [15] or represented as virtual relations in a distributed database [18, 3]. Many of these attributes will have scalar values: e.g., temperature and light levels, soil moisture conditions, etc. In these systems, we argue, one natural way to query for events of interest will be to use multi-dimensional range queries on these attributes. For example, scientists analyzing the growth of marine microorganisms might be interested in events that occurred within certain temperature and light conditions: List all events that have temperatures between 50◦ F and 60◦ F, and light levels between 10 and 20. Such range queries can be used in two distinct ways. They can help users efficiently drill-down their search for events of interest. The query described above illustrates this, where the scientist is presumably interested in discovering, and perhaps mapping the combined effect of temperature and light on the growth of marine micro-organisms. More importantly, they can be used by application software running within a sensor network for correlating events and triggering actions. For example, if in a habitat monitoring application, a bird alighting on its nest is indicated by a certain range of thermopile sensor readings, and a certain range of microphone readings, a multi-dimensional range query on those attributes enables higher confidence detection of the arrival of a flock of birds, and can trigger a system of cameras. In traditional database systems, such range queries are supported using pre-computed indices. Indices trade-off some initial pre-computation cost to achieve a significantly more efficient querying capability. For sensor networks, we assert that a centralized index for multi-dimensional range queries may not be feasible for energy-efficiency reasons (as well as the fact that the access bandwidth to this central index will be limited, particularly for queries emanating from within the network). Rather, we believe, there will be situations when it is more appropriate to build an innetwork distributed data structure for efficiently answering multi-dimensional range queries. In this paper, we present just such a data structure, that we call a DIM1 . DIMs are inspired by classical database indices, and are essentially embeddings of such indices within the sensor network. DIMs leverage two key ideas: in-network 1 Distributed Index for Multi-dimensional data. 63 data centric storage, and a novel locality-preserving geographic hash (Section 3). DIMs trace their lineage to datacentric storage systems [23]. The underlying mechanism in these systems allows nodes to consistently hash an event to some location within the network, which allows efficient retrieval of events. Building upon this, DIMs use a technique whereby events whose attribute values are close are likely to be stored at the same or nearby nodes. DIMs then use an underlying geographic routing algorithm (GPSR [16]) to route events and queries to their corresponding nodes in an entirely distributed fashion. We discuss the design of a DIM, presenting algorithms for event insertion and querying, for maintaining a DIM in the event of node failure, and for making DIMs robust to data or packet loss (Section 3). We then extensively evaluate DIMs using analysis (Section 4), simulation (Section 5), and actual implementation (Section 6). Our analysis reveals that, under reasonable assumptions about query distributions, DIMs scale quite well with network size (both insertion and query costs scale as O( √ N)). In detailed simulations, we show that in practice, the event insertion and querying costs of other alternatives are sometimes an order of magnitude the costs of DIMs, even for moderately sized network. Experiments on a small scale testbed validate the feasibility of DIMs (Section 6). Much work remains, including efficient support for skewed data distributions, existential queries, and node heterogeneity. We believe that DIMs will be an essential, but perhaps not necessarily the only, distributed data structure supporting efficient queries in sensor networks. DIMs will be part of a suite of such systems that enable feature extraction [7], simple range querying [10], exact-match queries [23], or continuous queries [15, 18]. All such systems will likely be integrated to a sensor network database system such as TinyDB [17]. Application designers could then choose the appropriate method of information access. For instance, a fire tracking application would use DIM to detect the hotspots, and would then use mechanisms that enable continuous queries [15, 18] to track the spatio-temporal progress of the hotspots. Finally, we note that DIMs are applicable not just to sensor networks, but to other deeply distributed systems (embedded networks for home and factory automation) as well. 2. RELATED WORK The basic problem that this paper addresses - multidimensional range queries - is typically solved in database systems using indexing techniques. The database community has focused mostly on centralized indices, but distributed indexing has received some attention in the literature. Indexing techniques essentially trade-off some data insertion cost to enable efficient querying. Indexing has, for long, been a classical research problem in the database community [5, 2]. Our work draws its inspiration from the class of multi-key constant branching index structures, exemplified by k-d trees [2], where k represents the dimensionality of the data space. Our approach essentially represents a geographic embedding of such structures in a sensor field. There is one important difference. The classical indexing structures are data-dependent (as are some indexing schemes that use locality preserving hashes, and developed in the theory literature [14, 8, 13]). The index structure is decided not only by the data, but also by the order in which data is inserted. Our current design is not data dependent. Finally, tangentially related to our work is the class of spatial indexing systems [21, 6, 11]. While there has been some work on distributed indexing, the problem has not been extensively explored. There exist distributed indices of a restricted kind-those that allow exact match or partial prefix match queries. Examples of such systems, of course, are the Internet Domain Name System, and the class of distributed hash table (DHT) systems exemplified by Freenet[4], Chord[24], and CAN[19]. Our work is superficially similar to CAN in that both construct a zone-based overlay atop of the underlying physical network. The underlying details make the two systems very different: CAN"s overlay is purely logical while our overlay is consistent with the underlying physical topology. More recent work in the Internet context has addressed support for range queries in DHT systems [1, 12], but it is unclear if these directly translate to the sensor network context. Several research efforts have expressed the vision of a database interface to sensor networks [9, 3, 18], and there are examples of systems that contribute to this vision [18, 3, 17]. Our work is similar in spirit to this body of literature. In fact, DIMs could become an important component of a sensor network database system such as TinyDB [17]. Our work departs from prior work in this area in two significant respects. Unlike these approaches, in our work the data generated at a node are hashed (in general) to different locations. This hashing is the key to scaling multi-dimensional range searches. In all the other systems described above, queries are flooded throughout the network, and can dominate the total cost of the system. Our work avoids query flooding by an appropriate choice of hashing. Madden et al. [17] also describe a distributed index, called Semantic Routing Trees (SRT). This index is used to direct queries to nodes that have detected relevant data. Our work differs from SRT in three key aspects. First, SRT is built on single attributes while DIM supports mulitple attributes. Second, SRT constructs a routing tree based on historical sensor readings, and therefore works well only for slowlychanging sensor values. Finally, in SRT queries are issued from a fixed node while in DIM queries can be issued from any node. A similar differentiation applies with respect to work on data-centric routing in sensor networks [15, 25], where data generated at a node is assumed to be stored at the node, and queries are either flooded throughout the network [15], or each source sets up a network-wide overlay announcing its presence so that mobile sinks can rendezvous with sources at the nearest node on the overlay [25]. These approaches work well for relatively long-lived queries. Finally, our work is most close related to data-centric storage [23] systems, which include geographic hash-tables (GHTs) [20], DIMENSIONS [7], and DIFS [10].In a GHT, data is hashed by name to a location within the network, enabling highly efficient rendezvous. GHTs are built upon the GPSR [16] protocol and leverage some interesting properties of that protocol, such as the ability to route to a node nearest to a given location. We also leverage properties in GPSR (as we describe later), but we use a locality-preserving hash to store data, enabling efficient multi-dimensional range queries. DIMENSIONS and DIFS can be thought of as using the same set of primitives as GHT (storage using consistent hashing), but for different ends: DIMENSIONS allows drill64 down search for features within a sensor network, while DIFS allows range queries on a single key in addition to other operations. 3. THE DESIGN OF DIMS Most sensor networks are deployed to collect data from the environment. In these networks, nodes (either individually or collaboratively) will generate events. An event can generally be described as a tuple of attribute values, A1, A2, · · · , Ak , where each attribute Ai represents a sensor reading, or some value corresponding to a detection (e.g., a confidence level). The focus of this paper is the design of systems to efficiently answer multi-dimensional range queries of the form: x1 − y1, x2 − y2, · · · , xk − yk . Such a query returns all events whose attribute values fall into the corresponding ranges. Notice that point queries, i.e., queries that ask for events with specified values for each attribute, are a special case of range queries. As we have discussed in Section 1, range queries can enable efficient correlation and triggering within the network. It is possible to implement range queries by flooding a query within the network. However, as we show in later sections, this alternative can be inefficient, particularly as the system scales, and if nodes within the network issue such queries relatively frequently. The other alternative, sending all events to an external storage node results in the access link being a bottleneck, especially if nodes within the network issue queries. Shenker et al. [23] also make similar arguments with respect to data-centric storage schemes in general; DIMs are an instance of such schemes. The system we present in this paper, the DIM, relies upon two foundations: a locality-preserving geographic hash, and an underlying geographic routing scheme. The key to resolving range queries efficiently is data locality: i.e., events with comparable attribute values are stored nearby. The basic insight underlying DIM is that data locality can be obtained by a locality-preserving geographic hash function. Our geographic hash function finds a localitypreserving mapping from the multi-dimensional space (described by the set of attributes) to a 2-d geographic space; this mapping is inspired by k-d trees [2] and is described later. Moreover, each node in the network self-organizes to claim part of the attribute space for itself (we say that each node owns a zone), so events falling into that space are routed to and stored at that node. Having established the mapping, and the zone structure, DIMs use a geographic routing algorithm previously developed in the literature to route events to their corresponding nodes, or to resolve queries. This algorithm, GPSR [16], essentially enables the delivery of a packet to a node at a specified location. The routing mechanism is simple: when a node receives a packet destined to a node at location X, it forwards the packet to the neighbor closest to X. In GPSR, this is called greedy-mode forwarding. When no such neighbor exists (as when there exists a void in the network), the node starts the packet on a perimeter mode traversal, using the well known right-hand rule to circumnavigate voids. GPSR includes efficient techniques for perimeter traversal that are based on graph planarization algorithms amenable to distributed implementation. For all of this to work, DIMs make two assumptions that are consistent with the literature [23]. First, all nodes know the approximate geographic boundaries of the network. These boundaries may either be configured in nodes at the time of deployment, or may be discovered using a simple protocol. Second, each node knows its geographic location. Node locations can be automatically determined by a localization system or by other means. Although the basic idea of DIMs may seem straightforward, it is challenging to design a completely distributed data structure that must be robust to packet losses and node failures, yet must support efficient query distribution and deal with communication voids and obstacles. We now describe the complete design of DIMs. 3.1 Zones The key idea behind DIMs, as we have discussed, is a geographic locality-preserving hash that maps a multi-attribute event to a geographic zone. Intuitively, a zone is a subdivision of the geographic extent of a sensor field. A zone is defined by the following constructive procedure. Consider a rectangle R on the x-y plane. Intuitively, R is the bounding rectangle that contains all sensors withing the network. We call a sub-rectangle Z of R a zone, if Z is obtained by dividing R k times, k ≥ 0, using a procedure that satisfies the following property: After the i-th division, 0 ≤ i ≤ k, R is partitioned into 2i equal sized rectangles. If i is an odd (even) number, the i-th division is parallel to the y-axis (x-axis). That is, the bounding rectangle R is first sub-divided into two zones at level 0 by a vertical line that splits R into two equal pieces, each of these sub-zones can be split into two zones at level 1 by a horizontal line, and so on. We call the non-negative integer k the level of zone Z, i.e. level(Z) = k. A zone can be identified either by a zone code code(Z) or by an address addr(Z). The code code(Z) is a 0-1 bit string of length level(Z), and is defined as follows. If Z lies in the left half of R, the first (from the left) bit of code(Z) is 0, else 1. If Z lies in the bottom half of R, the second bit of code(Z) is 0, else 1. The remaining bits of code(Z) are then recursively defined on each of the four quadrants of R. This definition of the zone code matches the definition of zones given above, encoding divisions of the sensor field geography by bit strings. Thus, in Figure 2, the zone in the top-right corner of the rectangle R has a zone code of 1111. Note that the zone codes collectively define a zone tree such that individual zones are at the leaves of this tree. The address of a zone Z, addr(Z), is defined to be the centroid of the rectangle defined by Z. The two representations of a zone (its code and its address) can each be computed from the other, assuming the level of the zone is known. Two zones are called sibling zones if their zone codes are the same except for the last bit. For example, if code(Z1) = 01101 and code(Z2) = 01100, then Z1 and Z2 are sibling zones. The sibling subtree of a zone is the subtree rooted at the left or right sibling of the zone in the zone tree. We uniquely define a backup zone for each zone as follows: if the sibling subtree of the zone is on the left, the backup zone is the right-most zone in the sibling subtree; otherwise, the backup zone is the left-most zone in the sibling subtree. For a zone Z, let p be the first level(Z) − 1 digits of code(Z). Let backup(Z) be the backup zone of zone Z. If code(Z) = p1, code(backup(Z)) = p01∗ with the most number of trailing 1"s (∗ means 0 or 1 occurrences). If 65 code(Z) = p0, code(backup(Z)) = p10∗ with the most number of trailing 0"s. 3.2 Associating Zones with Nodes Our definition of a zone is independent of the actual distribution of nodes in the sensor field, and only depends upon the geographic extent (the bounding rectangle) of the sensor field. Now we describe how zones are mapped to nodes. Conceptually, the sensor field is logically divided into zones and each zone is assigned to a single node. If the sensor network were deployed in a grid-like (i.e., very regular) fashion, then it is easy to see that there exists a k such that each node maps into a distinct level-k zone. In general, however, the node placements within a sensor field are likely to be less regular than the grid. For some k, some zones may be empty and other zones might have more than one node situated within them. One alternative would have been to choose a fixed k for the overall system, and then associate nodes with the zones they are in (and if a zone is empty, associate the nearest node with it, for some definition of nearest). Because it makes our overall query routing system simpler, we allow nodes in a DIM to map to different-sized zones. To precisely understand the associations between zones and nodes, we define the notion of zone ownership. For any given placement of network nodes, consider a node A. Let ZA to be the largest zone that includes only node A and no other node. Then, we say that A owns ZA. Notice that this definition of ownership may leave some sections of the sensor field un-associated with a node. For example, in Figure 2, the zone 110 does not contain any nodes and would not have an owner. To remedy this, for any empty zone Z, we define the owner to be the owner of backup(Z). In our example, that empty zone"s owner would also be the node that owns 1110, its backup zone. Having defined the association between nodes and zones, the next problem we tackle is: given a node placement, does there exist a distributed algorithm that enables each node to determine which zones it owns, knowing only the overall boundary of the sensor network? In principle, this should be relatively straightforward, since each node can simply determine the location of its neighbors, and apply simple geometric methods to determine the largest zone around it such that no other node resides in that zone. In practice, however, communication voids and obstacles make the algorithm much more challenging. In particular, resolving the ownership of zones that do not contain any nodes is complicated. Equally complicated is the case where the zone of a node is larger than its communication radius and the node cannot determine the boundaries of its zone by local communication alone. Our distributed zone building algorithm defers the resolution of such zones until when either a query is initiated, or when an event is inserted. The basic idea behind our algorithm is that each node tentatively builds up an idea of the zone it resides in just by communicating with its neighbors (remembering which boundaries of the zone are undecided because there is no radio neighbor that can help resolve that boundary). These undecided boundaries are later resolved by a GPSR perimeter traversal when data messages are actually routed. We now describe the algorithm, and illustrate it using examples. In our algorithm, each node uses an array bound[0..3] to maintain the four boundaries of the zone it owns (rememFigure 1: A network, where circles represent sensor nodes and dashed lines mark the network boundary. 1111 011 00 110 100 101 1110 010 Figure 2: The zone code and boundaries. 0 1 0 1 10 10 1 1 10 00 Figure 3: The Corresponding Zone Tree ber that in this algorithm, the node only tries to determine the zone it resides in, not the other zones it might own because those zones are devoid of nodes). When a node starts up, each node initializes this array to be the network boundary, i.e., initially each node assumes its zone contains the whole network. The zone boundary algorithm now relies upon GPSR"s beacon messages to learn the locations of neighbors within radio range. Upon hearing of such a neighbor, the node calls the algorithm in Figure 4 to update its zone boundaries and its code accordingly. In this algorithm, we assume that A is the node at which the algorithm is executed, ZA is its zone, and a is a newly discovered neighbor of A. (Procedure Contain(ZA, a) is used to decide if node a is located within the current zone boundaries of node A). Using this algorithm, then, each node can independently and asynchronously decide its own tentative zone based on the location of its neighbors. Figure 2 illustrates the results of applying this algorithm for the network in Figure 1. Figure 3 describes the corresponding zone tree. Each zone resides at a leaf node and the code of a zone is the path from the root to the zone if we represent the branch to the left 66 Build-Zone(a) 1 while Contain(ZA, a) 2 do if length(code(ZA)) mod 2 = 0 3 then new bound ← (bound[0] + bound[1])/2 4 if A.x < new bound 5 then bound[1] ← new bound 6 else bound[0] ← new bound 7 else new bound ← (bound[2] + bound[3])/2 8 if A.y < new bound 9 then bound[3] ← new bound 10 else bound[2] ← new bound 11 Update zone code code(ZA) Figure 4: Zone Boundary Determination, where A.x and A.y represent the geographic coordinate of node A. Insert-Event(e) 1 c ← Encode(e) 2 if Contain(ZA, c) = true and is Internal() = true 3 then Store e and exit 4 Send-Message(c, e) Send-Message(c, m) 1 if ∃ neighbor Y, Closer(Y, owner(m), m) = true 2 then addr(m) ← addr(Y ) 3 else if length(c) > length(code(m)) 4 then Update code(m) and addr(m) 5 source(m) ← caller 6 if is Owner(msg) = true 7 then owner(m) ← caller"s code 8 Send(m) Figure 5: Inserting an event in a DIM. Procedure Closer(A, B, m) returns true if code(A) is closer to code(m) than code(B). source(m) is used to set the source address of message m. child by 0 and the branch to the right child by 1. This binary tree forms the index that we will use in the following event and query processing procedures. We see that the zone sizes are different and depend on the local densities and so are the lengths of zone codes for different nodes. Notice that in Figure 2, there is an empty zone whose code should be 110. In this case, if the node in zone 1111 can only hear the node in zone 1110, it sets its boundary with the empty zone to undecided, because it did not hear from any neighboring nodes from that direction. As we have mentioned before, the undecided boundaries are resolved using GPSR"s perimeter mode when an event is inserted, or a query sent. We describe event insertion in the next step. Finally, this description does not describe how a node"s zone codes are adjusted when neighboring nodes fail, or new nodes come up. We return to this in Section 3.5. 3.3 Inserting an Event In this section, we describe how events are inserted into a DIM. There are two algorithms of interest: a consistent hashing technique for mapping an event to a zone, and a routing algorithm for storing the event at the appropriate zone. As we shall see, these two algorithms are inter-related. 3.3.1 Hashing an Event to a Zone In Section 3.1, we described a recursive tessellation of the geographic extent of a sensor field. We now describe a consistent hashing scheme for a DIM that supports range queries on m distinct attributes2 Let us denote these attributes A1 . . . Am. For simplicity, assume for now that the depth of every zone in the network is k, k is a multiple of m, and that this value of k is known to every node. We will relax this assumption shortly. Furthermore, for ease of discussion, we assume that all attribute values have been normalized to be between 0 and 1. Our hashing scheme assigns a k bit zone code to an event as follows. For i between 1 and m, if Ai < 0.5, the i-th bit of the zone code is assigned 0, else 1. For i between m + 1 and 2m, if Ai−m < 0.25 or Ai−m ∈ [0.5, 0.75), the i-th bit of the zone is assigned 0, else 1, because the next level divisions are at 0.25 and 0.75 which divide the ranges to [0, 0.25), [0.25, 0.5), [0.5, 0.75), and [0.75, 1). We repeat this procedure until all k bits have been assigned. As an example, consider event E = 0.3, 0.8 . For this event, the 5-bit zone code is code(ZA) = 01110. Essentially, our hashing scheme uses the values of the attributes in round-robin fashion on the zone tree (such as the one in Figure 3), in order to map an m-attribute event to a zone code. This is reminiscent of k-d trees [2], but is quite different from that data structure: zone trees are spatial embeddings and do not incorporate the re-balancing algorithms in k-d trees. In our design of DIMs, we do not require nodes to have zone codes of the same length, nor do we expect a node to know the zone codes of other nodes. Rather, suppose the encoding node is A and its own zone code is of length kA. Then, given an event E, node A only hashes E to a zone code of length kA. We denote the zone code assigned to an event E by code(E). As we describe below, as the event is routed, code(E) is refined by intermediate nodes. This lazy evaluation of zone codes allows different nodes to use different length zone codes without any explicit coordination. 3.3.2 Routing an Event to its Owner The aim of hashing an event to a zone code is to store the event at the node within the network node that owns that zone. We call this node the owner of the event. Consider an event E that has just been generated at a node A. After encoding event E, node A compares code(E) with code(A). If the two are identical, node A store event E locally; otherwise, node A will attempt to route the event to its owner. To do this, note that code(E) corresponds to some zone Z , which is A"s current guess for the zone at which event E should be stored. A now invokes GPSR to send a message to addr(Z ) (the centroid of Z , Section 3.1). The message contains the event E, code(E), and the target geographic location for storing the event. In the message, A also marks itself as the owner of event E. As we will see later, the guessed zone Z , the address addr(Z ), and the owner of E, all of them contained in the message, will be refined by intermediate forwarding nodes. GPSR now delivers this message to the next hop towards addr(Z ) from A. This next hop node (call it B) does not immediately forward the message. Rather, it attempts to com2 DIM does not assume that all nodes are homogeneous in terms of the sensors they have. Thus, in an m dimensional DIM, a node that does not possess all m sensors can use NULL values for the corresponding readings. DIM treats NULL as an extreme value for range comparisons. As an aside, a network may have many DIM instances running concurrently. 67 pute a new zone code for E to get a new code codenew(E). B will update the code contained in the message (and also the geographic destination of the message) if codenew(E) is longer than the event code in the message. In this manner, as the event wends its way to its owner, its zone code gets refined. Now, B compares its own code code(B) against the owner code owner(E) contained in the incoming message. If code(B) has a longer match with code(E) than the current owner owner(E), then B sets itself to be the current owner of E, meaning that if nobody is eligible to store E, then B will store the event (we shall see how this happens next). If B"s zone code does not exactly match code(E), B will invoke GPSR to deliver E to the next hop. 3.3.3 Resolving undecided zone boundaries during insertion Suppose that some node, say C, finds itself to be the destination (or eventual owner) of an event E. It does so by noticing that code code(C) equals code(E) after locally recomputing a code for E. In that case, C stores E locally, but only if all four of C"s zone boundaries are decided. When this condition holds, C knows for sure that no other nodes have overlapped zones with it. In this case, we call C an internal node. Recall, though, that because the zone discovery algorithm Section 3.2 only uses information from immediate neighbors, one or more of C"s boundaries may be undecided. If so, C assumes that some other nodes have a zone that overlaps with its own, and sets out to resolve this overlap. To do this, C now sets itself to be the owner of E and continues forwarding the message. Here we rely on GPSR"s perimeter mode routing to probe around the void that causes the undecided boundary. Since the message starts from C and is destined for a geographic location near C, GPSR guarantees that the message will be delivered back to C if no other nodes will update the information in the message. If the message comes back to C with itself to be the owner, C infers that it must be the true owner of the zone and stores E locally. If this does not happen, there are two possibilities. The first is that as the event traverses the perimeter, some intermediate node, say B whose zone overlaps with C"s marks itself to be the owner of the event, but otherwise does not change the event"s zone code. This node also recognizes that its own zone overlaps with C"s and initiates a message exchange which causes each of them to appropriately shrink their zone. Figures 6 through 8 show an example of this data-driven zone shrinking. Initially, both node A and node B have claimed the same zone 0 because they are out of radio range of each other. Suppose that A inserts an event E = 0.4, 0.8, 0.9 . A encodes E to 0 and claims itself to be the owner of E. Since A is not an internal node, it sends out E, looking for other owner candidates of E. Once E gets to node B, B will see in the message"s owner field A"s code that is the same as its own. B then shrinks its zone from 0 to 01 according to A"s location which is also recorded in the message and send a shrink request to A. Upon receiving this request, A also shrinks its zone from 0 to 00. A second possibility is if some intermediate node changes the destination code of E to a more specific value (i.e., longer zone code). Let us label this node D. D now tries to initiate delivery to the centroid of the new zone. This A B 0 0 110 100 1111 1110 101 Figure 6: Nodes A and B have claimed the same zone. A B <0.4,0.8,0.9> Figure 7: An event/query message (filled arrows) triggers zone shrinking (hollow arrows). A B 01 00 110 100 1111 1110 101 Figure 8: The zone layout after shrinking. Now node A and B have been mapped to different zones. might result in a new perimeter walk that returns to D (if, for example, D happens to be geographically closest to the centroid of the zone). However, D would not be the owner of the event, which would still be C. In routing to the centroid of this zone, the message may traverse the perimeter and return to D. Now D notices that C was the original owner, so it encapsulates the event and directs it to C. In case that there indeed is another node, say X, that owns an overlapped zone with C, X will notice this fact by finding in the message the same prefix of the code of one of its zones, but with a different geographic location from its own. X will shrink its zone to resolve the overlap. If X"s zone is smaller than or equal to C"s zone, X will also send a shrink request to C. Once C receives a shrink request, it will reduce its zone appropriately and fix its undecided boundary. In this manner, the zone formation process is resolved on demand in a data-driven way. 68 There are several interesting effects with respect to perimeter walking that arise in our algorithm. The first is that there are some cases where an event insertion might cause the entire outer perimeter of the network to be traversed3 . Figure 6 also works as an example where the outer perimeter is traversed. Event E inserted by A will eventually be stored in node B. Before node B stores event E, if B"s nominal radio range does not intersect the network boundary, it needs to send out E again as A did, because B in this case is not an internal node. But if B"s nominal radio range intersects the network boundary, it then has two choices. It can assume that there will not be any nodes outside the network boundary and so B is an internal node. This is an aggressive approach. On the other hand, B can also make a conservative decision assuming that there might be some other nodes it have not heard of yet. B will then force the message walking another perimeter before storing it. In some situations, especially for large zones where the node that owns a zone is far away from the centroid of the owned zone, there might exist a small perimeter around the destination that does not include the owner of the zone. The event will end up being stored at a different node than the real owner. In order to deal with this problem, we add an extra operation in event forwarding, called efficient neighbor discovery. Before invoking GPSR, a node needs to check if there exists a neighbor who is eligible to be the real owner of the event. To do this, a node C, say, needs to know the zone codes of its neighboring nodes. We deploy GPSR"s beaconing message to piggyback the zone codes for nodes. So by simply comparing the event"s code and neighbor"s code, a node can decide whether there exists a neighbor Y which is more likely to be the owner of event E. C delivers E to Y , which simply follows the decision making procedure discussed above. 3.3.4 Summary and Pseudo-code In summary, our event insertion procedure is designed to nicely interact with the zone discovery mechanism, and the event hashing mechanism. The latter two mechanisms are kept simple, while the event insertion mechanism uses lazy evaluation at each hop to refine the event"s zone code, and it leverages GPSR"s perimeter walking mechanism to fix undecided zone boundaries. In Section 3.5, we address robustness of event insertion to packet loss or to node failures. Figure 5 shows the pseudo-code for inserting and forwarding an event e. In this pseudo code, we have omitted a description of the zone shrinking procedure. In the pseudo code, procedure is Internal() is used to determine if the caller is an internal node and procedure is Owner() is used to determine if the caller is more eligible to be the owner of the event than is currently claimed owner as recorded in the message. Procedure Send-Message is used to send either an event message or a query message. If the message destination address has been changed, the packet source address needs also to be changed in order to avoid being dropped by GPSR, since GPSR does not allow a node to see the same packet in greedy mode twice. 3 This happens less frequently than for GHTs, where inserting an event to a location outside the actual (but inside the nominal) boundary of the network will always invoke an external perimeter walk. 3.4 Resolving and Routing Queries DIMs support both point queries4 and range queries. Routing a point query is identical to routing an event. Thus, the rest of this section details how range queries are routed. The key challenge in routing zone queries is brought out by the following strawman design. If the entire network was divided evenly into zones of depth k (for some pre-defined constant k), then the querier (the node issuing the query) could subdivide a given range query into the relevant subzones and route individual requests to each of the zones. This can be inefficient for large range queries and also hard to implement in our design where zone sizes are not predefined. Accordingly, we use a slightly different technique where a range query is initially routed to a zone corresponding to the entire range, and is then progressively split into smaller subqueries. We describe this algorithm here. The first step of the algorithm is to map a range query to a zone code prefix. Conceptually, this is easy; in a zone tree (Figure 3), there exists some node which contains the entire range query in its sub-tree, and none of its children in the tree do. The initial zone code we choose for the query is the zone code corresponding to that tree node, and is a prefix of the zone codes of all zones (note that these zones may not be geographically contiguous) in the subtree. The querier computes the zone code of Q, denoted by code(Q) and then starts routing a query to addr(code(Q)). Upon receiving a range query Q, a node A (where A is any node on the query propagation path) divides it into multiple smaller sized subqueries if there is an overlap between the zone of A, zone(A) and the zone code associated with Q, code(Q). Our approach to split a query Q into subqueries is as follows. If the range of Q"s first attribute contains the value 0.5, A divides Q into two sub-queries one of whose first attribute ranges from 0 to 0.5, and the other from 0.5 to 1. Then A decides the half that overlaps with its own zone. Let"s call it QA. If QA does not exist, then A stops splitting; otherwise, it continues splitting (using the second attribute range) and recomputing QA until QA is small enough so that it completely falls into zone(A) and hence A can now resolve it. For example, suppose that node A, whose code is 0110, is to split a range query Q = 0.3 − 0.8, 0.6 − 0.9 . The splitting steps is shown in Figure 2. After splitting, we obtain three smaller queries q0 = 0.3 − 0.5, 0.6 − 0.75 , q1 = 0.3 − 0.5, 0.75 − 0.9 , and q2 = 0.5 − 0.8, 0.6 − 0.9 . This splitting procedure is illustrated in Figure 9 which also shows the codes of each subquery after splitting. A then replies to subquery q0 with data stored locally and sends subqueries q1 and q2 using the procedure outlined above. More generally, if node A finds itself to be inside the zone subtree that maximally covers Q, it will send the subqueries that resulted from the split. Otherwise, if there is no overlap between A and Q, then A forwards Q as is (in this case Q is either the original query, or a product of an earlier split). Figure 10 describes the pseudo-code for the zone splitting algorithm. As shown in the above algorithm, once a subquery has been recognized as belonging to the caller"s zone, procedure Resolve is invoked to resolve the subquery and send a reply to the querier. Every query message contains 4 By point queries, we mean the equality condition on all indexed keys. DIM index attributes are not necessarily primary keys. 69 the geographic location of its initiator, so the corresponding reply message can be delivered directly back to the initiator. Finally, in the process of query resolution, zones might shrink similar to shrinkage during inserting. We omit this in the pseudo code. 3.5 Robustness Until now, we have not discussed the impact of node failures and packet losses, or node arrivals and departures on our algorithms. Packet losses can affect query and event insertion, and node failures can result in lost data, while node arrivals and departures can impact the zone structure. We now discuss how DIMs can be made robust to these kinds of dynamics. 3.5.1 Maintaining Zones In previous sections, we described how the zone discovery algorithm could leave zone boundaries undecided. These undecided boundaries are resolved during insertion or querying, using the zone shrinking procedure describe above. When a new node joins the network, the zone discovery mechanism (Section 3.2) will cause neighboring zones to appropriately adjust their zone boundaries. At this time, those zones can also transfer to the new node those events they store but which should belong to the new node. Before a node turns itself off (if this is indeed possible), it knows that its backup node (Section 3.1) will take over its zone, and will simply send all its events to its backup node. Node deletion may also cause zone expansion. In order to keep the mapping between the binary zone tree"s leaf nodes and zones, we allow zone expansion to only occur among sibling zones (Section 3.1). The rule is: if zone(A)"s sibling zone becomes empty, then A can expand its own zone to include its sibling zone. Now, we turn our attention to node failures. Node failures are just like node deletions except that a failed node does not have a chance to move its events to another node. But how does a node decide if its sibling has failed? If the sibling is within radio range, the absence of GPSR beaconing messages can detect this. Once it detects this, the node can expand its zone. A different approach is needed for detecting siblings who are not within radio range. These are the cases where two nodes own their zones after exchanging a shrink message; they do not periodically exchange messages thereafter to maintain this zone relationship. In this case, we detect the failure in a data-driven fashion, with obvious efficiency benefits compared to periodic keepalives. Once a node B has failed, an event or query message that previously should have been owned by the failed node will now be delivered to the node A that owns the empty zone left by node B. A can see this message because A stands right around the empty area left by B and is guaranteed to be visited in a GPSR perimeter traversal. A will set itself to be the owner of the message, and any node which would have dropped this message due to a perimeter loop will redirect the message to A instead. If A"s zone happens to be the sibling of B"s zone, A can safely expand its own zone and notify its expanded zone to its neighbors via GPSR beaconing messages. 3.5.2 Preventing Data Loss from Node Failure The algorithms described above are robust in terms of zone formation, but node failure can erase data. To avoid this, DIMs can employ two kinds of replication: local replication to be resilient to random node failures, and mirror replication for resilience to concurrent failure of geographically contiguous nodes. Mirror replication is conceptually easy. Suppose an event E has a zone code code(E). Then, the node that inserts E would store two copies of E; one at the zone denoted by code(E), and the other at the zone corresponding to the one"s complement of code(E). This technique essentially creates a mirror DIM. A querier would need, in parallel, to query both the original DIM and its mirror since there is no way of knowing if a collection of nodes has failed. Clearly, the trade-off here is an approximate doubling of both insertion and query costs. There exists a far cheaper technique to ensure resilience to random node failures. Our local replication technique rests on the observation that, for each node A, there exists a unique node which will take over its zone when A fails. This node is defined as the node responsible for A"s zone"s backup zone (see Section 3.1). The basic idea is that A replicates each data item it has in this node. We call this node A"s local replica. Let A"s local replica be B. Often B will be a radio neighbor of A and can be detected from GPSR beacons. Sometimes, however, this is not the case, and B will have to be explicitly discovered. We use an explicit message for discovering the local replica. Discovering the local replica is data-driven, and uses a mechanism similar to that of event insertion. Node A sends a message whose geographic destination is a random nearby location chosen by A. The location is close enough to A such that GPSR will guarantee that the message will delivered back to A. In addition, the message has three fields, one for the zone code of A, code(A), one for the owner owner(A) of zone(A) which is set to be empty, and one for the geographic location of owner(A). Then the packet will be delivered in GPSR perimeter mode. Each node that receives this message will compare its zone code and code(A) in the message, and if it is more eligible to be the owner of zone(A) than the current owner(A) recorded in the message, it will update the field owner(A) and the corresponding geographic location. Once the packet comes back to A, it will know the location of its local replica and can start to send replicas. In a dense sensor network, the local replica of a node is usually very near to the node, either its direct neighbor or 1-2 hops away, so the cost of sending replicas to local replication will not dominate the network traffic. However, a node"s local replica itself may fail. There are two ways to deal with this situation; periodic refreshes, or repeated datadriven discovery of local replicas. The former has higher overhead, but more quickly discovers failed replicas. 3.5.3 Robustness to Packet Loss Finally, the mechanisms for querying and event insertion can be easily made resilient to packet loss. For event insertion, a simple ACK scheme suffices. Of course, queries and responses can be lost as well. In this case, there exists an efficient approach for error recovery. This rests on the observation that the querier knows which zones fall within its query and should have responded (we assume that a node that has no data matching a query, but whose zone falls within the query, responds with a negative acknowledgment). After a conservative timeout, the querier can re-issue the queries selectively to these zones. If DIM cannot get any answers (positive or negative) from 70 <0.3-0.8, 0.6-0.9> <0.5-0.8, 0.6-0.9><0.3-0.5, 0.6-0.9> <0.3-0.5, 0.6-0.9> <0.3-0.5, 0.6-0.9> <0.3-0.5, 0.6-0.75> <0.3-0.5, 0.75-0.9> 0 0 1 1 1 1 Figure 9: An example of range query splitting Resolve-Range-Query(Q) 1 Qsub ← nil 2 q0, Qsub ← Split-Query(Q) 3 if q0 = nil 4 then c ← Encode(Q) 5 if Contain(c, code(A)) = true 6 then go to step 12 7 else Send-Message(c, q0) 8 else Resolve(q0) 9 if is Internal() = true 10 then Absorb (q0) 11 else Append q0 to Qsub 12 if Qsub = nil 13 then for each subquery q ∈ Qsub 14 do c ← Encode(q) 15 Send-Message(c, q) Figure 10: Query resolving algorithm certain zones after repeated timeouts, it can at least return the partial query results to the application together with the information about the zones from which data is missing. 4. DIMS: AN ANALYSIS In this section, we present a simple analytic performance evaluation of DIMs, and compare their performance against other possible approaches for implementing multi-dimensional range queries in sensor networks. In the next section, we validate these analyses using detailed packet-level simulations. Our primary metrics for the performance of a DIM are: Average Insertion Cost measures the average number of messages required to insert an event into the network. Average Query Delivery Cost measures the average number of messages required to route a query message to all the relevant nodes in the network. It does not measure the number of messages required to transmit responses to the querier; this latter number depends upon the precise data distribution and is the same for many of the schemes we compare DIMs against. In DIMs, event insertion essentially uses geographic routing. In a dense N-node network where the likelihood of traversing perimeters is small, the average event insertion cost proportional to √ N [23]. On the other hand, the query delivery cost depends upon the size of ranges specified in the query. Recall that our query delivery mechanism is careful about splitting a query into sub-queries, doing so only when the query nears the zone that covers the query range. Thus, when the querier is far from the queried zone, there are two components to the query delivery cost. The first, which is proportional to √ N, is the cost to deliver the query near the covering zone. If within this covering zone, there are M nodes, the message delivery cost of splitting the query is proportional to M. The average cost of query delivery depends upon the distribution of query range sizes. Now, suppose that query sizes follow some density function f(x), then the average cost of resolve a query can be approximated by Ê N 1 xf(x)dx. To give some intuition for the performance of DIMs, we consider four different forms for f(x): the uniform distribution where a query range encompassing the entire network is as likely as a point query; a bounded uniform distribution where all sizes up to a bound B are equally likely; an algebraic distribution in which most queries are small, but large queries are somewhat likely; and an exponential distribution where most queries are small and large queries are unlikely. In all our analyses, we make the simplifying assumption that the size of a query is proportional to the number of nodes that can answer that query. For the uniform distribution P(x) ∝ c for some constant c. If each query size from 1 . . . N is equally likely, the average query delivery cost of uniformly distributed queries is O(N). Thus, for uniformly distributed queries, the performance of DIMs is comparable to that of flooding. However, for the applications we envision, where nodes within the network are trying to correlate events, the uniform distribution is highly unrealistic. Somewhat more realistic is a situation where all query sizes are bounded by a constant B. In this case, the average cost for resolving such a query is approximately Ê B 1 xf(x)dx = O(B). Recall now that all queries have to pay an approximate cost of O( √ N) to deliver the query near the covering zone. Thus, if DIM limited queries to a size proportional to√ N, the average query cost would be O( √ N). The algebraic distribution, where f(x) ∝ x−k , for some constant k between 1 and 2, has an average query resolution cost given by Ê N 1 xf(x)dx = O(N2−k ). In this case, if k > 1.5, the average cost of query delivery is dominated by the cost to deliver the query to near the covering zone, given by O( √ N). Finally, for the exponential distribution, f(x) = ce−cx for some constant c, and the average cost is just the mean of the corresponding distribution, i.e., O(1) for large N. Asymptotically, then, the cost of the query for the exponential distribution is dominated by the cost to deliver the query near the covering zone (O( √ N)). Thus, we see that if queries follow either the bounded uniform distribution, the algebraic distribution, or the exponential distribution, the query cost scales as the insertion cost (for appropriate choice of constants for the bounded uniform and the algebraic distributions). How well does the performance of DIMs compare against alternative choices for implementing multi-dimensional queries? A simple alternative is called external storage [23], where all events are stored centrally in a node outside the sensor network. This scheme incurs an insertion cost of O( √ N), and a zero query cost. However, as [23] points out, such systems may be impractical in sensor networks since the access link to the external node becomes a hotspot. A second alternative implementation would store events at the node where they are generated. Queries are flooded 71 throughout the network, and nodes that have matching data respond. Examples of systems that can be used for this (although, to our knowledge, these systems do not implement multi-dimensional range queries) are Directed Diffusion [15] and TinyDB [17]. The flooding scheme incurs a zero insertion cost, but an O(N) query cost. It is easy to show that DIMs outperform flooding as long as the ratio of the number of insertions to the number of queries is less than √ N. A final alternative would be to use a geographic hash table (GHT [20]). In this approach, attribute values are assumed to be integers (this is actually quite a reasonable assumption since attribute values are often quantized), and events are hashed on some (say, the first) attribute. A range query is sub-divided into several sub-queries, one for each integer in the range of the first attribute. Each sub-query is then hashed to the appropriate location. The nodes that receive a sub-query only return events that match all other attribute ranges. In this approach, which we call GHT-R (GHT"s for range queries) the insertion cost is O( √ N). Suppose that the range of the first attribute contains r discrete values. Then the cost to deliver queries is O(r √ N). Thus, asymptotically, GHT-R"s perform similarly to DIMs. In practice, however, the proportionality constants are significantly different, and DIMs outperform GHT-Rs, as we shall show using detailed simulations. 5. DIMS: SIMULATION RESULTS Our analysis gives us some insight into the asymptotic behavior of various approaches for multi-dimensional range queries. In this section, we use simulation to compare DIMs against flooding and GHT-R; this comparison gives us a more detailed understanding of these approaches for moderate size networks, and gives us a nuanced view of the mechanistic differences between some of these approaches. 5.1 Simulation Methodology We use ns-2 for our simulations. Since DIMs are implemented on top of GPSR, we first ported an earlier GPSR implementation to the latest version of ns-2. We modified the GPSR module to call our DIM implementation when it receives any data message in transit or when it is about to drop a message because that message traversed the entire perimeter. This allows a DIM to modify message zone codes in flight (Section 3), and determine the actual owner of an event or query. In addition, to this, we implemented in ns-2 most of the DIM mechanisms described in Section 3. Of those mechanisms, the only one we did not implement is mirror replication. We have implemented selective query retransmission for resiliency to packet loss, but have left the evaluation of this mechanism to future work. Our DIM implementation in ns-2 is 2800 lines of code. Finally, we implemented GHT-R, our GHT-based multidimensional range query mechanism in ns-2. This implementation was relatively straightforward, given that we had ported GPSR, and modified GPSR to detect the completion of perimeter mode traversals. Using this implementation, we conducted a fairly extensive evaluation of DIM and two alternatives (flooding, and our GHT-R). For all our experiments, we use uniformly placed sensor nodes with network sizes ranging from 50 nodes to 300 nodes. Each node has a radio range of 40m. For the results presented here, each node has on average 20 nodes within its nominal radio range. We have conducted experiments at other node densities; they are in agreement with the results presented here. In all our experiments, each node first generates 3 events5 on average (more precisely, for a topology of size N, we have 3N events, and each node is equally likely to generate an event). We have conducted experiments for three different event value distributions. Our uniform event distribution generates 2-dimensional events and, for each dimension, every attribute value is equally likely. Our normal event distribution generates 2-dimensional events and, for each dimension, the attribute value is normally distributed with a mean corresponding to the mid-point of the attribute value range. The normal event distribution represents a skewed data set. Finally, our trace event distribution is a collection of 4-dimensional events obtained from a habitat monitoring network. As we shall see, this represents a fairly skewed data set. Having generated events, for each simulation we generate queries such that, on average, each node generates 2 queries. The query sizes are determined using the four size distributions we discussed in Section 4: uniform, boundeduniform, algebraic and exponential. Once a query size has been determined, the location of the query (i.e., the actual boundaries of the zone) are uniformly distributed. For our GHT-R experiments, the dynamic range of the attributes had 100 discrete values, but we restricted the query range for any one attribute to 50 discrete values to allow those simulations to complete in reasonable time. Finally, using one set of simulations we evaluate the efficacy of local replication by turning off random fractions of nodes and measuring the fidelity of the returned results. The primary metrics for our simulations are the average query and insertion costs, as defined in Section 4. 5.2 Results Although we have examined almost all the combinations of factors described above, we discuss only the most salient ones here, for lack of space. Figure 11 plots the average insertion costs for DIM and GHT-R (for flooding, of course, the insertion costs are zero). DIM incurs less per event overhead in inserting events (regardless of the actual event distribution; Figure 11 shows the cost for uniformly distributed events). The reason for this is interesting. In GHT-R, storing almost every event incurs a perimeter traversal, and storing some events require traversing the outer perimeter of the network [20]. By contrast, in DIM, storing an event incurs a perimeter traversal only when a node"s boundaries are undecided. Furthermore, an insertion or a query in a DIM can traverse the outer perimeter (Section 3.3), but less frequently than in GHTs. Figure 13 plots the average query cost for a bounded uniform query size distribution. For this graph (and the next) we use a uniform event distribution, since the event distribution does not affect the query delivery cost. For this simulation, our bound was 1 4 th the size of the largest possible 5 Our metrics are chosen so that the exact number of events and queries is unimportant for our discussion. Of course, the overall performance of the system will depend on the relative frequency of events and queries, as we discuss in Section 4. Since we don"t have realistic ratios for these, we focus on the microscopic costs, rather than on the overall system costs. 72 0 2 4 6 8 10 12 14 16 18 20 50 100 150 200 250 300 AverageCostperInsertion Network Size DIM GHT-R Figure 11: Average insertion cost for DIM and GHT. 0.4 0.5 0.6 0.7 0.8 0.9 1 5 10 15 20 25 30 Fractionofrepliescomparedwithnon-failurecase Fraction of failed nodes (%) No Replication Local Replication Figure 12: Local replication performance. query (e.g., a query of the form 0 − 0.5, 0 − 0.5 . Even for this generous query size, DIMs perform quite well (almost a third the cost of flooding). Notice, however, that GHTRs incur high query cost since almost any query requires as many subqueries as the width of the first attribute"s range. Figure 14 plots the average query cost for the exponential distribution (the average query size for this distribution was set to be 1 16 th the largest possible query). The superior scaling of DIMs is evident in these graphs. Clearly, this is the regime in which one might expect DIMs to perform best, when most of the queries are small and large queries are relatively rare. This is also the regime in which one would expect to use multi-dimensional range queries: to perform relatively tight correlations. As with the bounded uniform distribution, GHT query cost is dominated by the cost of sending sub-queries; for DIMs, the query splitting strategy works quite well in keep overall query delivery costs low. Figure 12 describes the efficacy of local replication. To obtain this figure, we conducted the following experiment. On a 100-node network, we inserted a number of events uniformly distributed throughout the network, then issued a query covering the entire network and recorded the answers. Knowing the expected answers for this query, we then successively removed a fraction f of nodes randomly, and re-issued the same query. The figure plots the fraction of expected responses actually received, with and without replication. As the graph shows, local replication performs well for random failures, returning almost 90% of the responses when up to 30% of the nodes have failed simultaneously 6 .In the absence of local replication, of course, when 6 In practice, the performance of local replication is likely to 0 100 200 300 400 500 600 700 50 100 150 200 250 300 AverageCostperQueryinBoundedUnifDistribution Network Size DIM flooding GHT-R Figure 13: Average query cost with a bounded uniform query distribution 0 50 100 150 200 250 300 350 400 450 50 100 150 200 250 300 AverageCostperQueryinExponentialDistribution Network Size DIM flooding GHT-R Figure 14: Average query cost with an exponential query distribution 30% of the nodes fail, the response rate is only 70% as one would expect. We note that DIMs (as currently designed) are not perfect. When the data is highly skewed-as it was for our trace data set from the habitat monitoring application where most of the event values fell into within 10% of the attribute"s range-a few DIM nodes will clearly become the bottleneck. This is depicted in Figure 15, which shows that for DIMs, and GHT-Rs, the maximum number of transmissions at any network node (the hotspots) is rather high. (For less skewed data distributions, and reasonable query size distributions, the hotspot curves for all three schemes are comparable.) This is a standard problem that the database indices have dealt with by tree re-balancing. In our case, simpler solutions might be possible (and we discuss this in Section 7). However, our use of the trace data demonstrates that DIMs work for events which have more than two dimensions. Increasing the number of dimensions does not noticeably degrade DIMs query cost (omitted for lack of space). Also omitted are experiments examining the impact of several other factors, as they do not affect our conclusions in any way. As we expected, DIMs are comparable in performance to flooding when all sizes of queries are equally likely. For an algebraic distribution of query sizes, the relative performance is close to that for the exponential distribution. For normally distributed events, the insertion costs be much better than this. Assuming a node and its replica don"t simultaneously fail often, a node will almost always detect a replica failure and re-replicate, leading to near 100% response rates. 73 0 2000 4000 6000 8000 10000 12000 50 100 150 200 250 300 MaximumHotspotonTraceDataSet Network Size DIM flooding GHT-R Figure 15: Hotspot usage DIM Zone Manager Query Router Query Processor Event Manager Event Router GPSR interface(Event driven/Thread based) update useuse update GPSR Upper interface(Event driven/Thread based) Lower interface(Event driven/Thread based) Greedy Forwarding Perimeter Forwarding Beaconing Neighbor List Manager update use MoteNIC (MicaRadio) IP Socket (802.11b/Ethernet) Figure 16: Software architecture of DIM over GPSR are comparable to that for the uniform distribution. Finally, we note that in all our evaluations we have only used list queries (those that request all events matching the specified range). We expect that for summary queries (those that expect an aggregate over matching events), the overall cost of DIMs could be lower because the matching data are likely to be found in one or a small number of zones. We leave an understanding of this to future work. Also left to future work is a detailed understanding of the impact of location error on DIM"s mechanisms. Recent work [22] has examined the impact of imprecise location information on other data-centric storage mechanisms such as GHTs, and found that there exist relatively simple fixes to GPSR that ameliorate the effects of location error. 6. IMPLEMENTATION We have implemented DIMs on a Linux platform suitable for experimentation on PDAs and PC-104 class machines. To implement DIMs, we had to develop and test an independent implementation of GPSR. Our GPSR implementation is full-featured, while our DIM implementation has most of the algorithms discussed in Section 3; some of the robustness extensions have only simpler variants implemented. The software architecture of DIM/GPSR system is shown in Figure 16. The entire system (about 5000 lines of code) is event-driven and multi-threaded. The DIM subsystem consists of six logical components: zone management, event maintenance, event routing, query routing, query processing, and GPSR interactions. The GPSR system is implemented as user-level daemon process. Applications are executed as clients. For the DIM subsystem, the GPSR module 0 2 4 6 8 10 12 14 16 0.25x0.25 0.50x0.50 0.75x0.75 1.0x1.0 Query size Average#ofreceivedresponses perquery Figure 17: Number of events received for different query sizes 0 2 4 6 8 10 12 14 16 0.25x0.25 0.50x0.50 0.75x0.75 1.0x1.0 Query sizeTotalnumberofmessages onlyforsendingthequery Figure 18: Query distribution cost provides several extensions: it exports information about neighbors, and provides callbacks during packet forwarding and perimeter-mode termination. We tested our implementation on a testbed consisting of 8 PC-104 class machines. Each of these boxes runs Linux and uses a Mica mote (attached through a serial cable) for communication. These boxes are laid out in an office building with a total spatial separation of over a hundred feet. We manually measured the locations of these nodes relative to some coordinate system and configured the nodes with their location. The network topology is approximately a chain. On this testbed, we inserted queries and events from a single designated node. Our events have two attributes which span all combinations of the four values [0, 0.25, 0.75, 1] (sixteen events in all). Our queries span four sizes, returning 1, 4, 9 and 16 events respectively. Figure 17 plots the number of events received for different sized queries. It might appear that we received fewer events than expected, but this graph doesn"t count the events that were already stored at the querier. With that adjustment, the number of responses matches our expectation. Finally, Figure 18 shows the total number of messages required for different query sizes on our testbed. While these experiments do not reveal as much about the performance range of DIMs as our simulations do, they nevertheless serve as proof-of-concept for DIMs. Our next step in the implementation is to port DIMs to the Mica motes, and integrate them into the TinyDB [17] sensor database engine on motes. 74 7. CONCLUSIONS In this paper, we have discussed the design and evaluation of a distributed data structure called DIM for efficiently resolving multi-dimensional range queries in sensor networks. Our design of DIMs relies upon a novel locality-preserving hash inspired by early work in database indexing, and is built upon GPSR. We have a working prototype, both of GPSR and DIM, and plan to conduct larger scale experiments in the future. There are several interesting future directions that we intend to pursue. One is adaptation to skewed data distributions, since these can cause storage and transmission hotspots. Unlike traditional database indices that re-balance trees upon data insertion, in sensor networks it might be feasible to re-structure the zones on a much larger timescale after obtaining a rough global estimate of the data distribution. Another direction is support for node heterogeneity in the zone construction process; nodes with larger storage space assert larger-sized zones for themselves. A third is support for efficient resolution of existential queries-whether there exists an event matching a multi-dimensional range. Acknowledgments This work benefited greatly from discussions with Fang Bian, Hui Zhang and other ENL lab members, as well as from comments provided by the reviewers and our shepherd Feng Zhao. 8. REFERENCES [1] J. Aspnes and G. Shah. Skip Graphs. In Proceedings of 14th Annual ACM-SIAM Symposium on Discrete Algorithms (SODA), Baltimore, MD, January 2003. [2] J. L. Bentley. Multidimensional Binary Search Trees Used for Associative Searching. Communicaions of the ACM, 18(9):475-484, 1975. [3] P. Bonnet, J. E. Gerhke, and P. Seshadri. Towards Sensor Database Systems. In Proceedings of the Second International Conference on Mobile Data Management, Hong Kong, January 2001. [4] I. Clarke, O. Sandberg, B. Wiley, and T. W. Hong. Freenet: A Distributed Anonymous Information Storage and Retrieval System. In Designing Privacy Enhancing Technologies: International Workshop on Design Issues in Anonymity and Unobservability. Springer, New York, 2001. [5] D. Comer. The Ubiquitous B-tree. ACM Computing Surveys, 11(2):121-137, 1979. [6] R. A. Finkel and J. L. Bentley. Quad Trees: A Data Structure for Retrieval on Composite Keys. Acta Informatica, 4:1-9, 1974. [7] D. Ganesan, D. Estrin, and J. Heidemann. DIMENSIONS: Why do we need a new Data Handling architecture for Sensor Networks? In Proceedings of the First Workshop on Hot Topics In Networks (HotNets-I), Princeton, NJ, October 2002. [8] A. Gionis, P. Indyk, and R. Motwani. Similarity Search in High Dimensions via Hashing. In Proceedings of the 25th VLDB conference, Edinburgh, Scotland, September 1999. [9] R. Govindan, J. Hellerstein, W. Hong, S. Madden, M. Franklin, and S. Shenker. The Sensor Network as a Database. Technical Report 02-771, Computer Science Department, University of Southern California, September 2002. [10] B. Greenstein, D. Estrin, R. Govindan, S. Ratnasamy, and S. Shenker. DIFS: A Distributed Index for Features in Sensor Networks. In Proceedings of 1st IEEE International Workshop on Sensor Network Protocols and Applications, Anchorage, AK, May 2003. [11] A. Guttman. R-trees: A Dynamic Index Structure for Spatial Searching. In Proceedings of the ACM SIGMOD, Boston, MA, June 1984. [12] M. Harren, J. M. Hellerstein, R. Huebsch, B. T. Loo, S. Shenker, and I. Stoica. Complex Queries in DHT-based Peer-to-Peer Networks. In P. Druschel, F. Kaashoek, and A. Rowstron, editors, Proceedings of 1st International Workshop on Peer-to-Peer Systems (IPTPS"02), volume 2429 of LNCS, page 242, Cambridge, MA, March 2002. Springer-Verlag. [13] P. Indyk and R. Motwani. Approximate Nearest Neighbors: Towards Removing the Curse of Dimensionality. In Proceedings of the 30th Annual ACM Symposium on Theory of Computing, Dallas, Texas, May 1998. [14] P. Indyk, R. Motwani, P. Raghavan, and S. Vempala. Locality-preserving Hashing in Multidimensional Spaces. In Proceedings of the 29th Annual ACM symposium on Theory of Computing, pages 618 - 625, El Paso, Texas, May 1997. ACM Press. [15] C. Intanagonwiwat, R. Govindan, and D. Estrin. Directed Diffusion: A Scalable and Robust Communication Paradigm for Sensor Networks. In Proceedings of the Sixth Annual ACM/IEEE International Conference on Mobile Computing and Networking (Mobicom 2000), Boston, MA, August 2000. [16] B. Karp and H. T. Kung. GPSR: Greedy Perimeter Stateless Routing for Wireless Networks. In Proceedings of the Sixth Annual ACM/IEEE International Conference on Mobile Computing and Networking (Mobicom 2000), Boston, MA, August 2000. [17] S. Madden, M. Franklin, J. Hellerstein, and W. Hong. The Design of an Acquisitional Query Processor for Sensor Networks. In Proceedings of ACM SIGCMOD, San Diego, CA, June 2003. [18] S. Madden, M. J. Franklin, J. M. Hellerstein, and W. Hong. TAG: a Tiny AGregation Service for ad-hoc Sensor Networks. In Proceedings of 5th Annual Symposium on Operating Systems Design and Implementation (OSDI), Boston, MA, December 2002. [19] S. Ratnasamy, P. Francis, M. Handley, R. Karp, and S. Shenker. A Scalable Content-Addressable Network. In Proceedings of the ACM SIGCOMM, San Diego, CA, August 2001. [20] S. Ratnasamy, B. Karp, L. Yin, F. Yu, D. Estrin, R. Govindan, and S. Shenker. GHT: A Geographic Hash Table for Data-Centric Storage. In Proceedings of the First ACM International Workshop on Wireless Sensor Networks and Applications, Atlanta, GA, September 2002. [21] H. Samet. Spatial Data Structures. In W. Kim, editor, Modern Database Systems: The Object Model, Interoperability and Beyond, pages 361-385. Addison Wesley/ACM, 1995. [22] K. Sead, A. Helmy, and R. Govindan. On the Effect of Localization Errors on Geographic Face Routing in Sensor Networks. In Under submission, 2003. [23] S. Shenker, S. Ratnasamy, B. Karp, R. Govindan, and D. Estrin. Data-Centric Storage in Sensornets. In Proc. ACM SIGCOMM Workshop on Hot Topics In Networks, Princeton, NJ, 2002. [24] I. Stoica, R. Morris, D. Karger, M. F. Kaashoek, and H. Balakrishnan. Chord: A Scalable Peer-To-Peer Lookup Service for Internet Applications. In Proceedings of the ACM SIGCOMM, San Diego, CA, August 2001. [25] F. Ye, H. Luo, J. Cheng, S. Lu, and L. Zhang. A Two-Tier Data Dissemination Model for Large-scale Wireless Sensor Networks. In Proceedings of the Eighth Annual ACM/IEEE International Conference on Mobile Computing and Networking (Mobicom"02), Atlanta, GA, September 2002. 75

datacentric storage system;multi-dimensional range query;multidimensional range query;event insertion;querying cost;query flooding;indexing technique;dim;distributed datum structure;asymptotic behavior;locality-preserving geographic hash;sensor network;geographic routing;centralized index;normal event distribution;efficient correlation;distributed index

train_C-49

Evaluating Opportunistic Routing Protocols with Large Realistic Contact Traces

Traditional mobile ad-hoc network (MANET) routing protocols assume that contemporaneous end-to-end communication paths exist between data senders and receivers. In some mobile ad-hoc networks with a sparse node population, an end-to-end communication path may break frequently or may not exist at any time. Many routing protocols have been proposed in the literature to address the problem, but few were evaluated in a realistic opportunistic network setting. We use simulation and contact traces (derived from logs in a production network) to evaluate and compare five existing protocols: direct-delivery, epidemic, random, PRoPHET, and Link-State, as well as our own proposed routing protocol. We show that the direct delivery and epidemic routing protocols suffer either low delivery ratio or high resource usage, and other protocols make tradeoffs between delivery ratio and resource usage.

1. INTRODUCTION Mobile opportunistic networks are one kind of delay-tolerant network (DTN) [6]. Delay-tolerant networks provide service despite long link delays or frequent link breaks. Long link delays happen in networks with communication between nodes at a great distance, such as interplanetary networks [2]. Link breaks are caused by nodes moving out of range, environmental changes, interference from other moving objects, radio power-offs, or failed nodes. For us, mobile opportunistic networks are those DTNs with sparse node population and frequent link breaks caused by power-offs and the mobility of the nodes. Mobile opportunistic networks have received increasing interest from researchers. In the literature, these networks include mobile sensor networks [25], wild-animal tracking networks [11], pocketswitched networks [8], and transportation networks [1, 14]. We expect to see more opportunistic networks when the one-laptopper-child (OLPC) project [18] starts rolling out inexpensive laptops with wireless networking capability for children in developing countries, where often no infrastructure exits. Opportunistic networking is one promising approach for those children to exchange information. One fundamental problem in opportunistic networks is how to route messages from their source to their destination. Mobile opportunistic networks differ from the Internet in that disconnections are the norm instead of the exception. In mobile opportunistic networks, communication devices can be carried by people [4], vehicles [1] or animals [11]. Some devices can form a small mobile ad-hoc network when the nodes move close to each other. But a node may frequently be isolated from other nodes. Note that traditional Internet routing protocols and ad-hoc routing protocols, such as AODV [20] or DSDV [19], assume that a contemporaneous endto-end path exists, and thus fail in mobile opportunistic networks. Indeed, there may never exist an end-to-end path between two given devices. In this paper, we study protocols for routing messages between wireless networking devices carried by people. We assume that people send messages to other people occasionally, using their devices; when no direct link exists between the source and the destination of the message, other nodes may relay the message to the destination. Each device represents a unique person (it is out of the scope of this paper when a device maybe carried by multiple people). Each message is destined for a specific person and thus for a specific node carried by that person. Although one person may carry multiple devices, we assume that the sender knows which device is the best to receive the message. We do not consider multicast or geocast in this paper. Many routing protocols have been proposed in the literature. Few of them were evaluated in realistic network settings, or even in realistic simulations, due to the lack of any realistic people mobility model. Random walk or random way-point mobility models are often used to evaluate the performance of those routing protocols. Although these synthetic mobility models have received extensive interest by mobile ad-hoc network researchers [3], they do not reflect people"s mobility patterns [9]. Realising the limitations of using random mobility models in simulations, a few researchers have studied routing protocols in mobile opportunistic networks with realistic mobility traces. Chaintreau et al. [5] theoretically analyzed the impact of routing algorithms over a model derived from a realistic mobility data set. Su et al. [22] simulated a set of routing 35 protocols in a small experimental network. Those studies help researchers better understand the theoretical limits of opportunistic networks, and the routing protocol performance in a small network (20-30 nodes). Deploying and experimenting large-scale mobile opportunistic networks is difficult, we too resort to simulation. Instead of using a complex mobility model to mimic people"s mobility patterns, we used mobility traces collected in a production wireless network at Dartmouth College to drive our simulation. Our messagegeneration model, however, was synthetic. To the best of our knowledge, we are the first to simulate the effect of routing protocols in a large-scale mobile opportunistic network, using realistic contact traces derived from real traces of a production network with more than 5, 000 users. Using realistic contact traces, we evaluate the performance of three naive routing protocols (direct-delivery, epidemic, and random) and two prediction-based routing protocols, PRoPHET [16] and Link-State [22]. We also propose a new prediction-based routing protocol, and compare it to the above in our evaluation. 2. ROUTING PROTOCOL A routing protocol is designed for forwarding messages from one node (source) to another node (destination). Any node may generate messages for any other node, and may carry messages destined for other nodes. In this paper, we consider only messages that are unicast (single destination). DTN routing protocols could be described in part by their transfer probability and replication probability; that is, when one node meets another node, what is the probability that a message should be transfered and if so, whether the sender should retain its copy. Two extremes are the direct-delivery protocol and the epidemic protocol. The former transfers with probability 1 when the node meets the destination, 0 for others, and no replication. The latter uses transfer probability 1 for all nodes and unlimited replication. Both these protocols have their advantages and disadvantages. All other protocols are between the two extremes. First, we define the notion of contact between two nodes. Then we describe five existing protocols before presenting our own proposal. A contact is defined as a period of time during which two nodes have the opportunity to communicate. Although we are aware that wireless technologies differ, we assume that a node can reliably detect the beginning and end time of a contact with nearby nodes. A node may be in contact with several other nodes at the same time. The contact history of a node is a sequence of contacts with other nodes. Node i has a contact history Hi(j), for each other node j, which denotes the historical contacts between node i and node j. We record the start and end time for each contact; however, the last contacts in the node"s contact history may not have ended. 2.1 Direct Delivery Protocol In this simple protocol, a message is transmitted only when the source node can directly communicate with the destination node of the message. In mobile opportunistic networks, however, the probability for the sender to meet the destination may be low, or even zero. 2.2 Epidemic Routing Protocol The epidemic routing protocol [23] floods messages into the network. The source node sends a copy of the message to every node that it meets. The nodes that receive a copy of the message also send a copy of the message to every node that they meet. Eventually, a copy of the message arrives at the destination of the message. This protocol is simple, but may use significant resources; excessive communication may drain each node"s battery quickly. Moreover, since each node keeps a copy of each message, storage is not used efficiently, and the capacity of the network is limited. At a minimum, each node must expire messages after some amount of time or stop forwarding them after a certain number of hops. After a message expires, the message will not be transmitted and will be deleted from the storage of any node that holds the message. An optimization to reduce the communication cost is to transfer index messages before transferring any data message. The index messages contain IDs of messages that a node currently holds. Thus, by examining the index messages, a node only transfers messages that are not yet contained on the other nodes. 2.3 Random Routing An obvious approach between the above two extremes is to select a transfer probability between 0 and 1 to forward messages at each contact. We use a simple replication strategy that allows only the source node to make replicas, and limits the replication to a specific number of copies. The message has some chance of being transferred to a highly mobile node, and thus may have a better chance to reach its destination before the message expires. 2.4 PRoPHET Protocol PRoPHET [16] is a Probabilistic Routing Protocol using History of past Encounters and Transitivity to estimate each node"s delivery probability for each other node. When node i meets node j, the delivery probability of node i for j is updated by pij = (1 − pij)p0 + pij, (1) where p0 is an initial probability, a design parameter for a given network. Lindgren et al. [16] chose 0.75, as did we in our evaluation. When node i does not meet j for some time, the delivery probability decreases by pij = αk pij, (2) where α is the aging factor (α < 1), and k is the number of time units since the last update. The PRoPHET protocol exchanges index messages as well as delivery probabilities. When node i receives node j"s delivery probabilities, node i may compute the transitive delivery probability through j to z with piz = piz + (1 − piz)pijpjzβ, (3) where β is a design parameter for the impact of transitivity; we used β = 0.25 as did Lindgren [16]. 2.5 Link-State Protocol Su et al. [22] use a link-state approach to estimate the weight of each path from the source of a message to the destination. They use the median inter-contact duration or exponentially aged intercontact duration as the weight on links. The exponentially aged inter-contact duration of node i and j is computed by wij = αwij + (1 − α)I, (4) where I is the new inter-contact duration and α is the aging factor. Nodes share their link-state weights when they can communicate with each other, and messages are forwarded to the node that have the path with the lowest link-state weight. 36 3. TIMELY-CONTACT PROBABILITY We also use historical contact information to estimate the probability of meeting other nodes in the future. But our method differs in that we estimate the contact probability within a period of time. For example, what is the contact probability in the next hour? Neither PRoPHET nor Link-State considers time in this way. One way to estimate the timely-contact probability is to use the ratio of the total contact duration to the total time. However, this approach does not capture the frequency of contacts. For example, one node may have a long contact with another node, followed by a long non-contact period. A third node may have a short contact with the first node, followed by a short non-contact period. Using the above estimation approach, both examples would have similar contact probability. In the second example, however, the two nodes have more frequent contacts. We design a method to capture the contact frequency of mobile nodes. For this purpose, we assume that even short contacts are sufficient to exchange messages.1 The probability for node i to meet node j is computed by the following procedure. We divide the contact history Hi(j) into a sequence of n periods of ΔT starting from the start time (t0) of the first contact in history Hi(j) to the current time. We number each of the n periods from 0 to n − 1, then check each period. If node i had any contact with node j during a given period m, which is [t0 + mΔT, t0 + (m + 1)ΔT), we set the contact status Im to be 1; otherwise, the contact status Im is 0. The probability p (0) ij that node i meets node j in the next ΔT can be estimated as the average of the contact status in prior intervals: p (0) ij = 1 n n−1X m=0 Im. (5) To adapt to the change of contact patterns, and reduce the storage space for contact histories, a node may discard old history contacts; in this situation, the estimate would be based on only the retained history. The above probability is the direct contact probability of two nodes. We are also interested in the probability that we may be able to pass a message through a sequence of k nodes. We define the k-order probability inductively, p (k) ij = p (0) ij + X α p (0) iα p (k−1) αj , (6) where α is any node other than i or j. 3.1 Our Routing Protocol We first consider the case of a two-hop path, that is, with only one relay node. We consider two approaches: either the receiving neighbor decides whether to act as a relay, or the source decides which neighbors to use as relay. 3.1.1 Receiver Decision Whenever a node meets other nodes, they exchange all their messages (or as above, index messages). If the destination of a message is the receiver itself, the message is delivered. Otherwise, if the probability of delivering the message to its destination through this receiver node within ΔT is greater than or equal to a certain threshold, the message is stored in the receiver"s storage to forward 1 In our simulation, however, we accurately model the communication costs and some short contacts will not succeed in transfer of all messages. to the destination. If the probability is less than the threshold, the receiver discards the message. Notice that our protocol replicates the message whenever a good-looking relay comes along. 3.1.2 Sender Decision To make decisions, a sender must have the information about its neighbors" contact probability with a message"s destination. Therefore, meta-data exchange is necessary. When two nodes meet, they exchange a meta-message, containing an unordered list of node IDs for which the sender of the metamessage has a contact probability greater than the threshold. After receiving a meta-message, a node checks whether it has any message that destined to its neighbor, or to a node in the node list of the neighbor"s meta-message. If it has, it sends a copy of the message. When a node receives a message, if the destination of the message is the receiver itself, the message is delivered. Otherwise, the message is stored in the receiver"s storage for forwarding to the destination. 3.1.3 Multi-node Relay When we use more than two hops to relay a message, each node needs to know the contact probabilities along all possible paths to the message destination. Every node keeps a contact probability matrix, in which each cell pij is a contact probability between to nodes i and j. Each node i computes its own contact probabilities (row i) with other nodes using Equation (5) whenever the node ends a contact with other nodes. Each row of the contact probability matrix has a version number; the version number for row i is only increased when node i updates the matrix entries in row i. Other matrix entries are updated through exchange with other nodes when they meet. When two nodes i and j meet, they first exchange their contact probability matrices. Node i compares its own contact matrix with node j"s matrix. If node j"s matrix has a row l with a higher version number, then node i replaces its own row l with node j"s row l. Likewise node j updates its matrix. After the exchange, the two nodes will have identical contact probability matrices. Next, if a node has a message to forward, the node estimates its neighboring node"s order-k contact probability to contact the destination of the message using Equation (6). If p (k) ij is above a threshold, or if j is the destination of the message, node i will send a copy of the message to node j. All the above effort serves to determine the transfer probability when two nodes meet. The replication decision is orthogonal to the transfer decision. In our implementation, we always replicate. Although PRoPHET [16] and Link-State [22] do no replication, as described, we added replication to those protocols for better comparison to our protocol. 4. EVALUATION RESULTS We evaluate and compare the results of direct delivery, epidemic, random, PRoPHET, Link-State, and timely-contact routing protocols. 4.1 Mobility traces We use real mobility data collected at Dartmouth College. Dartmouth College has collected association and disassociation messages from devices on its wireless network wireless users since spring 2001 [13]. Each message records the wireless card MAC address, the time of association/disassociation, and the name of the access point. We treat each unique MAC address as a node. For 37 more information about Dartmouth"s network and the data collection, see previous studies [7, 12]. Our data are not contacts in a mobile ad-hoc network. We can approximate contact traces by assuming that two users can communicate with each other whenever they are associated with the same access point. Chaintreau et al. [5] used Dartmouth data traces and made the same assumption to theoretically analyze the impact of human mobility on opportunistic forwarding algorithms. This assumption may not be accurate,2 but it is a good first approximation. In our simulation, we imagine the same clients and same mobility in a network with no access points. Since our campus has full WiFi coverage, we assume that the location of access points had little impact on users" mobility. We simulated one full month of trace data (November 2003) taken from CRAWDAD [13], with 5, 142 users. Although predictionbased protocols require prior contact history to estimate each node"s delivery probability, our preliminary results show that the performance improvement of warming-up over one month of trace was marginal. Therefore, for simplicity, we show the results of all protocols without warming-up. 4.2 Simulator We developed a custom simulator.3 Since we used contact traces derived from real mobility data, we did not need a mobility model and omitted physical and link-layer details for node discovery. We were aware that the time for neighbor discovery in different wireless technologies vary from less than one seconds to several seconds. Furthermore, connection establishment also takes time, such as DHCP. In our simulation, we assumed the nodes could discover and connect each other instantly when they were associated with a same AP. To accurately model communication costs, however, we simulated some MAC-layer behaviors, such as collision. The default settings of the network of our simulator are listed in Table 1, using the values recommended by other papers [22, 16]. The message probability was the probability of generating messages, as described in Section 4.3. The default transmission bandwidth was 11 Mb/s. When one node tried to transmit a message, it first checked whether any nearby node was transmitting. If it was, the node backed off a random number of slots. Each slot was 1 millisecond, and the maximum number of backoff slots was 30. The size of messages was uniformly distributed between 80 bytes and 1024 bytes. The hop count limit (HCL) was the maximum number of hops before a message should stop forwarding. The time to live (TTL) was the maximum duration that a message may exist before expiring. The storage capacity was the maximum space that a node can use for storing messages. For our routing method, we used a default prediction window ΔT of 10 hours and a probability threshold of 0.01. The replication factor r was not limited by default, so the source of a message transferred the messages to any other node that had a contact probability with the message destination higher than the probability threshold. 4.3 Message generation After each contact event in the contact trace, we generated a message with a given probability; we choose a source node and a des2 Two nodes may not have been able to directly communicate while they were at two far sides of an access point, or two nodes may have been able to directly communicate if they were between two adjacent access points. 3 We tried to use a general network simulator (ns2), which was extremely slow when simulating a large number of mobile nodes (in our case, more than 5000 nodes), and provided unnecessary detail in modeling lower-level network protocols. Table 1: Default Settings of the Simulation Parameter Default value message probability 0.001 bandwidth 11 Mb/s transmission slot 1 millisecond max backoff slots 30 message size 80-1024 bytes hop count limit (HCL) unlimited time to live (TTL) unlimited storage capacity unlimited prediction window ΔT 10 hours probability threshold 0.01 contact history length 20 replication always aging factor α 0.9 (0.98 PRoPHET) initial probability p0 0.75 (PRoPHET) transitivity impact β 0.25 (PRoPHET) 0 20000 40000 60000 80000 100000 120000 0 5 10 15 20 Numberofoccurrence hour movements contacts Figure 1: Movements and contacts duration each hour tination node randomly using a uniform distribution across nodes seen in the contact trace up to the current time. When there were more contacts during a certain period, there was a higher likelihood that a new message was generated in that period. This correlation is not unreasonable, since there were more movements during the day than during the night, and so the number of contacts. Figure 1 shows the statistics of the numbers of movements and the numbers of contacts during each hour of the day, summed across all users and all days. The plot shows a clear diurnal activity pattern. The activities reached lowest around 5am and peaked between 4pm and 5pm. We assume that in some applications, network traffic exhibits similar patterns, that is, people send more messages during the day, too. Messages expire after a TTL. We did not use proactive methods to notify nodes the delivery of messages, so that the messages can be removed from storage. 4.4 Metrics We define a set of metrics that we use in evaluating routing protocols in opportunistic networks: • delivery ratio, the ratio of the number of messages delivered to the number of total messages generated. • message transmissions, the total number of messages transmitted during the simulation across all nodes. 38 • meta-data transmissions, the total number of meta-data units transmitted during the simulation across all nodes. • message duplications, the number of times a message copy occurred, due to replication. • delay, the duration between a message"s generation time and the message"s delivery time. • storage usage, the max and mean of maximum storage (bytes) used across all nodes. 4.5 Results Here we compare simulation results of the six routing protocols. 0.001 0.01 0.1 1 unlimited 100 24 10 1 Deliveryratio Message time-to-live (TTL) (hour) direct random prediction state prophet epidemic Figure 2: Delivery ratio (log scale). The direct and random protocols for one-hour TTL had delivery ratios that were too low to be visible in the plot. Figure 2 shows the delivery ratio of all the protocols, with different TTLs. (In all the plots in the paper, prediction stands for our method, state stands for the Link-State protocol, and prophet represents PRoPHET.) Although we had 5,142 users in the network, the direct-delivery and random protocols had low delivery ratios (note the log scale). Even for messages with an unlimited lifetime, only 59 out of 2077 messages were delivered during this one-month simulation. The delivery ratio of epidemic routing was the best. The three prediction-based approaches had low delivery ratio, compared to epidemic routing. Although our method was slightly better than the other two, the advantage was marginal. The high delivery ratio of epidemic routing came with a price: excessive transmissions. Figure 3 shows the number of message data transmissions. The number of message transmissions of epidemic routing was more than 10 times higher than for the predictionbased routing protocols. Obviously, the direct delivery protocol had the lowest number of message transmissions - the number of message delivered. Among the three prediction-based methods, the PRoPHET transmitted fewer messages, but had comparable delivery-ratio as seen in Figure 2. Figure 4 shows that epidemic and all prediction-based methods had substantial meta-data transmissions, though epidemic routing had relatively more, with shorter TTLs. Because epidemic protocol transmitted messages at every contact, in turn, more nodes had messages that required meta-data transmission during contact. The direct-delivery and random protocols had no meta-data transmissions. In addition to its message transmissions and meta-data transmissions, the epidemic routing protocol also had excessive message 1 10 100 1000 10000 100000 1e+06 1e+07 1e+08 unlimited 100 24 10 1 Numberofmessagetransmitted Message time-to-live (TTL) (hour) direct random prediction state prophet epidemic Figure 3: Message transmissions (log scale) 1 10 100 1000 10000 100000 1e+06 1e+07 1e+08 unlimited 100 24 10 1 Numberofmeta-datatransmissions Message time-to-live (TTL) (hour) direct random prediction state prophet epidemic Figure 4: Meta-data transmissions (log scale). Direct and random protocols had no meta-data transmissions. duplications, spreading replicas of messages over the network. Figure 5 shows that epidemic routing had one or two orders more duplication than the prediction-based protocols. Recall that the directdelivery and random protocols did not replicate, thus had no data duplications. Figure 6 shows both the median and mean delivery delays. All protocols show similar delivery delays in both mean and median measures for medium TTLs, but differ for long and short TTLs. With a 100-hour TTL, or unlimited TTL, epidemic routing had the shortest delays. The direct-delivery had the longest delay for unlimited TTL, but it had the shortest delay for the one-hour TTL. The results seem contrary to our intuition: the epidemic routing protocol should be the fastest routing protocol since it spreads messages all over the network. Indeed, the figures show only the delay time for delivered messages. For direct delivery, random, and the probability-based routing protocols, relatively few messages were delivered for short TTLs, so many messages expired before they could reach their destination; those messages had infinite delivery delay and were not included in the median or mean measurements. For longer TTLs, more messages were delivered even for the directdelivery protocol. The statistics of longer TTLs for comparison are more meaningful than those of short TTLs. Since our message generation rate was low, the storage usage was also low in our simulation. Figure 7 shows the maximum and average of maximum volume (in KBytes) of messages stored 39 1 10 100 1000 10000 100000 1e+06 1e+07 1e+08 unlimited 100 24 10 1 Numberofmessageduplications Message time-to-live (TTL) (hour) direct random prediction state prophet epidemic Figure 5: Message duplications (log scale). Direct and random protocols had no message duplications. 1 10 100 1000 10000 unlimited100 24 10 1 unlimited100 24 10 1 Delay(minute) Message time-to-live (TTL) (hour) direct random prediction state prophet epidemic Mean delayMedian delay Figure 6: Median and mean delays (log scale). in each node. The epidemic routing had the most storage usage. The message time-to-live parameter was the big factor affecting the storage usage for epidemic and prediction-based routing protocols. We studied the impact of different parameters of our predictionbased routing protocol. Our prediction-based protocol was sensitive to several parameters, such as the probability threshold and the prediction window ΔT. Figure 8 shows the delivery ratios when we used different probability thresholds. (The leftmost value 0.01 is the value used for the other plots.) A higher probability threshold limited the transfer probability, so fewer messages were delivered. It also required fewer transmissions as shown in Figure 9. With a larger prediction window, we got a higher contact probability. Thus, for the same probability threshold, we had a slightly higher delivery ratio as shown in Figure 10, and a few more transmissions as shown in Figure 11. 5. RELATED WORK In addition to the protocols that we evaluated in our simulation, several other opportunistic network routing protocols have been proposed in the literature. We did not implement and evaluate these routing protocols, because either they require domain-specific information (location information) [14, 15], assume certain mobility patterns [17], present orthogonal approaches [10, 24] to other routing protocols. 0.1 1 10 100 1000 10000 unlimited100 24 10 1 unlimited100 24 10 1 Storageusage(KB) Message time-to-live (TTL) (hour) direct random prediction state prophet epidemic Mean of maximumMax of maximum Figure 7: Max and mean of maximum storage usage across all nodes (log scale). 0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1 Deliveryratio Probability threshold Figure 8: Probability threshold impact on delivery ratio of timely-contact routing. LeBrun et al. [14] propose a location-based delay-tolerant network routing protocol. Their algorithm assumes that every node knows its own position, and the destination is stationary at a known location. A node forwards data to a neighbor only if the neighbor is closer to the destination than its own position. Our protocol does not require knowledge of the nodes" locations, and learns their contact patterns. Leguay et al. [15] use a high-dimensional space to represent a mobility pattern, then routes messages to nodes that are closer to the destination node in the mobility pattern space. Location information of nodes is required to construct mobility patterns. Musolesi et al. [17] propose an adaptive routing protocol for intermittently connected mobile ad-hoc networks. They use a Kalman filter to compute the probability that a node delivers messages. This protocol assumes group mobility and cloud connectivity, that is, nodes move as a group, and among this group of nodes a contemporaneous end-to-end connection exists for every pair of nodes. When two nodes are in the same connected cloud, DSDV [19] routing is used. Network coding also draws much interest from DTN research. Erasure-coding [10, 24] explores coding algorithms to reduce message replicas. The source node replicates a message m times, then uses a coding scheme to encode them in one big message. After replicas are encoded, the source divides the big message into k 40 0 0.5 1 1.5 2 2.5 3 3.5 0 0.2 0.4 0.6 0.8 1 Numberofmessagetransmitted(million) Probability threshold Figure 9: Probability threshold impact on message transmission of timely-contact routing. 0 0.2 0.4 0.6 0.8 1 0.01 0.1 1 10 100 Deliveryratio Prediction window (hour) Figure 10: Prediction window impact on delivery ratio of timely-contact routing (semi-log scale). blocks of the same size, and transmits a block to each of the first k encountered nodes. If m of the blocks are received at the destination, the message can be restored, where m < k. In a uniformly distributed mobility scenario, the delivery probability increases because the probability that the destination node meets m relays is greater than it meets k relays, given m < k. 6. SUMMARY We propose a prediction-based routing protocol for opportunistic networks. We evaluate the performance of our protocol using realistic contact traces, and compare to five existing routing protocols. Our simulation results show that direct delivery had the lowest delivery ratio, the fewest data transmissions, and no meta-data transmission or data duplication. Direct delivery is suitable for devices that require an extremely low power consumption. The random protocol increased the chance of delivery for messages otherwise stuck at some low mobility nodes. Epidemic routing delivered the most messages. The excessive transmissions, and data duplication, however, consume more resources than portable devices may be able to provide. None of these protocols (direct-delivery, random and epidemic routing) are practical for real deployment of opportunistic networks, 0 0.5 1 1.5 2 2.5 3 3.5 0.01 0.1 1 10 100 Numberofmessagetransmitted(million) Prediction window (hour) Figure 11: Prediction window impact on message transmission of timely-contact routing (semi-log scale). because they either had an extremely low delivery ratio, or had an extremely high resource consumption. The prediction-based routing protocols had a delivery ratio more than 10 times better than that for direct-delivery and random routing, and fewer transmissions and less storage usage than epidemic routing. They also had fewer data duplications than epidemic routing. All the prediction-based routing protocols that we have evaluated had similar performance. Our method had a slightly higher delivery ratio, but more transmissions and higher storage usage. There are many parameters for prediction-based routing protocols, however, and different parameters may produce different results. Indeed, there is an opportunity for some adaptation; for example, high priority messages may be given higher transfer and replication probabilities to increase the chance of delivery and reduce the delay, or a node with infrequent contact may choose to raise its transfer probability. We only studied the impact of predicting peer-to-peer contact probability for routing in unicast messages. In some applications, context information (such as location) may be available for the peers. One may also consider other messaging models, for example, where messages are sent to a location, such that every node at that location will receive a copy of the message. Location prediction [21] may be used to predict nodes" mobility, and to choose as relays those nodes moving toward the destined location. Research on routing in opportunistic networks is still in its early stage. Many other issues of opportunistic networks, such as security and privacy, are mainly left open. We anticipate studying these issues in future work. 7. ACKNOWLEDGEMENT This research is a project of the Center for Mobile Computing and the Institute for Security Technology Studies at Dartmouth College. It was supported by DoCoMo Labs USA, the CRAWDAD archive at Dartmouth College (funded by NSF CRI Award 0454062), NSF Infrastructure Award EIA-9802068, and by Grant number 2005-DD-BX-1091 awarded by the Bureau of Justice Assistance. Points of view or opinions in this document are those of the authors and do not represent the official position or policies of any sponsor. 8. REFERENCES [1] John Burgess, Brian Gallagher, David Jensen, and Brian Neil Levine. MaxProp: routing for vehicle-based 41 disruption-tolerant networks. In Proceedings of the 25th IEEE International Conference on Computer Communications (INFOCOM), April 2006. [2] Scott Burleigh, Adrian Hooke, Leigh Torgerson, Kevin Fall, Vint Cerf, Bob Durst, Keith Scott, and Howard Weiss. Delay-tolerant networking: An approach to interplanetary Internet. IEEE Communications Magazine, 41(6):128-136, June 2003. [3] Tracy Camp, Jeff Boleng, and Vanessa Davies. A survey of mobility models for ad-hoc network research. Wireless Communication & Mobile Computing (WCMC): Special issue on Mobile ad-hoc Networking: Research, Trends and Applications, 2(5):483-502, 2002. [4] Andrew Campbell, Shane Eisenman, Nicholas Lane, Emiliano Miluzzo, and Ronald Peterson. People-centric urban sensing. In IEEE Wireless Internet Conference, August 2006. [5] Augustin Chaintreau, Pan Hui, Jon Crowcroft, Christophe Diot, Richard Gass, and James Scott. Impact of human mobility on the design of opportunistic forwarding algorithms. In Proceedings of the 25th IEEE International Conference on Computer Communications (INFOCOM), April 2006. [6] Kevin Fall. A delay-tolerant network architecture for challenged internets. In Proceedings of the 2003 Conference on Applications, Technologies, Architectures, and Protocols for Computer Communications (SIGCOMM), August 2003. [7] Tristan Henderson, David Kotz, and Ilya Abyzov. The changing usage of a mature campus-wide wireless network. In Proceedings of the 10th Annual International Conference on Mobile Computing and Networking (MobiCom), pages 187-201, September 2004. [8] Pan Hui, Augustin Chaintreau, James Scott, Richard Gass, Jon Crowcroft, and Christophe Diot. Pocket switched networks and human mobility in conference environments. In ACM SIGCOMM Workshop on Delay Tolerant Networking, pages 244-251, August 2005. [9] Ravi Jain, Dan Lelescu, and Mahadevan Balakrishnan. Model T: an empirical model for user registration patterns in a campus wireless LAN. In Proceedings of the 11th Annual International Conference on Mobile Computing and Networking (MobiCom), pages 170-184, 2005. [10] Sushant Jain, Mike Demmer, Rabin Patra, and Kevin Fall. Using redundancy to cope with failures in a delay tolerant network. In Proceedings of the 2005 Conference on Applications, Technologies, Architectures, and Protocols for Computer Communications (SIGCOMM), pages 109-120, August 2005. [11] Philo Juang, Hidekazu Oki, Yong Wang, Margaret Martonosi, Li-Shiuan Peh, and Daniel Rubenstein. Energy-efficient computing for wildlife tracking: Design tradeoffs and early experiences with ZebraNet. In the Tenth International Conference on Architectural Support for Programming Languages and Operating Systems, October 2002. [12] David Kotz and Kobby Essien. Analysis of a campus-wide wireless network. Wireless Networks, 11:115-133, 2005. [13] David Kotz, Tristan Henderson, and Ilya Abyzov. CRAWDAD data set dartmouth/campus. http://crawdad.cs.dartmouth.edu/dartmouth/campus, December 2004. [14] Jason LeBrun, Chen-Nee Chuah, Dipak Ghosal, and Michael Zhang. Knowledge-based opportunistic forwarding in vehicular wireless ad-hoc networks. In IEEE Vehicular Technology Conference, pages 2289-2293, May 2005. [15] Jeremie Leguay, Timur Friedman, and Vania Conan. Evaluating mobility pattern space routing for DTNs. In Proceedings of the 25th IEEE International Conference on Computer Communications (INFOCOM), April 2006. [16] Anders Lindgren, Avri Doria, and Olov Schelen. Probabilistic routing in intermittently connected networks. In Workshop on Service Assurance with Partial and Intermittent Resources (SAPIR), pages 239-254, 2004. [17] Mirco Musolesi, Stephen Hailes, and Cecilia Mascolo. Adaptive routing for intermittently connected mobile ad-hoc networks. In IEEE International Symposium on a World of Wireless Mobile and Multimedia Networks, pages 183-189, June 2005. extended version. [18] OLPC. One laptop per child project. http://laptop.org. [19] C. E. Perkins and P. Bhagwat. Highly dynamic destination-sequenced distance-vector routing (DSDV) for mobile computers. Computer Communication Review, pages 234-244, October 1994. [20] C. E. Perkins and E. M. Royer. ad-hoc on-demand distance vector routing. In IEEE Workshop on Mobile Computing Systems and Applications, pages 90-100, February 1999. [21] Libo Song, David Kotz, Ravi Jain, and Xiaoning He. Evaluating next-cell predictors with extensive Wi-Fi mobility data. IEEE Transactions on Mobile Computing, 5(12):1633-1649, December 2006. [22] Jing Su, Ashvin Goel, and Eyal de Lara. An empirical evaluation of the student-net delay tolerant network. In International Conference on Mobile and Ubiquitous Systems (MobiQuitous), July 2006. [23] Amin Vahdat and David Becker. Epidemic routing for partially-connected ad-hoc networks. Technical Report CS-2000-06, Duke University, July 2000. [24] Yong Wang, Sushant Jain, Margaret Martonosia, and Kevin Fall. Erasure-coding based routing for opportunistic networks. In ACM SIGCOMM Workshop on Delay Tolerant Networking, pages 229-236, August 2005. [25] Yu Wang and Hongyi Wu. DFT-MSN: the delay fault tolerant mobile sensor network for pervasive information gathering. In Proceedings of the 25th IEEE International Conference on Computer Communications (INFOCOM), April 2006. 42

contact trace;opportunistic network;route;epidemic protocol;frequent link break;end-to-end path;prophet;transfer probability;replication strategy;mobile opportunistic network;delay-tolerant network;random mobility model;direct-delivery protocol;simulation;realistic mobility trace;past encounter and transitivity history;history of past encounter and transitivity;unicast;routing protocol

train_C-50

CenWits: A Sensor-Based Loosely Coupled Search and Rescue System Using Witnesses

This paper describes the design, implementation and evaluation of a search and rescue system called CenWits. CenWits uses several small, commonly-available RF-based sensors, and a small number of storage and processing devices. It is designed for search and rescue of people in emergency situations in wilderness areas. A key feature of CenWits is that it does not require a continuously connected sensor network for its operation. It is designed for an intermittently connected network that provides only occasional connectivity. It makes a judicious use of the combined storage capability of sensors to filter, organize and store important information, combined battery power of sensors to ensure that the system remains operational for longer time periods, and intermittent network connectivity to propagate information to a processing center. A prototype of CenWits has been implemented using Berkeley Mica2 motes. The paper describes this implementation and reports on the performance measured from it.

1. INTRODUCTION Search and rescue of people in emergency situation in a timely manner is an extremely important service. It has been difficult to provide such a service due to lack of timely information needed to determine the current location of a person who may be in an emergency situation. With the emergence of pervasive computing, several systems [12, 19, 1, 5, 6, 4, 11] have been developed over the last few years that make use of small devices such as cell phones, sensors, etc. All these systems require a connected network via satellites, GSM base stations, or mobile devices. This requirement severely limits their applicability, particularly in remote wilderness areas where maintaining a connected network is very difficult. For example, a GSM transmitter has to be in the range of a base station to transmit. As a result, it cannot operate in most wilderness areas. While a satellite transmitter is the only viable solution in wilderness areas, it is typically expensive and cumbersome. Furthermore, a line of sight is required to transmit to satellite, and that makes it infeasible to stay connected in narrow canyons, large cities with skyscrapers, rain forests, or even when there is a roof or some other obstruction above the transmitter, e.g. in a car. An RF transmitter has a relatively smaller range of transmission. So, while an in-situ sensor is cheap as a single unit, it is expensive to build a large network that can provide connectivity over a large wilderness area. In a mobile environment where sensors are carried by moving people, power-efficient routing is difficult to implement and maintain over a large wilderness area. In fact, building an adhoc sensor network using only the sensors worn by hikers is nearly impossible due to a relatively small number of sensors spread over a large wilderness area. In this paper, we describe the design, implementation and evaluation of a search and rescue system called CenWits (Connection-less Sensor-Based Tracking System Using Witnesses). CenWits is comprised of mobile, in-situ sensors that are worn by subjects (people, wild animals, or in-animate objects), access points (AP) that collect information from these sensors, and GPS receivers and location points (LP) that provide location information to the sensors. A subject uses GPS receivers (when it can connect to a satellite) and LPs to determine its current location. The key idea of CenWits is that it uses a concept of witnesses to convey a subject"s movement and location information to the outside world. This averts a need for maintaining a connected network to transmit location information to the outside world. In particular, there is no need for expensive GSM or satellite transmitters, or maintaining an adhoc network of in-situ sensors in CenWits. 180 CenWits employs several important mechanisms to address the key problem of resource constraints (low signal strength, low power and limited memory) in sensors. In particular, it makes a judicious use of the combined storage capability of sensors to filter, organize and store important information, combined battery power of sensors to ensure that the system remains operational for longer time periods, and intermittent network connectivity to propagate information to a processing center. The problem of low signal strengths (short range RF communication) is addressed by avoiding a need for maintaining a connected network. Instead, CenWits propagates the location information of sensors using the concept of witnesses through an intermittently connected network. As a result, this system can be deployed in remote wilderness areas, as well as in large urban areas with skyscrapers and other tall structures. Also, this makes CenWits cost-effective. A subject only needs to wear light-weight and low-cost sensors that have GPS receivers but no expensive GSM or satellite transmitters. Furthermore, since there is no need for a connected sensor network, there is no need to deploy sensors in very large numbers. The problem of limited battery life and limited memory of a sensor is addressed by incorporating the concepts of groups and partitions. Groups and partitions allow sensors to stay in sleep or receive modes most of the time. Using groups and partitions, the location information collected by a sensor can be distributed among several sensors, thereby reducing the amount of memory needed in one sensor to store that information. In fact, CenWits provides an adaptive tradeoff between memory and power consumption of sensors. Each sensor can dynamically adjust its power and memory consumption based on its remaining power or available memory. It has amply been noted that the strength of sensor networks comes from the fact that several sensor nodes can be distributed over a relatively large area to construct a multihop network. This paper demonstrates that important large-scale applications can be built using sensors by judiciously integrating the storage, communication and computation capabilities of sensors. The paper describes important techniques to combine memory, transmission and battery power of many sensors to address resource constraints in the context of a search and rescue application. However, these techniques are quite general. We discuss several other sensor-based applications that can employ these techniques. While CenWits addresses the general location tracking and reporting problem in a wide-area network, there are two important differences from the earlier work done in this area. First, unlike earlier location tracking solutions, CenWits does not require a connected network. Second, unlike earlier location tracking solutions, CenWits does not aim for a very high accuracy of localization. Instead, the main goal is to provide an approximate, small area where search and rescue efforts can be concentrated. The rest of this paper is organized as follows. In Section 2, we overview some of the recent projects and technologies related to movement and location tracking, and search and rescue systems. In Section 3, we describe the overall architecture of CenWits, and provide a high-level description of its functionality. In the next section, Section 4, we discuss power and memory management in CenWits. To simplify our presentation, we will focus on a specific application of tracking lost/injured hikers in all these sections. In Section 6, we describe a prototype implementation of CenWits and present performance measured from this implementation. We discuss how the ideas of CenWits can be used to build several other applications in Section 7. Finally, in Section 8, we discuss some related issues and conclude the paper. 2. RELATED WORK A survey of location systems for ubiquitous computing is provided in [11]. A location tracking system for adhoc sensor networks using anchor sensors as reference to gain location information and spread it out to outer node is proposed in [17]. Most location tracking systems in adhoc sensor networks are for benefiting geographic-aware routing. They don"t fit well for our purposes. The well-known active badge system [19] lets a user carry a badge around. An infrared sensor in the room can detect the presence of a badge and determine the location and identification of the person. This is a useful system for indoor environment, where GPS doesn"t work. Locationing using 802.11 devices is probably the cheapest solution for indoor position tracking [8]. Because of the popularity and low cost of 802.11 devices, several business solutions based on this technology have been developed[1]. A system that combines two mature technologies and is viable in suburban area where a user can see clear sky and has GSM cellular reception at the same time is currently available[5]. This system receives GPS signal from a satellite and locates itself, draws location on a map, and sends location information through GSM network to the others who are interested in the user"s location. A very simple system to monitor children consists an RF transmitter and a receiver. The system alarms the holder of the receiver when the transmitter is about to run out of range [6]. Personal Locater Beacons (PLB) has been used for avalanche rescuing for years. A skier carries an RF transmitter that emits beacons periodically, so that a rescue team can find his/her location based on the strength of the RF signal. Luxury version of PLB combines a GPS receiver and a COSPASSARSAT satellite transmitter that can transmit user"s location in latitude and longitude to the rescue team whenever an accident happens [4]. However, the device either is turned on all the time resulting in fast battery drain, or must be turned on after the accident to function. Another related technology in widespread use today is the ONSTAR system [3], typically used in several luxury cars. In this system, a GPS unit provides position information, and a powerful transmitter relays that information via satellite to a customer service center. Designed for emergencies, the system can be triggered either by the user with the push of a button, or by a catastrophic accident. Once the system has been triggered, a human representative attempts to gain communication with the user via a cell phone built as an incar device. If contact cannot be made, emergency services are dispatched to the location provided by GPS. Like PLBs, this system has several limitations. First, it is heavy and expensive. It requires a satellite transmitter and a connected network. If connectivity with either the GPS network or a communication satellite cannot be maintained, the system fails. Unfortunately, these are common obstacles encountered in deep canyons, narrow streets in large cities, parking garages, and a number of other places. 181 The Lifetch system uses GPS receiver board combined with a GSM/GPRS transmitter and an RF transmitter in one wireless sensor node called Intelligent Communication Unit (ICU). An ICU first attempts to transmit its location to a control center through GSM/GPRS network. If that fails, it connects with other ICUs (adhoc network) to forward its location information until the information reaches an ICU that has GSM/GPRS reception. This ICU then transmits the location information of the original ICU via the GSM/GPRS network. ZebraNet is a system designed to study the moving patterns of zebras [13]. It utilizes two protocols: History-based protocol and flooding protocol. History-based protocol is used when the zebras are grazing and not moving around too much. While this might be useful for tracking zebras, it"s not suitable for tracking hikers because two hikers are most likely to meet each other only once on a trail. In the flooding protocol, a node dumps its data to a neighbor whenever it finds one and doesn"t delete its own copy until it finds a base station. Without considering routing loops, packet filtering and grouping, the size of data on a node will grow exponentially and drain the power and memory of a sensor node with in a short time. Instead, Cenwits uses a four-phase hand-shake protocol to ensure that a node transmits only as much information as the other node is willing to receive. While ZebraNet is designed for a big group of sensors moving together in the same direction with same speed, Cenwits is designed to be used in the scenario where sensors move in different directions at different speeds. Delay tolerant network architecture addresses some important problems in challenged (resource-constrained) networks [9]. While this work is mainly concerned with interoperability of challenged networks, some problems related to occasionally-connected networks are similar to the ones we have addressed in CenWits. Among all these systems, luxury PLB and Lifetch are designed for location tracking in wilderness areas. However, both of these systems require a connected network. Luxury PLB requires the user to transmit a signal to a satellite, while Lifetch requires connection to GSM/GPRS network. Luxury PLB transmits location information, only when an accident happens. However, if the user is buried in the snow or falls into a deep canyon, there is almost no chance for the signal to go through and be relayed to the rescue team. This is because satellite transmission needs line of sight. Furthermore, since there is no known history of user"s location, it is not possible for the rescue team to infer the current location of the user. Another disadvantage of luxury PLB is that a satellite transmitter is very expensive, costing in the range of $750. Lifetch attempts to transmit the location information by GSM/GPRS and adhoc sensor network that uses AODV as the routing protocol. However, having a cellular reception in remote areas in wilderness areas, e.g. American national parks is unlikely. Furthermore, it is extremely unlikely that ICUs worn by hikers will be able to form an adhoc network in a large wilderness area. This is because the hikers are mobile and it is very unlikely to have several ICUs placed dense enough to forward packets even on a very popular hike route. CenWits is designed to address the limitations of systems such as luxury PLB and Lifetch. It is designed to provide hikers, skiers, and climbers who have their activities mainly in wilderness areas a much higher chance to convey their location information to a control center. It is not reliant upon constant connectivity with any communication medium. Rather, it communicates information along from user to user, finally arriving at a control center. Unlike several of the systems discussed so far, it does not require that a user"s unit is constantly turned on. In fact, it can discover a victim"s location, even if the victim"s sensor was off at the time of accident and has remained off since then. CenWits solves one of the greatest problems plaguing modern search and rescue systems: it has an inherent on-site storage capability. This means someone within the network will have access to the last-known-location information of a victim, and perhaps his bearing and speed information as well. Figure 1: Hiker A and Hiker B are are not in the range of each other 3. CENWITS We describe CenWits in the context of locating lost/injured hikers in wilderness areas. Each hiker wears a sensor (MICA2 motes in our prototype) equipped with a GPS receiver and an RF transmitter. Each sensor is assigned a unique ID and maintains its current location based on the signal received by its GPS receiver. It also emits beacons periodically. When any two sensors are in the range of one another, they record the presence of each other (witness information), and also exchange the witness information they recorded earlier. The key idea here is that if two sensors come with in range of each other at any time, they become each other"s witnesses. Later on, if the hiker wearing one of these sensors is lost, the other sensor can convey the last known (witnessed) location of the lost hiker. Furthermore, by exchanging the witness information that each sensor recorded earlier, the witness information is propagated beyond a direct contact between two sensors. To convey witness information to a processing center or to a rescue team, access points are established at well-known locations that the hikers are expected to pass through, e.g. at the trail heads, trail ends, intersection of different trails, scenic view points, resting areas, and so on. Whenever a sensor node is in the vicinity of an access point, all witness information stored in that sensor is automatically dumped to the access point. Access points are connected to a processing center via satellite or some other network1 . The witness information is downloaded to the processing center from various access points at regular intervals. In case, connection to an access point is lost, the information from that 1 A connection is needed only between access points and a processing center. There is no need for any connection between different access points. 182 access point can be downloaded manually, e.g. by UAVs. To estimate the speed, location and direction of a hiker at any point in time, all witness information of that hiker that has been collected from various access points is processed. Figure 2: Hiker A and Hiker B are in the range of each other. A records the presence of B and B records the presence of A. A and B become each other"s witnesses. Figure 3: Hiker A is in the range of an access point. It uploads its recorded witness information and clears its memory. An example of how CenWits operates is illustrated in Figures 1, 2 and 3. First, hikers A and B are on two close trails, but out of range of each other (Figure 1). This is a very common scenario during a hike. For example, on a popular four-hour hike, a hiker might run into as many as 20 other hikers. This accounts for one encounter every 12 minutes on average. A slow hiker can go 1 mile (5,280 feet) per hour. Thus in 12 minutes a slow hiker can go as far as 1056 feet. This implies that if we were to put 20 hikers on a 4-hour, one-way hike evenly, the range of each sensor node should be at least 1056 feet for them to communicate with one another continuously. The signal strength starts dropping rapidly for two Mica2 nodes to communicate with each other when they are 180 feet away, and is completely lost when they are 230 feet away from each other[7]. So, for the sensors to form a sensor network on a 4-hour hiking trail, there should be at least 120 hikers scattered evenly. Clearly, this is extremely unlikely. In fact, in a 4-hour, less-popular hiking trail, one might only run into say five other hikers. CenWits takes advantage of the fact that sensors can communicate with one another and record their presence. Given a walking speed of one mile per hour (88 feet per minute) and Mica2 range of about 150 feet for non-line-of-sight radio transmission, two hikers have about 150/88 = 1.7 minutes to discover the presence of each other and exchange their witness information. We therefore design our system to have each sensor emit a beacon every one-and-a-half minute. In Figure 2, hiker B"s sensor emits a beacon when A is in range, this triggers A to exchange data with B. A communicates the following information to B: My ID is A; I saw C at 1:23 PM at (39◦ 49.3277655", 105◦ 39.1126776"), I saw E at 3:09 PM at (40◦ 49.2234879", 105◦ 20.3290168"). B then replies with My ID is B; I saw K at 11:20 AM at (39◦ 51.4531655", 105◦ 41.6776223"). In addition, A records I saw B at 4:17 PM at (41◦ 29.3177354", 105◦ 04.9106211") and B records I saw A at 4:17 PM at (41◦ 29.3177354", 105◦ 04.9106211"). B goes on his way to overnight camping while A heads back to trail head where there is an AP, which emits beacon every 5 seconds to avoid missing any hiker. A dumps all witness information it has collected to the access point. This is shown in Figure 3. 3.1 Witness Information: Storage A critical concern is that there is limited amount of memory available on motes (4 KB SDRAM memory, 128 KB flash memory, and 4-512 KB EEPROM). So, it is important to organize witness information efficiently. CenWits stores witness information at each node as a set of witness records (Format is shown in Figure 4. 1 B Node ID Record Time X, Y Location Time Hop Count 1 B 3 B 8 B 3 B Figure 4: Format of a witness record. When two nodes i and j encounter each other, each node generates a new witness record. In the witness record generated by i, Node ID is j, Record Time is the current time in i"s clock, (X,Y) are the coordinates of the location of i that i recorded most recently (either from satellite or an LP), Location Time is the time when the this location was recorded, and Hop Count is 0. Each node is assigned a unique Node Id when it enters a trail. In our current prototype, we have allocated one byte for Node Id, although this can be increased to two or more bytes if a large number of hikers are expected to be present at the same time. We can represent time in 17 bits to a second precision. So, we have allocated 3 bytes each for Record Time and Location Time. The circumference of the Earth is approximately 40,075 KM. If we use a 32-bit number to represent both longitude and latitude, the precision we get is 40,075,000/232 = 0.0093 meter = 0.37 inches, which is quite precise for our needs. So, we have allocated 4 bytes each for X and Y coordinates of the location of a node. In fact, a foot precision can be achieved by using only 27 bits. 3.2 Location Point and Location Inference Although a GPS receiver provides an accurate location information, it has it"s limitation. In canyons and rainy forests, a GPS receiver does not work. When there is a heavy cloud cover, GPS users have experienced inaccuracy in the reported location as well. Unfortunately, a lot of hiking trails are in dense forests and canyons, and it"s not that uncommon to rain after hikers start hiking. To address this, CenWits incorporates the idea of location points (LP). A location point can update a sensor node with its current location whenever the node is near that LP. LPs are placed at different locations in a wilderness area where GPS receivers don"t work. An LP is a very simple device that emits prerecorded location information at some regular time interval. It can be placed in difficult-to-reach places such as deep canyons and dense rain forests by simply dropping them from an airplane. LPs allow a sensor node to determine its current location more accurately. However, they are not 183 an essential requirement of CenWits. If an LP runs out of power, the CenWits will continue to work correctly. Figure 5: GPS receiver not working correctly. Sensors then have to rely on LP to provide coordination In Figure 5, B cannot get GPS reception due to bad weather. It then runs into A on the trail who doesn"t have GPS reception either. Their sensors record the presence of each other. After 10 minutes, A is in range of an LP that provides an accurate location information to A. When A returns to trail head and uploads its data (Figure 6), the system can draw a circle centered at the LP from which A fetched location information for the range of encounter location of A and B. By Overlapping this circle with the trail map, two or three possible location of encounter can be inferred. Thus when a rescue is required, the possible location of B can be better inferred (See Figures 7 and 8). Figure 6: A is back to trail head, It reports the time of encounter with B to AP, but no location information to AP Figure 7: B is still missing after sunset. CenWits infers the last contact point and draws the circle of possible current locations based on average hiking speed CenWits requires that the clocks of different sensor nodes be loosely synchronized with one another. Such a synchronization is trivial when GPS coverage is available. In addition, sensor nodes in CenWits synchronize their clocks whenever they are in the range of an AP or an LP. The Figure 8: Based on overlapping landscape, B might have hiked to wrong branch and fallen off a cliff. Hot rescue areas can thus be determined synchronization accuracy Cenwits needs is of the order of a second or so. As long as the clocks are synchronized with in one second range, whether A met B at 12:37"45 or 12:37"46 doesn"t matter in the ordering of witness events and inferring the path. 4. MEMORY AND POWER MANAGEMENT CenWits employs several important mechanisms to conserve power and memory. It is important to note while current sensor nodes have limited amount of memory, future sensor nodes are expected to have much more memory. With this in mind, the main focus in our design is to provide a tradeoff between the amount of memory available and amount of power consumption. 4.1 Memory Management The size of witness information stored at a node can get very large. This is because the node may come across several other nodes during a hike, and may end up accumulating a large amount of witness information over time. To address this problem, CenWits allows a node to pro-actively free up some parts of its memory periodically. This raises an interesting question of when and which witness record should be deleted from the memory of a node? CenWits uses three criteria to determine this: record count, hop count, and record gap. Record count refers to the number of witness records with same node id that a node has stored in its memory. A node maintains an integer parameter MAX RECORD COUNT. It stores at most MAX RECORD COUNT witness records of any node. Every witness record has a hop count field that stores the number times (hops) this record has been transferred since being created. Initially this field is set to 0. Whenever a node receives a witness record from another node, it increments the hop count of that record by 1. A node maintains an integer parameter called MAX HOP COUNT. It keeps only those witness records in its memory, whose hop count is less than MAX HOP COUNT. The MAX HOP COUNT parameter provides a balance between two conflicting goals: (1) To ensure that a witness record has been propagated to and thus stored at as many nodes as possible, so that it has a high probability of being dumped at some AP as quickly as possible; and (2) To ensure that a witness record is stored only at a few nodes, so that it does not clog up too much of the combined memory of all sensor nodes. We chose to use hop count instead of time-to-live to decide when to drop a packet. The main reason for this is that the probability of a packet reaching an AP goes up as the hop count adds up. For example, when the hop count if 5 for a specific record, 184 the record is in at least 5 sensor nodes. On the other hand, if we discard old records, without considering hop count, there is no guarantee that the record is present in any other sensor node. Record gap refers to the time difference between the record times of two witness records with the same node id. To save memory, a node n ensures the the record gap between any two witness records with the same node id is at least MIN RECORD GAP. For each node id i, n stores the witness record with the most recent record time rti, the witness with most recent record time that is at least MIN RECORD GAP time units before rti, and so on until the record count limit (MAX RECORD COUNT) is reached. When a node is tight in memory, it adjusts the three parameters, MAX RECORD COUNT, MAX HOP COUNT and MIN RECORD GAP to free up some memory. It decrements MAX RECORD COUNT and MAX HOP COUNT, and increments MIN RECORD GAP. It then first erases all witness records whose hop count exceeds the reduced MAX HOP COUNT value, and then erases witness records to satisfy the record gap criteria. Also, when a node has extra memory space available, e.g. after dumping its witness information at an access point, it resets MAX RECORD COUNT, MAX HOP COUNT and MIN RECORD GAP to some predefined values. 4.2 Power Management An important advantage of using sensors for tracking purposes is that we can regulate the behavior of a sensor node based on current conditions. For example, we mentioned earlier that a sensor should emit a beacon every 1.7 minute, given a hiking speed of 1 mile/hour. However, if a user is moving 10 feet/sec, a beacon should be emitted every 10 seconds. If a user is not moving at all, a beacon can be emitted every 10 minutes. In the night, a sensor can be put into sleep mode to save energy, when a user is not likely to move at all for a relatively longer period of time. If a user is active for only eight hours in a day, we can put the sensor into sleep mode for the other 16 hours and thus save 2/3rd of the energy. In addition, a sensor node can choose to not send any beacons during some time intervals. For example, suppose hiker A has communicated its witness information to three other hikers in the last five minutes. If it is running low on power, it can go to receive mode or sleep mode for the next ten minutes. It goes to receive mode if it is still willing to receive additional witness information from hikers that it encounters in the next ten minutes. It goes to sleep mode if it is extremely low on power. The bandwidth and energy limitations of sensor nodes require that the amount of data transferred among the nodes be reduced to minimum. It has been observed that in some scenarios 3000 instructions could be executed for the same energy cost of sending a bit 100m by radio [15]. To reduce the amount of data transfer, CenWits employs a handshake protocol that two nodes execute when they encounter one another. The goal of this protocol is to ensure that a node transmits only as much witness information as the other node is willing to receive. This protocol is initiated when a node i receives a beacon containing the node ID of the sender node j and i has not exchanged witness information with j in the last δ time units. Assume that i < j. The protocol consists of four phases (See Figure 9): 1. Phase I: Node i sends its receive constraints and the number of witness records it has in its memory. 2. Phase II: On receiving this message from i, j sends its receive constraints and the number of witness records it has in its memory. 3. Phase III: On receiving the above message from j, i sends its witness information (filtered based on receive constraints received in phase II). 4. Phase IV: After receiving the witness records from i, j sends its witness information (filtered based on receive constraints received in phase I). j <Constaints, Witness info size> <Constaints, Witness info size> <Filtered Witness info> <Filtered Witness info> i j j j i i i Figure 9: Four-Phase Hand Shake Protocol (i < j) Receive constraints are a function of memory and power. In the most general case, they are comprised of the three parameters (record count, hop count and record gap) used for memory management. If i is low on memory, it specifies the maximum number of records it is willing to accept from j. Similarly, i can ask j to send only those records that have hop count value less than MAX HOP COUNT − 1. Finally, i can include its MIN RECORD GAP value in its receive constraints. Note that the handshake protocol is beneficial to both i and j. They save memory by receiving only as much information as they are willing to accept and conserve energy by sending only as many witness records as needed. It turns out that filtering witness records based on MIN RECORD GAP is complex. It requires that the witness records of any given node be arranged in an order sorted by their record time values. Maintaining this sorted order is complex in memory, because new witness records with the same node id can arrive later that may have to be inserted in between to preserve the sorted order. For this reason, the receive constraints in the current CenWits prototype do not include record gap. Suppose i specifies a hop count value of 3. In this case, j checks the hop count field of every witness record before sending them. If the hop count value is greater than 3, the record is not transmitted. 4.3 Groups and Partitions To further reduce communication and increase the lifetime of our system, we introduce the notion of groups. The idea is based on the concept of abstract regions presented in [20]. A group is a set of n nodes that can be defined in terms of radio connectivity, geographic location, or other properties of nodes. All nodes within a group can communicate directly with one another and they share information to maintain their view of the external world. At any point in time, a group has exactly one leader that communicates 185 with external nodes on behalf of the entire group. A group can be static, meaning that the group membership does not change over the period of time, or it could be dynamic in which case nodes can leave or join the group. To make our analysis simple and to explain the advantages of group, we first discuss static groups. A static group is formed at the start of a hiking trail or ski slope. Suppose there are five family members who want to go for a hike in the Rocky Mountain National Park. Before these members start their hike, each one of them is given a sensor node and the information is entered in the system that the five nodes form a group. Each group member is given a unique id and every group member knows about other members of the group. The group, as a whole, is also assigned an id to distinguish it from other groups in the system. Figure 10: A group of five people. Node 2 is the group leader and it is communicating on behalf of the group with an external node 17. All other (shown in a lighter shade) are in sleep mode. As the group moves through the trail, it exchanges information with other nodes or groups that it comes across. At any point in time, only one group member, called the leader, sends and receives information on behalf of the group and all other n − 1 group members are put in the sleep mode (See Figure 10). It is this property of groups that saves us energy. The group leadership is time-multiplexed among the group members. This is done to make sure that a single node does not run out of battery due to continuous exchange of information. Thus after every t seconds, the leadership is passed on to another node, called the successor, and the leader (now an ordinary member) is put to sleep. Since energy is dear, we do not implement an extensive election algorithm for choosing the successor. Instead, we choose the successor on the basis of node id. The node with the next highest id in the group is chosen as the successor. The last node, of course, chooses the node with the lowest id as its successor. We now discuss the data storage schemes for groups. Memory is a scarce resource in sensor nodes and it is therefore important that witness information be stored efficiently among group members. Efficient data storage is not a trivial task when it comes to groups. The tradeoff is between simplicity of the scheme and memory savings. A simpler scheme incurs lesser energy cost as compared to a more sophisticated scheme, but offers lesser memory savings as well. This is because in a more complicated scheme, the group members have to coordinate to update and store information. After considering a number of different schemes, we have come to a conclusion that there is no optimal storage scheme for groups. The system should be able to adapt according to its requirements. If group members are low on battery, then the group can adapt a scheme that is more energy efficient. Similarly, if the group members are running out of memory, they can adapt a scheme that is more memory efficient. We first present a simple scheme that is very energy efficient but does not offer significant memory savings. We then present an alternate scheme that is much more memory efficient. As already mentioned a group can receive information only through the group leader. Whenever the leader comes across an external node e, it receives information from that node and saves it. In our first scheme, when the timeslot for the leader expires, the leader passes this new information it received from e to its successor. This is important because during the next time slot, if the new leader comes across another external node, it should be able to pass information about all the external nodes this group has witnessed so far. Thus the information is fully replicated on all nodes to maintain the correct view of the world. Our first scheme does not offer any memory savings but is highly energy efficient and may be a good choice when the group members are running low on battery. Except for the time when the leadership is switched, all n − 1 members are asleep at any given time. This means that a single member is up for t seconds once every n∗t seconds and therefore has to spend approximately only 1/nth of its energy. Thus, if there are 5 members in a group, we save 80% energy, which is huge. More energy can be saved by increasing the group size. We now present an alternate data storage scheme that aims at saving memory at the cost of energy. In this scheme we divide the group into what we call partitions. Partitions can be thought of as subgroups within a group. Each partition must have at least two nodes in it. The nodes within a partition are called peers. Each partition has one peer designated as partition leader. The partition leader stays in receive mode at all times, while all others peers a partition stay in the sleep mode. Partition leadership is time-multiplexed among the peers to make sure that a single node does not run out of battery. Like before, a group has exactly one leader and the leadership is time-multiplexed among partitions. The group leader also serves as the partition leader for the partition it belongs to (See Figure 11). In this scheme, all partition leaders participate in information exchange. Whenever a group comes across an external node e, every partition leader receives all witness information, but it only stores a subset of that information after filtering. Information is filtered in such a way that each partition leader has to store only B/K bytes of data, where K is the number of partitions and B is the total number of bytes received from e. Similarly when a group wants to send witness information to e, each partition leader sends only B/K bytes that are stored in the partition it belongs to. However, before a partition leader can send information, it must switch from receive mode to send mode. Also, partition leaders must coordinate with one another to ensure that they do not send their witness information at the same time, i.e. their message do not collide. All this is achieved by having the group leader send a signal to every partition leader in turn. 186 Figure 11: The figure shows a group of eight nodes divided into four partitions of 2 nodes each. Node 1 is the group leader whereas nodes 2, 9, and 7 are partition leaders. All other nodes are in the sleep mode. Since the partition leadership is time-multiplexed, it is important that any information received by the partition leader, p1, be passed on to the next leader, p2. This has to be done to make sure that p2 has all the information that it might need to send when it comes across another external node during its timeslot. One way of achieving this is to wake p2 up just before p1"s timeslot expires and then have p1 transfer information only to p2. An alternate is to wake all the peers up at the time of leadership change, and then have p1 broadcast the information to all peers. Each peer saves the information sent by p1 and then goes back to sleep. In both cases, the peers send acknowledgement to the partition leader after receiving the information. In the former method, only one node needs to wake up at the time of leadership change, but the amount of information that has to be transmitted between the nodes increases as time passes. In the latter case, all nodes have to be woken up at the time of leadership change, but small piece of information has to be transmitted each time among the peers. Since communication is much more expensive than bringing the nodes up, we prefer the second method over the first one. A group can be divided into partitions in more than one way. For example, suppose we have a group of six members. We can divide this group into three partitions of two peers each, or two partitions with three peers each. The choice once again depends on the requirements of the system. A few big partitions will make the system more energy efficient. This is because in this configuration, a greater number of nodes will stay in sleep mode at any given point in time. On the other hand, several small partitions will make the system memory efficient, since each node will have to store lesser information (See Figure 12). A group that is divided into partitions must be able to readjust itself when a node leaves or runs out of battery. This is crucial because a partition must have at least two nodes at any point in time to tolerate failure of one node. For example, in figure 3 (a), if node 2 or node 5 dies, the partition is left with only one node. Later on, if that single node in the partition dies, all witness information stored in that partition will be lost. We have devised a very simple protocol to solve this problem. We first explain how partiFigure 12: The figure shows two different ways of partitioning a group of six nodes. In (a), a group is divided into three partitions of two nodes. Node 1 is the group leader, nodes 9 and 5 are partition leaders, and nodes 2, 3, and 6 are in sleep mode. In (b) the group is divided into two partitions of three nodes. Node 1 is the group leader, node 9 is the partition leader and nodes 2, 3, 5, and 6 are in sleep mode. tions are adjusted when a peer dies, and then explain what happens if a partition leader dies. Suppose node 2 in figure 3 (a) dies. When node 5, the partition leader, sends information to node 2, it does not receive an acknowledgement from it and concludes that node 2 has died2 . At this point, node 5 contacts other partition leaders (nodes 1 ... 9) using a broadcast message and informs them that one of its peers has died. Upon hearing this, each partition leader informs node 5 (i) the number of nodes in its partition, (ii) a candidate node that node 5 can take if the number of nodes in its partition is greater than 2, and (iii) the amount of witness information stored in its partition. Upon hearing from every leader, node 5 chooses the candidate node from the partition with maximum number (must be greater than 2) of peers, and sends a message back to all leaders. Node 5 then sends data to its new peer to make sure that the information is replicated within the partition. However, if all partitions have exactly two nodes, then node 5 must join another partition. It chooses the partition that has the least amount of witness information to join. It sends its witness information to the new partition leader. Witness information and membership update is propagated to all peers during the next partition leadership change. We now consider the case where the partition leader dies. If this happens, then we wait for the partition leadership to change and for the new partition leader to eventually find out that a peer has died. Once the new partition leader finds out that it needs more peers, it proceeds with the protocol explained above. However, in this case, we do lose information that the previous partition leader might have received just before it died. This problem can be solved by implementing a more rigorous protocol, but we have decided to give up on accuracy to save energy. Our current design uses time-division multiplexing to schedule wakeup and sleep modes in the sensor nodes. However, recent work on radio wakeup sensors [10] can be used to do this scheduling more efficiently. we plan to incorporate radio wakeup sensors in CenWits when the hardware is mature. 2 The algorithm to conclude that a node has died can be made more rigorous by having the partition leader query the suspected node a few times. 187 5. SYSTEM EVALUATION A sensor is constrained in the amount of memory and power. In general, the amount of memory needed and power consumption depends on a variety of factors such as node density, number of hiker encounters, and the number of access points. In this Section, we provide an estimate of how long the power of a MICA2 mote will last under certain assumtions. First, we assume that each sensor node carries about 100 witness records. On encountering another hiker, a sensor node transmits 50 witness records and receives 50 new witness records. Since, each record is 16 bytes long, it will take 0.34 seconds to transmit 50 records and another 0.34 seconds to receive 50 records over a 19200 bps link. The power consumption of MICA2 due to CPU processing, transmission and reception are approximately 8.0 mA, 7.0 mA and 8.5 mA per hour respectively [18], and the capacity of an alkaline battery is 2500mAh. Since the radio module of Mica2 is half-duplex and assuming that the CPU is always active when a node is awake, power consumption due to transmission is 8 + 8.5 = 16.5 mA per hour and due to reception is 8 + 7 = 15mA per hour. So, average power consumtion due to transmission and reception is (16.5 + 15)/2 = 15.75 mA per hour. Given that the capacity of an alkaline battery is 2500 mAh, a battery should last for 2500/15.75 = 159 hours of transmission and reception. An encounter between two hikers results in exchange of about 50 witness records that takes about 0.68 seconds as calculated above. Thus, a single alkaline battery can last for (159 ∗ 60 ∗ 60)/0.68 = 841764 hiker encounters. Assuming that a node emits a beacon every 90 seconds and a hiker encounter occurs everytime a beacon is emitted (worst-case scenario), a single alkaline battery will last for (841764 ∗ 90)/(30 ∗ 24 ∗ 60 ∗ 60) = 29 days. Since, a Mica2 is equipped with two batteries, a Mica2 sensor can remain operation for about two months. Notice that this calculation is preliminary, because it assumes that hikers are active 24 hours of the day and a hiker encounter occurs every 90 seconds. In a more realistic scenario, power is expected to last for a much longer time period. Also, this time period will significantly increase when groups of hikers are moving together. Finally, the lifetime of a sensor running on two batteries can definitely be increased significantly by using energy scavenging techniques and energy harvesting techniques [16, 14]. 6. PROTOTYPE IMPLEMENTATION We have implemented a prototype of CenWits on MICA2 sensor 900MHz running Mantis OS 0.9.1b. One of the sensor is equipped with MTS420CA GPS module, which is capable of barometric pressure and two-axis acceleration sensing in addition to GPS location tracking. We use SiRF, the serial communication protocol, to control GPS module. SiRF has a rich command set, but we record only X and Y coordinates. A witness record is 16 bytes long. When a node starts up, it stores its current location and emits a beacon periodicallyin the prototype, a node emits a beacon every minute. We have conducted a number of experiments with this prototype. A detailed report on these experiments with the raw data collected and photographs of hikers, access points etc. is available at http://csel.cs.colorado.edu/∼huangjh/ Cenwits.index.htm. Here we report results from three of them. In all these experiments, there are three access points (A, B and C) where nodes dump their witness information. These access points also provide location information to the nodes that come with in their range. We first show how CenWits can be used to determine the hiking trail a hiker is most likely on and the speed at which he is hiking, and identify hot search areas in case he is reported missing. Next, we show the results of power and memory management techniques of CenWits in conserving power and memory of a sensor node in one of our experiments. 6.1 Locating Lost Hikers The first experiment is called Direct Contact. It is a very simple experiment in which a single hiker starts from A, goes to B and then C, and finally returns to A (See Figure 13). The goal of this experiment is to illustrate that CenWits can deduce the trail a hiker takes by processing witness information. Figure 13: Direct Contact Experiment Node Id Record (X,Y) Location Hop Time Time Count 1 15 (12,7) 15 0 1 33 (31,17) 33 0 1 46 (12,23) 46 0 1 10 (12,7) 10 0 1 48 (12,23) 48 0 1 16 (12,7) 16 0 1 34 (31,17) 34 0 Table 1: Witness information collected in the direct contact experiment. The witness information dumped at the three access points was then collected and processed at a control center. Part of the witness information collected at the control center is shown in Table 1. The X,Y locations in this table correspond to the location information provided by access points A, B, and C. A is located at (12,7), B is located at (31,17) and C is located at (12,23). Three encounter points (between hiker 1 and the three access points) extracted from 188 this witness information are shown in Figure 13 (shown in rectangular boxes). For example, A,1 at 16 means 1 came in contact with A at time 16. Using this information, we can infer the direction in which hiker 1 was moving and speed at which he was moving. Furthermore, given a map of hiking trails in this area, it is clearly possible to identify the hiking trail that hiker 1 took. The second experiment is called Indirect Inference. This experiment is designed to illustrate that the location, direction and speed of a hiker can be inferred by CenWits, even if the hiker never comes in the range of any access point. It illustrates the importance of witness information in search and rescue applications. In this experiment, there are three hikers, 1, 2 and 3. Hiker 1 takes a trail that goes along access points A and B, while hiker 3 takes trail that goes along access points C and B. Hiker 2 takes a trail that does not come in the range of any access points. However, this hiker meets hiker 1 and 3 during his hike. This is illustrated in Figure 14. Figure 14: Indirect Inference Experiment Node Id Record (X,Y) Location Hop Time Time Count 2 16 (12,7) 6 0 2 15 (12,7) 6 0 1 4 (12,7) 4 0 1 6 (12,7) 6 0 1 29 (31,17) 29 0 1 31 (31,17) 31 0 Table 2: Witness information collected from hiker 1 in indirect inference experiment. Part of the witness information collected at the control center from access points A, B and C is shown in Tables 2 and 3. There are some interesting data in these tables. For example, the location time in some witness records is not the same as the record time. This means that the node that generated that record did not have its most up-to-date location at the encounter time. For example, when hikers 1 and 2 meet at time 16, the last recorded location time of Node Id Record (X,Y) Location Hop Time Time Count 3 78 (12,23) 78 0 3 107 (31,17) 107 0 3 106 (31,17) 106 0 3 76 (12,23) 76 0 3 79 (12,23) 79 0 2 94 (12,23) 79 0 1 16 (?,?) ? 1 1 15 (?,?) ? 1 Table 3: Witness information collected from hiker 3 in indirect inference experiment. hiker 1 is (12,7) recorded at time 6. So, node 1 generates a witness record with record time 16, location (12,7) and location time 6. In fact, the last two records in Table 3 have (?,?) as their location. This has happened because these witness records were generate by hiker 2 during his encounter with 1 at time 15 and 16. Until this time, hiker 2 hadn"t come in contact with any location points. Interestingly, a more accurate location information of 1 and 2 encounter or 2 and 3 encounter can be computed by process the witness information at the control center. It took 25 units of time for hiker 1 to go from A (12,7) to B (31,17). Assuming a constant hiking speed and a relatively straight-line hike, it can be computed that at time 16, hiker 1 must have been at location (18,10). Thus (18,10) is a more accurate location of encounter between 1 and 2. Finally, our third experiment called Identifying Hot Search Areas is designed to determine the trail a hiker has taken and identify hot search areas for rescue after he is reported missing. There are six hikers (1, 2, 3, 4, 5 and 6) in this experiment. Figure 15 shows the trails that hikers 1, 2, 3, 4 and 5 took, along with the encounter points obtained from witness records collected at the control center. For brevity, we have not shown the entire witness information collected at the control center. This information is available at http://csel.cs.colorado.edu/∼huangjh/Cenwits/index.htm. Figure 15: Identifying Hot Search Area Experiment (without hiker 6) 189 Now suppose hiker 6 is reported missing at time 260. To determine the hot search areas, the witness records of hiker 6 are processed to determine the trail he is most likely on, the speed at which he had been moving, direction in which he had been moving, and his last known location. Based on this information and the hiking trail map, hot search areas are identified. The hiking trail taken by hiker 6 as inferred by CenWits is shown by a dotted line and the hot search areas identified by CenWits are shown by dark lines inside the dotted circle in Figure 16. Figure 16: Identifying Hot Search Area Experiment (with hiker 6) 6.2 Results of Power and Memory Management The witness information shown in Tables 1, 2 and 3 has not been filtered using the three criteria described in Section 4.1. For example, the witness records generated by 3 at record times 76, 78 and 79 (see Table 3) have all been generated due a single contact between access point C and node 3. By applying the record gap criteria, two of these three records will be erased. Similarly, the witness records generated by 1 at record times 10, 15 and 16 (see Table 1) have all been generated due a single contact between access point A and node 1. Again, by applying the record gap criteria, two of these three records will be erased. Our experiments did not generate enough data to test the impact of record count or hop count criteria. To evaluate the impact of these criteria, we simulated CenWits to generate a significantly large number of records for a given number of hikers and access points. We generated witness records by having the hikers walk randomly. We applied the three criteria to measure the amount of memory savings in a sensor node. The results are shown in Table 4. The number of hikers in this simulation was 10 and the number of access points was 5. The number of witness records reported in this table is an average number of witness records a sensor node stored at the time of dump to an access point. These results show that the three memory management criteria significantly reduces the memory consumption of sensor nodes in CenWits. For example, they can reduce MAX MIN MAX # of RECORD RECORD HOP Witness COUNT GAP COUNT Records 5 5 5 628 4 5 5 421 3 5 5 316 5 10 5 311 5 20 5 207 5 5 4 462 5 5 3 341 3 20 3 161 Table 4: Impact of memory management techniques. the memory consumption by up to 75%. However, these results are premature at present for two reasons: (1) They are generated via simulation of hikers walking at random; and (2) It is not clear what impact the erasing of witness records has on the accuracy of inferred location/hot search areas of lost hikers. In our future work, we plan to undertake a major study to address these two concerns. 7. OTHER APPLICATIONS In addition to the hiking in wilderness areas, CenWits can be used in several other applications, e.g. skiing, climbing, wild life monitoring, and person tracking. Since CenWits relies only on intermittent connectivity, it can take advantage of the existing cheap and mature technologies, and thereby make tracking cheaper and fairly accurate. Since CenWits doesn"t rely on keeping track of a sensor holder all time, but relies on maintaining witnesses, the system is relatively cheaper and widely applicable. For example, there are some dangerous cliffs in most ski resorts. But it is too expensive for a ski resort to deploy a connected wireless sensor network through out the mountain. Using CenWits, we can deploy some sensors at the cliff boundaries. These boundary sensors emit beacons quite frequently, e.g. every second, and so can record presence of skiers who cross the boundary and fall off the cliff. Ski patrols can cruise the mountains every hour, and automatically query the boundary sensor when in range using PDAs. If a PDA shows that a skier has been close to the boundary sensor, the ski patrol can use a long range walkie-talkie to query control center at the resort base to check the witness record of the skier. If there is no witness record after the recorded time in the boundary sensor, there is a high chance that a rescue is needed. In wildlife monitoring, a very popular method is to attach a GPS receiver on the animals. To collect data, either a satellite transmitter is used, or the data collector has to wait until the GPS receiver brace falls off (after a year or so) and then search for the GPS receiver. GPS transmitters are very expensive, e.g. the one used in geese tracking is $3,000 each [2]. Also, it is not yet known if continuous radio signal is harmful to the birds. Furthermore, a GPS transmitter is quite bulky and uncomfortable, and as a result, birds always try to get rid of it. Using CenWits, not only can we record the presence of wildlife, we can also record the behavior of wild animals, e.g. lions might follow the migration of deers. CenWits does nor require any bulky and expensive satellite transmitters, nor is there a need to wait for a year and search for the braces. CenWits provides a very simple and cost-effective solution in this case. Also, access points 190 can be strategically located, e.g. near a water source, to increase chances of collecting up-to-date data. In fact, the access points need not be statically located. They can be placed in a low-altitude plane (e.g a UAV) and be flown over a wilderness area to collect data from wildlife. In large cities, CenWits can be used to complement GPS, since GPS doesn"t work indoor and near skyscrapers. If a person A is reported missing, and from the witness records we find that his last contacts were C and D, we can trace an approximate location quickly and quite efficiently. 8. DISCUSSION AND FUTURE WORK This paper presents a new search and rescue system called CenWits that has several advantages over the current search and rescue systems. These advantages include a looselycoupled system that relies only on intermittent network connectivity, power and storage efficiency, and low cost. It solves one of the greatest problems plaguing modern search and rescue systems: it has an inherent on-site storage capability. This means someone within the network will have access to the last-known-location information of a victim, and perhaps his bearing and speed information as well. It utilizes the concept of witnesses to propagate information, infer current possible location and speed of a subject, and identify hot search and rescue areas in case of emergencies. A large part of CenWits design focuses on addressing the power and memory limitations of current sensor nodes. In fact, power and memory constraints depend on how much weight (of sensor node) a hiker is willing to carry and the cost of these sensors. An important goal of CenWits is build small chips that can be implanted in hiking boots or ski jackets. This goal is similar to the avalanche beacons that are currently implanted in ski jackets. We anticipate that power and memory will continue to be constrained in such an environment. While the paper focuses on the development of a search and rescue system, it also provides some innovative, systemlevel ideas for information processing in a sensor network system. We have developed and experimented with a basic prototype of CenWits at present. Future work includes developing a more mature prototype addressing important issues such as security, privacy, and high availability. There are several pressing concerns regarding security, privacy, and high availability in CenWits. For example, an adversary can sniff the witness information to locate endangered animals, females, children, etc. He may inject false information in the system. An individual may not be comfortable with providing his/her location and movement information, even though he/she is definitely interested in being located in a timely manner at the time of emergency. In general, people in hiking community are friendly and usually trustworthy. So, a bullet-proof security is not really required. However, when CenWits is used in the context of other applications, security requirements may change. Since the sensor nodes used in CenWits are fragile, they can fail. In fact, the nature and level of security, privacy and high availability support needed in CenWits strongly depends on the application for which it is being used and the individual subjects involved. Accordingly, we plan to design a multi-level support for security, privacy and high availability in CenWits. So far, we have experimented with CenWits in a very restricted environment with a small number of sensors. Our next goal is to deploy this system in a much larger and more realistic environment. In particular, discussions are currenly underway to deploy CenWits in the Rocky Mountain and Yosemite National Parks. 9. REFERENCES [1] 802.11-based tracking system. http://www.pangonetworks.com/locator.htm. [2] Brent geese 2002. http://www.wwt.org.uk/brent/. [3] The onstar system. http://www.onstar.com. [4] Personal locator beacons with GPS receiver and satellite transmitter. http://www.aeromedix.com/. [5] Personal tracking using GPS and GSM system. http://www.ulocate.com/trimtrac.html. [6] Rf based kid tracking system. http://www.ion-kids.com/. [7] F. Alessio. Performance measurements with motes technology. MSWiM"04, 2004. [8] P. Bahl and V. N. Padmanabhan. RADAR: An in-building RF-based user location and tracking system. IEEE Infocom, 2000. [9] K. Fall. A delay-tolerant network architecture for challenged internets. In SIGCOMM, 2003. [10] L. Gu and J. Stankovic. Radio triggered wake-up capability for sensor networks. In Real-Time Applications Symposium, 2004. [11] J. Hightower and G. Borriello. Location systems for ubiquitous computing. IEEE Computer, 2001. [12] W. Jaskowski, K. Jedrzejek, B. Nyczkowski, and S. Skowronek. Lifetch life saving system. CSIDC, 2004. [13] P. Juang, H. Oki, Y. Wang, M. Martonosi, L. Peh, and D. Rubenstein. Energy-efficient computing for wildlife tracking: design tradeoffs and early experiences with ZebraNet. In ASPLOS, 2002. [14] K. Kansal and M. Srivastava. Energy harvesting aware power management. In Wireless Sensor Networks: A Systems Perspective, 2005. [15] G. J. Pottie and W. J. Kaiser. Embedding the internet: wireless integrated network sensors. Communications of the ACM, 43(5), May 2000. [16] S. Roundy, P. K. Wright, and J. Rabaey. A study of low-level vibrations as a power source for wireless sensor networks. Computer Communications, 26(11), 2003. [17] C. Savarese, J. M. Rabaey, and J. Beutel. Locationing in distributed ad-hoc wireless sensor networks. ICASSP, 2001. [18] V. Shnayder, M. Hempstead, B. Chen, G. Allen, and M. Welsh. Simulating the power consumption of large-scale sensor network applications. In Sensys, 2004. [19] R. Want and A. Hopper. Active badges and personal interactive computing objects. IEEE Transactions of Consumer Electronics, 1992. [20] M. Welsh and G. Mainland. Programming sensor networks using abstract regions. First USENIX/ACM Symposium on Networked Systems Design and Implementation (NSDI "04), 2004. 191

intermittent network connectivity;gp receiver;pervasive computing;connected network;satellite transmitter;emergency situation;search and rescue;witness;beacon;location tracking system;rf transmitter;group and partition;sensor network;hiker

train_C-52

Fairness in Dead-Reckoning based Distributed Multi-Player Games

In a distributed multi-player game that uses dead-reckoning vectors to exchange movement information among players, there is inaccuracy in rendering the objects at the receiver due to network delay between the sender and the receiver. The object is placed at the receiver at the position indicated by the dead-reckoning vector, but by that time, the real position could have changed considerably at the sender. This inaccuracy would be tolerable if it is consistent among all players; that is, at the same physical time, all players see inaccurate (with respect to the real position of the object) but the same position and trajectory for an object. But due to varying network delays between the sender and different receivers, the inaccuracy is different at different players as well. This leads to unfairness in game playing. In this paper, we first introduce an error measure for estimating this inaccuracy. Then we develop an algorithm for scheduling the sending of dead-reckoning vectors at a sender that strives to make this error equal at different receivers over time. This algorithm makes the game very fair at the expense of increasing the overall mean error of all players. To mitigate this effect, we propose a budget based algorithm that provides improved fairness without increasing the mean error thereby maintaining the accuracy of game playing. We have implemented both the scheduling algorithm and the budget based algorithm as part of BZFlag, a popular distributed multi-player game. We show through experiments that these algorithms provide fairness among players in spite of widely varying network delays. An additional property of the proposed algorithms is that they require less number of DRs to be exchanged (compared to the current implementation of BZflag) to achieve the same level of accuracy in game playing.

1. INTRODUCTION In a distributed multi-player game, players are normally distributed across the Internet and have varying delays to each other or to a central game server. Usually, in such games, the players are part of the game and in addition they may control entities that make up the game. During the course of the game, the players and the entities move within the game space. A player sends information about her movement as well as the movement of the entities she controls to the other players using a Dead-Reckoning (DR) vector. A DR vector contains information about the current position of the player/entity in terms of x, y and z coordinates (at the time the DR vector was sent) as well as the trajectory of the entity in terms of the velocity component in each of the dimensions. Each of the participating players receives such DR vectors from one another and renders the other players/entities on the local consoles until a new DR vector is received for that player/entity. In a peer-to-peer game, players send DR vectors directly to each other; in a client-server game, these DR vectors may be forwarded through a game server. The idea of DR is used because it is almost impossible for players/entities to exchange their current positions at every time unit. DR vectors are quantization of the real trajectory (which we refer to as real path) at a player. Normally, a new DR vector is computed and sent whenever the real path deviates from the path extrapolated using the previous DR vector (say, in terms of distance in the x, y, z plane) by some amount specified by a threshold. We refer to the trajectory that can be computed using the sequence of DR vectors as the exported path. Therefore, at the sending player, there is a deviation between the real path and the exported path. The error due to this deviation can be removed if each movement of player/entity is communicated to the other players at every time unit; that is a DR vector is generated at every time unit thereby making the real and exported paths the same. Given that it is not feasible to satisfy this due to bandwidth limitations, this error is not of practical interest. Therefore, the receiving players can, at best, follow the exported path. Because of the network delay between the sending and receiving players, when a DR vector is received and rendered at a player, the original trajectory of the player/entity may have already changed. Thus, in physical time, there is a deviation at the receiving player between the exported path and the rendered trajectory (which we refer to as placed path). We refer to this error as the export error. Note that the export error, in turn, results in a deviation between the real and the placed paths. The export error manifests itself due to the deviation between the exported path at the sender and the placed path at the receiver (i) 1 before the DR vector is received at the receiver (referred to as the before export error, and (ii) after the DR vector is received at the receiver (referred to as the after export error). In an earlier paper [1], we showed that by synchronizing the clocks at all the players and by using a technique based on time-stamping messages that carry the DR vectors, we can guarantee that the after export error is made zero. That is, the placed and the exported paths match after the DR vector is received. We also showed that the before export error can never be eliminated since there is always a non-zero network delay, but can be significantly reduced using our technique [1]. Henceforth we assume that the players use such a technique which results in unavoidable but small overall export error. In this paper we consider the problem of different and varying network delays between each sender-receiver pair of a DR vector, and consequently, the different and varying export errors at the receivers. Due to the difference in the export errors among the receivers, the same entity is rendered at different physical time at different receivers. This brings in unfairness in game playing. For instance a player with a large delay would always see an entity late in physical time compared to the other players and, therefore, her action on the entity would be delayed (in physical time) even if she reacted instantaneously after the entity was rendered. Our goal in this paper is to improve the fairness of these games in spite of the varying network delays by equalizing the export error at the players. We explore whether the time-average of the export errors (which is the cumulative export error over a period of time averaged over the time period) at all the players can be made the same by scheduling the sending of the DR vectors appropriately at the sender. We propose two algorithms to achieve this. Both the algorithms are based on delaying (or dropping) the sending of DR vectors to some players on a continuous basis to try and make the export error the same at all the players. At an abstract level, the algorithm delays sending DR vectors to players whose accumulated error so far in the game is smaller than others; this would mean that the export error due to this DR vector at these players will be larger than that of the other players, thereby making them the same. The goal is to make this error at least approximately equal at every DR vector with the deviation in the error becoming smaller as time progresses. The first algorithm (which we refer to as the scheduling algorithm) is based on estimating the delay between players and refining the sending of DR vectors by scheduling them to be sent to different players at different times at every DR generation point. Through an implementation of this algorithm using the open source game BZflag, we show that this algorithm makes the game very fair (we measure fairness in terms of the standard deviation of the error). The drawback of this algorithm is that it tends to push the error of all the players towards that of the player with the worst error (which is the error at the farthest player, in terms of delay, from the sender of the DR). To alleviate this effect, we propose a budget based algorithm which budgets how the DRs are sent to different players. At a high level, the algorithm is based on the idea of sending more DRs to players who are farther away from the sender compared to those who are closer. Experimental results from BZflag illustrates that the budget based algorithm follows a more balanced approach. It improves the fairness of the game but at the same time does so without pushing up the mean error of the players thereby maintaining the accuracy of the game. In addition, the budget based algorithm is shown to achieve the same level of accuracy of game playing as the current implementation of BZflag using much less number of DR vectors. 2. PREVIOUS WORK Earlier work on network games to deal with network latency has mostly focussed on compensation techniques for packet delay and loss [2, 3, 4]. These methods are aimed at making large delays and message loss tolerable for players but does not consider the problems that may be introduced by varying delays from the server to different players or from the players to one another. For example, the concept of local lag has been used in [3] where each player delays every local operation for a certain amount of time so that remote players can receive information about the local operation and execute the same operation at the about same time, thus reducing state inconsistencies. The online multi-player game MiMaze [2, 5, 6], for example, takes a static bucket synchronization approach to compensate for variable network delays. In MiMaze, each player delays all events by 100 ms regardless of whether they are generated locally or remotely. Players with a network delay larger than 100 ms simply cannot participate in the game. In general, techniques based on bucket synchronization depend on imposing a worst case delay on all the players. There have been a few papers which have studied the problem of fairness in a distributed game by more sophisticated message delivery mechanisms. But these works [7, 8] assume the existence of a global view of the game where a game server maintains a view (or state) of the game. Players can introduce objects into the game or delete objects that are already part of the game (for example, in a first-person shooter game, by shooting down the object). These additions and deletions are communicated to the game server using action messages. Based on these action messages, the state of the game is changed at the game server and these changes are communicated to the players using update messages. Fairness is achieved by ordering the delivery of action and update messages at the game server and players respectively based on the notion of a fair-order which takes into account the delays between the game server and the different players. Objects that are part of the game may move but how this information is communicated to the players seems to be beyond the scope of these works. In this sense, these works are very limited in scope and may be applicable only to firstperson shooter games and that too to only games where players are not part of the game. DR vectors can be exchanged directly among the players (peerto-peer model) or using a central server as a relay (client-server model). It has been shown in [9] that multi-player games that use DR vectors together with bucket synchronization are not cheatproof unless additional mechanisms are put in place. Both the scheduling algorithm and the budget-based algorithm described in our paper use DR vectors and hence are not cheat-proof. For example, a receiver could skew the delay estimate at the sender to make the sender believe that the delay between the sender and the receiver is high thereby gaining undue advantage. We emphasize that the focus of this paper is on fairness without addressing the issue of cheating. In the next section, we describe the game model that we use and illustrate how senders and receivers exchange DR vectors and how entities are rendered at the receivers based on the time-stamp augmented DR vector exchange as described in [1]. In Section 4, we describe the DR vector scheduling algorithm that aims to make the export error equal across the players with varying delays from the sender of a DR vector, followed by experimental results obtained from instrumentation of the scheduling algorithm on the open source game BZFlag. Section 5, describes the budget based algorithm that achieves improved fairness but without reducing the level accuracy of game playing. Conclusions are presented in Section 6. 2 3. GAME MODEL The game architecture is based on players distributed across the Internet and exchanging DR vectors to each other. The DR vectors could either be sent directly from one player to another (peerto-peer model) or could be sent through a game server which receives the DR vector from a player and forwards it to other players (client-server model). As mentioned before, we assume synchronized clocks among the participating players. Each DR vector sent from one player to another specifies the trajectory of exactly one player/entity. We assume a linear DR vector in that the information contained in the DR vector is only enough at the receiving player to compute the trajectory and render the entity in a straight line path. Such a DR vector contains information about the starting position and velocity of the player/entity where the velocity is constant1 . Thus, the DR vectors sent by a player specifies the current time at the player when the DR vector is computed (not the time at which this DR vector is sent to the other players as we will explain later), the current position of the player/entity in terms of the x, y, z coordinates and the velocity vector in the direction of x, y and z coordinates. Specifically, the ith DR vector sent by player j about the kth entity is denoted by DRj ik and is represented by the following tuple (Tj ik, xj ik, yj ik, zj ik, vxj ik, vyj ik, vzj ik). Without loss of generality, in the rest of the discussion, we consider a sequence of DR vectors sent by only one player and for only one entity. For simplicity, we consider a two dimensional game space rather than a three dimensional one. Hence we use DRi to denote the ith such DR vector represented as the tuple (Ti, xi, yi, vxi, vyi). The receiving player computes the starting position for the entity based on xi, yi and the time difference between when the DR vector is received and the time Ti at which it was computed. Note that the computation of time difference is feasible since all the clocks are synchronized. The receiving player then uses the velocity components to project and render the trajectory of the entity. This trajectory is followed until a new DR vector is received which changes the position and/or velocity of the entity. timeT1 Real Exported Placed dt1 A B C D DR1 = (T1, x1, y1, vx1, vy1) computed at time T1 and sent to the receiver DR0 = (T0, x0, y0, vx0, vy0) computed at time T0 and sent to the receiver T0 dt0 Placed E Figure 1: Trajectories and deviations. Based on this model, Figure 1 illustrates the sending and receiv1 Other type of DR vectors include quadratic DR vectors which specify the acceleration of the entity and cubic spline DR vectors that consider the starting position and velocity and the ending position and velocity of the entity. ing of DR vectors and the different errors that are encountered. The figure shows the reception of DR vectors at a player (henceforth called the receiver). The horizontal axis shows the time which is synchronized among all the players. The vertical axis tries to conceptually capture the two-dimensional position of an entity. Assume that at time T0 a DR vector DR0 is computed by the sender and immediately sent to the receiver. Assume that DR0 is received at the receiver after a delay of dt0 time units. The receiver computes the initial position of the entity as (x0 + vx0 × dt0, y0 + vy0 × dt0) (shown as point E). The thick line EBD represents the projected and rendered trajectory at the receiver based on the velocity components vx0 and vy0 (placed path). At time T1 a DR vector DR1 is computed for the same entity and immediately sent to the receiver2 . Assume that DR1 is received at the receiver after a delay of dt1 time units. When this DR vector is received, assume that the entity is at point D. A new position for the entity is computed as (x1 + vx1 × dt1, y1 + vy0 × dt1) and the entity is moved to this position (point C). The velocity components vx1 and vy1 are used to project and render this entity further. Let us now consider the error due to network delay. Although DR1 was computed at time T1 and sent to the receiver, it did not reach the receiver until time T1 + dt1. This means, although the exported path based on DR1 at the sender at time T1 is the trajectory AC, until time T1 + dt1, at the receiver, this entity was being rendered at trajectory BD based on DR0. Only at time T1 + dt1 did the entity get moved to point C from which point onwards the exported and the placed paths are the same. The deviation between the exported and placed paths creates an error component which we refer to as the export error. A way to represent the export error is to compute the integral of the distance between the two trajectories over the time when they are out of sync. We represent the integral of the distances between the placed and exported paths due to some DR DRi over a time interval [t1, t2] as Err(DRi, t1, t2). In the figure, the export error due to DR1 is computed as the integral of the distance between the trajectories AC and BD over the time interval [T1, T1 + dt1]. Note that there could be other ways of representing this error as well, but in this paper, we use the integral of the distance between the two trajectories as a measure of the export error. Note that there would have been an export error created due to the reception of DR0 at which time the placed path would have been based on a previous DR vector. This is not shown in the figure but it serves to remind the reader that the export error is cumulative when a sequence of DR vectors are received. Starting from time T1 onwards, there is a deviation between the real and the exported paths. As we discussed earlier, this export error is unavoidable. The above figure and example illustrates one receiver only. But in reality, DR vectors DR0 and DR1 are sent by the sender to all the participating players. Each of these players receives DR0 and DR1 after varying delays thereby creating different export error values at different players. The goal of the DR vector scheduling algorithm to be described in the next section is to make this (cumulative) export error equal at every player independently for each of the entities that make up the game. 4. SCHEDULING ALGORITHM FORSENDING DR VECTORS In Section 3 we showed how delay from the sender of a new DR 2 Normally, DR vectors are not computed on a periodic basis but on an on-demand basis where the decision to compute a new DR vector is based on some threshold being exceeded on the deviation between the real path and the path exported by the previous DR vector. 3 vector to the receiver of the DR vector could lead to export error because of the deviation of the placed path from the exported path at the receiver until this new DR vector is received. We also mentioned that the goal of the DR vector scheduling algorithm is to make the export error equal at all receivers over a period of time. Since the game is played in a distributed environment, it makes sense for the sender of an entity to keep track of all the errors at the receivers and try to make them equal. However, the sender cannot know the actual error at a receiver till it gets some information regarding the error back from the receiver. Our algorithm estimates the error to compute a schedule to send DR vectors to the receivers and corrects the error when it gets feedbacks from the receivers. In this section we provide motivations for the algorithm and describe the steps it goes through. Throughout this section, we will use the following example to illustrate the algorithm. timeT1 Exported path Placed path at receiver 2 dt1 A B C D E F T0 G2 G1 dt2 DR1 sent to receiver 1 DR1 sent to receiver 2 T1 1 T1 2 da1 da2 G H I J K L N M DR1 estimated to be received by receiver 2 DR1 estimated to be received by receiver 1 DR1 actually received by receiver 1 DR1 actually received by receiver 2 DR0 sent to both receivers DR1 computed by sender Placed path at receiver 1 Figure 2: DR vector flow between a sender and two receivers and the evolution of estimated and actual placed paths at the receivers. DR0 = (T0, T0, x0, y0, vx0, vy0), sent at time T0 to both receivers. DR1 = (T1, T1 1 , x1, y1, vx1, vy1) sent at time T1 1 = T1+δ1 to receiver 1 and DR1 = (T1, T2 1 , x1, y1, vx1, vy1) sent at time T2 1 = T1 + δ2 to receiver 2. Consider the example in Figure 2. The figure shows a single sender sending DR vectors for an entity to two different receivers 1 and 2. DR0 computed at T0 is sent and received by the receivers sometime between T0 and T1 at which time they move the location of the entity to match the exported path. Thus, the path of the entity is shown only from the point where the placed path matches the exported path for DR0. Now consider DR1. At time T1, DR1 is computed by the sender but assume that it is not immediately sent to the receivers and is only sent after time δ1 to receiver 1 (at time T1 1 = T1 + δ1) and after time δ2 to receiver 2 (at time T2 1 = T1 + δ2). Note that the sender includes the sending timestamp with the DR vector as shown in the figure. Assume that the sender estimates (it will be clear shortly why the sender has to estimate the delay) that after a delay of dt1, receiver 1 will receive it, will use the coordinate and velocity parameters to compute the entity"s current location and move it there (point C) and from this time onwards, the exported and the placed paths will become the same. However, in reality, receiver 1 receives DR1 after a delay of da1 (which is less than sender"s estimates of dt1), and moves the corresponding entity to point H. Similarly, the sender estimates that after a delay of dt2, receiver 2 will receive DR1, will compute the current location of the entity and move it to that point (point E), while in reality it receives DR1 after a delay of da2 > dt2 and moves the entity to point N. The other points shown on the placed and exported paths will be used later in the discussion to describe different error components. 4.1 Computation of Relative Export Error Referring back to the discussion from Section 3, from the sender"s perspective, the export error at receiver 1 due to DR1 is given by Err(DR1, T1, T1 + δ1 + dt1) (the integral of the distance between the trajectories AC and DB over the time interval [T1, T1 + δ1 + dt1]) of Figure 2. This is due to the fact that the sender uses the estimated delay dt1 to compute this error. Similarly, the export error from the sender"s perspective at received 2 due to DR1 is given by Err(DR1, T1, T1 + δ2 + dt2) (the integral of the distance between the trajectories AE and DF over the time interval [T1, T1 + δ2 + dt2]). Note that the above errors from the sender"s perspective are only estimates. In reality, the export error will be either smaller or larger than the estimated value, based on whether the delay estimate was larger or smaller than the actual delay that DR1 experienced. This difference between the estimated and the actual export error is the relative export error (which could either be positive or negative) which occurs for every DR vector that is sent and is accumulated at the sender. The concept of relative export error is illustrated in Figure 2. Since the actual delay to receiver 1 is da1, the export error induced by DR1 at receiver 1 is Err(DR1, T1, T1 + δ1 + da1). This means, there is an error in the estimated export error and the sender can compute this error only after it gets a feedback from the receiver about the actual delay for the delivery of DR1, i.e., the value of da1. We propose that once receiver 1 receives DR1, it sends the value of da1 back to the sender. The receiver can compute this information as it knows the time at which DR1 was sent (T1 1 = T1 + δ1, which is appended to the DR vector as shown in Figure 2) and the local receiving time (which is synchronized with the sender"s clock). Therefore, the sender computes the relative export error for receiver 1, represented using R1 as R1 = Err(DR1, T1, T1 + δ1 + dt1) − Err(DR1, T1, T1 + δ1 + da1) = Err(DR1, T1 + δ1 + dt1, T1 + δ1 + da1) Similarly the relative export error for receiver 2 is computed as R2 = Err(DR1, T1, T1 + δ2 + dt2) − Err(DR1, T1, T1 + δ2 + da2) = Err(DR1, T1 + δ2 + dt2, T1 + δ2 + da2) Note that R1 > 0 as da1 < dt1, and R2 < 0 as da2 > dt2. Relative export errors are computed by the sender as and when it receives the feedback from the receivers. This example shows the 4 relative export error values after DR1 is sent and the corresponding feedbacks are received. 4.2 Equalization of Error Among Receivers We now explain what we mean by making the errors equal at all the receivers and how this can be achieved. As stated before the sender keeps estimates of the delays to the receivers, dt1 and dt2 in the example of Figure 2. This says that at time T1 when DR1 is computed, the sender already knows how long it may take messages carrying this DR vector to reach the receivers. The sender uses this information to compute the export errors, which are Err(DR1, T1, T1 + δ1 + dt1) and Err(DR1, T1, T1 + δ2 + dt2) for receivers 1 and 2, respectively. Note that the areas of these error components are a function of δ1 and δ2 as well as the network delays dt1 and dt2. If we are to make the exports errors due to DR1 the same at both receivers, the sender needs to choose δ1 and δ2 such that Err(DR1, T1, T1 + δ1 + dt1) = Err(DR1, T1, T1 + δ2 + dt2). But when T1 was computed there could already have been accumulated relative export errors due to previous DR vectors (DR0 and the ones before). Let us represent the accumulated relative error up to DRi for receiver j as Ri j. To accommodate these accumulated relative errors, the sender should now choose δ1 and δ2 such that R0 1 + Err(DR1, T1, T1 + δ1 + dt1) = R0 2 + Err(DR1, T1, T1 + δ2 + dt2) The δi determines the scheduling instant of the DR vector at the sender for receiver i. This method of computation of δ"s ensures that the accumulated export error (i.e., total actual error) for each receiver equalizes at the transmission of each DR vector. In order to establish this, assume that the feedback for DR vector Di from a receiver comes to the sender before schedule for Di+1 is computed. Let Si m and Ai m denote the estimated error for receiver m used for computing schedule for Di and accumulated error for receiver m computed after receiving feedback for Di, respectively. Then Ri m = Ai m −Si m. In order to compute the schedule instances (i.e., δ"s) for Di, for any pair of receivers m and n, we do Ri−1 m + Si m = Ri−1 n + Si n. The following theorem establishes the fact that the accumulated export error is equalized at every scheduling instant. THEOREM 4.1. When the schedule instances for sending Di are computed for any pair of receivers m and n, the following condition is satisfied: i−1 k=1 Ak m + Si m = i−1 k=1 Ak n + Si n. Proof: By induction. Assume that the premise holds for some i. We show that it holds for i+1. The base case for i = 1 holds since initially R0 m = R0 n = 0, and the S1 m = S1 n is used to compute the scheduling instances. In order to compute the schedule for Di+1, the we first compute the relative errors as Ri m = Ai m − Si m, and Ri n = Ai n − Si n. Then to compute δ"s we execute Ri m + Si+1 m = Ri n + Si+1 n Ai m − Si m + Si+1 m = Ai n − Si n + Si+1 n . Adding the condition of the premise on both sides we get, i k=1 Ak m + Si+1 m = i k=1 Ak n + Si+1 n . 4.3 Computation of the Export Error Let us now consider how the export errors can be computed. From the previous section, to find δ1 and δ2 we need to find Err(DR1, T1, T1 +δ1 +dt1) and Err(DR1, T1, T1 +δ2 +dt2). Note that the values of R0 1 and R0 2 are already known at the sender. Consider the computation of Err(DR1, T1, T1 +δ1 +dt1). This is the integral of the distance between the trajectories AC due to DR1 and BD due to DR0. From DR0 and DR1, point A is (X1, Y1) = (x1, y1) and point B is (X0, Y0) = (x0 + (T1 − T0) × vx0, y0 + (T1 − T0) × vy0). The trajectory AC can be represented as a function of time as (X1(t), Y1(t) = (X1 + vx1 × t, Y1 + vy1 × t) and the trajectory of BD can be represented as (X0(t), Y0(t) = (X0 + vx0 × t, Y0 + vy0 × t). The distance between the two trajectories as a function of time then becomes, dist(t) = (X1(t) − X0(t))2 + (Y1(t) − Y0(t))2 = ((X1 − X0) + (vx1 − vx0)t)2 +((Y1 − Y0) + (vy1 − vy0)t)2 = ((vx1 − vx0)2 + (vy1 − vy0)2)t2 +2((X1 − X0)(vx1 − vx0) +(Y1 − Y0)(vy1 − vy0))t +(X1 − X0)2 + (Y1 − Y0)2 Let a = (vx1 − vx0)2 + (vy1 − vy0)2 b = 2((X1 − X0)(vx1 − vx0) +(Y1 − Y0)(vy1 − vy0)) c = (X1 − X0)2 + (Y1 − Y0)2 Then dist(t) can be written as dist(t) = a × t2 + b × t + c. Then Err(DR1, t1, t2) for some time interval [t1, t2] becomes t2 t1 dist(t) dt = t2 t1 a × t2 + b × t + c dt. A closed form solution for the indefinite integral a × t2 + b × t + c dt = (2at + b) √ at2 + bt + c 4a + 1 2 ln 1 2b + at √ a + at2 + bt + c c 1 √ a − 1 8 ln 1 2b + at √ a + at2 + bt + c b2 a− 3 2 Err(DR1, T1, T1 +δ1 +dt1) and Err(DR1, T1, T1 +δ2 +dt2) can then be calculated by applying the appropriate limits to the above solution. In the next section, we consider the computation of the δ"s for N receivers. 5 4.4 Computation of Scheduling Instants We again look at the computation of δ"s by referring to Figure 2. The sender chooses δ1 and δ2 such that R0 1 + Err(DR1, T1, T1 + δ1 +dt1) = R0 2 +Err(DR1, T1, T1 +δ2 +dt2). If R0 1 and R0 2 both are zero, then δ1 and δ2 should be chosen such that Err(DR1, T1, T1+ δ1 +dt1) = Err(DR1, T1, T1 +δ2 +dt2). This equality will hold if δ1 + dt1 = δ2 + dt2. Thus, if there is no accumulated relative export error, all that the sender needs to do is to choose the δ"s in such a way that they counteract the difference in the delay to the two receivers, so that they receive the DR vector at the same time. As discussed earlier, because the sender is not able to a priori learn the delay, there will always be an accumulated relative export error from a previous DR vector that does have to be taken into account. To delve deeper into this, consider the computation of the export error as illustrated in the previous section. To compute the δ"s we require that R0 1 + Err(DR1, T1, T1 + δ1 + dt1) = R0 2 + Err(DR1, T1, T1 + δ2 + dt2). That is, R0 1 + T1+δ1+dt1 T1 dist(t) dt = R0 2 + T1+δ2+dt2 T1 dist(t) dt. That is R0 1 + T1+dt1 T1 dist(t) dt + T1+dt1+δ1 T1+dt1 dist(t) dt = R0 2 + T1+dt2 T1 dist(t) dt + T1+dt2+δ2 T1+dt2 dist(t) dt. The components R0 1, R0 2, are already known to (or estimated by) the sender. Further, the error components T1+dt1 T1 dist(t) dt and T1+dt2 T1 dist(t) dt can be a priori computed by the sender using estimated values of dt1 and dt2. Let us use E1 to denote R0 1 + T1+dt1 T1 dist(t) dt and E2 to denote R0 2 + T1+dt2 T1 dist(t) dt. Then, we require that E1 + T1+dt1+δ1 T1+dt1 dist(t) dt = E2 + T1+dt2+δ2 T1+dt2 dist(t) dt. Assume that E1 > E2. Then, for the above equation to hold, we require that T1+dt1+δ1 T1+dt1 dist(t) dt < T1+dt2+δ2 T1+dt2 dist(t) dt. To make the game as fast as possible within this framework, the δ values should be made as small as possible so that DR vectors are sent to the receivers as soon as possible subject to the fairness requirement. Given this, we would choose δ1 to be zero and compute δ2 from the equation E1 = E2 + T1+dt2+δ2 T1+dt2 dist(t) dt. In general, if there are N receivers 1, . . . , N, when a sender generates a DR vector and decides to schedule them to be sent, it first computes the Ei values for all of them from the accumulated relative export errors and estimates of delays. Then, it finds the smallest of these values. Let Ek be the smallest value. The sender makes δk to be zero and computes the rest of the δ"s from the equality Ei + T1+dti+δi T1+dti dist(t) dt = Ek, ∀i 1 ≤ i ≤ N, i = k. (1) The δ"s thus obtained gives the scheduling instants of the DR vector for the receivers. 4.5 Steps of the Scheduling Algorithm For the purpose of the discussion below, as before let us denote the accumulated relative export error at a sender for receiver k up until DRi to be Ri k. Let us denote the scheduled delay at the sender before DRi is sent to receiver k as δi k. Given the above discussion, the algorithm steps are as follows: 1. The sender computes DRi at (say) time Ti and then computes δi k, and Ri−1 k , ∀k, 1 ≤ k ≤ N based on the estimation of delays dtk, ∀k, 1 ≤ k ≤ N as per Equation (1). It schedules, DRi to be sent to receiver k at time Ti + δi k. 2. The DR vectors are sent to the receivers at the scheduled times which are received after a delay of dak, ∀k, 1 ≤ k ≤ N where dak ≤ or > dtk. The receivers send the value of dak back to the sender (the receiver can compute this value based on the time stamps on the DR vector as described earlier). 3. The sender computes Ri k as described earlier and illustrated in Figure 2. The sender also recomputes (using exponential averaging method similar to round-trip time estimation by TCP [10]) the estimate of delay dtk from the new value of dak for receiver k. 4. Go back to Step 1 to compute DRi+1 when it is required and follow the steps of the algorithm to schedule and send this DR vector to the receivers. 4.6 Handling Cases in Practice So far we implicity assumed that DRi is sent out to all receivers before a decision is made to compute the next DR vector DRi+1, and the receivers send the value of dak corresponding to DRi and this information reaches the sender before it computes DRi+1 so that it can compute Ri+1 k and then use it in the computation of δi+1 k . Two issues need consideration with respect to the above algorithm when it is used in practice. • It may so happen that a new DR vector is computed even before the previous DR vector is sent out to all receivers. How will this situation be handled? • What happens if the feedback does not arrive before DRi+1 is computed and scheduled to be sent? Let us consider the first scenario. We assume that DRi has been scheduled to be sent and the scheduling instants are such that δi 1 < δi 2 < · · · < δi N . Assume that DRi+1 is to be computed (because the real path has deviated exceeding a threshold from the path exported by DRi) at time Ti+1 where Ti + δi k < Ti+1 < Ti + δi k+1. This means, DRi has been sent only to receivers up to k in the scheduled order. In our algorithm, in this case, the scheduled delay ordering queue is flushed which means DRi is not sent to receivers still queued to receive it, but a new scheduling order is computed for all the receivers to send DRi+1. For those receivers who have been sent DRi, assume for now that daj, 1 ≤ j ≤ k has been received from all receivers (the scenario where daj has not been received will be considered as a part of the second scenario later). For these receivers, Ei j, 1 ≤ j ≤ k can be computed. For those receivers j, k + 1 ≤ j ≤ N to whom DRi was not sent Ei j does not apply. Consider a receiver j, k + 1 ≤ j ≤ N to whom DRi was not sent. Refer to Figure 3. For such a receiver j, when DRi+1 is to be scheduled and 6 timeTi Exported path dtj A B C D Ti-1 Gi j DRi+1 computed by sender and DRi for receiver k+1 to N is removed from queue DRi+1 scheduled for receiver k+1 Ti+1 G H E F DRi scheduled for receiver j DRi computed by sender Placed path at receiver k+1 Gi+1 j Figure 3: Schedule computation when DRi is not sent to receiver j, k + 1 ≤ j ≤ N. δi+1 j needs to be computed, the total export error is the accumulated relative export error at time Ti when schedule for DRi was computed, plus the integral of the distance between the two trajectories AC and BD of Figure 3 over the time interval [Ti, Ti+1 + δi+1 j + dtj]. Note that this integral is given by Err(DRi, Ti, Ti+1) + Err(DRi+1, Ti+1, Ti+1 + δi+1 j + dtj). Therefore, instead of Ei j of Equation (1), we use the value Ri−1 j + Err(DRi, Ti, Ti+1) + Err(DRi+1, Ti+1, Ti+1 + δi+1 j + dtj) where Ri−1 j is relative export error used when the schedule for DRi was computed. Now consider the second scenario. Here the feedback dak corresponding to DRi has not arrived before DRi+1 is computed and scheduled. In this case, Ri k cannot be computed. Thus, in the computation of δk for DRi+1, this will be assumed to be zero. We do assume that a reliable mechanism is used to send dak back to the sender. When this information arrives at a later time, Ri k will be computed and accumulated to future relative export errors (for example Ri+1 k if dak is received before DRi+2 is computed) and used in the computation of δk when a future DR vector is to be scheduled (for example DRi+2). 4.7 Experimental Results In order to evaluate the effectiveness and quantify benefits obtained through the use of the scheduling algorithm, we implemented the proposed algorithm in BZFlag (Battle Zone Flag) [11] game. It is a first-person shooter game where the players in teams drive tanks and move within a battle field. The aim of the players is to navigate and capture flags belonging to the other team and bring them back to their own area. The players shoot each other"s tanks using shooting bullets. The movement of the tanks as well as that of the shots are exchanged among the players using DR vectors. We have modified the implementation of BZFlag to incorporate synchronized clocks among the players and the server and exchange time-stamps with the DR vector. We set up a testbed with four players running the instrumented version of BZFlag, with one as a sender and the rest as receivers. The scheduling approach and the base case where each DR vector was sent to all the receivers concurrently at every trigger point were implemented in the same run by tagging the DR vectors according to the type of approach used to send the DR vector. NISTNet [12] was used to introduce delays across the sender and the three receivers. Mean delays of 800ms, 500ms and 200ms were introduced between the sender and first, second and the third receiver, respectively. We introduce a variance of 100 msec (to the mean delay of each receiver) to model variability in delay. The sender logged the errors of each receiver every 100 milliseconds for both the scheduling approach and the base case. The sender also calculated the standard deviation and the mean of the accumulated export error of all the receivers every 100 milliseconds. Figure 4 plots the mean and standard deviation of the accumulated export error of all the receivers in the scheduling case against the base case. Note that the x-axis of these graphs (and the other graphs that follow) represents the system time when the snapshot of the game was taken. Observe that the standard deviation of the error with scheduling is much lower as compared to the base case. This implies that the accumulated errors of the receivers in the scheduling case are closer to one another. This shows that the scheduling approach achieves fairness among the receivers even if they are at different distances (i.e, latencies) from the sender. Observe that the mean of the accumulated error increased multifold with scheduling in comparison to the base case. Further exploration for the reason for the rise in the mean led to the conclusion that every time the DR vectors are scheduled in a way to equalize the total error, it pushes each receivers total error higher. Also, as the accumulated error has an estimated component, the schedule is not accurate to equalize the errors for the receivers, leading to the DR vector reaching earlier or later than the actual schedule. In either case, the error is not equalized and if the DR vector reaches late, it actually increases the error for a receiver beyond the highest accumulated error. This means that at the next trigger, this receiver will be the one with highest error and every other receiver"s error will be pushed to this error value. This flip-flop effect leads to the increase in the accumulated error for all the receivers. The scheduling for fairness leads to the decrease in standard deviation (i.e., increases the fairness among different players), but it comes at the cost of higher mean error, which may not be a desirable feature. This led us to explore different ways of equalizing the accumulated errors. The approach discussed in the following section is a heuristic approach based on the following idea. Using the same amount of DR vectors over time as in the base case, instead of sending the DR vectors to all the receivers at the same frequency as in the base case, if we can increase the frequency of sending the DR vectors to the receiver with higher accumulated error and decrease the frequency of sending DR vectors to the receiver with lower accumulated error, we can equalize the export error of all receivers over time. At the same time we wish to decrease the error of the receiver with the highest accumulated error in the base case (of course, this receiver would be sent more DR vectors than in the base case). We refer to such an algorithm as a budget based algorithm. 5. BUDGET BASED ALGORITHM In a game, the sender of an entity sends DR vectors to all the receivers every time a threshold is crossed by the entity. Lower the threshold, more DR vectors are generated during a given time period. Since the DR vectors are sent to all the receivers and the network delay between the sender-receiver pairs cannot be avoided, the before export error 3 with the most distant player will always 3 Note that after export error is eliminated by using synchronized clock among the players. 7 0 1000 2000 3000 4000 5000 15950 16000 16050 16100 16150 16200 16250 16300 MeanAccumulatedError Time in Seconds Base Case Scheduling Algorithm #1 0 50 100 150 200 250 300 350 400 450 500 15950 16000 16050 16100 16150 16200 16250 16300 StandardDeviationofAccumulatedError Time in Seconds Base Case Scheduling Algorithm #1 Figure 4: Mean and standard deviation of error with scheduling and without (i.e., base case). be higher than the rest. In order to mitigate the imbalance in the error, we propose to send DR vectors selectively to different players based on the accumulated errors of these players. The budget based algorithm is based on this idea and there are two variations of it. One is a probabilistic budget based scheme and the other, a deterministic budget base scheme. 5.1 Probabilistic budget based scheme The probabilistic budget based scheme has three main steps: a) lower the dead reckoning threshold but at the same time keep the total number of DRs sent the same as the base case, b) at every trigger, probabilistically pick a player to send the DR vector to, and c) send the DR vector to the chosen player. These steps are described below. The lowering of DR threshold is implemented as follows. Lowering the threshold is equivalent to increasing the number of trigger points where DR vectors are generated. Suppose the threshold is such that the number of triggers caused by it in the base case is t and at each trigger n DR vectors sent by the sender, which results in a total of nt DR vectors. Our goal is to keep the total number of DR vectors sent by the sender fixed at nt, but lower the number of DR vectors sent at each trigger (i.e., do not send the DR vector to all the receivers). Let n and t be the number of DR vectors sent at each trigger and number of triggers respectively in the modified case. We want to ensure n t = nt. Since we want to increase the number of trigger points, i.e, t > t, this would mean that n < n. That is, not all receivers will be sent the DR vector at every trigger. In the probabilistic budget based scheme, at each trigger, a probability is calculated for each receiver to be sent a DR vector and only one receiver is sent the DR (n = 1). This probability is based on the relative weights of the receivers" accumulated errors. That is, a receiver with a higher accumulated error will have a higher probability of being sent the DR vector. Consider that the accumulated error for three players are a1, a2 and a3 respectively. Then the probability of player 1 receiving the DR vector would be a1 a1+a2+a3 . Similarly for the other players. Once the player is picked, the DR vector is sent to that player. To compare the probabilistic budget based algorithm with the base case, we needed to lower the threshold for the base case (for fair comparison). As the dead reckoning threshold in the base case was already very fine, it was decided that instead of lowering the threshold, the probabilistic budget based approach would be compared against a modified base case that would use the normal threshold as the budget based algorithm but the base case was modified such that every third trigger would be actually used to send out a DR vector to all the three receivers used in our experiments. This was called as the 1/3 base case as it resulted in 1/3 number of DR vectors being sent as compared to the base case. The budget per trigger for the probability based approach was calculated as one DR vector at each trigger as compared to three DR vectors at every third trigger in the 1/3 base case; thus the two cases lead to the same number of DR vectors being sent out over time. In order to evaluate the effectiveness of the probabilistic budget based algorithm, we instrumented the BZFlag game to use this approach. We used the same testbed consisting of one sender and three receivers with delays of 800ms, 500ms and 200ms from the sender and with low delay variance (100ms) and moderate delay variance (180ms). The results are shown in Figures 5 and 6. As mentioned earlier, the x-axis of these graphs represents the system time when the snapshot of the game was taken. Observe from the figures that the standard deviation of the accumulated error among the receivers with the probabilistic budget based algorithm is less than the 1/3 base case and the mean is a little higher than the 1/3 base case. This implies that the game is fairer as compared to the 1/3 base case at the cost of increasing the mean error by a small amount as compared to the 1/3 base case. The increase in mean error in the probabilistic case compared to the 1/3 base case can be attributed to the fact that the even though the probabilistic approach on average sends the same number of DR vectors as the 1/3 base case, it sometimes sends DR vectors to a receiver less frequently and sometimes more frequently than the 1/3 base case due to its probabilistic nature. When a receiver does not receive a DR vector for a long time, the receiver"s trajectory is more and more off of the sender"s trajectory and hence the rate of buildup of the error at the receiver is higher. At times when a receiver receives DR vectors more frequently, it builds up error at a lower rate but there is no way of reversing the error that was built up when it did not receive a DR vector for a long time. This leads the receivers to build up more error in the probabilistic case as compared to the 1/3 base case where the receivers receive a DR vector almost periodically. 8 0 200 400 600 800 1000 15950 16000 16050 16100 16150 16200 16250 16300 MeanAccumulatedError Time in Seconds 1/3 Base Case Deterministic Algorithm Probabilistic Algorithm 0 50 100 150 200 250 300 350 400 450 500 15950 16000 16050 16100 16150 16200 16250 16300 StandardDeviationofAccumulatedError Time in Seconds 1/3 Base Case Deterministic Algorithm Probabilistic Algorithm Figure 5: Mean and standard deviation of error for different algorithms (including budget based algorithms) for low delay variance. 0 200 400 600 800 1000 16960 16980 17000 17020 17040 17060 17080 17100 17120 17140 17160 17180 MeanAccumulatedError Time in Seconds 1/3 Base Case Deterministic Algorithm Probabilistic Algorithm 0 50 100 150 200 250 300 16960 16980 17000 17020 17040 17060 17080 17100 17120 17140 17160 17180 StandardDeviationofAccumulatedError Time in Seconds 1/3 Base Case Deterministic Algorithm Probabilistic Algorithm Figure 6: Mean and standard deviation of error for different algorithms (including budget based algorithms) for moderate delay variance. 5.2 Deterministic budget based scheme To bound the increase in mean error we decided to modify the budget based algorithm to be deterministic. The first two steps of the algorithm are the same as in the probabilistic algorithm; the trigger points are increased to lower the threshold and accumulated errors are used to compute the probability that a receiver will receiver a DR vector. Once these steps are completed, a deterministic schedule for the receiver is computed as follows: 1. If there is any receiver(s) tagged to receive a DR vector at the current trigger, the sender sends out the DR vector to the respective receiver(s). If at least one receiver was sent a DR vector, the sender calculates the probabilities of each receiver receiving a DR vector as explained before and follows steps 2 to 6, else it does not do anything. 2. For each receiver, the probability value is multiplied with the budget available at each trigger (which is set to 1 as explained below) to give the frequency of sending the DR vector to each receiver. 3. If any of the receiver"s frequency after multiplying with the budget goes over 1, the receiver"s frequency is set as 1 and the surplus amount is equally distributed to all the receivers by adding the amount to their existing frequencies. This process is repeated until all the receivers have a frequency of less than or equal to 1. This is due to the fact that at a trigger we cannot send more than one DR vector to the respective receiver. That will be wastage of DR vectors by sending redundant information. 4. (1/frequency) gives us the schedule at which the sender should send DR vectors to the respective receiver. Credit obtained previously (explained in step 5) if any is subtracted from the schedule. Observe that the resulting value of the schedule might not be an integer; hence, the value is rounded off by taking the ceiling of the schedule. For example, if the frequency is 1/3.5, this implies that we would like to have a DR vector sent every 3.5 triggers. However, we are constrained to send it at the 4th trigger giving us a credit of 0.5. When we do send the DR vector next time, we would be able to send it 9 on the 3rd trigger because of the 0.5 credit. 5. The difference between the schedule and the ceiling of the schedule is the credit that the receiver has obtained which is remembered for the future and used at the next time as explained in step 4. 6. For each of those receivers who were sent a DR vector at the current trigger, the receivers are tagged to receive the next DR vector at the trigger that happens exactly schedule (the ceiling of the schedule) number of times away from the current trigger. Observe that no other receiver"s schedule is modified at this point as they all are running a schedule calculated at some previous point of time. Those schedules will be automatically modified at the trigger when they are scheduled to receive the next DR vector. At the first trigger, the sender sends the DR vector to all the receivers and uses a relative probability of 1/n for each receiver and follows the steps 2 to 6 to calculate the next schedule for each receiver in the same way as mentioned for other triggers. This algorithm ensures that every receiver has a guaranteed schedule of receiving DR vectors and hence there is no irregularity in sending the DR vector to any receiver as was observed in the budget based probabilistic algorithm. We used the testbed described earlier (three receivers with varying delays) to evaluate the deterministic algorithm using the budget of 1 DR vector per trigger so as to use the same number of DR vectors as in the 1/3 base case. Results from our experiments are shown in Figures 5 and 6. It can be observed that the standard deviation of error in the deterministic budget based algorithm is less than the 1/3 base case and also has the same mean error as the 1/3 base case. This indicates that the deterministic algorithm is more fair than the 1/3 base case and at the same time does not increase the mean error thereby leading to a better game quality compared to the probabilistic algorithm. In general, when comparing the deterministic approach to the probabilistic approach, we found that the mean accumulated error was always less in the deterministic approach. With respect to standard deviation of the accumulated error, we found that in the fixed or low variance cases, the deterministic approach was generally lower, but in higher variance cases, it was harder to draw conclusions as the probabilistic approach was sometimes better than the deterministic approach. 6. CONCLUSIONS AND FUTURE WORK In distributed multi-player games played across the Internet, object and player trajectory within the game space are exchanged in terms of DR vectors. Due to the variable delay between players, these DR vectors reach different players at different times. There is unfair advantage gained by receivers who are closer to the sender of the DR as they are able to render the sender"s position more accurately in real time. In this paper, we first developed a model for estimating the error in rendering player trajectories at the receivers. We then presented an algorithm based on scheduling the DR vectors to be sent to different players at different times thereby equalizing the error at different players. This algorithm is aimed at making the game fair to all players, but tends to increase the mean error of the players. To counter this effect, we presented budget based algorithms where the DR vectors are still scheduled to be sent at different players at different times but the algorithm balances the need for fairness with the requirement that the error of the worst case players (who are furthest from the sender) are not increased compared to the base case (where all DR vectors are sent to all players every time a DR vector is generated). We presented two variations of the budget based algorithms and through experimentation showed that the algorithms reduce the standard deviation of the error thereby making the game more fair and at the same time has comparable mean error to the base case. 7. REFERENCES [1] S.Aggarwal, H. Banavar, A. Khandelwal, S. Mukherjee, and S. Rangarajan, Accuracy in Dead-Reckoning based Distributed Multi-Player Games, Proceedings of ACM SIGCOMM 2004 Workshop on Network and System Support for Games (NetGames 2004), Aug. 2004. [2] L. Gautier and C. Diot, Design and Evaluation of MiMaze, a Multiplayer Game on the Internet, in Proc. of IEEE Multimedia (ICMCS"98), 1998. [3] M. Mauve, Consistency in Replicated Continuous Interactive Media, in Proc. of the ACM Conference on Computer Supported Cooperative Work (CSCW"00), 2000, pp. 181-190. [4] S.K. Singhal and D.R. Cheriton, Exploiting Position History for Efficient Remote Rendering in Networked Virtual Reality, Presence: Teleoperators and Virtual Environments, vol. 4, no. 2, pp. 169-193, 1995. [5] C. Diot and L. Gautier, A Distributed Architecture for Multiplayer Interactive Applications on the Internet, in IEEE Network Magazine, 1999, vol. 13, pp. 6-15. [6] L. Pantel and L.C. Wolf, On the Impact of Delay on Real-Time Multiplayer Games, in Proc. of ACM NOSSDAV"02, May 2002. [7] Y. Lin, K. Guo, and S. Paul, Sync-MS: Synchronized Messaging Service for Real-Time Multi-Player Distributed Games, in Proc. of 10th IEEE International Conference on Network Protocols (ICNP), Nov 2002. [8] K. Guo, S. Mukherjee, S. Rangarajan, and S. Paul, A Fair Message Exchange Framework for Distributed Multi-Player Games, in Proc. of NetGames2003, May 2003. [9] N. E. Baughman and B. N. Levine, Cheat-Proof Playout for Centralized and Distributed Online Games, in Proc. of IEEE INFOCOM"01, April 2001. [10] M. Allman and V. Paxson, On Estimating End-to-End Network Path Properties, in Proc. of ACM SIGCOMM"99, Sept. 1999. [11] BZFlag Forum, BZFlag Game, URL: http://www.bzflag.org. [12] Nation Institute of Standards and Technology, NIST Net, URL: http://snad.ncsl.nist.gov/nistnet/. 10

fairness;dead-reckoning vector;export error;network delay;budget based algorithm;clock synchronization;distribute multi-player game;bucket synchronization;mean error;distributed multi-player game;quantization;dead-reckon;scheduling algorithm;accuracy

train_C-53

Globally Synchronized Dead-Reckoning with Local Lag for Continuous Distributed Multiplayer Games

Dead-Reckoning (DR) is an effective method to maintain consistency for Continuous Distributed Multiplayer Games (CDMG). Since DR can filter most unnecessary state updates and improve the scalability of a system, it is widely used in commercial CDMG. However, DR cannot maintain high consistency, and this constrains its application in highly interactive games. With the help of global synchronization, DR can achieve higher consistency, but it still cannot eliminate before inconsistency. In this paper, a method named Globally Synchronized DR with Local Lag (GS-DR-LL), which combines local lag and Globally Synchronized DR (GS-DR), is presented. Performance evaluation shows that GS-DR-LL can effectively decrease before inconsistency, and the effects increase with the lag.

1. INTRODUCTION Nowadays, many distributed multiplayer games adopt replicated architectures. In such games, the states of entities are changed not only by the operations of players, but also by the passing of time [1, 2]. These games are referred to as Continuous Distributed Multiplayer Games (CDMG). Like other distributed applications, CDMG also suffer from the consistency problem caused by network transmission delay. Although new network techniques (e.g. QoS) can reduce or at least bound the delay, they can not completely eliminate it, as there exists the physical speed limitation of light, for instance, 100 ms is needed for light to propagate from Europe to Australia [3]. There are many studies about the effects of network transmission delay in different applications [4, 5, 6, 7]. In replication based games, network transmission delay makes the states of local and remote sites to be inconsistent, which can cause serious problems, such as reducing the fairness of a game and leading to paradoxical situations etc. In order to maintain consistency for distributed systems, many different approaches have been proposed, among which local lag and Dead-Reckoning (DR) are two representative approaches. Mauve et al [1] proposed local lag to maintain high consistency for replicated continuous applications. It synchronizes the physical clocks of all sites in a system. After an operation is issued at local site, it delays the execution of the operation for a short time. During this short time period the operation is transmitted to remote sites, and all sites try to execute the operation at a same physical time. In order to tackle the inconsistency caused by exceptional network transmission delay, a time warp based mechanism is proposed to repair the state. Local lag can achieve significant high consistency, but it is based on operation transmission, which forwards every operation on a shared entity to remote sites. Since operation transmission mechanism requests that all operations should be transmitted in a reliable way, message filtering is difficult to be deployed and the scalability of a system is limited. DR is based on state transmission mechanism. In addition to the high fidelity model that maintains the accurate states of its own entities, each site also has a DR model that estimates the states of all entities (including its own entities). After each update of its own entities, a site compares the accurate state with the estimated one. If the difference exceeds a pre-defined threshold, a state update would be transmitted to all sites and all DR models would be corrected. Through state estimation, DR can not only maintain consistency but also decrease the number of transmitted state updates. Compared with aforementioned local lag, DR cannot maintain high consistency. Due to network transmission delay, when a remote site receives a state update of an entity the state of the entity might have changed at the site sending the state update. In order to make DR maintain high consistency, Aggarwal et al [8] proposed Globally Synchronized DR (GS-DR), which synchronizes the physical clocks of all sites in a system and adds time stamps to transmitted state updates. Detailed description of GS-DR can be found in Section 3. When a state update is available, GS-DR immediately updates the state of local site and then transmits the state update to remote sites, which causes the states of local site and remote sites to be inconsistent in the transmission procedure. Thus with the synchronization of physical clocks, GS-DR can eliminate after inconsistency, but it cannot tackle before inconsistency [8]. In this paper, we propose a new method named globally synchronized DR with Local Lag (GS-DR-LL), which combines local lag and GS-DR. By delaying the update to local site, GS-DR-LL can achieve higher consistency than GS-DR. The rest of this paper is organized as follows: Section 2 gives the definition of consistency and corresponding metrics; the cause of the inconsistency of DR is analyzed in Section 3; Section 4 describes how GS-DR-LL works; performance evaluation is presented in Section 5; Section 6 concludes the paper. 2. CONSISTENCY DEFINITIONS AND METRICS The consistency of replicated applications has already been well defined in discrete domain [9, 10, 11, 12], but few related work has been done in continuous domain. Mauve et al [1] have given a definition of consistency for replicated applications in continuous domain, but the definition is based on operation transmission and it is difficult for the definition to describe state transmission based methods (e.g. DR). Here, we present an alternative definition of consistency in continuous domain, which suits state transmission based methods well. Given two distinct sites i and j, which have replicated a shared entity e, at a given time t, the states of e at sites i and j are Si(t) and Sj(t). DEFINITION 1: the states of e at sites i and j are consistent at time t, iff: De(i, j, t) = |Si(t) - Sj(t)| = 0 (1) DEFINITION 2: the states of e at sites i and j are consistent between time t1 and t2 (t1 < t2), iff: De(i, j, t1, t2) = dt|)t(S)t(S| t2 t1 ji = 0 (2) In this paper, formulas (1) and (2) are used to determine whether the states of shared entities are consistent between local and remote sites. Due to network transmission delay, it is difficult to maintain the states of shared entities absolutely consistent. Corresponding metrics are needed to measure the consistency of shared entities between local and remote sites. De(i, j, t) can be used as a metric to measure the degree of consistency at a certain time point. If De(i, j, t1) > De(i, j, t2), it can be stated that between sites i and j, the consistency of the states of entity e at time point t1 is lower than that at time point t2. If De(i, j, t) > De(l, k, t), it can be stated that, at time point t, the consistency of the states of entity e between sites i and j is lower than that between sites l and k. Similarly, De(i, j, t1, t2) can been used as a metric to measure the degree of consistency in a certain time period. If De(i, j, t1, t2) > De(i, j, t3, t4) and |t1 - t2| = |t3 - t4|, it can be stated that between sites i and j, the consistency of the states of entity e between time points t1 and t2 is lower than that between time points t3 and t4. If De(i, j, t1, t2) > De(l, k, t1, t2), it can be stated that between time points t1 and t2, the consistency of the states of entity e between sites i and j is lower than that between sites l and k. In DR, the states of entities are composed of the positions and orientations of entities and some prediction related parameters (e.g. the velocities of entities). Given two distinct sites i and j, which have replicated a shared entity e, at a given time point t, the positions of e at sites i and j are (xit, yit, zit) and (xjt, yjt, zjt), De(i, j, t) and D (i, j, t1, t2) could be calculated as: De(i, j, t) = )zz()yy()xx( jtit 2 jtit 2 jtit 2 (3) De(i, j, t1, t2) = dt)zz()yy()xx( 2t 1t jtit 2 jtit 2 jtit 2 (4) In this paper, formulas (3) and (4) are used as metrics to measure the consistency of shared entities between local and remote sites. 3. INCONSISTENCY IN DR The inconsistency in DR can be divided into two sections by the time point when a remote site receives a state update. The inconsistency before a remote site receives a state update is referred to as before inconsistency, and the inconsistency after a remote site receives a state update is referred to as after inconsistency. Before inconsistency and after inconsistency are similar with the terms before export error and after export error [8]. After inconsistency is caused by the lack of synchronization between the physical clocks of all sites in a system. By employing physical clock synchronization, GS-DR can accurately calculate the states of shared entities after receiving state updates, and it can eliminate after inconsistency. Before inconsistency is caused by two reasons. The first reason is the delay of sending state updates, as local site does not send a state update unless the difference between accurate state and the estimated one is larger than a predefined threshold. The second reason is network transmission delay, as a shared entity can be synchronized only after remote sites receiving corresponding state update. Figure 1. The paths of a shared entity by using GS-DR. For example, it is assumed that the velocity of a shared entity is the only parameter to predict the entity"s position, and current position of the entity can be calculated by its last position and current velocity. To simplify the description, it is also assumed that there are only two sites i and j in a game session, site i acts as 2 The 5th Workshop on Network & System Support for Games 2006 - NETGAMES 2006 local site and site j acts as remote site, and t1 is the time point the local site updates the state of the shared entity. Figure 1 illustrates the paths of the shared entity at local site and remote site in x axis by using GS-DR. At the beginning, the positions of the shared entity are the same at sites i and j and the velocity of the shared entity is 0. Before time point t0, the paths of the shared entity at sites i and j in x coordinate are exactly the same. At time point t0, the player at site i issues an operation, which changes the velocity in x axis to v0. Site i first periodically checks whether the difference between the accurate position of the shared entity and the estimated one, 0 in this case, is larger than a predefined threshold. At time point t1, site i finds that the difference is larger than the threshold and it sends a state update to site j. The state update contains the position and velocity of the shared entity at time point t1 and time point t1 is also attached as a timestamp. At time point t2, the state update reaches site j, and the received state and the time deviation between time points t1 and t2 are used to calculate the current position of the shared entity. Then site j updates its replicated entity"s position and velocity, and the paths of the shared entity at sites i and j overlap again. From Figure 1, it can be seen that the after inconsistency is 0, and the before consistency is composed of two parts, D1 and D2. D1 is De(i, j, t0, t1) and it is caused by the state filtering mechanism of DR. D2 is De(i, j, t1, t2) and it is caused by network transmission delay. 4. GLOBALLY SYNCHRONIZED DR WITH LOCAL LAG From the analysis in Section 3, It can be seen that GS-DR can eliminate after inconsistency, but it cannot effectively tackle before inconsistency. In order to decrease before inconsistency, we propose GS-DR-LL, which combines GS-DR with local lag and can effectively decrease before inconsistency. In GS-DR-LL, the state of a shared entity at a certain time point t is notated as S = (t, pos, par 1, par 2, ……, par n), in which pos means the position of the entity and par 1 to par n means the parameters to calculate the position of the entity. In order to simplify the description of GS-DR-LL, it is assumed that there are only one shared entity and one remote site. At the beginning of a game session, the states of the shared entity are the same at local and remote sites, with the same position p0 and parameters pars0 (pars represents all the parameters). Local site keeps three states: the real state of the entity Sreal, the predicted state at remote site Sp-remote, and the latest state updated to remote site Slate. Remote site keep only one state Sremote, which is the real state of the entity at remote site. Therefore, at the beginning of a game session Sreal = Sp-remote = Slate = Sremote = (t0, p0, pars0). In GS-DR-LL, it is assumed that the physical clocks of all sites are synchronized with a deviation of less than 50 ms (using NTP or GPS clock). Furthermore, it is necessary to make corrections to a physical clock in a way that does not result in decreasing the value of the clock, for example by slowing down or halting the clock for a period of time. Additionally it is assumed that the game scene is updated at a fixed frequency and T stands for the time interval between two consecutive updates, for example, if the scene update frequency is 50 Hz, T would be 20 ms. n stands for the lag value used by local lag, and t stands for current physical time. After updating the scene, local site waits for a constant amount of time T. During this time period, local site receives the operations of the player and stores them in a list L. All operations in L are sorted by their issue time. At the end of time period T, local site executes all stored operations, whose issue time is between t - T and t, on Slate to get the new Slate, and it also executes all stored operations, whose issue time is between t - (n + T) and t - n, on Sreal to get the new Sreal. Additionally, local site uses Sp-remote and corresponding prediction methods to estimate the new Sp-remote. After new Slate, Sreal, and Sp-remote are calculated, local site compares whether the difference between the new Slate and Spremote exceeds the predefined threshold. If YES, local site sends new Slate to remote site and Sp-remote is updated with new Slate. Note that the timestamp of the sent state update is t. After that, local site uses Sreal to update local scene and deletes the operations, whose issue time is less than t - n, from L. After updating the scene, remote site waits for a constant amount of time T. During this time period, remote site stores received state update(s) in a list R. All state updates in R are sorted by their timestamps. At the end of time period T, remote site checks whether R contains state updates whose timestamps are less than t - n. Note that t is current physical time and it increases during the transmission of state updates. If YES, it uses these state updates and corresponding prediction methods to calculate the new Sremote, else they use Sremote and corresponding prediction methods to estimate the new Sremote. After that, local site uses Sremote to update local scene and deletes the sate updates, whose timestamps are less than t - n, from R. From the above description, it can been see that the main difference between GS-DR and GS-DR-LL is that GS-DR-LL uses the operations, whose issue time is less than t - n, to calculate Sreal. That means that the scene seen by local player is the results of the operations issued a period of time (i.e. n) ago. Meanwhile, if the results of issued operations make the difference between Slate and Sp-remote exceed a predefined threshold, corresponding state updates are sent to remote sites immediately. The aforementioned is the basic mechanism of GS-DR-LL. In the case with multiple shared entities and remote sites, local site calculates Slate, Sreal, and Sp-remote for different shared entities respectively, if there are multiple Slate need to be transmitted, local site packets them in one state update and then send it to all remote sites. Figure 2 illustrates the paths of a shared entity at local site and remote site while using GS-DR and GS-DR-LL. All conditions are the same with the conditions used in the aforementioned example describing GS-DR. Compared with t1, t2, and n, T (i.e. the time interval between two consecutive updates) is quite small and it is ignored in the following description. At time point t0, the player at site i issues an operation, which changes the velocity of the shared entity form 0 to v0. By using GS-DR-LL, the results of the operation are updated to local scene at time point t0 + n. However the operation is immediately used to calculate Slate, thus in spite of GS-DR or GS-DR-LL, at time point t1 site i finds that the difference between accurate position and the estimated one is larger than the threshold and it sends a state update to site j. At time point t2, the state update is received by remote site j. Assuming that the timestamp of the state update is less than t - n, site j uses it to update local scene immediately. The 5th Workshop on Network & System Support for Games 2006 - NETGAMES 2006 3 With GS-DR, the time period of before inconsistency is (t2 - t1) + (t1 - t0), whereas it decreases to (t2 - t1 - n) + (t1 - t0) with the help of GS-DR-LL. Note that t2 - t1 is caused by network transmission delay and t1 - t0 is caused by the state filtering mechanism of DR. If n is larger than t2 - t1, GS-DR-LL can eliminate the before inconsistency caused by network transmission delay, but it cannot eliminate the before inconsistency caused by the state filtering mechanism of DR (unless the threshold is set to 0). In highly interactive games, which request high consistency and GS-DR-LL might be employed, the results of operations are quite difficult to be estimated and a small threshold must be used. Thus, in practice, most before inconsistency is caused by network transmission delay and GS-DR-LL has the capability to eliminate such before inconsistency. Figure 2. The paths of a shared entity by using GS-DR and GS-DR-LL. To GS-DR-LL, the selection of lag value n is very important, and both network transmission delay and the effects of local lag on interaction should be considered. According to the results of HCI related researches, humans cannot perceive the delay imposed on a system when it is smaller than a specific value, and the specific value depends on both the system and the task. For example, in a graphical user interface a delay of approximately 150 ms cannot be noticed for keyboard interaction and the threshold increases to 195 ms for mouse interaction [13], and a delay of up to 50 ms is uncritical for a car-racing game [5]. Thus if network transmission delay is less than the specific value of a game system, n can be set to the specific value. Else n can be set in terms of the effects of local lag on the interaction of a system [14]. In the case that a large n must be used, some HCI methods (e.g. echo [15]) can be used to relieve the negative effects of the large lag. In the case that n is larger than the network transmission delay, GS-DR-LL can eliminate most before inconsistency. Traditional local lag requests that the lag value must be larger than typical network transmission delay, otherwise state repairs would flood the system. However GS-DR-LL allows n to be smaller than typical network transmission delay. In this case, the before inconsistency caused by network transmission delay still exists, but it can be decreased. 5. PERFORMANCE EVALUATION In order to evaluate GS-DR-LL and compare it with GS-DR in a real application, we had implemented both two methods in a networked game named spaceship [1]. Spaceship is a very simple networked computer game, in which players can control their spaceships to accelerate, decelerate, turn, and shoot spaceships controlled by remote players with laser beams. If a spaceship is hit by a laser beam, its life points decrease one. If the life points of a spaceship decrease to 0, the spaceship is removed from the game and the player controlling the spaceship loses the game. In our practical implementation, GS-DR-LL and GS-DR coexisted in the game system, and the test bed was composed of two computers connected by 100 M switched Ethernet, with one computer acted as local site and the other acted as remote site. In order to simulate network transmission delay, a specific module was developed to delay all packets transmitted between the two computers in terms of a predefined delay value. The main purpose of performance evaluation is to study the effects of GS-DR-LL on decreasing before inconsistency in a particular game system under different thresholds, lags, and network transmission delays. Two different thresholds were used in the evaluation, one is 10 pixels deviation in position or 15 degrees deviation in orientation, and the other is 4 pixels or 5 degrees. Six different combinations of lag and network transmission delay were used in the evaluation and they could be divided into two categories. In one category, the lag was fixed at 300 ms and three different network transmission delays (100 ms, 300 ms, and 500 ms) were used. In the other category, the network transmission delay was fixed at 800 ms and three different lags (100 ms, 300 ms, and 500 ms) were used. Therefore the total number of settings used in the evaluation was 12 (2 × 6). The procedure of performance evaluation was composed of three steps. In the first step, two participants were employed to play the game, and the operation sequences were recorded. Based on the records, a sub operation sequence, which lasted about one minute and included different operations (e.g. accelerate, decelerate, and turn), was selected. In the second step, the physical clocks of the two computers were synchronized first. Under different settings and consistency maintenance approaches, the selected sub operation sequence was played back on one computer, and it drove the two spaceships, one was local and the other was remote, to move. Meanwhile, the tracks of the spaceships on the two computers were recorded separately and they were called as a track couple. Since there are 12 settings and 2 consistency maintenance approaches, the total number of recorded track couples was 24. In the last step, to each track couple, the inconsistency between them was calculated, and the unit of inconsistency was pixel. Since the physical clocks of the two computers were synchronized, the calculation of inconsistency was quite simple. The inconsistency at a particular time point was the distance between the positions of the two spaceships at that time point (i.e. formula (3)). In order to show the results of inconsistency in a clear way, only parts of the results, which last about 7 seconds, are used in the following figures, and the figures show almost the same parts of the results. Figures 3, 4, and 5 show the results of inconsistency when the lag is fixed at 300 ms and the network transmission delays are 100, 300, and 500 ms. It can been seen that inconsistency does exist, but in most of the time it is 0. Additionally, inconsistency increases with the network transmission delay, but decreases with the threshold. Compared with GS-DR, GS-DR-LL can decrease more inconsistency, and it eliminates most inconsistency when the network transmission delay is 100 ms and the threshold is 4 pixels or 5 degrees. 4 The 5th Workshop on Network & System Support for Games 2006 - NETGAMES 2006 According to the prediction and state filtering mechanisms of DR, inconsistency cannot be completely eliminated if the threshold is not 0. With the definitions of before inconsistency and after inconsistency, it can be indicated that GS-DR and GS-DR-LL both can eliminate after inconsistency, and GS-DR-LL can effectively decrease before inconsistency. It can be foreseen that with proper lag and threshold (e.g. the lag is larger than the network transmission delay and the threshold is 0), GS-DR-LL even can eliminate before inconsistency. 0 10 20 30 40 0.0 1.5 3.1 4.6 6.1 Time (seconds) Inconsistency(pixels) GS-DR-LL GS-DR The threshold is 10 pixels or 15degrees 0 10 20 30 40 0.0 1.5 3.1 4.6 6.1 Time (seconds) Inconsistency(pixels) GS-DR-LL GS-DR The threshold is 4 pixels or 5degrees Figure 3. Inconsistency when the network transmission delay is 100 ms and the lag is 300 ms. 0 10 20 30 40 0.0 1.5 3.1 4.6 6.1 Time (seconds) Inconsistency(pixels) GS-DR-LL GS-DR The threshold is 10 pixels or 15degrees 0 10 20 30 40 0.0 1.5 3.1 4.6 6.1 Time (seconds) Inconsistency(pixels) GS-DR-LL GS-DR The threshold is 4 pixels or 5degrees Figure 4. Inconsistency when the network transmission delay is 300 ms and the lag is 300 ms. 0 10 20 30 40 0.0 1.5 3.1 4.6 6.2 Time (seconds) Inconsistency(pixels) GS-DR-LL GS-DR The threshold is 10 pixels or 15degrees 0 10 20 30 40 0.0 1.5 3.1 4.6 6.1 Time (seconds) Inconsistency(pixels) GS-DR-LL GS-DR The threshold is 4 pixels or 5degrees Figure 5. Inconsistency when the network transmission delay is 500 ms and the lag is 300 ms. Figures 6, 7, and 8 show the results of inconsistency when the network transmission delay is fixed at 800 ms and the lag are 100, 300, and 500 ms. It can be seen that with GS-DR-LL before inconsistency decreases with the lag. In traditional local lag, the lag must be set to a value larger than typical network transmission delay, otherwise the state repairs would flood the system. From the above results it can be seen that there does not exist any constraint on the selection of the lag, with GS-DR-LL a system would work fine even if the lag is much smaller than the network transmission delay. The 5th Workshop on Network & System Support for Games 2006 - NETGAMES 2006 5 From all above results, it can be indicated that GS-DR and GSDR-LL both can eliminate after inconsistency, and GS-DR-LL can effectively decrease before inconsistency, and the effects increase with the lag. 0 10 20 30 40 0.0 1.5 3.1 4.7 6.2 Time (seconds) Inconsistency(pixels) GS-DR-LL GS-DR The threshold is 10 pixels or 15degrees 0 10 20 30 40 0.0 1.5 3.1 4.6 6.1 Time (seconds) Inconsistency(pixels) GS-DR-LL GS-DR The threshold is 4 pixels or 5degrees Figure 6. Inconsistency when the network transmission delay is 800 ms and the lag is 100 ms. 0 10 20 30 40 0.0 1.5 3.1 4.6 6.1 Time (seconds) Inconsistency(pixels) GS-DR-LL GS-DR The threshold is 10 pixels or 15degrees 0 10 20 30 40 0.0 1.5 3.1 4.6 6.1 Time (seconds) Inconsistency(pixels) GS-DR-LL GS-DR The threshold is 4 pixels or 5degrees Figure 7. Inconsistency when the network transmission delay is 800 ms and the lag is 300 ms. 0 10 20 30 40 0.0 1.5 3.1 4.6 6.1 Time (seconds) Inconsistency(pixels) GS-DR-LL GS-DR The threshold is 10 pixels or 15degrees 0 10 20 30 40 0.0 1.5 3.1 4.6 6.1 Time (seconds) Inconsistency(pixels) GS-DR-LL GS-DR The threshold is 4 pixels or 5degrees Figure 8. Inconsistency when the network transmission delay is 800 ms and the lag is 500 ms. 6. CONCLUSIONS Compared with traditional DR, GS-DR can eliminate after inconsistency through the synchronization of physical clocks, but it cannot tackle before inconsistency, which would significantly influence the usability and fairness of a game. In this paper, we proposed a method named GS-DR-LL, which combines local lag and GS-DR, to decrease before inconsistency through delaying updating the execution results of local operations to local scene. Performance evaluation indicates that GS-DR-LL can effectively decrease before inconsistency, and the effects increase with the lag. GS-DR-LL has significant implications to consistency maintenance approaches. First, GS-DR-LL shows that improved DR can not only eliminate after inconsistency but also decrease 6 The 5th Workshop on Network & System Support for Games 2006 - NETGAMES 2006 before inconsistency, with proper lag and threshold, it would even eliminate before inconsistency. As a result, the application of DR can be greatly broadened and it could be used in the systems which request high consistency (e.g. highly interactive games). Second, GS-DR-LL shows that by combining local lag and GSDR, the constraint on selecting lag value is removed and a lag, which is smaller than typical network transmission delay, could be used. As a result, the application of local lag can be greatly broadened and it could be used in the systems which have large typical network transmission delay (e.g. Internet based games). 7. REFERENCES [1] Mauve, M., Vogel, J., Hilt, V., and Effelsberg, W. Local-Lag and Timewarp: Providing Consistency for Replicated Continuous Applications. IEEE Transactions on Multimedia, Vol. 6, No.1, 2004, 47-57. [2] Li, F.W., Li, L.W., and Lau, R.W. Supporting Continuous Consistency in Multiplayer Online Games. In Proc. of ACM Multimedia, 2004, 388-391. [3] Pantel, L. and Wolf, L. On the Suitability of Dead Reckoning Schemes for Games. In Proc. of NetGames, 2002, 79-84. [4] Alhalabi, M.O., Horiguchi, S., and Kunifuji, S. An Experimental Study on the Effects of Network Delay in Cooperative Shared Haptic Virtual Environment. Computers and Graphics, Vol. 27, No. 2, 2003, 205-213. [5] Pantel, L. and Wolf, L.C. On the Impact of Delay on RealTime Multiplayer Games. In Proc. of NOSSDAV, 2002, 2329. [6] Meehan, M., Razzaque, S., Whitton, M.C., and Brooks, F.P. Effect of Latency on Presence in Stressful Virtual Environments. In Proc. of IEEE VR, 2003, 141-148. [7] Bernier, Y.W. Latency Compensation Methods in Client/Server In-Game Protocol Design and Optimization. In Proc. of Game Developers Conference, 2001. [8] Aggarwal, S., Banavar, H., and Khandelwal, A. Accuracy in Dead-Reckoning based Distributed Multi-Player Games. In Proc. of NetGames, 2004, 161-165. [9] Raynal, M. and Schiper, A. From Causal Consistency to Sequential Consistency in Shared Memory Systems. In Proc. of Conference on Foundations of Software Technology and Theoretical Computer Science, 1995, 180-194. [10] Ahamad, M., Burns, J.E., Hutto, P.W., and Neiger, G. Causal Memory. In Proc. of International Workshop on Distributed Algorithms, 1991, 9-30. [11] Herlihy, M. and Wing, J. Linearizability: a Correctness Condition for Concurrent Objects. ACM Transactions on Programming Languages and Systems, Vol. 12, No. 3, 1990, 463-492. [12] Misra, J. Axioms for Memory Access in Asynchronous Hardware Systems. ACM Transactions on Programming Languages and Systems, Vol. 8, No. 1, 1986, 142-153. [13] Dabrowski, J.R. and Munson, E.V. Is 100 Milliseconds too Fast. In Proc. of SIGCHI Conference on Human Factors in Computing Systems, 2001, 317-318. [14] Chen, H., Chen, L., and Chen, G.C. Effects of Local-Lag Mechanism on Cooperation Performance in a Desktop CVE System. Journal of Computer Science and Technology, Vol. 20, No. 3, 2005, 396-401. [15] Chen, L., Chen, H., and Chen, G.C. Echo: a Method to Improve the Interaction Quality of CVEs. In Proc. of IEEE VR, 2005, 269-270. The 5th Workshop on Network & System Support for Games 2006 - NETGAMES 2006 7

local lag;physical clock;time warp;usability and fairness;continuous replicate application;network transmission delay;distribute multi-player game;accurate state;gs-dr-ll;dead-reckon;multiplayer game;consistency;correction

train_C-54

Remote Access to Large Spatial Databases

Enterprises in the public and private sectors have been making their large spatial data archives available over the Internet. However, interactive work with such large volumes of online spatial data is a challenging task. We propose two efficient approaches to remote access to large spatial data. First, we introduce a client-server architecture where the work is distributed between the server and the individual clients for spatial query evaluation, data visualization, and data management. We enable the minimization of the requirements for system resources on the client side while maximizing system responsiveness as well as the number of connections one server can handle concurrently. Second, for prolonged periods of access to large online data, we introduce APPOINT (an Approach for Peer-to-Peer Offloading the INTernet). This is a centralized peer-to-peer approach that helps Internet users transfer large volumes of online data efficiently. In APPOINT, active clients of the clientserver architecture act on the server"s behalf and communicate with each other to decrease network latency, improve service bandwidth, and resolve server congestions.

1. INTRODUCTION In recent years, enterprises in the public and private sectors have provided access to large volumes of spatial data over the Internet. Interactive work with such large volumes of online spatial data is a challenging task. We have been developing an interactive browser for accessing spatial online databases: the SAND (Spatial and Non-spatial Data) Internet Browser. Users of this browser can interactively and visually manipulate spatial data remotely. Unfortunately, interactive remote access to spatial data slows to a crawl without proper data access mechanisms. We developed two separate methods for improving the system performance, together, form a dynamic network infrastructure that is highly scalable and provides a satisfactory user experience for interactions with large volumes of online spatial data. The core functionality responsible for the actual database operations is performed by the server-based SAND system. SAND is a spatial database system developed at the University of Maryland [12]. The client-side SAND Internet Browser provides a graphical user interface to the facilities of SAND over the Internet. Users specify queries by choosing the desired selection conditions from a variety of menus and dialog boxes. SAND Internet Browser is Java-based, which makes it deployable across many platforms. In addition, since Java has often been installed on target computers beforehand, our clients can be deployed on these systems with little or no need for any additional software installation or customization. The system can start being utilized immediately without any prior setup which can be extremely beneficial in time-sensitive usage scenarios such as emergencies. There are two ways to deploy SAND. First, any standard Web browser can be used to retrieve and run the client piece (SAND Internet Browser) as a Java application or an applet. This way, users across various platforms can continuously access large spatial data on a remote location with little or 15 no need for any preceding software installation. The second option is to use a stand-alone SAND Internet Browser along with a locally-installed Internet-enabled database management system (server piece). In this case, the SAND Internet Browser can still be utilized to view data from remote locations. However, frequently accessed data can be downloaded to the local database on demand, and subsequently accessed locally. Power users can also upload large volumes of spatial data back to the remote server using this enhanced client. We focused our efforts in two directions. We first aimed at developing a client-server architecture with efficient caching methods to balance local resources on one side and the significant latency of the network connection on the other. The low bandwidth of this connection is the primary concern in both cases. The outcome of this research primarily addresses the issues of our first type of usage (i.e., as a remote browser application or an applet) for our browser and other similar applications. The second direction aims at helping users that wish to manipulate large volumes of online data for prolonged periods. We have developed a centralized peerto-peer approach to provide the users with the ability to transfer large volumes of data (i.e., whole data sets to the local database) more efficiently by better utilizing the distributed network resources among active clients of a clientserver architecture. We call this architecture APPOINTApproach for Peer-to-Peer Offloading the INTernet. The results of this research addresses primarily the issues of the second type of usage for our SAND Internet Browser (i.e., as a stand-alone application). The rest of this paper is organized as follows. Section 2 describes our client-server approach in more detail. Section 3 focuses on APPOINT, our peer-to-peer approach. Section 4 discusses our work in relation to existing work. Section 5 outlines a sample SAND Internet Browser scenario for both of our remote access approaches. Section 6 contains concluding remarks as well as future research directions. 2. THE CLIENT-SERVER APPROACH Traditionally, Geographic Information Systems (GIS) such as ArcInfo from ESRI [2] and many spatial databases are designed to be stand-alone products. The spatial database is kept on the same computer or local area network from where it is visualized and queried. This architecture allows for instantaneous transfer of large amounts of data between the spatial database and the visualization module so that it is perfectly reasonable to use large-bandwidth protocols for communication between them. There are however many applications where a more distributed approach is desirable. In these cases, the database is maintained in one location while users need to work with it from possibly distant sites over the network (e.g., the Internet). These connections can be far slower and less reliable than local area networks and thus it is desirable to limit the data flow between the database (server) and the visualization unit (client) in order to get a timely response from the system. Our client-server approach (Figure 1) allows the actual database engine to be run in a central location maintained by spatial database experts, while end users acquire a Javabased client component that provides them with a gateway into the SAND spatial database engine. Our client is more than a simple image viewer. Instead, it operates on vector data allowing the client to execute many operations such as zooming or locational queries locally. In Figure 1: SAND Internet Browser - Client-Server architecture. essence, a simple spatial database engine is run on the client. This database keeps a copy of a subset of the whole database whose full version is maintained on the server. This is a concept similar to ‘caching". In our case, the client acts as a lightweight server in that given data, it evaluates queries and provides the visualization module with objects to be displayed. It initiates communication with the server only in cases where it does not have enough data stored locally. Since the locally run database is only updated when additional or newer data is needed, our architecture allows the system to minimize the network traffic between the client and the server when executing the most common user-side operations such as zooming and panning. In fact, as long as the user explores one region at a time (i.e., he or she is not panning all over the database), no additional data needs to be retrieved after the initial population of the client-side database. This makes the system much more responsive than the Web mapping services. Due to the complexity of evaluating arbitrary queries (i.e., more complex queries than window queries that are needed for database visualization), we do not perform user-specified queries on the client. All user queries are still evaluated on the server side and the results are downloaded onto the client for display. However, assuming that the queries are selective enough (i.e., there are far fewer elements returned from the query than the number of elements in the database), the response delay is usually within reasonable limits. 2.1 Client-Server Communication As mentioned above, the SAND Internet Browser is a client piece of the remotely accessible spatial database server built around the SAND kernel. In order to communicate with the server, whose application programming interface (API) is a Tcl-based scripting language, a servlet specifically designed to interface the SAND Internet Browser with the SAND kernel is required on the server side. This servlet listens on a given port of the server for incoming requests from the client. It translates these requests into the SAND-Tcl language. Next, it transmits these SAND-Tcl commands or scripts to the SAND kernel. After results are provided by the kernel, the servlet fetches and processes them, and then sends those results back to the originating client. Once the Java servlet is launched, it waits for a client to initiate a connection. It handles both requests for the actual client Java code (needed when the client is run as an applet) and the SAND traffic. When the client piece is launched, it connects back to the SAND servlet, the communication is driven by the client piece; the server only responds to the client"s queries. The client initiates a transaction by 6 sending a query. The Java servlet parses the query and creates a corresponding SAND-Tcl expression or script in the SAND kernel"s native format. It is then sent to the kernel for evaluation or execution. The kernel"s response naturally depends on the query and can be a boolean value, a number or a string representing a value (e.g., a default color) or, a whole tuple (e.g., in response to a nearest tuple query). If a script was sent to the kernel (e.g., requesting all the tuples matching some criteria), then an arbitrary amount of data can be returned by the SAND server. In this case, the data is first compressed before it is sent over the network to the client. The data stream gets decompressed at the client before the results are parsed. Notice, that if another spatial database was to be used instead of the SAND kernel, then only a simple modification to the servlet would need to be made in order for the SAND Internet Browser to function properly. In particular, the queries sent by the client would need to be recoded into another query language which is native to this different spatial database. The format of the protocol used for communication between the servlet and the client is unaffected. 3. THE PEER-TO-PEER APPROACH Many users may want to work on a complete spatial data set for a prolonged period of time. In this case, making an initial investment of downloading the whole data set may be needed to guarantee a satisfactory session. Unfortunately, spatial data tends to be large. A few download requests to a large data set from a set of idle clients waiting to be served can slow the server to a crawl. This is due to the fact that the common client-server approach to transferring data between the two ends of a connection assumes a designated role for each one of the ends (i.e, some clients and a server). We built APPOINT as a centralized peer-to-peer system to demonstrate our approach for improving the common client-server systems. A server still exists. There is a central source for the data and a decision mechanism for the service. The environment still functions as a client-server environment under many circumstances. Yet, unlike many common client-server environments, APPOINT maintains more information about the clients. This includes, inventories of what each client downloads, their availabilities, etc. When the client-server service starts to perform poorly or a request for a data item comes from a client with a poor connection to the server, APPOINT can start appointing appropriate active clients of the system to serve on behalf of the server, i.e., clients who have already volunteered their services and can take on the role of peers (hence, moving from a client-server scheme to a peer-to-peer scheme). The directory service for the active clients is still performed by the server but the server no longer serves all of the requests. In this scheme, clients are used mainly for the purpose of sharing their networking resources rather than introducing new content and hence they help offload the server and scale up the service. The existence of a server is simpler in terms of management of dynamic peers in comparison to pure peerto-peer approaches where a flood of messages to discover who is still active in the system should be used by each peer that needs to make a decision. The server is also the main source of data and under regular circumstances it may not forward the service. Data is assumed to be formed of files. A single file forms the atomic means of communication. APPOINT optimizes requests with respect to these atomic requests. Frequently accessed data sets are replicated as a byproduct of having been requested by a large number of users. This opens up the potential for bypassing the server in future downloads for the data by other users as there are now many new points of access to it. Bypassing the server is useful when the server"s bandwidth is limited. Existence of a server assures that unpopular data is also available at all times. The service depends on the availability of the server. The server is now more resilient to congestion as the service is more scalable. Backups and other maintenance activities are already being performed on the server and hence no extra administrative effort is needed for the dynamic peers. If a peer goes down, no extra precautions are taken. In fact, APPOINT does not require any additional resources from an already existing client-server environment but, instead, expands its capability. The peers simply get on to or get off from a table on the server. Uploading data is achieved in a similar manner as downloading data. For uploads, the active clients can again be utilized. Users can upload their data to a set of peers other than the server if the server is busy or resides in a distant location. Eventually the data is propagated to the server. All of the operations are performed in a transparent fashion to the clients. Upon initial connection to the server, they can be queried as to whether or not they want to share their idle networking time and disk space. The rest of the operations follow transparently after the initial contact. APPOINT works on the application layer but not on lower layers. This achieves platform independence and easy deployment of the system. APPOINT is not a replacement but an addition to the current client-server architectures. We developed a library of function calls that when placed in a client-server architecture starts the service. We are developing advanced peer selection schemes that incorporate the location of active clients, bandwidth among active clients, data-size to be transferred, load on active clients, and availability of active clients to form a complete means of selecting the best clients that can become efficient alternatives to the server. With APPOINT we are defining a very simple API that could be used within an existing client-server system easily. Instead of denial of service or a slow connection, this API can be utilized to forward the service appropriately. The API for the server side is: start(serverPortNo) makeFileAvailable(file,location,boolean) callback receivedFile(file,location) callback errorReceivingFile(file,location,error) stop() Similarly the API for the client side is: start(clientPortNo,serverPortNo,serverAddress) makeFileAvailable(file,location,boolean) receiveFile(file,location) sendFile(file,location) stop() The server, after starting the APPOINT service, can make all of the data files available to the clients by using the makeFileAvailable method. This will enable APPOINT to treat the server as one of the peers. The two callback methods of the server are invoked when a file is received from a client, or when an error is encountered while receiving a file from a client. APPOINT guar7 Figure 2: The localization operation in APPOINT. antees that at least one of the callbacks will be called so that the user (who may not be online anymore) can always be notified (i.e., via email). Clients localizing large data files can make these files available to the public by using the makeFileAvailable method on the client side. For example, in our SAND Internet Browser, we have the localization of spatial data as a function that can be chosen from our menus. This functionality enables users to download data sets completely to their local disks before starting their queries or analysis. In our implementation, we have calls to the APPOINT service both on the client and the server sides as mentioned above. Hence, when a localization request comes to the SAND Internet Browser, the browser leaves the decisions to optimally find and localize a data set to the APPOINT service. Our server also makes its data files available over APPOINT. The mechanism for the localization operation is shown with more details from the APPOINT protocols in Figure 2. The upload operation is performed in a similar fashion. 4. RELATED WORK There has been a substantial amount of research on remote access to spatial data. One specific approach has been adopted by numerous Web-based mapping services (MapQuest [5], MapsOnUs [6], etc.). The goal in this approach is to enable remote users, typically only equipped with standard Web browsers, to access the company"s spatial database server and retrieve information in the form of pictorial maps from them. The solution presented by most of these vendors is based on performing all the calculations on the server side and transferring only bitmaps that represent results of user queries and commands. Although the advantage of this solution is the minimization of both hardware and software resources on the client site, the resulting product has severe limitations in terms of available functionality and response time (each user action results in a new bitmap being transferred to the client). Work described in [9] examines a client-server architecture for viewing large images that operates over a lowbandwidth network connection. It presents a technique based on wavelet transformations that allows the minimization of the amount of data needed to be transferred over the network between the server and the client. In this case, while the server holds the full representation of the large image, only a limited amount of data needs to be transferred to the client to enable it to display a currently requested view into the image. On the client side, the image is reconstructed into a pyramid representation to speed up zooming and panning operations. Both the client and the server keep a common mask that indicates what parts of the image are available on the client and what needs to be requested. This also allows dropping unnecessary parts of the image from the main memory on the server. Other related work has been reported in [16] where a client-server architecture is described that is designed to provide end users with access to a server. It is assumed that this data server manages vast databases that are impractical to be stored on individual clients. This work blends raster data management (stored in pyramids [22]) with vector data stored in quadtrees [19, 20]. For our peer-to-peer transfer approach (APPOINT), Napster is the forefather where a directory service is centralized on a server and users exchange music files that they have stored on their local disks. Our application domain, where the data is already freely available to the public, forms a prime candidate for such a peer-to-peer approach. Gnutella is a pure (decentralized) peer-to-peer file exchange system. Unfortunately, it suffers from scalability issues, i.e., floods of messages between peers in order to map connectivity in the system are required. Other systems followed these popular systems, each addressing a different flavor of sharing over the Internet. Many peer-to-peer storage systems have also recently emerged. PAST [18], Eternity Service [7], CFS [10], and OceanStore [15] are some peer-to-peer storage systems. Some of these systems have focused on anonymity while others have focused on persistence of storage. Also, other approaches, like SETI@Home [21], made other resources, such as idle CPUs, work together over the Internet to solve large scale computational problems. Our goal is different than these approaches. With APPOINT, we want to improve existing client-server systems in terms of performance by using idle networking resources among active clients. Hence, other issues like anonymity, decentralization, and persistence of storage were less important in our decisions. Confirming the authenticity of the indirectly delivered data sets is not yet addressed with APPOINT. We want to expand our research, in the future, to address this issue. From our perspective, although APPOINT employs some of the techniques used in peer-to-peer systems, it is also closely related to current Web caching architectures. Squirrel [13] forms the middle ground. It creates a pure peer-topeer collaborative Web cache among the Web browser caches of the machines in a local-area network. Except for this recent peer-to-peer approach, Web caching is mostly a wellstudied topic in the realm of server/proxy level caching [8, 11, 14, 17]. Collaborative Web caching systems, the most relevant of these for our research, focus on creating either a hierarchical, hash-based, central directory-based, or multicast-based caching schemes. We do not compete with these approaches. In fact, APPOINT can work in tandem with collaborative Web caching if they are deployed together. We try to address the situation where a request arrives at a server, meaning all the caches report a miss. Hence, the point where the server is reached can be used to take a central decision but then the actual service request can be forwarded to a set of active clients, i.e., the down8 load and upload operations. Cache misses are especially common in the type of large data-based services on which we are working. Most of the Web caching schemes that are in use today employ a replacement policy that gives a priority to replacing the largest sized items over smaller-sized ones. Hence, these policies would lead to the immediate replacement of our relatively large data files even though they may be used frequently. In addition, in our case, the user community that accesses a certain data file may also be very dispersed from a network point of view and thus cannot take advantage of any of the caching schemes. Finally, none of the Web caching methods address the symmetric issue of large data uploads. 5. A SAMPLE APPLICATION FedStats [1] is an online source that enables ordinary citizens access to official statistics of numerous federal agencies without knowing in advance which agency produced them. We are using a FedStats data set as a testbed for our work. Our goal is to provide more power to the users of FedStats by utilizing the SAND Internet Browser. As an example, we looked at two data files corresponding to Environmental Protection Agency (EPA)-regulated facilities that have chlorine and arsenic, respectively. For each file, we had the following information available: EPA-ID, name, street, city, state, zip code, latitude, longitude, followed by flags to indicate if that facility is in the following EPA programs: Hazardous Waste, Wastewater Discharge, Air Emissions, Abandoned Toxic Waste Dump, and Active Toxic Release. We put this data into a SAND relation where the spatial attribute ‘location" corresponds to the latitude and longitude. Some queries that can be handled with our system on this data include: 1. Find all EPA-regulated facilities that have arsenic and participate in the Air Emissions program, and: (a) Lie in Georgia to Illinois, alphabetically. (b) Lie within Arkansas or 30 miles within its border. (c) Lie within 30 miles of the border of Arkansas (i.e., both sides of the border). 2. For each EPA-regulated facility that has arsenic, find all EPA-regulated facilities that have chlorine and: (a) That are closer to it than to any other EPAregulated facility that has arsenic. (b) That participate in the Air Emissions program and are closer to it than to any other EPAregulated facility which has arsenic. In order to avoid reporting a particular facility more than once, we use our ‘group by EPA-ID" mechanism. Figure 3 illustrates the output of an example query that finds all arsenic sites within a given distance of the border of Arkansas. The sites are obtained in an incremental manner with respect to a given point. This ordering is shown by using different color shades. With this example data, it is possible to work with the SAND Internet Browser online as an applet (connecting to a remote server) or after localizing the data and then opening it locally. In the first case, for each action taken, the client-server architecture will decide what to ask for from the server. In the latter case, the browser will use the peerto-peer APPOINT architecture for first localizing the data. 6. CONCLUDING REMARKS An overview of our efforts in providing remote access to large spatial data has been given. We have outlined our approaches and introduced their individual elements. Our client-server approach improves the system performance by using efficient caching methods when a remote server is accessed from thin-clients. APPOINT forms an alternative approach that improves performance under an existing clientserver system by using idle client resources when individual users want work on a data set for longer periods of time using their client computers. For the future, we envision development of new efficient algorithms that will support large online data transfers within our peer-to-peer approach using multiple peers simultaneously. We assume that a peer (client) can become unavailable at any anytime and hence provisions need to be in place to handle such a situation. To address this, we will augment our methods to include efficient dynamic updates. Upon completion of this step of our work, we also plan to run comprehensive performance studies on our methods. Another issue is how to access data from different sources in different formats. In order to access multiple data sources in real time, it is desirable to look for a mechanism that would support data exchange by design. The XML protocol [3] has emerged to become virtually a standard for describing and communicating arbitrary data. GML [4] is an XML variant that is becoming increasingly popular for exchange of geographical data. We are currently working on making SAND XML-compatible so that the user can instantly retrieve spatial data provided by various agencies in the GML format via their Web services and then explore, query, or process this data further within the SAND framework. This will turn the SAND system into a universal tool for accessing any spatial data set as it will be deployable on most platforms, work efficiently given large amounts of data, be able to tap any GML-enabled data source, and provide an easy to use graphical user interface. This will also convert the SAND system from a research-oriented prototype into a product that could be used by end users for accessing, viewing, and analyzing their data efficiently and with minimum effort. 7. REFERENCES [1] Fedstats: The gateway to statistics from over 100 U.S. federal agencies. http://www.fedstats.gov/, 2001. [2] Arcinfo: Scalable system of software for geographic data creation, management, integration, analysis, and dissemination. http://www.esri.com/software/ arcgis/arcinfo/index.html, 2002. [3] Extensible markup language (xml). http://www.w3.org/XML/, 2002. [4] Geography markup language (gml) 2.0. http://opengis.net/gml/01-029/GML2.html, 2002. [5] Mapquest: Consumer-focused interactive mapping site on the web. http://www.mapquest.com, 2002. [6] Mapsonus: Suite of online geographic services. http://www.mapsonus.com, 2002. [7] R. Anderson. The Eternity Service. In Proceedings of the PRAGOCRYPT"96, pages 242-252, Prague, Czech Republic, September 1996. [8] L. Breslau, P. Cao, L. Fan, G. Phillips, and S. Shenker. Web caching and Zipf-like distributions: 9 Figure 3: Sample output from the SAND Internet Browser - Large dark dots indicate the result of a query that looks for all arsenic sites within a given distance from Arkansas. Different color shades are used to indicate ranking order by the distance from a given point. Evidence and implications. In Proceedings of the IEEE Infocom"99, pages 126-134, New York, NY, March 1999. [9] E. Chang, C. Yap, and T. Yen. Realtime visualization of large images over a thinwire. In R. Yagel and H. Hagen, editors, Proceedings IEEE Visualization"97 (Late Breaking Hot Topics), pages 45-48, Phoenix, AZ, October 1997. [10] F. Dabek, M. F. Kaashoek, D. Karger, R. Morris, and I. Stoica. Wide-area cooperative storage with CFS. In Proceedings of the ACM SOSP"01, pages 202-215, Banff, AL, October 2001. [11] A. Dingle and T. Partl. Web cache coherence. Computer Networks and ISDN Systems, 28(7-11):907-920, May 1996. [12] C. Esperan¸ca and H. Samet. Experience with SAND/Tcl: a scripting tool for spatial databases. Journal of Visual Languages and Computing, 13(2):229-255, April 2002. [13] S. Iyer, A. Rowstron, and P. Druschel. Squirrel: A decentralized peer-to-peer Web cache. Rice University/Microsoft Research, submitted for publication, 2002. [14] D. Karger, A. Sherman, A. Berkheimer, B. Bogstad, R. Dhanidina, K. Iwamoto, B. Kim, L. Matkins, and Y. Yerushalmi. Web caching with consistent hashing. Computer Networks, 31(11-16):1203-1213, May 1999. [15] J. Kubiatowicz, D. Bindel, Y. Chen, S. Czerwinski, P. Eaton, D. Geels, R. Gummadi, S. Rhea, H. Weatherspoon, W. Weimer, C. Wells, and B. Zhao. OceanStore: An architecture for global-scale persistent store. In Proceedings of the ACM ASPLOS"00, pages 190-201, Cambridge, MA, November 2000. [16] M. Potmesil. Maps alive: viewing geospatial information on the WWW. Computer Networks and ISDN Systems, 29(8-13):1327-1342, September 1997. Also Hyper Proceedings of the 6th International World Wide Web Conference, Santa Clara, CA, April 1997. [17] M. Rabinovich, J. Chase, and S. Gadde. Not all hits are created equal: Cooperative proxy caching over a wide-area network. Computer Networks and ISDN Systems, 30(22-23):2253-2259, November 1998. [18] A. Rowstron and P. Druschel. Storage management and caching in PAST, a large-scale, persistent peer-to-peer storage utility. In Proceedings of the ACM SOSP"01, pages 160-173, Banff, AL, October 2001. [19] H. Samet. Applications of Spatial Data Structures: Computer Graphics, Image Processing, and GIS. Addison-Wesley, Reading, MA, 1990. [20] H. Samet. The Design and Analysis of Spatial Data Structures. Addison-Wesley, Reading, MA, 1990. [21] SETI@Home. http://setiathome.ssl.berkeley.edu/, 2001. [22] L. J. Williams. Pyramidal parametrics. Computer Graphics, 17(3):1-11, July 1983. Also Proceedings of the SIGGRAPH"83 Conference, Detroit, July 1983. 10

large spatial datum;sand;datum visualization;remote access;internet;datum management;web browser;gi;dynamic network infrastructure;client/server;spatial query evaluation;client-server architecture;peer-to-peer;centralized peer-to-peer approach;internet-enabled database management system;network latency

train_C-55

Context Awareness for Group Interaction Support

In this paper, we present an implemented system for supporting group interaction in mobile distributed computing environments. First, an introduction to context computing and a motivation for using contextual information to facilitate group interaction is given. We then present the architecture of our system, which consists of two parts: a subsystem for location sensing that acquires information about the location of users as well as spatial proximities between them, and one for the actual context-aware application, which provides services for group interaction.

1. INTRODUCTION Today"s computing environments are characterized by an increasing number of powerful, wirelessly connected mobile devices. Users can move throughout an environment while carrying their computers with them and having remote access to information and services, anytime and anywhere. New situations appear, where the user"s context - for example his current location or nearby people - is more dynamic; computation does not occur at a single location and in a single context any longer, but comprises a multitude of situations and locations. This development leads to a new class of applications, which are aware of the context in which they run in and thus bringing virtual and real worlds together. Motivated by this and the fact, that only a few studies have been done for supporting group communication in such computing environments [12], we have developed a system, which we refer to as Group Interaction Support System (GISS). It supports group interaction in mobile distributed computing environments in a way that group members need not to at the same place any longer in order to interact with each other or just to be aware of the others situation. In the following subchapters, we will give a short overview on context aware computing and motivate its benefits for supporting group interaction. A software framework for developing contextsensitive applications is presented, which serves as middleware for GISS. Chapter 2 presents the architecture of GISS, and chapter 3 and 4 discuss the location sensing and group interaction concepts of GISS in more detail. Chapter 5 gives a final summary of our work. 1.1 What is Context Computing? According to Merriam-Webster"s Online Dictionary1 , context is defined as the interrelated conditions in which something exists or occurs. Because this definition is very general, many approaches have been made to define the notion of context with respect to computing environments. Most definitions of context are done by enumerating examples or by choosing synonyms for context. The term context-aware has been introduced first in [10] where context is referred to as location, identities of nearby people and objects, and changes to those objects. In [2], context is also defined by an enumeration of examples, namely location, identities of the people around the user, the time of the day, season, temperature etc. [9] defines context as the user"s location, environment, identity and time. Here we conform to a widely accepted and more formal definition, which defines context as any information than can be used to characterize the situation of an entity. An entity is a person, place, or object that is considered relevant to the interaction between a user and an application, including the user and applications themselves. [4] [4] identifies four primary types of context information (sometimes referred to as context dimensions), that are - with respect to characterizing the situation of an entity - more important than others. These are location, identity, time and activity, which can also be used to derive other sources of contextual information (secondary context types). For example, if we know a person"s identity, we can easily derive related information about this person from several data sources (e.g. day of birth or e-mail address). According to this definition, [4] defines a system to be contextaware if it uses context to provide relevant information and/or services to the user, where relevancy depends on the user"s task. [4] also gives a classification of features for context-aware applications, which comprises presentation of information and services to a user, automatic execution of a service and tagging of context to information for later retrieval. Figure 1. Layers of a context-aware system Context computing is based on two major issues, namely identifying relevant context (identity, location, time, activity) and using obtained context (automatic execution, presentation, tagging). In order to do this, there are a few layers between (see Figure 1). First, the obtained low-level context information has to be transformed, aggregated and interpreted (context transformation) and represented in an abstract context world model (context representation), either centralized or decentralized. Finally, the stored context information is used to trigger certain context events (context triggering). [7] 1.2 Group Interaction in Context After these abstract and formal definitions about what context and context computing is, we will now focus on the main goal of this work, namely how the interaction of mobile group members can be supported by using context information. In [6] we have identified organizational systems to be crucial for supporting mobile groups (see Figure 2). First, there has to be an Information and Knowledge Management System, which is capable of supporting a team with its information processing- and knowledge gathering needs. The next part is the Awareness System, which is dedicated to the perceptualisation of the effects of team activity. It does this by communicating work context, agenda and workspace information to the users. The Interaction Systems provide support for the communication among team members, either synchronous or asynchronous, and for the shared access to artefacts, such as documents. Mobility Systems deploy mechanisms to enable any-place access to team memory as well as the capturing and delivery of awareness information from and to any places. Finally yet importantly, the organisational innovation system integrates aspects of the team itself, like roles, leadership and shared facilities. With respect to these five aspects of team support, we focus on interaction and partly cover mobility- and awareness-support. Group interaction includes all means that enable group members to communicate freely with all the other members. At this point, the question how context information can be used for supporting group interaction comes up. We believe that information about the current situation of a person provides a surplus value to existing group interaction systems. Context information facilitates group interaction by allowing each member to be aware of the availability status or the current location of each other group member, which again makes it possible to form groups dynamically, to place virtual post-its in the real world or to determine which people are around. Figure 2. Support for Mobile Groups [6] Most of today"s context-aware applications use location and time only, and location is referred to as a crucial type of context information [3]. We also see the importance of location information in mobile and ubiquitous environments, wherefore a main focus of our work is on the utilization of location information and information about users in spatial proximity. Nevertheless, we believe that location, as the only used type of context information, is not sufficient to support group interaction, wherefore we also take advantage of the other three primary types, namely identity, time and activity. This provides a comprehensive description of a user"s current situation and thus enabling numerous means for supporting group interaction, which are described in detail in chapter 4.4. When we look at the types of context information stated above, we can see that all of them are single user-centred, taking into account only the context of the user itself. We believe, that for the support of group interaction, the status of the group itself has also be taken into account. Therefore, we have added a fifth contextdimension group-context, which comprises more than the sum of the individual member"s contexts. Group context includes any information about the situation of a whole group, for example how many members a group currently has or if a certain group meets right now. 1.3 Context Middleware The Group Interaction Support System (GISS) uses the softwareframework introduced in [1], which serves as a middleware for developing context-sensitive applications. This so-called Context Framework is based on a distributed communication architecture and it supports different kinds of transport protocols and message coding mechanisms. 89 A main feature of the framework is the abstraction of context information retrieval via various sensors and its delivery to a level where no difference appears, for the application designer, between these different kinds of context retrieval mechanisms; the information retrieval is hidden from the application developer. This is achieved by so-called entities, which describe objectse.g. a human user - that are important for a certain context scenario. Entities express their functionality by the use of so-called attributes, which can be loaded into the entity. These attributes are complex pieces of software, which are implemented as Java classes. Typical attributes are encapsulations of sensors, but they can also be used to implement context services, for example to notify other entities about location changes of users. Each entity can contain a collection of such attributes, where an entity itself is an attribute. The initial set of attributes an entity contains can change dynamically at runtime, if an entity loads or unloads attributes from the local storage or over the network. In order to load and deploy new attributes, an entity has to reference a class loader and a transport and lookup layer, which manages the lookup mechanism for discovering other entities and the transport. XML configuration files specify which initial set of entities should be loaded and which attributes these entities own. The communication between entities and attributes is based on context events. Each attribute is able to trigger events, which are addressed to other attributes and entities respectively, independently on which physical computer they are running. Among other things, and event contains the name of the event and a list of parameters delivering information about the event itself. Related with this event-based architecture is the use of ECA (Event-Condition-Action)-rules for defining the behaviour of the context system. Therefore, every entity has a rule-interpreter, which catches triggered events, checks conditions associated with them and causes certain actions. These rules are referenced by the entity"s XML configuration. A rule itself is even able to trigger the insertion of new rules or the unloading of existing rules at runtime in order to change the behaviour of the context system dynamically. To sum up, the context framework provides a flexible, distributed architecture for hiding low-level sensor data from high-level applications and it hides external communication details from the application developer. Furthermore, it is able to adapt its behaviour dynamically by loading attributes, entities or ECArules at runtime. 2. ARCHITECTURE OVERVIEW As GISS uses the Context Framework described in chapter 1.3 as middleware, every user is represented by an entity, as well as the central server, which is responsible for context transformation, context representation and context triggering (cf. Figure 1). A main part of our work is about the automated acquisition of position information and its sensor-independent provision at application level. We do not only sense the current location of users, but also determine spatial proximities between them. Developing the architecture, we focused on keeping the client as simple as possible and reducing the communication between client and server to a minimum. Each client may have various location and/or proximity sensors attached, which are encapsulated by respective Context Framework-attributes (Sensor Encapsulation). These attributes are responsible for integrating native sensor-implementations into the Context Framework and sending sensor-dependent position information to the server. We consider it very important to support different types of sensors even at the same time, in order to improve location accuracy on the one hand, while providing a pervasive location-sensing environment with seamless transition between different location sensing techniques on the other hand. All location- and proximity-sensors supported are represented by server-side context-attributes, which correspond to the client-side sensor encapsulation-attributes and abstract the sensor-dependent position information received from all users via the wireless network (sensor abstraction). This requires a context repository, where the mapping of diverse physical positions to standardized locations is stored. The standardized location- and proximity-information of each user is then passed to the so-called Sensor Fusion-attributes, one for symbolic locations and a second one for spatial proximities. Their job is to merge location- and proximityinformation of clients, respectively, which is described in detail in Chapter 3.3. Every time the symbolic location of a user or the spatial proximity between two users changes, the Sensor Fusion-attributes notify the GISS Core-attribute, which controls the application. Because of the abstraction of sensor-dependent position information, the system can easily be extended by additional sensors, just by implementing the (typically two) attributes for encapsulating sensors (some sensors may not need a client-side part), abstracting physical positions and observing the interface to GISS Core. Figure 3. Architecture of the Group Interaction Support System (GISS) The GISS Core-attribute is the central coordinator of the application as it shows to the user. It not only serves as an interface to the location-sensing subsystem, but also collects further context information in other dimensions (time, identity or activity). 90 Every time a change in the context of one or more users is detected, GISS Core evaluates the effect of these changes on the user, on the groups he belongs to and on the other members of these groups. Whenever necessary, events are thrown to the affected clients to trigger context-aware activities, like changing the presentation of awareness information or the execution of services. The client-side part of the application is kept as simple as possible. Furthermore, modular design was not only an issue on the sensor side but also when designing the user interface architecture. Thus, the complete user interface can be easily exchanged, if all of the defined events are taken into account and understood by the new interface-attribute. The currently implemented user interface is split up in two parts, which are also represented by two attributes. The central attribute on client-side is the so-called Instant Messenger Encapsulation, which on the one hand interacts with the server through events and on the other hand serves as a proxy for the external application the user interface is built on. As external application, we use an existing open source instant messenger - the ICQ2 -compliant Simple Instant Messenger (SIM)3 . We have chosen and instant messenger as front-end because it provides a well-known interface for most users and facilitates a seamless integration of group interaction support, thus increasing acceptance and ease of use. As the basic functionality of the instant messenger - to serve as a client in an instant messenger network - remains fully functional, our application is able to use the features already provided by the messenger. For example, the contexts activity and identity are derived from the messenger network as it is described later. The Instant Messenger Encapsulation is also responsible for supporting group communication. Through the interface of the messenger, it provides means of synchronous and asynchronous communication as well as a context-aware reminder system and tools for managing groups and the own availability status. The second part of the user interface is a visualisation of the user"s locations, which is implemented in the attribute Viewer. The current implementation provides a two-dimensional map of the campus, but it can easily be replaced by other visualisations, a three-dimensional VRML-model for example. Furthermore, this visualisation is used to show the artefacts for asynchronous communication. Based on a floor plan-view of the geographical area the user currently resides in, it gives a quick overview of which people are nearby, their state and provides means to interact with them. In the following chapters 3 and 4, we describe the location sensing-backend and the application front-end for supporting group interaction in more detail. 3. LOCATION SENSING In the following chapter, we will introduce a location model, which is used for representing locations; afterwards, we will describe the integration of location- and proximity-sensors in 2 http://www.icq.com/ 3 http://sim-icq.sourceforge.net more detail. Finally, we will have a closer look on the fusion of location- and proximity-information, acquired by various sensors. 3.1 Location Model A location model (i.e. a context representation for the contextinformation location) is needed to represent the locations of users, in order to be able to facilitate location-related queries like given a location, return a list of all the objects there or given an object, return its current location. In general, there are two approaches [3,5]: symbolic models, which represent location as abstract symbols, and a geometric model, which represent location as coordinates. We have chosen a symbolic location model, which refers to locations as abstract symbols like Room P111 or Physics Building, because we do not require geometric location data. Instead, abstract symbols are more convenient for human interaction at application level. Furthermore, we use a symbolic location containment hierarchy similar to the one introduced in [11], which consists of top-level regions, which contain buildings, which contain floors, and the floors again contain rooms. We also distinguish four types, namely region (e.g. a whole campus), section (e.g. a building or an outdoor section), level (e.g. a certain floor in a building) and area (e.g. a certain room). We introduce a fifth type of location, which we refer to as semantic. These socalled semantic locations can appear at any level in the hierarchy and they can be nested, but they do not necessarily have a geographic representation. Examples for such semantic locations are tagged objects within a room (e.g. a desk and a printer on this desk) or the name of a department, which contains certain rooms. Figure 4. Symbolic Location Containment Hierarchy The hierarchy of symbolic locations as well as the type of each position is stored in the context repository. 3.2 Sensors Our architecture supports two different kinds of sensors: location sensors, which acquire location information, and proximity sensors, which detect spatial proximities between users. As described above, each sensor has a server- and in most cases a corresponding client-side-implementation, too. While the clientattributes (Sensor Abstraction) are responsible for acquiring low-level sensor-data and transmitting it to the server, the corresponding Sensor Encapsulation-attributes transform them into a uniform and sensor-independent format, namely symbolic locations and IDs of users in spatial proximity, respectively. 91 Afterwards, the respective attribute Sensor Fusion is being triggered with this sensor-independent information of a certain user, detected by a particular sensor. Such notifications are performed every time the sensor acquired new information. Accordingly, Sensor Abstraction-attributes are responsible to detect when a certain sensor is no longer available on the client side (e.g. if it has been unplugged by the user) or when position respectively proximity could not be determined any longer (e.g. RFID reader cannot detect tags) and notify the corresponding sensor fusion about this. 3.2.1 Location Sensors In order to sense physical positions, the Sensor Encapsulationattributes asynchronously transmit sensor-dependent position information to the server. The corresponding location Sensor Abstraction-attributes collect these physical positions delivered by the sensors of all users, and perform a repository-lookup in order to get the associated symbolic location. This requires certain tables for each sensor, which map physical positions to symbolic locations. One physical position may have multiple symbolic locations at different accuracy-levels in the location hierarchy assigned to, for example if a sensor covers several rooms. If such a mapping could be found, an event is thrown in order to notify the attribute Location Sensor Fusion about the symbolic locations a certain sensor of a particular user determined. We have prototypically implemented three kinds of location sensors, which are based on WLAN (IEEE 802.11), Bluetooth and RFID (Radio Frequency Identification). We have chosen these three completely different sensors because of their differences concerning accuracy, coverage and administrative effort, in order to evaluate the flexibility of our system (see Table 1). The most accurate one is an RFID sensor, which is based on an active RFID-reader. As soon as the reader is plugged into the client, it scans for active RFID tags in range and transmits their serial numbers to the server, where they are mapped to symbolic locations. We also take into account RSSI (Radio Signal Strength Information), which provides position accuracy of few centimetres and thus enables us to determine which RFID-tag is nearest. Due to this high accuracy, RFID is used for locating users within rooms. The administration is quite simple; once a new RFID tag is placed, its serial number simply has to be assigned to a single symbolic location. A drawback is the poor availability, which can be traced back to the fact that RFID readers are still very expensive. The second one is an 802.11 WLAN sensor. Therefore, we integrated a purely software-based, commercial WLAN positioning system for tracking clients on the university campuswide WLAN infrastructure. The reached position accuracy is in the range of few meters and thus is suitable for location sensing at the granularity of rooms. A big disadvantage is that a map of the whole area has to be calibrated with measuring points at a distance of 5 meters each. Because most mobile computers are equipped with WLAN technology and the positioning-system is a software-only solution, nearly everyone is able to use this kind of sensor. Finally, we have implemented a Bluetooth sensor, which detects Bluetooth tags (i.e. Bluetooth-modules with known position) in range and transmits them to the server that maps to symbolic locations. Because of the fact that we do not use signal strengthinformation in the current implementation, the accuracy is above 10 meters and therefore a single Bluetooth MAC address is associated with several symbolic locations, according to the physical locations such a Bluetooth module covers. This leads to the disadvantage that the range of each Bluetooth-tag has to be determined and mapped to symbolic locations within this range. Table 1. Comparison of implemented sensors Sensor Accuracy Coverage Administration RFID < 10 cm poor easy WLAN 1-4 m very well very timeconsuming Bluetooth ~ 10 m well time-consuming 3.2.2 Proximity Sensors Any sensor that is able to detect whether two users are in spatial proximity is referred to as proximity sensor. Similar to the location sensors, the Proximity Sensor Abstraction-attributes collect physical proximity information of all users and transform them to mappings of user-IDs. We have implemented two types of proximity-sensors, which are based on Bluetooth on the one hand and on fused symbolic locations (see chapter 3.3.1) on the other hand. The Bluetooth-implementation goes along with the implementation of the Bluetooth-based location sensor. The already determined Bluetooth MAC addresses in range of a certain client are being compared with those of all other clients, and each time the attribute Bluetooth Sensor Abstraction detects congruence, it notifies the proximity sensor fusion about this. The second sensor is based on symbolic locations processed by Location Sensor Fusion, wherefore it does not need a client-side implementation. Each time the fused symbolic location of a certain user changes, it checks whether he is at the same symbolic location like another user and again notifies the proximity sensor fusion about the proximity between these two users. The range can be restricted to any level of the location containment hierarchy, for example to room granularity. A currently unresolved issue is the incomparable granularity of different proximity sensors. For example, the symbolic locations at same level in the location hierarchy mostly do not cover the same geographic area. 3.3 Sensor Fusion Core of the location sensing subsystem is the sensor fusion. It merges data of various sensors, while coping with differences concerning accuracy, coverage and sample-rate. According to the two kinds of sensors described in chapter 3.2, we distinguish between fusion of location sensors on the one hand, and fusion of proximity sensors on the other hand. The fusion of symbolic locations as well as the fusion of spatial proximities operates on standardized information (cf. Figure 3). This has the advantage, that additional position- and proximitysensors can be added easily or the fusion algorithms can be replaced by ones that are more sophisticated. 92 Fusion is performed for each user separately and takes into account the measurements at a single point in time only (i.e. no history information is used for determining the current location of a certain user). The algorithm collects all events thrown by the Sensor Abstraction-attributes, performs fusion and triggers the GISS Core-attribute if the symbolic location of a certain user or the spatial proximity between users changed. An important feature is the persistent storage of location- and proximity-history in a database in order to allow future retrieval. This enables applications to visualize the movement of users for example. 3.3.1 Location Sensor Fusion Goal of the fusion of location information is to improve precision and accuracy by merging the set of symbolic locations supplied by various location sensors, in order to reduce the number of these locations to a minimum, ideally to a single symbolic location per user. This is quite difficult, because different sensors may differ in accuracy and sample rate as well. The Location Sensor Fusion-attribute is triggered by events, which are thrown by the Location Sensor Abstractionattributes. These events contain information about the identity of the user concerned, his current location and the sensor by which the location has been determined. If the attribute Location Sensor Fusion receives such an event, it checks if the amount of symbolic locations of the user concerned has changed (compared with the last event). If this is the case, it notifies the GISS Core-attribute about all symbolic locations this user is currently associated with. However, this information is not very useful on its own if a certain user is associated with several locations. As described in chapter 3.2.1, a single location sensor may deliver multiple symbolic locations. Moreover, a certain user may have several location sensors, which supply symbolic locations differing in accuracy (i.e. different levels in the location containment hierarchy). To cope with this challenge, we implemented a fusion algorithm in order to reduce the number of symbolic locations to a minimum (ideally to a single location). In a first step, each symbolic location is associated with its number of occurrences. A symbolic location may occur several times if it is referred to by more than one sensor or if a single sensor detects multiple tags, which again refer to several locations. Furthermore, this number is added to the previously calculated number of occurrences of each symbolic location, which is a child-location of the considered one in the location containment hierarchy. For example, if - in Figure 4 - room2 occurs two times and desk occurs a single time, the value 2 of room2 is added to the value 1 of desk, whereby desk finally gets the value 3. In a final step, only those symbolic locations are left which are assigned with the highest number of occurrences. A further reduction can be achieved by assigning priorities to sensors (based on accuracy and confidence) and cumulating these priorities for each symbolic location instead of just counting the number of occurrences. If the remaining fused locations have changed (i.e. if they differ from the fused locations the considered user is currently associated with), they are provided with the current timestamp, written to the database and the GISS-attribute is notified about where the user is probably located. Finally, the most accurate, common location in the location hierarchy is calculated (i.e. the least upper bound of these symbolic locations) in order to get a single symbolic location. If it changes, the GISS Core-attribute is triggered again. 3.3.2 Proximity Sensor Fusion Proximity sensor fusion is much simpler than the fusion of symbolic locations. The corresponding proximity sensor fusionattribute is triggered by events, which are thrown by the Proximity Sensor Abstraction-attributes. These special events contain information about the identity of the two users concerned, if they are currently in spatial proximity or if proximity no longer persists, and by which proximity-sensor this has been detected. If the sensor fusion-attribute is notified by a certain Proximity Sensor Abstraction-attribute about an existing spatial proximity, it first checks if these two users are already known to be in proximity (detected either by another user or by another proximity-sensor of the user, which caused the event). If not, this change in proximity is written to the context repository with current timestamp. Similarly, if the attribute Proximity Fusion is notified about an ended proximity, it checks if the users are still known to be in proximity, and writes this change to the repository if not. Finally, if spatial proximity between the two users actually changed, an event is thrown to notify the GISS Core-attribute about this. 4. CONTEXTSENSITIVE INTERACTION 4.1 Overview In most of today"s systems supporting interaction in groups, the provided means lack any awareness of the user"s current context, thus being unable to adapt to his needs. In our approach, we use context information to enhance interaction and provide further services, which offer new possibilities to the user. Furthermore, we believe that interaction in groups also has to take into account the current context of the group itself and not only the context of individual group members. For this reason, we also retrieve information about the group"s current context, derived from the contexts of the group members together with some sort of meta-information (see chapter 4.3). The sources of context used for our application correspond with the four primary context types given in chapter 1.1 - identity (I), location (L), time (T) and activity (A). As stated before, we also take into account the context of the group the user is interaction with, so that we could add a fifth type of context informationgroup awareness (G) - to the classification. Using this context information, we can trigger context-aware activities in all of the three categories described in chapter 1.1 - presentation of information (P), automatic execution of services (A) and tagging of context to information for later retrieval (T). Table 2 gives an overview of activities we have already implemented; they are described comprehensively in chapter 4.4. The table also shows which types of context information are used for each activity and the category the activity could be classified in. 93 Table 2. Classification of implemented context-aware activities Service L T I A G P A T Location Visualisation X X X Group Building Support X X X X Support for Synchronous Communication X X X X Support for Asynchronous Communication X X X X X X X Availability Management X X X Task Management Support X X X X Meeting Support X X X X X X Reasons for implementing these very features are to take advantage of all four types of context information in order to support group interaction by utilizing a comprehensive knowledge about the situation a single user or a whole group is in. A critical issue for the user acceptance of such a system is the usability of its interface. We have evaluated several ways of presenting context-aware means of interaction to the user, until we came to the solution we use right now. Although we think that the user interface that has been implemented now offers the best trade-off between seamless integration of features and ease of use, it would be no problem to extend the architecture with other user interfaces, even on different platforms. The chosen solution is based on an existing instant messenger, which offers several possibilities to integrate our system (see chapter 4.2). The biggest advantage of this approach is that the user is confronted with a graphical user interface he is already used to in most cases. Furthermore, our system uses an instant messenger account as an identifier, so that the user does not have to register a further account anywhere else (for example, the user can use his already existing ICQ2 -account). 4.2 Instant Messenger Integration Our system is based upon an existing instant messenger, the socalled Simple Instant Messenger (SIM)3 . The implementation of this messenger is carried out as a project at Sourceforge4 . SIM supports multiple messenger protocols such as AIM5 , ICQ2 and MSN6 . It also supports connections to multiple accounts at the same time. Furthermore, full support for SMS-notification (where provided from the used protocol) is given. SIM is based on a plug-in concept. All protocols as well as parts of the user-interface are implemented as plug-ins. Its architecture is also used to extend the application"s abilities to communicate with external applications. For this purpose, a remote control plug-in is provided, by which SIM can be controlled from external applications via socket connection. This remote control interface is extensively used by GISS for retrieving the contact list, setting the user"s availability-state or sending messages. The 4 http://sourceforge.net/ 5 http://www.aim.com/ 6 http://messenger.msn.com/ functionality of the plug-in was extended in several ways, for example to accept messages for an account (as if they would have been sent via the messenger network). The messenger, more exactly the contact list (i.e. a list of profiles of all people registered with the instant messenger, which is visualized by listing their names as it can be seen in Figure 5), is also used to display locations of other members of the groups a user belongs to. This provides location awareness without taking too much space or requesting the user"s full attention. A more comprehensive description of these features is given in chapter 4.4. 4.3 Sources of Context Information While the location-context of a user is obtained from our location sensing subsystem described in chapter 3, we consider further types of context than location relevant for the support of group interaction, too. Local time as a very important context dimension can be easily retrieved from the real time clock of the user"s system. Besides location and time, we also use context information of user"s activity and identity, where we exploit the functionality provided by the underlying instant messenger system. Identity (or more exactly, the mapping of IDs to names as well as additional information from the user"s profile) can be distilled out of the contents of the user"s contact list. Information about the activity or a certain user is only available in a very restricted area, namely the activity at the computer itself. Other activities like making a phone call or something similar, cannot be recognized with the current implementation of the activity sensor. The only context-information used is the instant messenger"s availability state, thus only providing a very coarse classification of the user"s activity (online, offline, away, busy etc.). Although this may not seem to be very much information, it is surely relevant and can be used to improve or even enable several services. Having collected the context information from all available users, it is now possible to distil some information about the context of a certain group. Information about the context of a group includes how many members the group currently has, if the group meets right now, which members are participating at a meeting, how many members have read which of the available posts from other team members and so on. Therefore, some additional information like a list of members for each group is needed. These lists can be assembled manually (by users joining and leaving groups) or retrieved automatically. The context of a group is secondary context and is aggregated from the available contexts of the group members. Every time the context of a single group member changes, the context of the whole group is changing and has to be recalculated. With knowledge about a user"s context and the context of the groups he belongs to, we can provide several context-aware services to the user, which enhance his interaction abilities. A brief description of these services is given in chapter 4.4. 94 4.4 Group Interaction Support 4.4.1 Visualisation of Location Information An important feature is the visualisation of location information, thus allowing users to be aware of the location of other users and members of groups he joined, respectively. As already described in chapter 2, we use two different forms of visualisation. The maybe more important one is to display location information in the contact list of the instant messenger, right beside the name, thus being always visible while not drawing the user"s attention on it (compared with a twodimensional view for example, which requires a own window for displaying a map of the environment). Due to the restricted space in the contact list, it has been necessary to implement some sort of level-of-detail concept. As we use a hierarchical location model, we are able to determine the most accurate common location of two users. In the contact list, the current symbolic location one level below the previously calculated common location is then displayed. If, for example, user A currently resides in room P121 at the first floor of a building and user B, which has to be displayed in the contact list of user A, is in room P304 at the third floor, the most accurate common location of these two users is the building they are in. For that reason, the floor (i.e. one level beyond the common location, namely the building) of user B is displayed in the contact list of user A. If both people reside on the same floor or even in the same room, the room would be taken. Figure 5 shows a screenshot of the Simple Instant Messenger3 , where the current location of those people, whose location is known by GISS, is displayed in brackets right beside their name. On top of the image, the heightened, integrated GISS-toolbar is shown, which currently contains the following, implemented functionality (from left to right): Asynchronous communication for groups (see chapter 4.4.4), context-aware reminders (see chapter 4.4.6), two-dimensional visualisation of locationinformation, forming and managing groups (see chapter 4.4.2), context-aware availability-management (see chapter 4.4.5) and finally a button for terminating GISS. Figure 5. GISS integration in Simple Instant Messenger3 As displaying just this short form of location may not be enough for the user, because he may want to see the most accurate position available, a fully qualified position is shown if a name in the contact-list is clicked (e.g. in the form of desk@room2@department1@1stfloor@building 1@campus). The second possible form of visualisation is a graphical one. We have evaluated a three-dimensional view, which was based on a VRML model of the respective area (cf. Figure 6). Due to lacks in navigational and usability issues, we decided to use a twodimensional view of the floor (it is referred to as level in the location hierarchy, cf. Figure 4). Other levels of granularity like section (e.g. building) and region (e.g. campus) are also provided. In this floor-plan-based view, the current locations are shown in the manner of ICQ2 contacts, which are placed at the currently sensed location of the respective person. The availability-status of a user, for example away if he is not on the computer right now, or busy if he does not want to be disturbed, is visualized by colour-coding the ICQ2 -flower left beside the name. Furthermore, the floor-plan-view shows so-called the virtual post-its, which are virtual counterparts of real-life post-its and serve as our means of asynchronous communication (more about virtual post-its can be found in chapter 4.4.4). Figure 6. 3D-view of the floor (VRML) Figure 7 shows the two-dimensional map of a certain floor, where several users are currently located (visualized by their name and the flower left beside). The location of the client, on which the map is displayed, is visualized by a green circle. Down to the right, two virtual post-its can be seen. Figure 7. 2D view of the floor Another feature of the 2D-view is the visualisation of locationhistory of users. As we store the complete history of a user"s locations together with a timestamp, we are able to provide information about the locations he has been back in time. When the mouse is moved over the name of a certain user in the 2Dview, footprints of a user, placed at the locations he has been, are faded out the stronger, the older the location information is. 95 4.4.2 Forming and Managing Groups To support interaction in groups, it is first necessary to form groups. As groups can have different purposes, we distinguish two types of groups. So-called static groups are groups, which are built up manually by people joining and leaving them. Static groups can be further divided into two subtypes. In open static groups, everybody can join and leave anytime, useful for example to form a group of lecture attendees of some sort of interest group. Closed static groups have an owner, who decides, which persons are allowed to join, although everybody could leave again at any time. Closed groups enable users for example to create a group of their friends, thus being able to communicate with them easily. In contrast to that, we also support the creation of dynamic groups. They are formed among persons, who are at the same location at the same time. The creation of dynamic groups is only performed at locations, where it makes sense to form groups, for example in lecture halls or meeting rooms, but not on corridors or outdoor. It would also be not very meaningful to form a group only of the people residing in the left front sector of a hall; instead, the complete hall should be considered. For these reasons, all the defined locations in the hierarchy are tagged, whether they allow the formation of groups or not. Dynamic groups are also not only formed granularity of rooms, but also on higher levels in the hierarchy, for example with the people currently residing in the area of a department. As the members of dynamic groups constantly change, it is possible to create an open static group out of them. 4.4.3 Synchronous Communication for Groups The most important form of synchronous communication on computers today is instant messaging; some people even see instant messaging to be the real killer application on the Internet. This has also motivated the decision to build GISS upon an instant messaging system. In today"s messenger systems, peer-to-peer-communication is extensively supported. However, when it comes to communication in groups, the support is rather poor most of the time. Often, only sending a message to multiple recipients is supported, lacking means to take into account the current state of the recipients. Furthermore, groups can only be formed of members in one"s contact list, thus being not able to send messages to a group, where not all of its members are known (which may be the case in settings, where the participants of a lecture form a group). Our approach does not have the mentioned restrictions. We introduce group-entries in the user"s contact list; enable him or his to send messages to this group easily, without knowing who exactly is currently a member of this group. Furthermore, group messages are only delivered to persons, who are currently not busy, thus preventing a disturbance by a message, which is possibly unimportant for the user. These features cannot be carried out in the messenger network itself, so whenever a message to a group account is sent, we intercept it and route it through our system to all the recipients, which are available at a certain time. Communication via a group account is also stored centrally, enabling people to query missed messages or simply viewing the message history. 4.4.4 Asynchronous Communication for Groups Asynchronous communication in groups is not a new idea. The goal of this approach is not to reinvent the wheel, as email is maybe the most widely used form of asynchronous communication on computers and is broadly accepted and standardized. In out work, we aim at the combination of asynchronous communication with location awareness. For this reason, we introduce the concept of so-called virtual postits (cp. [13]), which are messages that are bound to physical locations. These virtual post-its could be either visible for all users that are passing by or they can be restricted to be visible for certain groups of people only. Moreover, a virtual post-it can also have an expiry date after which it is dropped and not displayed anymore. Virtual post-its can also be commented by others, thus providing some from of forum-like interaction, where each post-it forms a thread. Virtual post-its are displayed automatically, whenever a user (available) passes by the first time. Afterwards, post-its can be accessed via the 2D-viewer, where all visible post-its are shown. All readers of a post-it are logged and displayed when viewing it, providing some sort of awareness about the group members" activities in the past. 4.4.5 Context-aware Availability Management Instant messengers in general provide some kind of availability information about a user. Although this information can be only defined in a very coarse granularity, we have decided to use these means of gathering activity context, because the introduction of an additional one would strongly decrease the usability of the system. To support the user managing his availability, we provide an interface that lets the user define rules to adapt his availability to the current context. These rules follow the form on event (E) if condition (C) then action (A), which is directly supported by the ECA-rules of the Context Framework described in chapter 1.3. The testing of conditions is periodically triggered by throwing events (whenever the context of a user changes). The condition itself is defined by the user, who can demand the change of his availability status as the action in the rule. As a condition, the user can define his location, a certain time (also triggering daily, every week or every month) or any logical combination of these criteria. 4.4.6 Context-Aware Reminders Reminders [14] are used to give the user the opportunity of defining tasks and being reminded of those, when certain criteria are fulfilled. Thus, a reminder can be seen as a post-it to oneself, which is only visible in certain cases. Reminders can be bound to a certain place or time, but also to spatial proximity of users or groups. These criteria can be combined with Boolean operators, thus providing a powerful means to remind the user of tasks that he wants to carry out when a certain context occurs. A reminder will only pop up the first time the actual context meets the defined criterion. On showing up the reminder, the user has the chance to resubmit it to be reminded again, for example five minutes later or the next time a certain user is in spatial proximity. 96 4.4.7 Context-Aware Recognition and Notification of Group Meetings With the available context information, we try to recognize meetings of a group. The determination of the criteria, when the system recognizes a group having a meeting, is part of the ongoing work. In a first approach, we use the location- and activity-context of the group members to determine a meeting. Whenever more than 50 % of the members of a group are available at a location, where a meeting is considered to make sense (e.g. not on a corridor), a meeting minutes post-it is created at this location and all absent group members are notified of the meeting and the location it takes place. During the meeting, the comment-feature of virtual post-its provides a means to take notes for all of the participants. When members are joining or leaving the meeting, this is automatically added as a note to the list of comments. Like the recognition of the beginning of a meeting, the recognition of its end is still part of ongoing work. If the end of the meeting is recognized, all group members get the complete list of comments as a meeting protocol at the end of the meeting. 5. CONCLUSIONS This paper discussed the potentials of support for group interaction by using context information. First, we introduced the notions of context and context computing and motivated their value for supporting group interaction. An architecture is presented to support context-aware group interaction in mobile, distributed environments. It is built upon a flexible and extensible framework, thus enabling an easy adoption to available context sources (e.g. by adding additional sensors) as well as the required form of representation. We have prototypically developed a set of services, which enhance group interaction by taking into account the current context of the users as well as the context of groups itself. Important features are dynamic formation of groups, visualization of location on a two-dimensional map as well as unobtrusively integrated in an instant-messenger, asynchronous communication by virtual post-its, which are bound to certain locations, and a context-aware availability-management, which adapts the availability-status of a user to his current situation. To provide location information, we have implemented a subsystem for automated acquisition of location- and proximityinformation provided by various sensors, which provides a technology-independent presentation of locations and spatial proximities between users and merges this information using sensor-independent fusion algorithms. A history of locations as well as of spatial proximities is stored in a database, thus enabling context history-based services. 6. REFERENCES [1] Beer, W., Christian, V., Ferscha, A., Mehrmann, L. Modeling Context-aware Behavior by Interpreted ECA Rules. In Proceedings of the International Conference on Parallel and Distributed Computing (EUROPAR"03). (Klagenfurt, Austria, August 26-29, 2003). Springer Verlag, LNCS 2790, 1064-1073. [2] Brown, P.J., Bovey, J.D., Chen X. Context-Aware Applications: From the Laboratory to the Marketplace. IEEE Personal Communications, 4(5) (1997), 58-64. [3] Chen, H., Kotz, D. A Survey of Context-Aware Mobile Computing Research. Technical Report TR2000-381, Computer Science Department, Dartmouth College, Hanover, New Hampshire, November 2000. [4] Dey, A. Providing Architectural Support for Building Context-Aware Applications. Ph.D. Thesis, Department of Computer Science, Georgia Institute of Technology, Atlanta, November 2000. [5] Svetlana Domnitcheva. Location Modeling: State of the Art and Challenges. In Proceedings of the Workshop on Location Modeling for Ubiquitous Computing. (Atlanta, Georgia, United States, September 30, 2001). 13-19. [6] Ferscha, A. Workspace Awareness in Mobile Virtual Teams. In Proceedings of the IEEE 9th International Workshop on Enabling Technologies: Infrastructure for Collaborative Enterprises (WETICE"00). (Gaithersburg, Maryland, March 14-16, 2000). IEEE Computer Society Press, 272-277. [7] Ferscha, A. Coordination in Pervasive Computing Environments. In Proceedings of the Twelfth International IEEE Workshop on Enabling Technologies: Infrastructure for Collaborative Enterprises (WETICE-2003). (June 9-11, 2003). IEEE Computer Society Press, 3-9. [8] Leonhard, U. Supporting Location Awareness in Open Distributed Systems. Ph.D. Thesis, Department of Computing, Imperial College, London, May 1998. [9] Ryan, N., Pascoe, J., Morse, D. Enhanced Reality Fieldwork: the Context-Aware Archaeological Assistant. Gaffney, V., Van Leusen, M., Exxon, S. (eds.) Computer Applications in Archaeology (1997) [10] Schilit, B.N., Theimer, M. Disseminating Active Map Information to Mobile Hosts. IEEE Network, 8(5) (1994), 22-32. [11] Schilit, B.N. A System Architecture for Context-Aware Mobile Computing. Ph.D. Thesis, Columbia University, Department of Computer Science, May 1995. [12] Wang, B., Bodily, J., Gupta, S.K.S. Supporting Persistent Social Groups in Ubiquitous Computing Environments Using Context-Aware Ephemeral Group Service. In Proceedings of the Second IEEE International Conference on Pervasive Computing and Communications (PerCom"04). (March 14-17, 2004). IEEE Computer Society Press, 287-296. [13] Pascoe, J. The Stick-e Note Architecture: Extending the Interface Beyond the User. Proceedings of the 2nd International Conference of Intelligent User Interfaces (IUI"97). (Orlando, USA, 1997), 261-264. [14] Dey, A., Abowd, G. CybreMinder: A Context-Aware System for Supporting Re-minders. Proceedings of the 2nd International Symposium on Handheld and Ubiquitous Computing (HUC"00). (Bristol, UK, 2000), 172-186. 97

software framework;xml configuration file;context awareness;location sense;event-condition-action;fifth contextdimension group-context;contextaware;group interaction;sensor fusion;mobility system

train_C-56

A Hierarchical Process Execution Support for Grid Computing

Grid is an emerging infrastructure used to share resources among virtual organizations in a seamless manner and to provide breakthrough computing power at low cost. Nowadays there are dozens of academic and commercial products that allow execution of isolated tasks on grids, but few products support the enactment of long-running processes in a distributed fashion. In order to address such subject, this paper presents a programming model and an infrastructure that hierarchically schedules process activities using available nodes in a wide grid environment. Their advantages are automatic and structured distribution of activities and easy process monitoring and steering.

1. INTRODUCTION Grid computing is a model for wide-area distributed and parallel computing across heterogeneous networks in multiple administrative domains. This research field aims to promote sharing of resources and provides breakthrough computing power over this wide network of virtual organizations in a seamless manner [8]. Traditionally, as in Globus [6], Condor-G [9] and Legion [10], there is a minimal infrastructure that provides data resource sharing, computational resource utilization management, and distributed execution. Specifically, considering distributed execution, most of the existing grid infrastructures supports execution of isolated tasks, but they do not consider their task interdependencies as in processes (workflows) [12]. This deficiency restricts better scheduling algorithms, distributed execution coordination and automatic execution recovery. There are few proposed middleware infrastructures that support process execution over the grid. In general, they model processes by interconnecting their activities through control and data dependencies. Among them, WebFlow [1] emphasizes an architecture to construct distributed processes; Opera-G [3] provides execution recovering and steering, GridFlow [5] focuses on improved scheduling algorithms that take advantage of activity dependencies, and SwinDew [13] supports totally distributed execution on peer-to-peer networks. However, such infrastructures contain scheduling algorithms that are centralized by process [1, 3, 5], or completely distributed, but difficult to monitor and control [13]. In order to address such constraints, this paper proposes a structured programming model for process description and a hierarchical process execution infrastructure. The programming model employs structured control flow to promote controlled and contextualized activity execution. Complementary, the support infrastructure, which executes a process specification, takes advantage of the hierarchical structure of a specified process in order to distribute and schedule strong dependent activities as a unit, allowing a better execution performance and fault-tolerance and providing localized communication. The programming model and the support infrastructure, named X avantes, are under implementation in order to show the feasibility of the proposed model and to demonstrate its two major advantages: to promote widely distributed process execution and scheduling, but in a controlled, structured and localized way. Next Section describes the programming model, and Section 3, the support infrastructure for the proposed grid computing model. Section 4 demonstrates how the support infrastructure executes processes and distributes activities. Related works are presented and compared to the proposed model in Section 5. The last Section concludes this paper encompassing the advantages of the proposed hierarchical process execution support for the grid computing area and lists some future works. 87 Middleware 2004 Companion ProcessElement Process Activity Controller 1 * 1 * Figure 1: High-level framework of the programming model 2. PROGRAMMING MODEL The programming model designed for the grid computing architecture is very similar to the specified to the Business Process Execution Language (BPEL) [2]. Both describe processes in XML [4] documents, but the former specifies processes strictly synchronous and structured, and has more constructs for structured parallel control. The rationale behind of its design is the possibility of hierarchically distribute the process control and coordination based on structured constructs, differently from BPEL, which does not allow hierarchical composition of processes. In the proposed programming model, a process is a set of interdependent activities arranged to solve a certain problem. In detail, a process is composed of activities, subprocesses, and controllers (see Figure 1). Activities represent simple tasks that are executed on behalf of a process; subprocesses are processes executed in the context of a parent process; and controllers are control elements used to specify the execution order of these activities and subprocesses. Like structured languages, controllers can be nested and then determine the execution order of other controllers. Data are exchanged among process elements through parameters. They are passed by value, in case of simple objects, or by reference, if they are remote objects shared among elements of the same controller or process. External data can be accessed through data sources, such as relational databases or distributed objects. 2.1 Controllers Controllers are structured control constructs used to define the control flow of processes. There are sequential and parallel controllers. The sequential controller types are: block, switch, for and while. The block controller is a simple sequential construct, and the others mimic equivalent structured programming language constructs. Similarly, the parallel types are: par, parswitch, parfor and parwhile. They extend the respective sequential counterparts to allow parallel execution of process elements. All parallel controller types fork the execution of one or more process elements, and then, wait for each execution to finish. Indeed, they contain a fork and a join of execution. Aiming to implement a conditional join, all parallel controller types contain an exit condition, evaluated all time that an element execution finishes, in order to determine when the controller must end. The parfor and parwhile are the iterative versions of the parallel controller types. Both fork executions while the iteration condition is true. This provides flexibility to determine, at run-time, the number of process elements to execute simultaneously. When compared to workflow languages, the parallel controller types represent structured versions of the workflow control constructors, because they can nest other controllers and also can express fixed and conditional forks and joins, present in such languages. 2.2 Process Example This section presents an example of a prime number search application that receives a certain range of integers and returns a set of primes contained in this range. The whole computation is made by a process, which uses a parallel controller to start and dispatch several concurrent activities of the same type, in order to find prime numbers. The portion of the XML document that describes the process and activity types is shown below. <PROCESS_TYPE NAME="FindPrimes"> <IN_PARAMETER TYPE="int" NAME="min"/> <IN_PARAMETER TYPE="int" NAME="max"/> <IN_PARAMETER TYPE="int" NAME="numPrimes"/> <IN_PARAMETER TYPE="int" NAME="numActs"/> <BODY> <PRE_CODE> setPrimes(new RemoteHashSet()); parfor.setMin(getMin()); parfor.setMax(getMax()); parfor.setNumPrimes(getNumPrimes()); parfor.setNumActs(getNumActs()); parfor.setPrimes(getPrimes()); parfor.setCounterBegin(0); parfor.setCounterEnd(getNumActs()-1); </PRE_CODE> <PARFOR NAME="parfor"> <IN_PARAMETER TYPE="int" NAME="min"/> <IN_PARAMETER TYPE="int" NAME="max"/> <IN_PARAMETER TYPE="int" NAME="numPrimes"/> <IN_PARAMETER TYPE="int" NAME="numActs"/> <IN_PARAMETER TYPE="RemoteCollection" NAME="primes"/> <ITERATE> <PRE_CODE> int range= (getMax()-getMin()+1)/getNumActs(); int minNum = range*getCounter()+getMin(); int maxNum = minNum+range-1; if (getCounter() == getNumActs()-1) maxNum = getMax(); findPrimes.setMin(minNum); findPrimes.setMax(maxNum); findPrimes.setNumPrimes(getNumPrimes()); findPrimes.setPrimes(getPrimes()); </PRE_CODE> <ACTIVITY TYPE="FindPrimes" NAME="findPrimes"/> </ITERATE> </PARFOR> </BODY> <OUT_PARAMETER TYPE="RemoteCollection" NAME="primes"/> </PROCESS_TYPE> Middleware for Grid Computing 88 <ACTIVITY_TYPE NAME="FindPrimes"> <IN_PARAMETER TYPE="int" NAME="min"/> <IN_PARAMETER TYPE="int" NAME="max"/> <IN_PARAMETER TYPE="int" NAME="numPrimes"/> <IN_PARAMETER TYPE="RemoteCollection" NAME="primes"/> <CODE> for (int num=getMin(); num<=getMax(); num++) { // stop, required number of primes was found if (primes.size() >= getNumPrimes()) break; boolean prime = true; for (int i=2; i<num; i++) { if (num % i == 0) { prime = false; break; } } if (prime) { primes.add(new Integer(num)); } } </CODE> </ACTIVITY_TYPE> Firstly, a process type that finds prime numbers, named FindPrimes, is defined. It receives, through its input parameters, a range of integers in which prime numbers have to be found, the number of primes to be returned, and the number of activities to be executed in order to perform this work. At the end, the found prime numbers are returned as a collection through its output parameter. This process contains a PARFOR controller aiming to execute a determined number of parallel activities. It iterates from 0 to getNumActs() - 1, which determines the number of activities, starting a parallel activity in each iteration. In such case, the controller divides the whole range of numbers in subranges of the same size, and, in each iteration, starts a parallel activity that finds prime numbers in a specific subrange. These activities receive a shared object by reference in order to store the prime numbers just found and control if the required number of primes has been reached. Finally, it is defined the activity type, FindPrimes, used to find prime numbers in each subrange. It receives, through its input parameters, the range of numbers in which it has to find prime numbers, the total number of prime numbers to be found by the whole process, and, passed by reference, a collection object to store the found prime numbers. Between its CODE markers, there is a simple code to find prime numbers, which iterates over the specified range and verifies if the current integer is a prime. Additionally, in each iteration, the code verifies if the required number of primes, inserted in the primes collection by all concurrent activities, has been reached, and exits if true. The advantage of using controllers is the possibility of the support infrastructure determines the point of execution the process is in, allowing automatic recovery and monitoring, and also the capability of instantiating and dispatching process elements only when there are enough computing resources available, reducing unnecessary overhead. Besides, due to its structured nature, they can be easily composed and the support infrastructure can take advantage of this in order to distribute hierarchically the nested controllers to Group Server Group Java Virtual Machine RMI JDBC Group Manager Process Server Java Virtual Machine RMI JDBC Process Coordinator Worker Java Virtual Machine RMI Activity Manager Repository Figure 2: Infrastructure architecture different machines over the grid, allowing enhanced scalability and fault-tolerance. 3. SUPPORT INFRASTRUCTURE The support infrastructure comprises tools for specification, and services for execution and monitoring of structured processes in highly distributed, heterogeneous and autonomous grid environments. It has services to monitor availability of resources in the grid, to interpret processes and schedule activities and controllers, and to execute activities. 3.1 Infrastructure Architecture The support infrastructure architecture is composed of groups of machines and data repositories, which preserves its administrative autonomy. Generally, localized machines and repositories, such as in local networks or clusters, form a group. Each machine in a group must have a Java Virtual Machine (JVM) [11], and a Java Runtime Library, besides a combination of the following grid support services: group manager (GM), process coordinator (PC) and activity manager (AM). This combination determines what kind of group node it represents: a group server, a process server, or simply a worker (see Figure 2). In a group there are one or more group managers, but only one acts as primary and the others, as replicas. They are responsible to maintain availability information of group machines. Moreover, group managers maintain references to data resources of the group. They use group repositories to persist and recover the location of nodes and their availability. To control process execution, there are one or more process coordinators per group. They are responsible to instantiate and execute processes and controllers, select resources, and schedule and dispatch activities to workers. In order to persist and recover process execution and data, and also load process specification, they use group repositories. Finally, in several group nodes there is an activity manager. It is responsible to execute activities in the hosted machine on behalf of the group process coordinators, and to inform the current availability of the associated machine to group managers. They also have pendent activity queues, containing activities to be executed. 3.2 Inter-group Relationships In order to model real grid architecture, the infrastructure must comprise several, potentially all, local networks, like Internet does. Aiming to satisfy this intent, local groups are 89 Middleware 2004 Companion GM GM GM GM Figure 3: Inter-group relationships connected to others, directly or indirectly, through its group managers (see Figure 3). Each group manager deals with requests of its group (represented by dashed ellipses), in order to register local machines and maintain correspondent availability. Additionally, group managers communicate to group managers of other groups. Each group manager exports coarse availability information to group managers of adjacent groups and also receives requests from other external services to furnish detailed availability information. In this way, if there are resources available in external groups, it is possible to send processes, controllers and activities to these groups in order to execute them in external process coordinators and activity managers, respectively. 4. PROCESS EXECUTION In the proposed grid architecture, a process is specified in XML, using controllers to determine control flow; referencing other processes and activities; and passing objects to their parameters in order to define data flow. After specified, the process is compiled in a set of classes, which represent specific process, activity and controller types. At this time, it can be instantiated and executed by a process coordinator. 4.1 Dynamic Model To execute a specified process, it must be instantiated by referencing its type on a process coordinator service of a specific group. Also, the initial parameters must be passed to it, and then it can be started. The process coordinator carries out the process by executing the process elements included in its body sequentially. If the element is a process or a controller, the process coordinator can choose to execute it in the same machine or to pass it to another process coordinator in a remote machine, if available. Else, if the element is an activity, it passes to an activity manager of an available machine. Process coordinators request the local group manager to find available machines that contain the required service, process coordinator or activity manager, in order to execute a process element. Then, it can return a local machine, a machine in another group or none, depending on the availability of such resource in the grid. It returns an external worker (activity manager machine) if there are no available workers in the local group; and, it returns an external process server (process coordinator machine), if there are no available process servers or workers in the local group. Obeying this rule, group managers try to find process servers in the same group of the available workers. Such procedure is followed recursively by all process coGM FindPrimes Activity AM FindPrimes Activity AM FindPrimes Activity AM FindPrimes Process PC Figure 4: FindPrimes process execution ordinators that execute subprocesses or controllers of a process. Therefore, because processes are structured by nesting process elements, the process execution is automatically distributed hierarchically through one or more grid groups according to the availability and locality of computing resources. The advantage of this distribution model is wide area execution, which takes advantage of potentially all grid resources; and localized communication of process elements, because strong dependent elements, which are under the same controller, are placed in the same or near groups. Besides, it supports easy monitoring and steering, due to its structured controllers, which maintain state and control over its inner elements. 4.2 Process Execution Example Revisiting the example shown in Section 2.2, a process type is specified to find prime numbers in a certain range of numbers. In order to solve this problem, it creates a number of activities using the parfor controller. Each activity, then, finds primes in a determined part of the range of numbers. Figure 4 shows an instance of this process type executing over the proposed infrastructure. A FindPrimes process instance is created in an available process coordinator (PC), which begins executing the parfor controller. In each iteration of this controller, the process coordinator requests to the group manager (GM) an available activity manager (AM) in order to execute a new instance of the FindPrimes activity. If there is any AM available in this group or in an external one, the process coordinator sends the activity class and initial parameters to this activity manager and requests its execution. Else, if no activity manager is available, then the controller enters in a wait state until an activity manager is made available, or is created. In parallel, whenever an activity finishes, its result is sent back to the process coordinator, which records it in the parfor controller. Then, the controller waits until all activities that have been started are finished, and it ends. At this point, the process coordinator verifies that there is no other process element to execute and finishes the process. 5. RELATED WORK There are several academic and commercial products that promise to support grid computing, aiming to provide interfaces, protocols and services to leverage the use of widely Middleware for Grid Computing 90 distributed resources in heterogeneous and autonomous networks. Among them, Globus [6], Condor-G [9] and Legion [10] are widely known. Aiming to standardize interfaces and services to grid, the Open Grid Services Architecture (OGSA) [7] has been defined. The grid architectures generally have services that manage computing resources and distribute the execution of independent tasks on available ones. However, emerging architectures maintain task dependencies and automatically execute tasks in a correct order. They take advantage of these dependencies to provide automatic recovery, and better distribution and scheduling algorithms. Following such model, WebFlow [1] is a process specification tool and execution environment constructed over CORBA that allows graphical composition of activities and their distributed execution in a grid environment. Opera-G [3], like WebFlow, uses a process specification language similar to the data flow diagram and workflow languages, but furnishes automatic execution recovery and limited steering of process execution. The previously referred architectures and others that enact processes over the grid have a centralized coordination. In order to surpass this limitation, systems like SwinDew [13] proposed a widely distributed process execution, in which each node knows where to execute the next activity or join activities in a peer-to-peer environment. In the specific area of activity distribution and scheduling, emphasized in this work, GridFlow [5] is remarkable. It uses a two-level scheduling: global and local. In the local level, it has services that predict computing resource utilization and activity duration. Based on this information, GridFlow employs a PERT-like technique that tries to forecast the activity execution start time and duration in order to better schedule them to the available resources. The architecture proposed in this paper, which encompasses a programming model and an execution support infrastructure, is widely decentralized, differently from WebFlow and Opera-G, being more scalable and fault-tolerant. But, like the latter, it is designed to support execution recovery. Comparing to SwinDew, the proposed architecture contains widely distributed process coordinators, which coordinate processes or parts of them, differently from SwinDew where each node has a limited view of the process: only the activity that starts next. This makes easier to monitor and control processes. Finally, the support infrastructure breaks the process and its subprocesses for grid execution, allowing a group to require another group for the coordination and execution of process elements on behalf of the first one. This is different from GridFlow, which can execute a process in at most two levels, having the global level as the only responsible to schedule subprocesses in other groups. This can limit the overall performance of processes, and make the system less scalable. 6. CONCLUSION AND FUTURE WORK Grid computing is an emerging research field that intends to promote distributed and parallel computing over the wide area network of heterogeneous and autonomous administrative domains in a seamless way, similar to what Internet does to the data sharing. There are several products that support execution of independent tasks over grid, but only a few supports the execution of processes with interdependent tasks. In order to address such subject, this paper proposes a programming model and a support infrastructure that allow the execution of structured processes in a widely distributed and hierarchical manner. This support infrastructure provides automatic, structured and recursive distribution of process elements over groups of available machines; better resource use, due to its on demand creation of process elements; easy process monitoring and steering, due to its structured nature; and localized communication among strong dependent process elements, which are placed under the same controller. These features contribute to better scalability, fault-tolerance and control for processes execution over the grid. Moreover, it opens doors for better scheduling algorithms, recovery mechanisms, and also, dynamic modification schemes. The next work will be the implementation of a recovery mechanism that uses the execution and data state of processes and controllers to recover process execution. After that, it is desirable to advance the scheduling algorithm to forecast machine use in the same or other groups and to foresee start time of process elements, in order to use this information to pre-allocate resources and, then, obtain a better process execution performance. Finally, it is interesting to investigate schemes of dynamic modification of processes over the grid, in order to evolve and adapt long-term processes to the continuously changing grid environment. 7. ACKNOWLEDGMENTS We would like to thank Paulo C. Oliveira, from the State Treasury Department of Sao Paulo, for its deeply revision and insightful comments. 8. REFERENCES [1] E. Akarsu, G. C. Fox, W. Furmanski, and T. Haupt. WebFlow: High-Level Programming Environment and Visual Authoring Toolkit for High Performance Distributed Computing. In Proceedings of Supercom puting (SC98), 1998. [2] T. Andrews and F. Curbera. Specification: Business Process Execution Language for W eb Services V ersion 1.1. IBM DeveloperWorks, 2003. Available at http://www-106.ibm.com/developerworks/library/wsbpel. [3] W. Bausch. O PERA -G :A M icrokernelfor Com putationalG rids. PhD thesis, Swiss Federal Institute of Technology, Zurich, 2004. [4] T. Bray and J. Paoli. Extensible M arkup Language (X M L) 1.0. XML Core WG, W3C, 2004. Available at http://www.w3.org/TR/2004/REC-xml-20040204. [5] J. Cao, S. A. Jarvis, S. Saini, and G. R. Nudd. GridFlow: Workflow Management for Grid Computing. In Proceedings ofthe International Sym posium on Cluster Com puting and the G rid (CCG rid 2003), 2003. [6] I. Foster and C. Kesselman. Globus: A Metacomputing Infrastructure Toolkit. Intl.J. Supercom puter A pplications, 11(2):115-128, 1997. [7] I. Foster, C. Kesselman, J. M. Nick, and S. Tuecke. The Physiology ofthe G rid: A n O pen G rid Services A rchitecture for D istributed System s Integration. 91 Middleware 2004 Companion Open Grid Service Infrastructure WG, Global Grid Forum, 2002. [8] I. Foster, C. Kesselman, and S. Tuecke. The Anatomy of the Grid: Enabling Scalable Virtual Organization. The Intl.JournalofH igh Perform ance Com puting A pplications, 15(3):200-222, 2001. [9] J. Frey, T. Tannenbaum, M. Livny, I. Foster, and S. Tuecke. Condor-G: A Computational Management Agent for Multi-institutional Grids. In Proceedings of the Tenth Intl.Sym posium on H igh Perform ance D istributed Com puting (H PD C-10). IEEE, 2001. [10] A. S. Grimshaw and W. A. Wulf. Legion - A View from 50,000 Feet. In Proceedings ofthe Fifth Intl. Sym posium on H igh Perform ance D istributed Com puting. IEEE, 1996. [11] T. Lindholm and F. Yellin. The Java V irtualM achine Specification. Sun Microsystems, Second Edition edition, 1999. [12] B. R. Schulze and E. R. M. Madeira. Grid Computing with Active Services. Concurrency and Com putation: Practice and Experience Journal, 5(16):535-542, 2004. [13] J. Yan, Y. Yang, and G. K. Raikundalia. Enacting Business Processes in a Decentralised Environment with P2P-Based Workflow Support. In Proceedings of the Fourth Intl.Conference on W eb-Age Inform ation M anagem ent(W A IM 2003), 2003. Middleware for Grid Computing 92

distributed computing;parallel computing;distribute middleware;parallel execution;grid computing;process support;hierarchical process execution;grid architecture;process execution;distributed system;process description;distributed scheduling;scheduling algorithm;distributed application;distributed process;distributed execution

train_C-57

Congestion Games with Load-Dependent Failures: Identical Resources

We define a new class of games, congestion games with loaddependent failures (CGLFs), which generalizes the well-known class of congestion games, by incorporating the issue of resource failures into congestion games. In a CGLF, agents share a common set of resources, where each resource has a cost and a probability of failure. Each agent chooses a subset of the resources for the execution of his task, in order to maximize his own utility. The utility of an agent is the difference between his benefit from successful task completion and the sum of the costs over the resources he uses. CGLFs possess two novel features. It is the first model to incorporate failures into congestion settings, which results in a strict generalization of congestion games. In addition, it is the first model to consider load-dependent failures in such framework, where the failure probability of each resource depends on the number of agents selecting this resource. Although, as we show, CGLFs do not admit a potential function, and in general do not have a pure strategy Nash equilibrium, our main theorem proves the existence of a pure strategy Nash equilibrium in every CGLF with identical resources and nondecreasing cost functions.

1. INTRODUCTION We study the effects of resource failures in congestion settings. This study is motivated by a variety of situations in multi-agent systems with unreliable components, such as machines, computers etc. We define a model for congestion games with load-dependent failures (CGLFs) which provides simple and natural description of such situations. In this model, we are given a finite set of identical resources (service providers) where each element possesses a failure probability describing the probability of unsuccessful completion of its assigned tasks as a (nondecreasing) function of its congestion. There is a fixed number of agents, each having a task which can be carried out by any of the resources. For reliability reasons, each agent may decide to assign his task, simultaneously, to a number of resources. Thus, the congestion on the resources is not known in advance, but is strategy-dependent. Each resource is associated with a cost, which is a (nonnegative) function of the congestion experienced by this resource. The objective of each agent is to maximize his own utility, which is the difference between his benefit from successful task completion and the sum of the costs over the set of resources he uses. The benefits of the agents from successful completion of their tasks are allowed to vary across the agents. The resource cost function describes the cost suffered by an agent for selecting that resource, as a function of the number of agents who have selected it. Thus, it is natural to assume that these functions are nonnegative. In addition, in many real-life applications of our model the resource cost functions have a special structure. In particular, they can monotonically increase or decrease with the number of the users, depending on the context. The former case is motivated by situations where high congestion on a resource causes longer delay in its assigned tasks execution and as a result, the cost of utilizing this resource might be higher. A typical example of such situation is as follows. Assume we need to deliver an important package. Since there is no guarantee that a courier will reach the destination in time, we might send several couriers to deliver the same package. The time required by each courier to deliver the package increases with the congestion on his way. In addition, the payment to a courier is proportional to the time he spends in delivering the package. Thus, the payment to the courier increases when the congestion increases. The latter case (decreasing cost functions) describes situations where a group of agents using a particular resource have an opportunity to share its cost among the group"s members, or, the cost of 210 using a resource decreases with the number of users, according to some marketing policy. Our results We show that CGLFs and, in particular, CGLFs with nondecreasing cost functions, do not admit a potential function. Therefore, the CGLF model can not be reduced to congestion games. Nevertheless, if the failure probabilities are constant (do not depend on the congestion) then a potential function is guaranteed to exist. We show that CGLFs and, in particular, CGLFs with decreasing cost functions, do not possess pure strategy Nash equilibria. However, as we show in our main result, there exists a pure strategy Nash equilibrium in any CGLF with nondecreasing cost functions. Related work Our model extends the well-known class of congestion games [11]. In a congestion game, every agent has to choose from a finite set of resources, where the utility (or cost) of an agent from using a particular resource depends on the number of agents using it, and his total utility (cost) is the sum of the utilities (costs) obtained from the resources he uses. An important property of these games is the existence of pure strategy Nash equilibria. Monderer and Shapley [9] introduced the notions of potential function and potential game and proved that the existence of a potential function implies the existence of a pure strategy Nash equilibrium. They observed that Rosenthal [11] proved his theorem on congestion games by constructing a potential function (hence, every congestion game is a potential game). Moreover, they showed that every finite potential game is isomorphic to a congestion game; hence, the classes of finite potential games and congestion games coincide. Congestion games have been extensively studied and generalized. In particular, Leyton-Brown and Tennenholtz [5] extended the class of congestion games to the class of localeffect games. In a local-effect game, each agent"s payoff is effected not only by the number of agents who have chosen the same resources as he has chosen, but also by the number of agents who have chosen neighboring resources (in a given graph structure). Monderer [8] dealt with another type of generalization of congestion games, in which the resource cost functions are player-specific (PS-congestion games). He defined PS-congestion games of type q (q-congestion games), where q is a positive number, and showed that every game in strategic form is a q-congestion game for some q. Playerspecific resource cost functions were discussed for the first time by Milchtaich [6]. He showed that simple and strategysymmetric PS-congestion games are not potential games, but always possess a pure strategy Nash equilibrium. PScongestion games were generalized to weighted congestion games [6] (or, ID-congestion games [7]), in which the resource cost functions are not only player-specific, but also depend on the identity of the users of the resource. Ackermann et al. [1] showed that weighted congestion games admit pure strategy Nash equilibria if the strategy space of each player consists of the bases of a matroid on the set of resources. Much of the work on congestion games has been inspired by the fact that every such game has a pure strategy Nash equilibrium. In particular, Fabrikant et al. [3] studied the computational complexity of finding pure strategy Nash equilibria in congestion games. Intensive study has also been devoted to quantify the inefficiency of equilibria in congestion games. Koutsoupias and Papadimitriou [4] proposed the worst-case ratio of the social welfare achieved by a Nash equilibrium and by a socially optimal strategy profile (dubbed the price of anarchy) as a measure of the performance degradation caused by lack of coordination. Christodoulou and Koutsoupias [2] considered the price of anarchy of pure equilibria in congestion games with linear cost functions. Roughgarden and Tardos [12] used this approach to study the cost of selfish routing in networks with a continuum of users. However, the above settings do not take into consideration the possibility that resources may fail to execute their assigned tasks. In the computer science context of congestion games, where the alternatives of concern are machines, computers, communication lines etc., which are obviously prone to failures, this issue should not be ignored. Penn, Polukarov and Tennenholtz were the first to incorporate the issue of failures into congestion settings [10]. They introduced a class of congestion games with failures (CGFs) and proved that these games, while not being isomorphic to congestion games, always possess Nash equilibria in pure strategies. The CGF-model significantly differs from ours. In a CGF, the authors considered the delay associated with successful task completion, where the delay for an agent is the minimum of the delays of his successful attempts and the aim of each agent is to minimize his expected delay. In contrast with the CGF-model, in our model we consider the total cost of the utilized resources, where each agent wishes to maximize the difference between his benefit from a successful task completion and the sum of his costs over the resources he uses. The above differences imply that CGFs and CGLFs possess different properties. In particular, if in our model the resource failure probabilities were constant and known in advance, then a potential function would exist. This, however, does not hold for CGFs; in CGFs, the failure probabilities are constant but there is no potential function. Furthermore, the procedures proposed by the authors in [10] for the construction of a pure strategy Nash equilibrium are not valid in our model, even in the simple, agent-symmetric case, where all agents have the same benefit from successful completion of their tasks. Our work provides the first model of congestion settings with resource failures, which considers the sum of congestiondependent costs over utilized resources, and therefore, does not extend the CGF-model, but rather generalizes the classic model of congestion games. Moreover, it is the first model to consider load-dependent failures in the above context. 211 Organization The rest of the paper is organized as follows. In Section 2 we define our model. In Section 3 we present our results. In 3.1 we show that CGLFs, in general, do not have pure strategy Nash equilibria. In 3.2 we focus on CGLFs with nondecreasing cost functions (nondecreasing CGLFs). We show that these games do not admit a potential function. However, in our main result we show the existence of pure strategy Nash equilibria in nondecreasing CGLFs. Section 4 is devoted to a short discussion. Many of the proofs are omitted from this conference version of the paper, and will appear in the full version. 2. THE MODEL The scenarios considered in this work consist of a finite set of agents where each agent has a task that can be carried out by any element of a set of identical resources (service providers). The agents simultaneously choose a subset of the resources in order to perform their tasks, and their aim is to maximize their own expected payoff, as described in the sequel. Let N be a set of n agents (n ∈ N), and let M be a set of m resources (m ∈ N). Agent i ∈ N chooses a strategy σi ∈ Σi which is a (potentially empty) subset of the resources. That is, Σi is the power set of the set of resources: Σi = P(M). Given a subset S ⊆ N of the agents, the set of strategy combinations of the members of S is denoted by ΣS = ×i∈SΣi, and the set of strategy combinations of the complement subset of agents is denoted by Σ−S (Σ−S = ΣN S = ×i∈N SΣi). The set of pure strategy profiles of all the agents is denoted by Σ (Σ = ΣN ). Each resource is associated with a cost, c(·), and a failure probability, f(·), each of which depends on the number of agents who use this resource. We assume that the failure probabilities of the resources are independent. Let σ = (σ1, . . . , σn) ∈ Σ be a pure strategy profile. The (m-dimensional) congestion vector that corresponds to σ is hσ = (hσ e )e∈M , where hσ e = ˛ ˛{i ∈ N : e ∈ σi} ˛ ˛. The failure probability of a resource e is a monotone nondecreasing function f : {1, . . . , n} → [0, 1) of the congestion experienced by e. The cost of utilizing resource e is a function c : {1, . . . , n} → R+ of the congestion experienced by e. The outcome for agent i ∈ N is denoted by xi ∈ {S, F}, where S and F, respectively, indicate whether the task execution succeeded or failed. We say that the execution of agent"s i task succeeds if the task of agent i is successfully completed by at least one of the resources chosen by him. The benefit of agent i from his outcome xi is denoted by Vi(xi), where Vi(S) = vi, a given (nonnegative) value, and Vi(F) = 0. The utility of agent i from strategy profile σ and his outcome xi, ui(σ, xi), is the difference between his benefit from the outcome (Vi(xi)) and the sum of the costs of the resources he has used: ui(σ, xi) = Vi(xi) − X e∈σi c(hσ e ) . The expected utility of agent i from strategy profile σ, Ui(σ), is, therefore: Ui(σ) = 1 − Y e∈σi f(hσ e ) ! vi − X e∈σi c(hσ e ) , where 1 − Q e∈σi f(hσ e ) denotes the probability of successful completion of agent i"s task. We use the convention thatQ e∈∅ f(hσ e ) = 1. Hence, if agent i chooses an empty set σi = ∅ (does not assign his task to any resource), then his expected utility, Ui(∅, σ−i), equals zero. 3. PURE STRATEGY NASH EQUILIBRIA IN CGLFS In this section we present our results on CGLFs. We investigate the property of the (non-)existence of pure strategy Nash equilibria in these games. We show that this class of games does not, in general, possess pure strategy equilibria. Nevertheless, if the resource cost functions are nondecreasing then such equilibria are guaranteed to exist, despite the non-existence of a potential function. 3.1 Decreasing Cost Functions We start by showing that the class of CGLFs and, in particular, the subclass of CGLFs with decreasing cost functions, does not, in general, possess Nash equilibria in pure strategies. Consider a CGLF with two agents (N = {1, 2}) and two resources (M = {e1, e2}). The cost function of each resource is given by c(x) = 1 xx , where x ∈ {1, 2}, and the failure probabilities are f(1) = 0.01 and f(2) = 0.26. The benefits of the agents from successful task completion are v1 = 1.1 and v2 = 4. Below we present the payoff matrix of the game. ∅ {e1} {e2} {e1, e2} ∅ U1 = 0 U1 = 0 U1 = 0 U1 = 0 U2 = 0 U2 = 2.96 U2 = 2.96 U2 = 1.9996 {e1} U1 = 0.089 U1 = 0.564 U1 = 0.089 U1 = 0.564 U2 = 0 U2 = 2.71 U2 = 2.96 U2 = 2.7396 {e2} U1 = 0.089 U1 = 0.089 U1 = 0.564 U1 = 0.564 U2 = 0 U2 = 2.96 U2 = 2.71 U2 = 2.7396 {e1, e2} U1 = −0.90011 U1 = −0.15286 U1 = −0.15286 U1 = 0.52564 U2 = 0 U2 = 2.71 U2 = 2.71 U2 = 3.2296 Table 1: Example for non-existence of pure strategy Nash equilibria in CGLFs. It can be easily seen that for every pure strategy profile σ in this game there exist an agent i and a strategy σi ∈ Σi such that Ui(σ−i, σi) > Ui(σ). That is, every pure strategy profile in this game is not in equilibrium. However, if the cost functions in a given CGLF do not decrease in the number of users, then, as we show in the main result of this paper, a pure strategy Nash equilibrium is guaranteed to exist. 212 3.2 Nondecreasing Cost Functions This section focuses on the subclass of CGLFs with nondecreasing cost functions (henceforth, nondecreasing CGLFs). We show that nondecreasing CGLFs do not, in general, admit a potential function. Therefore, these games are not congestion games. Nevertheless, we prove that all such games possess pure strategy Nash equilibria. 3.2.1 The (Non-)Existence of a Potential Function Recall that Monderer and Shapley [9] introduced the notions of potential function and potential game, where potential game is defined to be a game that possesses a potential function. A potential function is a real-valued function over the set of pure strategy profiles, with the property that the gain (or loss) of an agent shifting to another strategy while the other agents" strategies are kept unchanged, equals to the corresponding increment of the potential function. The authors [9] showed that the classes of finite potential games and congestion games coincide. Here we show that the class of CGLFs and, in particular, the subclass of nondecreasing CGLFs, does not admit a potential function, and therefore is not included in the class of congestion games. However, for the special case of constant failure probabilities, a potential function is guaranteed to exist. To prove these statements we use the following characterization of potential games [9]. A path in Σ is a sequence τ = (σ0 → σ1 → · · · ) such that for every k ≥ 1 there exists a unique agent, say agent i, such that σk = (σk−1 −i , σi) for some σi = σk−1 i in Σi. A finite path τ = (σ0 → σ1 → · · · → σK ) is closed if σ0 = σK . It is a simple closed path if in addition σl = σk for every 0 ≤ l = k ≤ K − 1. The length of a simple closed path is defined to be the number of distinct points in it; that is, the length of τ = (σ0 → σ1 → · · · → σK ) is K. Theorem 1. [9] Let G be a game in strategic form with a vector U = (U1, . . . , Un) of utility functions. For a finite path τ = (σ0 → σ1 → · · · → σK ), let U(τ) = PK k=1[Uik (σk )− Uik (σk−1 )], where ik is the unique deviator at step k. Then, G is a potential game if and only if U(τ) = 0 for every simple closed path τ of length 4. Load-Dependent Failures Based on Theorem 1, we present the following counterexample that demonstrates the non-existence of a potential function in CGLFs. We consider the following agent-symmetric game G in which two agents (N = {1, 2}) wish to assign a task to two resources (M = {e1, e2}). The benefit from a successful task completion of each agent equals v, and the failure probability function strictly increases with the congestion. Consider the simple closed path of length 4 which is formed by α = (∅, {e2}) , β = ({e1}, {e2}) , γ = ({e1}, {e1, e2}) , δ = (∅, {e1, e2}) : {e2} {e1, e2} ∅ U1 = 0 U1 = 0 U2 = (1 − f(1)) v − c(1) U2 = ` 1 − f(1)2 ´ v − 2c(1) {e1} U1 = (1 − f(1)) v − c(1) U1 = (1 − f(2)) v − c(2) U2 = (1 − f(1)) v − c(1) U2 = (1 − f(1)f(2)) v − c(1) − c(2) Table 2: Example for non-existence of potentials in CGLFs. Therefore, U1(α) − U1(β) + U2(β) − U2(γ) + U1(γ) − U1(δ) +U2(δ) − U2(α) = v (1 − f(1)) (f(1) − f(2)) = 0. Thus, by Theorem 1, nondecreasing CGLFs do not admit potentials. As a result, they are not congestion games. However, as presented in the next section, the special case in which the failure probabilities are constant, always possesses a potential function. Constant Failure Probabilities We show below that CGLFs with constant failure probabilities always possess a potential function. This follows from the fact that the expected benefit (revenue) of each agent in this case does not depend on the choices of the other agents. In addition, for each agent, the sum of the costs over his chosen subset of resources, equals the payoff of an agent choosing the same strategy in the corresponding congestion game. Assume we are given a game G with constant failure probabilities. Let τ = (α → β → γ → δ → α) be an arbitrary simple closed path of length 4. Let i and j denote the active agents (deviators) in τ and z ∈ Σ−{i,j} be a fixed strategy profile of the other agents. Let α = (xi, xj, z), β = (yi, xj, z), γ = (yi, yj, z), δ = (xi, yj, z), where xi, yi ∈ Σi and xj, yj ∈ Σj. Then, U(τ) = Ui(xi, xj, z) − Ui(yi, xj, z) +Uj(yi, xj, z) − Uj(yi, yj, z) +Ui(yi, yj, z) − Ui(xi, yj, z) +Uj(xi, yj, z) − Uj(xi, xj, z) = 1 − f|xi| vi − X e∈xi c(h (xi,xj ,z) e ) − . . . − 1 − f|xj | vj + X e∈xj c(h (xi,xj ,z) e ) = » 1 − f|xi| vi − . . . − 1 − f|xj | vj − » X e∈xi c(h (xi,xj ,z) e ) − . . . − X e∈xj c(h (xi,xj ,z) e ) . Notice that » 1 − f|xi| vi − . . . − 1 − f|xj | vj = 0, as a sum of a telescope series. The remaining sum equals 0, by applying Theorem 1 to congestion games, which are known to possess a potential function. Thus, by Theorem 1, G is a potential game. 213 We note that the above result holds also for the more general settings with non-identical resources (having different failure probabilities and cost functions) and general cost functions (not necessarily monotone and/or nonnegative). 3.2.2 The Existence of a Pure Strategy Nash Equilibrium In the previous section, we have shown that CGLFs and, in particular, nondecreasing CGLFs, do not admit a potential function, but this fact, in general, does not contradict the existence of an equilibrium in pure strategies. In this section, we present and prove the main result of this paper (Theorem 2) which shows the existence of pure strategy Nash equilibria in nondecreasing CGLFs. Theorem 2. Every nondecreasing CGLF possesses a Nash equilibrium in pure strategies. The proof of Theorem 2 is based on Lemmas 4, 7 and 8, which are presented in the sequel. We start with some definitions and observations that are needed for their proofs. In particular, we present the notions of A-, D- and S-stability and show that a strategy profile is in equilibrium if and only if it is A-, D- and S- stable. Furthermore, we prove the existence of such a profile in any given nondecreasing CGLF. Definition 3. For any strategy profile σ ∈ Σ and for any agent i ∈ N, the operation of adding precisely one resource to his strategy, σi, is called an A-move of i from σ. Similarly, the operation of dropping a single resource is called a D-move, and the operation of switching one resource with another is called an S-move. Clearly, if agent i deviates from strategy σi to strategy σi by applying a single A-, D- or S-move, then max {|σi σi|, |σi σi|} = 1, and vice versa, if max {|σi σi|, |σi σi|} = 1 then σi is obtained from σi by applying exactly one such move. For simplicity of exposition, for any pair of sets A and B, let µ(A, B) = max {|A B|, |B A|}. The following lemma implies that any strategy profile, in which no agent wishes unilaterally to apply a single A-, Dor S-move, is a Nash equilibrium. More precisely, we show that if there exists an agent who benefits from a unilateral deviation from a given strategy profile, then there exists a single A-, D- or S-move which is profitable for him as well. Lemma 4. Given a nondecreasing CGLF, let σ ∈ Σ be a strategy profile which is not in equilibrium, and let i ∈ N such that ∃xi ∈ Σi for which Ui(σ−i, xi) > Ui(σ). Then, there exists yi ∈ Σi such that Ui(σ−i, yi) > Ui(σ) and µ(yi, σi) = 1. Therefore, to prove the existence of a pure strategy Nash equilibrium, it suffices to look for a strategy profile for which no agent wishes to unilaterally apply an A-, D- or S-move. Based on the above observation, we define A-, D- and Sstability as follows. Definition 5. A strategy profile σ is said to be A-stable (resp., D-stable, S-stable) if there are no agents with a profitable A- (resp., D-, S-) move from σ. Similarly, we define a strategy profile σ to be DS-stable if there are no agents with a profitable D- or S-move from σ. The set of all DS-stable strategy profiles is denoted by Σ0 . Obviously, the profile (∅, . . . , ∅) is DS-stable, so Σ0 is not empty. Our goal is to find a DS-stable profile for which no profitable A-move exists, implying this profile is in equilibrium. To describe how we achieve this, we define the notions of light (heavy) resources and (nearly-) even strategy profiles, which play a central role in the proof of our main result. Definition 6. Given a strategy profile σ, resource e is called σ-light if hσ e ∈ arg mine∈M hσ e and σ-heavy otherwise. A strategy profile σ with no heavy resources will be termed even. A strategy profile σ satisfying |hσ e − hσ e | ≤ 1 for all e, e ∈ M will be termed nearly-even. Obviously, every even strategy profile is nearly-even. In addition, in a nearly-even strategy profile, all heavy resources (if exist) have the same congestion. We also observe that the profile (∅, . . . , ∅) is even (and DS-stable), so the subset of even, DS-stable strategy profiles is not empty. Based on the above observations, we define two types of an A-move that are used in the sequel. Suppose σ ∈ Σ0 is a nearly-even DS-stable strategy profile. For each agent i ∈ N, let ei ∈ arg mine∈M σi hσ e . That is, ei is a lightest resource not chosen previously by i. Then, if there exists any profitable A-move for agent i, then the A-move with ei is profitable for i as well. This is since if agent i wishes to unilaterally add a resource, say a ∈ M σi, then Ui (σ−i, (σi ∪ {a})) > Ui(σ). Hence, 1 − Y e∈σi f(hσ e )f(hσ a + 1) ! vi − X e∈σi c(hσ e ) − c(hσ a + 1) > 1 − Y e∈σi f(hσ e ) ! vi − X e∈σi c(hσ e ) ⇒ vi Y e∈σi f(hσ e ) > c(hσ a + 1) 1 − f(hσ a + 1) ≥ c(hσ ei + 1) 1 − f(hσ ei + 1) ⇒ Ui (σ−i, (σi ∪ {ei})) > Ui(σ) . If no agent wishes to change his strategy in this manner, i.e. Ui(σ) ≥ Ui(σ−i, σi ∪{ei}) for all i ∈ N, then by the above Ui(σ) ≥ Ui(σ−i, σi ∪{a}) for all i ∈ N and a ∈ M σi. Hence, σ is A-stable and by Lemma 4, σ is a Nash equilibrium strategy profile. Otherwise, let N(σ) denote the subset of all agents for which there exists ei such that a unilateral addition of ei is profitable. Let a ∈ arg minei : i∈N(σ) hσ ei . Let also i ∈ N(σ) be the agent for which ei = a. If a is σ-light, then let σ = (σ−i, σi ∪ {a}). In this case we say that σ is obtained from σ by a one-step addition of resource a, and a is called an added resource. If a is σ-heavy then there exists a σ-light resource b and an agent j such that a ∈ σj and b /∈ σj. Then let σ = ` σ−{i,j}, σi ∪ {a}, (σj {a}) ∪ {b} ´ . In this case we say that σ is obtained from σ by a two-step addition of resource b, and b is called an added resource. We notice that, in both cases, the congestion of each resource in σ is the same as in σ, except for the added resource, for which its congestion in σ increased by 1. Thus, since the added resource is σ-light and σ is nearly-even, σ is nearly-even. Then, the following lemma implies the Sstability of σ . 214 Lemma 7. In a nondecreasing CGLF, every nearly-even strategy profile is S-stable. Coupled with Lemma 7, the following lemma shows that if σ is a nearly-even and DS-stable strategy profile, and σ is obtained from σ by a one- or two-step addition of resource a, then the only potential cause for a non-DS-stability of σ is the existence of an agent k ∈ N with σk = σk, who wishes to drop the added resource a. Lemma 8. Let σ be a nearly-even DS-stable strategy profile of a given nondecreasing CGLF, and let σ be obtained from σ by a one- or two-step addition of resource a. Then, there are no profitable D-moves for any agent i ∈ N with σi = σi. For an agent i ∈ N with σi = σi, the only possible profitable D-move (if exists) is to drop the added resource a. We are now ready to prove our main result - Theorem 2. Let us briefly describe the idea behind the proof. By Lemma 4, it suffices to prove the existence of a strategy profile which is A-, D- and S-stable. We start with the set of even and DS-stable strategy profiles which is obviously not empty. In this set, we consider the subset of strategy profiles with maximum congestion and maximum sum of the agents" utilities. Assuming on the contrary that every DSstable profile admits a profitable A-move, we show the existence of a strategy profile x in the above subset, such that a (one-step) addition of some resource a to x results in a DSstable strategy. Then by a finite series of one- or two-step addition operations we obtain an even, DS-stable strategy profile with strictly higher congestion on the resources, contradicting the choice of x. The full proof is presented below. Proof of Theorem 2: Let Σ1 ⊆ Σ0 be the subset of all even, DS-stable strategy profiles. Observe that since (∅, . . . , ∅) is an even, DS-stable strategy profile, then Σ1 is not empty, and minσ∈Σ0 ˛ ˛{e ∈ M : e is σ−heavy} ˛ ˛ = 0. Then, Σ1 could also be defined as Σ1 = arg min σ∈Σ0 ˛ ˛{e ∈ M : e is σ−heavy} ˛ ˛ , with hσ being the common congestion. Now, let Σ2 ⊆ Σ1 be the subset of Σ1 consisting of all those profiles with maximum congestion on the resources. That is, Σ2 = arg max σ∈Σ1 hσ . Let UN (σ) = P i∈N Ui(σ) denotes the group utility of the agents, and let Σ3 ⊆ Σ2 be the subset of all profiles in Σ2 with maximum group utility. That is, Σ3 = arg max σ∈Σ2 X i∈N Ui(σ) = arg max σ∈Σ2 UN (σ) . Consider first the simple case in which maxσ∈Σ1 hσ = 0. Obviously, in this case, Σ1 = Σ2 = Σ3 = {x = (∅, . . . , ∅)}. We show below that by performing a finite series of (onestep) addition operations on x, we obtain an even, DSstable strategy profile y with higher congestion, that is with hy > hx = 0, in contradiction to x ∈ Σ2 . Let z ∈ Σ0 be a nearly-even (not necessarily even) DS-stable profile such that mine∈M hz e = 0, and note that the profile x satisfies the above conditions. Let N(z) be the subset of agents for which a profitable A-move exists, and let i ∈ N(z). Obviously, there exists a z-light resource a such that Ui(z−i, zi ∪ {a}) > Ui(z) (otherwise, arg mine∈M hz e ⊆ zi, in contradiction to mine∈M hz e = 0). Consider the strategy profile z = (z−i, zi ∪ {a}) which is obtained from z by a (one-step) addition of resource a by agent i. Since z is nearly-even and a is z-light, we can easily see that z is nearly-even. Then, Lemma 7 implies that z is S-stable. Since i is the only agent using resource a in z , by Lemma 8, no profitable D-moves are available. Thus, z is a DS-stable strategy profile. Therefore, since the number of resources is finite, there is a finite series of one-step addition operations on x = (∅, . . . , ∅) that leads to strategy profile y ∈ Σ1 with hy = 1 > 0 = hx , in contradiction to x ∈ Σ2 . We turn now to consider the other case where maxσ∈Σ1 hσ ≥ 1. In this case we select from Σ3 a strategy profile x, as described below, and use it to contradict our contrary assumption. Specifically, we show that there exists x ∈ Σ3 such that for all j ∈ N, vjf(hx )|xj |−1 ≥ c(hx + 1) 1 − f(hx + 1) . (1) Let x be a strategy profile which is obtained from x by a (one-step) addition of some resource a ∈ M by some agent i ∈ N(x) (note that x is nearly-even). Then, (1) is derived from and essentially equivalent to the inequality Uj(x ) ≥ Uj(x−j, xj {a}), for all a ∈ xj. That is, after performing an A-move with a by i, there is no profitable D-move with a. Then, by Lemmas 7 and 8, x is DS-stable. Following the same lines as above, we construct a procedure that initializes at x and achieves a strategy profile y ∈ Σ1 with hy > hx , in contradiction to x ∈ Σ2 . Now, let us confirm the existence of x ∈ Σ3 that satisfies (1). Let x ∈ Σ3 and let M(x) be the subset of all resources for which there exists a profitable (one-step) addition. First, we show that (1) holds for all j ∈ N such that xj ∩M(x) = ∅, that is, for all those agents with one of their resources being desired by another agent. Let a ∈ M(x), and let x be the strategy profile that is obtained from x by the (one-step) addition of a by agent i. Assume on the contrary that there is an agent j with a ∈ xj such that vjf(hx )|xj |−1 < c(hx + 1) 1 − f(hx + 1) . Let x = (x−j, xj {a}). Below we demonstrate that x is a DS-stable strategy profile and, since x and x correspond to the same congestion vector, we conclude that x lies in Σ2 . In addition, we show that UN (x ) > UN (x), contradicting the fact that x ∈ Σ3 . To show that x ∈ Σ0 we note that x is an even strategy profile, and thus no S-moves may be performed for x . In addition, since hx = hx and x ∈ Σ0 , there are no profitable D-moves for any agent k = i, j. It remains to show that there are no profitable D-moves for agents i and j as well. 215 Since Ui(x ) > Ui(x), we get vif(hx )|xi| > c(hx + 1) 1 − f(hx + 1) ⇒ vif(hx )|xi |−1 = vif(hx )|xi| > c(hx + 1) 1 − f(hx + 1) > c(hx ) 1 − f(hx) = c(hx ) 1 − f(hx ) , which implies Ui(x ) > Ui(x−i, xi {b}), for all b ∈ xi . Thus, there are no profitable D-moves for agent i. By the DS-stability of x, for agent j and for all b ∈ xj, we have Uj(x) ≥ Uj(x−j, xj {b}) ⇒ vjf(hx )|xj |−1 ≥ c(hx ) 1 − f(hx) . Then, vjf(hx )|xj |−1 > vjf(hx )|xj | = vjf(hx )|xj |−1 ≥ c(hx ) 1 − f(hx) = c(hx ) 1 − f(hx ) ⇒ Uj(x ) > Uj(x−j, xj {b}), for all b ∈ xi. Therefore, x is DS-stable and lies in Σ2 . To show that UN (x ), the group utility of x , satisfies UN (x ) > UN (x), we note that hx = hx , and thus Uk(x ) = Uk(x), for all k ∈ N {i, j}. Therefore, we have to show that Ui(x ) + Uj(x ) > Ui(x) + Uj(x), or Ui(x ) − Ui(x) > Uj(x) − Uj(x ). Observe that Ui(x ) > Ui(x) ⇒ vif(hx )|xi| > c(hx + 1) 1 − f(hx + 1) and Uj(x ) < Uj(x ) ⇒ vjf(hx )|xj |−1 < c(hx + 1) 1 − f(hx + 1) , which yields vif(hx )|xi| > vjf(hx )|xj |−1 . Thus, Ui(x ) − Ui(x) = 1 − f(hx )|xi|+1 vi − (|xi| + 1) c(hx ) − h 1 − f(hx )|xi| vi − |xi|c(hx ) i = vif(hx )|xi| (1 − f(hx )) − c(hx ) > vjf(hx )|xj |−1 (1 − f(hx )) − c(hx ) = 1 − f(hx )|xj | vj − |xj|c(hx ) − h 1 − f(hx )|xj |−1 vj − (|xi| − 1) c(hx ) i = Uj(x) − Uj(x ) . Therefore, x lies in Σ2 and satisfies UN (x ) > UN (x), in contradiction to x ∈ Σ3 . Hence, if x ∈ Σ3 then (1) holds for all j ∈ N such that xj ∩M(x) = ∅. Now let us see that there exists x ∈ Σ3 such that (1) holds for all the agents. For that, choose an agent i ∈ arg mink∈N vif(hx )|xk| . If there exists a ∈ xi ∩ M(x) then i satisfies (1), implying by the choice of agent i, that the above obviously yields the correctness of (1) for any agent k ∈ N. Otherwise, if no resource in xi lies in M(x), then let a ∈ xi and a ∈ M(x). Since a ∈ xi, a /∈ xi, and hx a = hx a , then there exists agent j such that a ∈ xj and a /∈ xj. One can easily check that the strategy profile x = ` x−{i,j}, (xi {a}) ∪ {a }, (xj {a }) ∪ {a} ´ lies in Σ3 . Thus, x satisfies (1) for agent i, and therefore, for any agent k ∈ N. Now, let x ∈ Σ3 satisfy (1). We show below that by performing a finite series of one- and two-step addition operations on x, we can achieve a strategy profile y that lies in Σ1 , such that hy > hx , in contradiction to x ∈ Σ2 . Let z ∈ Σ0 be a nearly-even (not necessarily even), DS-stable strategy profile, such that vi Y e∈zi {b} f(hz e) ≥ c(hz b + 1) 1 − f(hz b + 1) , (2) for all i ∈ N and for all z-light resource b ∈ zi. We note that for profile x ∈ Σ3 ⊆ Σ1 , with all resources being x-light, conditions (2) and (1) are equivalent. Let z be obtained from z by a one- or two-step addition of a z-light resource a. Obviously, z is nearly-even. In addition, hz e ≥ hz e for all e ∈ M, and mine∈M hz e ≥ mine∈M hz e. To complete the proof we need to show that z is DS-stable, and, in addition, that if mine∈M hz e = mine∈M hz e then z has property (2). The DS-stability of z follows directly from Lemmas 7 and 8, and from (2) with respect to z. It remains to prove property (2) for z with mine∈M hz e = mine∈M hz e. Using (2) with respect to z, for any agent k with zk = zk and for any zlight resource b ∈ zk, we get vk Y e∈zk {b} f(hz e ) ≥ vk Y e∈zk {b} f(hz e) ≥ c(hz b + 1) 1 − f(hz b + 1) = c(hz b + 1) 1 − f(hz b + 1) , as required. Now let us consider the rest of the agents. Assume z is obtained by the one-step addition of a by agent i. In this case, i is the only agent with zi = zi. The required property for agent i follows directly from Ui(z ) > Ui(z). In the case of a two-step addition, let z = ` z−{i,j}, zi ∪ {b}, (zj {b}) ∪ {a}), where b is a z-heavy resource. For agent i, from Ui(z−i, zi ∪ {b}) > Ui(z) we get 1 − Y e∈zi f(hz e)f(hz b + 1) ! vi − X e∈zi c(hz e) − c(hz b + 1) > 1 − Y e∈zi f(hz e) ! vi − X e∈zi c(hz e) ⇒ vi Y e∈zi f(hz e) > c(hz b + 1) 1 − f(hz b + 1) , (3) and note that since hz b ≥ hz e for all e ∈ M and, in particular, for all z -light resources, then c(hz b + 1) 1 − f(hz b + 1) ≥ c(hz e + 1) 1 − f(hz e + 1) , (4) for any z -light resource e . 216 Now, since hz e ≥ hz e for all e ∈ M and b is z-heavy, then vi Y e∈zi {e } f(hz e ) ≥ vi Y e∈zi {e } f(hz e) = vi Y e∈(zi∪{b}) {e } f(hz e) ≥ vi Y e∈zi f(hz e) , for any z -light resource e . The above, coupled with (3) and (4), yields the required. For agent j we just use (2) with respect to z and the equality hz b = hz a . For any z -light resource e , vj Y e∈zj {e } f(hz e ) ≥ vi Y e∈zi {e } f(hz e) ≥ c(hz e + 1) 1 − f(hz e + 1) = c(hz e + 1) 1 − f(hz e + 1) . Thus, since the number of resources is finite, there is a finite series of one- and two-step addition operations on x that leads to strategy profile y ∈ Σ1 with hy > hx , in contradiction to x ∈ Σ2 . This completes the proof. 4. DISCUSSION In this paper, we introduce and investigate congestion settings with unreliable resources, in which the probability of a resource"s failure depends on the congestion experienced by this resource. We defined a class of congestion games with load-dependent failures (CGLFs), which generalizes the wellknown class of congestion games. We study the existence of pure strategy Nash equilibria and potential functions in the presented class of games. We show that these games do not, in general, possess pure strategy equilibria. Nevertheless, if the resource cost functions are nondecreasing then such equilibria are guaranteed to exist, despite the non-existence of a potential function. The CGLF-model can be modified to the case where the agents pay only for non-faulty resources they selected. Both the model discussed in this paper and the modified one are reasonable. In the full version we will show that the modified model leads to similar results. In particular, we can show the existence of a pure strategy equilibrium for nondecreasing CGLFs also in the modified model. In future research we plan to consider various extensions of CGLFs. In particular, we plan to consider CGLFs where the resources may have different costs and failure probabilities, as well as CGLFs in which the resource failure probabilities are mutually dependent. In addition, it is of interest to develop an efficient algorithm for the computation of pure strategy Nash equilibrium, as well as discuss the social (in)efficiency of the equilibria. 5. REFERENCES [1] H. Ackermann, H. R¨oglin, and B. V¨ocking. Pure nash equilibria in player-specific and weighted congestion games. In WINE-06, 2006. [2] G. Christodoulou and E. Koutsoupias. The price of anarchy of finite congestion games. In Proceedings of the 37th Annual ACM Symposium on Theory and Computing (STOC-05), 2005. [3] A. Fabrikant, C. Papadimitriou, and K. Talwar. The complexity of pure nash equilibria. In STOC-04, pages 604-612, 2004. [4] E. Koutsoupias and C. Papadimitriou. Worst-case equilibria. In Proceedings of the 16th Annual Symposium on Theoretical Aspects of Computer Science, pages 404-413, 1999. [5] K. Leyton-Brown and M. Tennenholtz. Local-effect games. In IJCAI-03, 2003. [6] I. Milchtaich. Congestion games with player-specific payoff functions. Games and Economic Behavior, 13:111-124, 1996. [7] D. Monderer. Solution-based congestion games. Advances in Mathematical Economics, 8:397-407, 2006. [8] D. Monderer. Multipotential games. In IJCAI-07, 2007. [9] D. Monderer and L. Shapley. Potential games. Games and Economic Behavior, 14:124-143, 1996. [10] M. Penn, M. Polukarov, and M. Tennenholtz. Congestion games with failures. In Proceedings of the 6th ACM Conference on Electronic Commerce (EC-05), pages 259-268, 2005. [11] R. Rosenthal. A class of games possessing pure-strategy nash equilibria. International Journal of Game Theory, 2:65-67, 1973. [12] T. Roughgarden and E. Tardos. How bad is selfish routing. Journal of the ACM, 49(2):236-259, 2002. 217

potential function;nash equilibrium;nondecreasing cost function;resource cost function;pure strategy nash equilibrium;load-dependent failure;load-dependent resource failure;identical resource;real-valued function;failure probability;localeffect game;congestion game

train_C-58

A Scalable Distributed Information Management System∗

We present a Scalable Distributed Information Management System (SDIMS) that aggregates information about large-scale networked systems and that can serve as a basic building block for a broad range of large-scale distributed applications by providing detailed views of nearby information and summary views of global information. To serve as a basic building block, a SDIMS should have four properties: scalability to many nodes and attributes, flexibility to accommodate a broad range of applications, administrative isolation for security and availability, and robustness to node and network failures. We design, implement and evaluate a SDIMS that (1) leverages Distributed Hash Tables (DHT) to create scalable aggregation trees, (2) provides flexibility through a simple API that lets applications control propagation of reads and writes, (3) provides administrative isolation through simple extensions to current DHT algorithms, and (4) achieves robustness to node and network reconfigurations through lazy reaggregation, on-demand reaggregation, and tunable spatial replication. Through extensive simulations and micro-benchmark experiments, we observe that our system is an order of magnitude more scalable than existing approaches, achieves isolation properties at the cost of modestly increased read latency in comparison to flat DHTs, and gracefully handles failures.

1. INTRODUCTION The goal of this research is to design and build a Scalable Distributed Information Management System (SDIMS) that aggregates information about large-scale networked systems and that can serve as a basic building block for a broad range of large-scale distributed applications. Monitoring, querying, and reacting to changes in the state of a distributed system are core components of applications such as system management [15, 31, 37, 42], service placement [14, 43], data sharing and caching [18, 29, 32, 35, 46], sensor monitoring and control [20, 21], multicast tree formation [8, 9, 33, 36, 38], and naming and request routing [10, 11]. We therefore speculate that a SDIMS in a networked system would provide a distributed operating systems backbone and facilitate the development and deployment of new distributed services. For a large scale information system, hierarchical aggregation is a fundamental abstraction for scalability. Rather than expose all information to all nodes, hierarchical aggregation allows a node to access detailed views of nearby information and summary views of global information. In a SDIMS based on hierarchical aggregation, different nodes can therefore receive different answers to the query find a [nearby] node with at least 1 GB of free memory or find a [nearby] copy of file foo. A hierarchical system that aggregates information through reduction trees [21, 38] allows nodes to access information they care about while maintaining system scalability. To be used as a basic building block, a SDIMS should have four properties. First, the system should be scalable: it should accommodate large numbers of participating nodes, and it should allow applications to install and monitor large numbers of data attributes. Enterprise and global scale systems today might have tens of thousands to millions of nodes and these numbers will increase over time. Similarly, we hope to support many applications, and each application may track several attributes (e.g., the load and free memory of a system"s machines) or millions of attributes (e.g., which files are stored on which machines). Second, the system should have flexibility to accommodate a broad range of applications and attributes. For example, readdominated attributes like numCPUs rarely change in value, while write-dominated attributes like numProcesses change quite often. An approach tuned for read-dominated attributes will consume high bandwidth when applied to write-dominated attributes. Conversely, an approach tuned for write-dominated attributes will suffer from unnecessary query latency or imprecision for read-dominated attributes. Therefore, a SDIMS should provide mechanisms to handle different types of attributes and leave the policy decision of tuning replication to the applications. Third, a SDIMS should provide administrative isolation. In a large system, it is natural to arrange nodes in an organizational or an administrative hierarchy. A SDIMS should support administraSession 10: Distributed Information Systems 379 tive isolation in which queries about an administrative domain"s information can be satisfied within the domain so that the system can operate during disconnections from other domains, so that an external observer cannot monitor or affect intra-domain queries, and to support domain-scoped queries efficiently. Fourth, the system must be robust to node failures and disconnections. A SDIMS should adapt to reconfigurations in a timely fashion and should also provide mechanisms so that applications can tradeoff the cost of adaptation with the consistency level in the aggregated results when reconfigurations occur. We draw inspiration from two previous works: Astrolabe [38] and Distributed Hash Tables (DHTs). Astrolabe [38] is a robust information management system. Astrolabe provides the abstraction of a single logical aggregation tree that mirrors a system"s administrative hierarchy. It provides a general interface for installing new aggregation functions and provides eventual consistency on its data. Astrolabe is robust due to its use of an unstructured gossip protocol for disseminating information and its strategy of replicating all aggregated attribute values for a subtree to all nodes in the subtree. This combination allows any communication pattern to yield eventual consistency and allows any node to answer any query using local information. This high degree of replication, however, may limit the system"s ability to accommodate large numbers of attributes. Also, although the approach works well for read-dominated attributes, an update at one node can eventually affect the state at all nodes, which may limit the system"s flexibility to support write-dominated attributes. Recent research in peer-to-peer structured networks resulted in Distributed Hash Tables (DHTs) [18, 28, 29, 32, 35, 46]-a data structure that scales with the number of nodes and that distributes the read-write load for different queries among the participating nodes. It is interesting to note that although these systems export a global hash table abstraction, many of them internally make use of what can be viewed as a scalable system of aggregation trees to, for example, route a request for a given key to the right DHT node. Indeed, rather than export a general DHT interface, Plaxton et al."s [28] original application makes use of hierarchical aggregation to allow nodes to locate nearby copies of objects. It seems appealing to develop a SDIMS abstraction that exposes this internal functionality in a general way so that scalable trees for aggregation can be a basic system building block alongside the DHTs. At a first glance, it might appear to be obvious that simply fusing DHTs with Astrolabe"s aggregation abstraction will result in a SDIMS. However, meeting the SDIMS requirements forces a design to address four questions: (1) How to scalably map different attributes to different aggregation trees in a DHT mesh? (2) How to provide flexibility in the aggregation to accommodate different application requirements? (3) How to adapt a global, flat DHT mesh to attain administrative isolation property? and (4) How to provide robustness without unstructured gossip and total replication? The key contributions of this paper that form the foundation of our SDIMS design are as follows. 1. We define a new aggregation abstraction that specifies both attribute type and attribute name and that associates an aggregation function with a particular attribute type. This abstraction paves the way for utilizing the DHT system"s internal trees for aggregation and for achieving scalability with both nodes and attributes. 2. We provide a flexible API that lets applications control the propagation of reads and writes and thus trade off update cost, read latency, replication, and staleness. 3. We augment an existing DHT algorithm to ensure path convergence and path locality properties in order to achieve administrative isolation. 4. We provide robustness to node and network reconfigurations by (a) providing temporal replication through lazy reaggregation that guarantees eventual consistency and (b) ensuring that our flexible API allows demanding applications gain additional robustness by using tunable spatial replication of data aggregates or by performing fast on-demand reaggregation to augment the underlying lazy reaggregation or by doing both. We have built a prototype of SDIMS. Through simulations and micro-benchmark experiments on a number of department machines and PlanetLab [27] nodes, we observe that the prototype achieves scalability with respect to both nodes and attributes through use of its flexible API, inflicts an order of magnitude lower maximum node stress than unstructured gossiping schemes, achieves isolation properties at a cost of modestly increased read latency compared to flat DHTs, and gracefully handles node failures. This initial study discusses key aspects of an ongoing system building effort, but it does not address all issues in building a SDIMS. For example, we believe that our strategies for providing robustness will mesh well with techniques such as supernodes [22] and other ongoing efforts to improve DHTs [30] for further improving robustness. Also, although splitting aggregation among many trees improves scalability for simple queries, this approach may make complex and multi-attribute queries more expensive compared to a single tree. Additional work is needed to understand the significance of this limitation for real workloads and, if necessary, to adapt query planning techniques from DHT abstractions [16, 19] to scalable aggregation tree abstractions. In Section 2, we explain the hierarchical aggregation abstraction that SDIMS provides to applications. In Sections 3 and 4, we describe the design of our system for achieving the flexibility, scalability, and administrative isolation requirements of a SDIMS. In Section 5, we detail the implementation of our prototype system. Section 6 addresses the issue of adaptation to the topological reconfigurations. In Section 7, we present the evaluation of our system through large-scale simulations and microbenchmarks on real networks. Section 8 details the related work, and Section 9 summarizes our contribution. 2. AGGREGATION ABSTRACTION Aggregation is a natural abstraction for a large-scale distributed information system because aggregation provides scalability by allowing a node to view detailed information about the state near it and progressively coarser-grained summaries about progressively larger subsets of a system"s data [38]. Our aggregation abstraction is defined across a tree spanning all nodes in the system. Each physical node in the system is a leaf and each subtree represents a logical group of nodes. Note that logical groups can correspond to administrative domains (e.g., department or university) or groups of nodes within a domain (e.g., 10 workstations on a LAN in CS department). An internal non-leaf node, which we call virtual node, is simulated by one or more physical nodes at the leaves of the subtree for which the virtual node is the root. We describe how to form such trees in a later section. Each physical node has local data stored as a set of (attributeType, attributeName, value) tuples such as (configuration, numCPUs, 16), (mcast membership, session foo, yes), or (file stored, foo, myIPaddress). The system associates an aggregation function ftype with each attribute type, and for each level-i subtree Ti in the system, the system defines an aggregate value Vi,type,name for each (at380 tributeType, attributeName) pair as follows. For a (physical) leaf node T0 at level 0, V0,type,name is the locally stored value for the attribute type and name or NULL if no matching tuple exists. Then the aggregate value for a level-i subtree Ti is the aggregation function for the type, ftype computed across the aggregate values of each of Ti"s k children: Vi,type,name = ftype(V0 i−1,type,name,V1 i−1,type,name,...,Vk−1 i−1,type,name). Although SDIMS allows arbitrary aggregation functions, it is often desirable that these functions satisfy the hierarchical computation property [21]: f(v1,...,vn)= f(f(v1,...,vs1 ), f(vs1+1,...,vs2 ), ..., f(vsk+1,...,vn)), where vi is the value of an attribute at node i. For example, the average operation, defined as avg(v1,...,vn) = 1/n.∑n i=0 vi, does not satisfy the property. Instead, if an attribute stores values as tuples (sum,count), the attribute satisfies the hierarchical computation property while still allowing the applications to compute the average from the aggregate sum and count values. Finally, note that for a large-scale system, it is difficult or impossible to insist that the aggregation value returned by a probe corresponds to the function computed over the current values at the leaves at the instant of the probe. Therefore our system provides only weak consistency guarantees - specifically eventual consistency as defined in [38]. 3. FLEXIBILITY A major innovation of our work is enabling flexible aggregate computation and propagation. The definition of the aggregation abstraction allows considerable flexibility in how, when, and where aggregate values are computed and propagated. While previous systems [15, 29, 38, 32, 35, 46] implement a single static strategy, we argue that a SDIMS should provide flexible computation and propagation to efficiently support wide variety of applications with diverse requirements. In order to provide this flexibility, we develop a simple interface that decomposes the aggregation abstraction into three pieces of functionality: install, update, and probe. This definition of the aggregation abstraction allows our system to provide a continuous spectrum of strategies ranging from lazy aggregate computation and propagation on reads to aggressive immediate computation and propagation on writes. In Figure 1, we illustrate both extreme strategies and an intermediate strategy. Under the lazy Update-Local computation and propagation strategy, an update (or write) only affects local state. Then, a probe (or read) that reads a level-i aggregate value is sent up the tree to the issuing node"s level-i ancestor and then down the tree to the leaves. The system then computes the desired aggregate value at each layer up the tree until the level-i ancestor that holds the desired value. Finally, the level-i ancestor sends the result down the tree to the issuing node. In the other extreme case of the aggressive Update-All immediate computation and propagation on writes [38], when an update occurs, changes are aggregated up the tree, and each new aggregate value is flooded to all of a node"s descendants. In this case, each level-i node not only maintains the aggregate values for the level-i subtree but also receives and locally stores copies of all of its ancestors" level- j ( j > i) aggregation values. Also, a leaf satisfies a probe for a level-i aggregate using purely local data. In an intermediate Update-Up strategy, the root of each subtree maintains the subtree"s current aggregate value, and when an update occurs, the leaf node updates its local state and passes the update to its parent, and then each successive enclosing subtree updates its aggregate value and passes the new value to its parent. This strategy satisfies a leaf"s probe for a level-i aggregate value by sending the probe up to the level-i ancestor of the leaf and then sending the aggregate value down to the leaf. Finally, notice that other strategies exist. In general, an Update-Upk-Downj strategy aggregates up to parameter description optional attrType Attribute Type aggrfunc Aggregation Function up How far upward each update is sent (default: all) X down How far downward each aggregate is sent (default: none) X domain Domain restriction (default: none) X expTime Expiry Time Table 1: Arguments for the install operation the kth level and propagates the aggregate values of a node at level l (s.t. l ≤ k) downward for j levels. A SDIMS must provide a wide range of flexible computation and propagation strategies to applications for it to be a general abstraction. An application should be able to choose a particular mechanism based on its read-to-write ratio that reduces the bandwidth consumption while attaining the required responsiveness and precision. Note that the read-to-write ratio of the attributes that applications install vary extensively. For example, a read-dominated attribute like numCPUs rarely changes in value, while a writedominated attribute like numProcesses changes quite often. An aggregation strategy like Update-All works well for read-dominated attributes but suffers high bandwidth consumption when applied for write-dominated attributes. Conversely, an approach like UpdateLocal works well for write-dominated attributes but suffers from unnecessary query latency or imprecision for read-dominated attributes. SDIMS also allows non-uniform computation and propagation across the aggregation tree with different up and down parameters in different subtrees so that applications can adapt with the spatial and temporal heterogeneity of read and write operations. With respect to spatial heterogeneity, access patterns may differ for different parts of the tree, requiring different propagation strategies for different parts of the tree. Similarly with respect to temporal heterogeneity, access patterns may change over time requiring different strategies over time. 3.1 Aggregation API We provide the flexibility described above by splitting the aggregation API into three functions: Install() installs an aggregation function that defines an operation on an attribute type and specifies the update strategy that the function will use, Update() inserts or modifies a node"s local value for an attribute, and Probe() obtains an aggregate value for a specified subtree. The install interface allows applications to specify the k and j parameters of the Update-Upk-Downj strategy along with the aggregation function. The update interface invokes the aggregation of an attribute on the tree according to corresponding aggregation function"s aggregation strategy. The probe interface not only allows applications to obtain the aggregated value for a specified tree but also allows a probing node to continuously fetch the values for a specified time, thus enabling an application to adapt to spatial and temporal heterogeneity. The rest of the section describes these three interfaces in detail. 3.1.1 Install The Install operation installs an aggregation function in the system. The arguments for this operation are listed in Table 1. The attrType argument denotes the type of attributes on which this aggregation function is invoked. Installed functions are soft state that must be periodically renewed or they will be garbage collected at expTime. The arguments up and down specify the aggregate computation 381 Update Strategy On Update On Probe for Global Aggregate Value On Probe for Level-1 Aggregate Value Update-Local Update-Up Update-All Figure 1: Flexible API parameter description optional attrType Attribute Type attrName Attribute Name mode Continuous or One-shot (default: one-shot) X level Level at which aggregate is sought (default: at all levels) X up How far up to go and re-fetch the value (default: none) X down How far down to go and reaggregate (default: none) X expTime Expiry Time Table 2: Arguments for the probe operation and propagation strategy Update-Upk-Downj. The domain argument, if present, indicates that the aggregation function should be installed on all nodes in the specified domain; otherwise the function is installed on all nodes in the system. 3.1.2 Update The Update operation takes three arguments attrType, attrName, and value and creates a new (attrType, attrName, value) tuple or updates the value of an old tuple with matching attrType and attrName at a leaf node. The update interface meshes with installed aggregate computation and propagation strategy to provide flexibility. In particular, as outlined above and described in detail in Section 5, after a leaf applies an update locally, the update may trigger re-computation of aggregate values up the tree and may also trigger propagation of changed aggregate values down the tree. Notice that our abstraction associates an aggregation function with only an attrType but lets updates specify an attrName along with the attrType. This technique helps achieve scalability with respect to nodes and attributes as described in Section 4. 3.1.3 Probe The Probe operation returns the value of an attribute to an application. The complete argument set for the probe operation is shown in Table 2. Along with the attrName and the attrType arguments, a level argument specifies the level at which the answers are required for an attribute. In our implementation we choose to return results at all levels k < l for a level-l probe because (i) it is inexpensive as the nodes traversed for level-l probe also contain level k aggregates for k < l and as we expect the network cost of transmitting the additional information to be small for the small aggregates which we focus and (ii) it is useful as applications can efficiently get several aggregates with a single probe (e.g., for domain-scoped queries as explained in Section 4.2). Probes with mode set to continuous and with finite expTime enable applications to handle spatial and temporal heterogeneity. When node A issues a continuous probe at level l for an attribute, then regardless of the up and down parameters, updates for the attribute at any node in A"s level-l ancestor"s subtree are aggregated up to level l and the aggregated value is propagated down along the path from the ancestor to A. Note that continuous mode enables SDIMS to support a distributed sensor-actuator mechanism where a sensor monitors a level-i aggregate with a continuous mode probe and triggers an actuator upon receiving new values for the probe. The up and down arguments enable applications to perform ondemand fast re-aggregation during reconfigurations, where a forced re-aggregation is done for the corresponding levels even if the aggregated value is available, as we discuss in Section 6. When present, the up and down arguments are interpreted as described in the install operation. 3.1.4 Dynamic Adaptation At the API level, the up and down arguments in install API can be regarded as hints, since they suggest a computation strategy but do not affect the semantics of an aggregation function. A SDIMS implementation can dynamically adjust its up/down strategies for an attribute based on its measured read/write frequency. But a virtual intermediate node needs to know the current up and down propagation values to decide if the local aggregate is fresh in order to answer a probe. This is the key reason why up and down need to be statically defined at the install time and can not be specified in the update operation. In dynamic adaptation, we implement a leasebased mechanism where a node issues a lease to a parent or a child denoting that it will keep propagating the updates to that parent or child. We are currently evaluating different policies to decide when to issue a lease and when to revoke a lease. 4. SCALABILITY Our design achieves scalability with respect to both nodes and attributes through two key ideas. First, it carefully defines the aggregation abstraction to mesh well with its underlying scalable DHT system. Second, it refines the basic DHT abstraction to form an Autonomous DHT (ADHT) to achieve the administrative isolation properties that are crucial to scaling for large real-world systems. In this section, we describe these two ideas in detail. 4.1 Leveraging DHTs In contrast to previous systems [4, 15, 38, 39, 45], SDIMS"s aggregation abstraction specifies both an attribute type and attribute name and associates an aggregation function with a type rather than just specifying and associating a function with a name. Installing a single function that can operate on many different named attributes matching a type improves scalability for sparse attribute types with large, sparsely-filled name spaces. For example, to construct a file location service, our interface allows us to install a single function that computes an aggregate value for any named file. A subtree"s aggregate value for (FILELOC, name) would be the ID of a node in the subtree that stores the named file. Conversely, Astrolabe copes with sparse attributes by having aggregation functions compute sets or lists and suggests that scalability can be improved by representing such sets with Bloom filters [6]. Supporting sparse names within a type provides at least two advantages. First, when the value associated with a name is updated, only the state associ382 001 010100 000 011 101 111 110 011 111 001 101 000 100 110010 L0 L1 L2 L3 Figure 2: The DHT tree corresponding to key 111 (DHTtree111) and the corresponding aggregation tree. ated with that name needs to be updated and propagated to other nodes. Second, splitting values associated with different names into different aggregation values allows our system to leverage Distributed Hash Tables (DHTs) to map different names to different trees and thereby spread the function"s logical root node"s load and state across multiple physical nodes. Given this abstraction, scalably mapping attributes to DHTs is straightforward. DHT systems assign a long, random ID to each node and define an algorithm to route a request for key k to a node rootk such that the union of paths from all nodes forms a tree DHTtreek rooted at the node rootk. Now, as illustrated in Figure 2, by aggregating an attribute along the aggregation tree corresponding to DHTtreek for k =hash(attribute type, attribute name), different attributes will be aggregated along different trees. In comparison to a scheme where all attributes are aggregated along a single tree, aggregating along multiple trees incurs lower maximum node stress: whereas in a single aggregation tree approach, the root and the intermediate nodes pass around more messages than leaf nodes, in a DHT-based multi-tree, each node acts as an intermediate aggregation point for some attributes and as a leaf node for other attributes. Hence, this approach distributes the onus of aggregation across all nodes. 4.2 Administrative Isolation Aggregation trees should provide administrative isolation by ensuring that for each domain, the virtual node at the root of the smallest aggregation subtree containing all nodes of that domain is hosted by a node in that domain. Administrative isolation is important for three reasons: (i) for security - so that updates and probes flowing in a domain are not accessible outside the domain, (ii) for availability - so that queries for values in a domain are not affected by failures of nodes in other domains, and (iii) for efficiency - so that domain-scoped queries can be simple and efficient. To provide administrative isolation to aggregation trees, a DHT should satisfy two properties: 1. Path Locality: Search paths should always be contained in the smallest possible domain. 2. Path Convergence: Search paths for a key from different nodes in a domain should converge at a node in that domain. Existing DHTs support path locality [18] or can easily support it by using the domain nearness as the distance metric [7, 17], but they do not guarantee path convergence as those systems try to optimize the search path to the root to reduce response latency. For example, Pastry [32] uses prefix routing in which each node"s routing table contains one row per hexadecimal digit in the nodeId space where the ith row contains a list of nodes whose nodeIds differ from the current node"s nodeId in the ith digit with one entry for each possible digit value. Given a routing topology, to route a packet to an arbitrary destination key, a node in Pastry forwards a packet to the node with a nodeId prefix matching the key in at least one more digit than the current node. If such a node is not known, the current node uses an additional data structure, the leaf set containing 110XX 010XX 011XX 100XX 101XX univ dep1 dep2 key = 111XX 011XX 100XX 101XX 110XX 010XX L1 L0 L2 Figure 3: Example shows how isolation property is violated with original Pastry. We also show the corresponding aggregation tree. 110XX 010XX 011XX 100XX 101XX univ dep1 dep2 key = 111XX X 011XX 100XX 101XX 110XX 010XX L0 L1 L2 Figure 4: Autonomous DHT satisfying the isolation property. Also the corresponding aggregation tree is shown. L immediate higher and lower neighbors in the nodeId space, and forwards the packet to a node with an identical prefix but that is numerically closer to the destination key in the nodeId space. This process continues until the destination node appears in the leaf set, after which the message is routed directly. Pastry"s expected number of routing steps is logn, where n is the number of nodes, but as Figure 3 illustrates, this algorithm does not guarantee path convergence: if two nodes in a domain have nodeIds that match a key in the same number of bits, both of them can route to a third node outside the domain when routing for that key. Simple modifications to Pastry"s route table construction and key-routing protocols yield an Autonomous DHT (ADHT) that satisfies the path locality and path convergence properties. As Figure 4 illustrates, whenever two nodes in a domain share the same prefix with respect to a key and no other node in the domain has a longer prefix, our algorithm introduces a virtual node at the boundary of the domain corresponding to that prefix plus the next digit of the key; such a virtual node is simulated by the existing node whose id is numerically closest to the virtual node"s id. Our ADHT"s routing table differs from Pastry"s in two ways. First, each node maintains a separate leaf set for each domain of which it is a part. Second, nodes use two proximity metrics when populating the routing tables - hierarchical domain proximity is the primary metric and network distance is secondary. Then, to route a packet to a global root for a key, ADHT routing algorithm uses the routing table and the leaf set entries to route to each successive enclosing domain"s root (the virtual or real node in the domain matching the key in the maximum number of digits). Additional details about the ADHT algorithm are available in an extended technical report [44]. Properties. Maintaining a different leaf set for each administrative hierarchy level increases the number of neighbors that each node tracks to (2b)∗lgb n+c.l from (2b)∗lgb n+c in unmodified Pastry, where b is the number of bits in a digit, n is the number of nodes, c is the leaf set size, and l is the number of domain levels. Routing requires O(lgbn + l) steps compared to O(lgbn) steps in Pastry; also, each routing hop may be longer than in Pastry because the modified algorithm"s routing table prefers same-domain nodes over nearby nodes. We experimentally quantify the additional routing costs in Section 7. In a large system, the ADHT topology allows domains to im383 A1 A2 B1 ((B1.B.,1), (B.,1),(.,1)) ((B1.B.,1), (B.,1),(.,1)) L2 L1 L0 ((B1.B.,1), (B.,1),(.,3)) ((A1.A.,1), (A.,2),(.,2)) ((A1.A.,1), (A.,1),(.,1)) ((A2.A.,1), (A.,1),(.,1)) Figure 5: Example for domain-scoped queries prove security for sensitive attribute types by installing them only within a specified domain. Then, aggregation occurs entirely within the domain and a node external to the domain can neither observe nor affect the updates and aggregation computations of the attribute type. Furthermore, though we have not implemented this feature in the prototype, the ADHT topology would also support domainrestricted probes that could ensure that no one outside of a domain can observe a probe for data stored within the domain. The ADHT topology also enhances availability by allowing the common case of probes for data within a domain to depend only on a domain"s nodes. This, for example, allows a domain that becomes disconnected from the rest of the Internet to continue to answer queries for local data. Aggregation trees that provide administrative isolation also enable the definition of simple and efficient domain-scoped aggregation functions to support queries like what is the average load on machines in domain X? For example, consider an aggregation function to count the number of machines in an example system with three machines illustrated in Figure 5. Each leaf node l updates attribute NumMachines with a value vl containing a set of tuples of form (Domain, Count) for each domain of which the node is a part. In the example, the node A1 with name A1.A. performs an update with the value ((A1.A.,1),(A.,1),(.,1)). An aggregation function at an internal virtual node hosted on node N with child set C computes the aggregate as a set of tuples: for each domain D that N is part of, form a tuple (D,∑c∈C(count|(D,count) ∈ vc)). This computation is illustrated in the Figure 5. Now a query for NumMachines with level set to MAX will return the aggregate values at each intermediate virtual node on the path to the root as a set of tuples (tree level, aggregated value) from which it is easy to extract the count of machines at each enclosing domain. For example, A1 would receive ((2, ((B1.B.,1),(B.,1),(.,3))), (1, ((A1.A.,1),(A.,2),(.,2))), (0, ((A1.A.,1),(A.,1),(.,1)))). Note that supporting domain-scoped queries would be less convenient and less efficient if aggregation trees did not conform to the system"s administrative structure. It would be less efficient because each intermediate virtual node will have to maintain a list of all values at the leaves in its subtree along with their names and it would be less convenient as applications that need an aggregate for a domain will have to pick values of nodes in that domain from the list returned by a probe and perform computation. 5. PROTOTYPE IMPLEMENTATION The internal design of our SDIMS prototype comprises of two layers: the Autonomous DHT (ADHT) layer manages the overlay topology of the system and the Aggregation Management Layer (AML) maintains attribute tuples, performs aggregations, stores and propagates aggregate values. Given the ADHT construction described in Section 4.2, each node implements an Aggregation Management Layer (AML) to support the flexible API described in Section 3. In this section, we describe the internal state and operation of the AML layer of a node in the system. local MIB MIBs ancestor reduction MIB (level 1)MIBs ancestor MIB from child 0X... MIB from child 0X... Level 2 Level 1 Level 3 Level 0 1XXX... 10XX... 100X... From parents0X.. To parent 0X... −− aggregation functions From parents To parent 10XX... 1X.. 1X.. 1X.. To parent 11XX... Node Id: (1001XXX) 1001X.. 100X.. 10X.. 1X.. Virtual Node Figure 6: Example illustrating the data structures and the organization of them at a node. We refer to a store of (attribute type, attribute name, value) tuples as a Management Information Base or MIB, following the terminology from Astrolabe [38] and SNMP [34]. We refer an (attribute type, attribute name) tuple as an attribute key. As Figure 6 illustrates, each physical node in the system acts as several virtual nodes in the AML: a node acts as leaf for all attribute keys, as a level-1 subtree root for keys whose hash matches the node"s ID in b prefix bits (where b is the number of bits corrected in each step of the ADHT"s routing scheme), as a level-i subtree root for attribute keys whose hash matches the node"s ID in the initial i ∗ b bits, and as the system"s global root for attribute keys whose hash matches the node"s ID in more prefix bits than any other node (in case of a tie, the first non-matching bit is ignored and the comparison is continued [46]). To support hierarchical aggregation, each virtual node at the root of a level-i subtree maintains several MIBs that store (1) child MIBs containing raw aggregate values gathered from children, (2) a reduction MIB containing locally aggregated values across this raw information, and (3) an ancestor MIB containing aggregate values scattered down from ancestors. This basic strategy of maintaining child, reduction, and ancestor MIBs is based on Astrolabe [38], but our structured propagation strategy channels information that flows up according to its attribute key and our flexible propagation strategy only sends child updates up and ancestor aggregate results down as far as specified by the attribute key"s aggregation function. Note that in the discussion below, for ease of explanation, we assume that the routing protocol is correcting single bit at a time (b = 1). Our system, built upon Pastry, handles multi-bit correction (b = 4) and is a simple extension to the scheme described here. For a given virtual node ni at level i, each child MIB contains the subset of a child"s reduction MIB that contains tuples that match ni"s node ID in i bits and whose up aggregation function attribute is at least i. These local copies make it easy for a node to recompute a level-i aggregate value when one child"s input changes. Nodes maintain their child MIBs in stable storage and use a simplified version of the Bayou log exchange protocol (sans conflict detection and resolution) for synchronization after disconnections [26]. Virtual node ni at level i maintains a reduction MIB of tuples with a tuple for each key present in any child MIB containing the attribute type, attribute name, and output of the attribute type"s aggregate functions applied to the children"s tuples. A virtual node ni at level i also maintains an ancestor MIB to store the tuples containing attribute key and a list of aggregate values at different levels scattered down from ancestors. Note that the 384 list for a key might contain multiple aggregate values for a same level but aggregated at different nodes (see Figure 4). So, the aggregate values are tagged not only with level information, but are also tagged with ID of the node that performed the aggregation. Level-0 differs slightly from other levels. Each level-0 leaf node maintains a local MIB rather than maintaining child MIBs and a reduction MIB. This local MIB stores information about the local node"s state inserted by local applications via update() calls. We envision various sensor programs and applications insert data into local MIB. For example, one program might monitor local configuration and perform updates with information such as total memory, free memory, etc., A distributed file system might perform update for each file stored on the local node. Along with these MIBs, a virtual node maintains two other tables: an aggregation function table and an outstanding probes table. An aggregation function table contains the aggregation function and installation arguments (see Table 1) associated with an attribute type or an attribute type and name. Each aggregate function is installed on all nodes in a domain"s subtree, so the aggregate function table can be thought of as a special case of the ancestor MIB with domain functions always installed up to a root within a specified domain and down to all nodes within the domain. The outstanding probes table maintains temporary information regarding in-progress probes. Given these data structures, it is simple to support the three API functions described in Section 3.1. Install The Install operation (see Table 1) installs on a domain an aggregation function that acts on a specified attribute type. Execution of an install operation for function aggrFunc on attribute type attrType proceeds in two phases: first the install request is passed up the ADHT tree with the attribute key (attrType, null) until it reaches the root for that key within the specified domain. Then, the request is flooded down the tree and installed on all intermediate and leaf nodes. Update When a level i virtual node receives an update for an attribute from a child below: it first recomputes the level-i aggregate value for the specified key, stores that value in its reduction MIB and then, subject to the function"s up and domain parameters, passes the updated value to the appropriate parent based on the attribute key. Also, the level-i (i ≥ 1) virtual node sends the updated level-i aggregate to all its children if the function"s down parameter exceeds zero. Upon receipt of a level-i aggregate from a parent, a level k virtual node stores the value in its ancestor MIB and, if k ≥ i−down, forwards this aggregate to its children. Probe A Probe collects and returns the aggregate value for a specified attribute key for a specified level of the tree. As Figure 1 illustrates, the system satisfies a probe for a level-i aggregate value using a four-phase protocol that may be short-circuited when updates have previously propagated either results or partial results up or down the tree. In phase 1, the route probe phase, the system routes the probe up the attribute key"s tree to either the root of the level-i subtree or to a node that stores the requested value in its ancestor MIB. In the former case, the system proceeds to phase 2 and in the latter it skips to phase 4. In phase 2, the probe scatter phase, each node that receives a probe request sends it to all of its children unless the node"s reduction MIB already has a value that matches the probe"s attribute key, in which case the node initiates phase 3 on behalf of its subtree. In phase 3, the probe aggregation phase, when a node receives values for the specified key from each of its children, it executes the aggregate function on these values and either (a) forwards the result to its parent (if its level is less than i) or (b) initiates phase 4 (if it is at level i). Finally, in phase 4, the aggregate routing phase the aggregate value is routed down to the node that requested it. Note that in the extreme case of a function installed with up = down = 0, a level-i probe can touch all nodes in a level-i subtree while in the opposite extreme case of a function installed with up = down = ALL, probe is a completely local operation at a leaf. For probes that include phases 2 (probe scatter) and 3 (probe aggregation), an issue is how to decide when a node should stop waiting for its children to respond and send up its current aggregate value. A node stops waiting for its children when one of three conditions occurs: (1) all children have responded, (2) the ADHT layer signals one or more reconfiguration events that mark all children that have not yet responded as unreachable, or (3) a watchdog timer for the request fires. The last case accounts for nodes that participate in the ADHT protocol but that fail at the AML level. At a virtual node, continuous probes are handled similarly as one-shot probes except that such probes are stored in the outstanding probe table for a time period of expTime specified in the probe. Thus each update for an attribute triggers re-evaluation of continuous probes for that attribute. We implement a lease-based mechanism for dynamic adaptation. A level-l virtual node for an attribute can issue the lease for levell aggregate to a parent or a child only if up is greater than l or it has leases from all its children. A virtual node at level l can issue the lease for level-k aggregate for k > l to a child only if down≥ k −l or if it has the lease for that aggregate from its parent. Now a probe for level-k aggregate can be answered by level-l virtual node if it has a valid lease, irrespective of the up and down values. We are currently designing different policies to decide when to issue a lease and when to revoke a lease and are also evaluating them with the above mechanism. Our current prototype does not implement access control on install, update, and probe operations but we plan to implement Astrolabe"s [38] certificate-based restrictions. Also our current prototype does not restrict the resource consumption in executing the aggregation functions; but, ‘techniques from research on resource management in server systems and operating systems [2, 3] can be applied here. 6. ROBUSTNESS In large scale systems, reconfigurations are common. Our two main principles for robustness are to guarantee (i) read availability - probes complete in finite time, and (ii) eventual consistency - updates by a live node will be visible to probes by connected nodes in finite time. During reconfigurations, a probe might return a stale value for two reasons. First, reconfigurations lead to incorrectness in the previous aggregate values. Second, the nodes needed for aggregation to answer the probe become unreachable. Our system also provides two hooks that applications can use for improved end-to-end robustness in the presence of reconfigurations: (1) Ondemand re-aggregation and (2) application controlled replication. Our system handles reconfigurations at two levels - adaptation at the ADHT layer to ensure connectivity and adaptation at the AML layer to ensure access to the data in SDIMS. 6.1 ADHT Adaptation Our ADHT layer adaptation algorithm is same as Pastry"s adaptation algorithm [32] - the leaf sets are repaired as soon as a reconfiguration is detected and the routing table is repaired lazily. Note that maintaining extra leaf sets does not degrade the fault-tolerance property of the original Pastry; indeed, it enhances the resilience of ADHTs to failures by providing additional routing links. Due to redundancy in the leaf sets and the routing table, updates can be routed towards their root nodes successfully even during failures. 385 Reconfig reconfig notices DHT partial DHT complete DHT ends Lazy Time Data 3 7 81 2 4 5 6starts Lazy Data starts Lazy Data starts Lazy Data repairrepair reaggr reaggr reaggr reaggr happens Figure 7: Default lazy data re-aggregation time line Also note that the administrative isolation property satisfied by our ADHT algorithm ensures that the reconfigurations in a level i domain do not affect the probes for level i in a sibling domain. 6.2 AML Adaptation Broadly, we use two types of strategies for AML adaptation in the face of reconfigurations: (1) Replication in time as a fundamental baseline strategy, and (2) Replication in space as an additional performance optimization that falls back on replication in time when the system runs out of replicas. We provide two mechanisms for replication in time. First, lazy re-aggregation propagates already received updates to new children or new parents in a lazy fashion over time. Second, applications can reduce the probability of probe response staleness during such repairs through our flexible API with appropriate setting of the down parameter. Lazy Re-aggregation: The DHT layer informs the AML layer about reconfigurations in the network using the following three function calls - newParent, failedChild, and newChild. On newParent(parent, prefix), all probes in the outstanding-probes table corresponding to prefix are re-evaluated. If parent is not null, then aggregation functions and already existing data are lazily transferred in the background. Any new updates, installs, and probes for this prefix are sent to the parent immediately. On failedChild(child, prefix), the AML layer marks the child as inactive and any outstanding probes that are waiting for data from this child are re-evaluated. On newChild(child, prefix), the AML layer creates space in its data structures for this child. Figure 7 shows the time line for the default lazy re-aggregation upon reconfiguration. Probes initiated between points 1 and 2 and that are affected by reconfigurations are reevaluated by AML upon detecting the reconfiguration. Probes that complete or start between points 2 and 8 may return stale answers. On-demand Re-aggregation: The default lazy aggregation scheme lazily propagates the old updates in the system. Additionally, using up and down knobs in the Probe API, applications can force on-demand fast re-aggregation of updates to avoid staleness in the face of reconfigurations. In particular, if an application detects or suspects an answer as stale, then it can re-issue the probe increasing the up and down parameters to force the refreshing of the cached data. Note that this strategy will be useful only after the DHT adaptation is completed (Point 6 on the time line in Figure 7). Replication in Space: Replication in space is more challenging in our system than in a DHT file location application because replication in space can be achieved easily in the latter by just replicating the root node"s contents. In our system, however, all internal nodes have to be replicated along with the root. In our system, applications control replication in space using up and down knobs in the Install API; with large up and down values, aggregates at the intermediate virtual nodes are propagated to more nodes in the system. By reducing the number of nodes that have to be accessed to answer a probe, applications can reduce the probability of incorrect results occurring due to the failure of nodes that do not contribute to the aggregate. For example, in a file location application, using a non-zero positive down parameter ensures that a file"s global aggregate is replicated on nodes other than the root. 0.1 1 10 100 1000 10000 0.0001 0.01 1 100 10000 Avg.numberofmessagesperoperation Read to Write ratio Update-All Up=ALL, Down=9 Up=ALL, Down=6 Update-Up Update-Local Up=2, Down=0 Up=5, Down=0 Figure 8: Flexibility of our approach. With different UP and DOWN values in a network of 4096 nodes for different readwrite ratios. Probes for the file location can then be answered without accessing the root; hence they are not affected by the failure of the root. However, note that this technique is not appropriate in some cases. An aggregated value in file location system is valid as long as the node hosting the file is active, irrespective of the status of other nodes in the system; whereas an application that counts the number of machines in a system may receive incorrect results irrespective of the replication. If reconfigurations are only transient (like a node temporarily not responding due to a burst of load), the replicated aggregate closely or correctly resembles the current state. 7. EVALUATION We have implemented a prototype of SDIMS in Java using the FreePastry framework [32] and performed large-scale simulation experiments and micro-benchmark experiments on two real networks: 187 machines in the department and 69 machines on the PlanetLab [27] testbed. In all experiments, we use static up and down values and turn off dynamic adaptation. Our evaluation supports four main conclusions. First, flexible API provides different propagation strategies that minimize communication resources at different read-to-write ratios. For example, in our simulation we observe Update-Local to be efficient for read-to-write ratios below 0.0001, Update-Up around 1, and Update-All above 50000. Second, our system is scalable with respect to both nodes and attributes. In particular, we find that the maximum node stress in our system is an order lower than observed with an Update-All, gossiping approach. Third, in contrast to unmodified Pastry which violates path convergence property in upto 14% cases, our system conforms to the property. Fourth, the system is robust to reconfigurations and adapts to failures with in a few seconds. 7.1 Simulation Experiments Flexibility and Scalability: A major innovation of our system is its ability to provide flexible computation and propagation of aggregates. In Figure 8, we demonstrate the flexibility exposed by the aggregation API explained in Section 3. We simulate a system with 4096 nodes arranged in a domain hierarchy with branching factor (bf) of 16 and install several attributes with different up and down parameters. We plot the average number of messages per operation incurred for a wide range of read-to-write ratios of the operations for different attributes. Simulations with other sizes of networks with different branching factors reveal similar results. This graph clearly demonstrates the benefit of supporting a wide range of computation and propagation strategies. Although having a small UP 386 1 10 100 1000 10000 100000 1e+06 1e+07 1 10 100 1000 10000 100000 MaximumNodeStress Number of attributes installed Gossip 256 Gossip 4096 Gossip 65536 DHT 256 DHT 4096 DHT 65536 Figure 9: Max node stress for a gossiping approach vs. ADHT based approach for different number of nodes with increasing number of sparse attributes. value is efficient for attributes with low read-to-write ratios (write dominated applications), the probe latency, when reads do occur, may be high since the probe needs to aggregate the data from all the nodes that did not send their aggregate up. Conversely, applications that wish to improve probe overheads or latencies can increase their UP and DOWN propagation at a potential cost of increase in write overheads. Compared to an existing Update-all single aggregation tree approach [38], scalability in SDIMS comes from (1) leveraging DHTs to form multiple aggregation trees that split the load across nodes and (2) flexible propagation that avoids propagation of all updates to all nodes. Figure 9 demonstrates the SDIMS"s scalability with nodes and attributes. For this experiment, we build a simulator to simulate both Astrolabe [38] (a gossiping, Update-All approach) and our system for an increasing number of sparse attributes. Each attribute corresponds to the membership in a multicast session with a small number of participants. For this experiment, the session size is set to 8, the branching factor is set to 16, the propagation mode for SDIMS is Update-Up, and the participant nodes perform continuous probes for the global aggregate value. We plot the maximum node stress (in terms of messages) observed in both schemes for different sized networks with increasing number of sessions when the participant of each session performs an update operation. Clearly, the DHT based scheme is more scalable with respect to attributes than an Update-all gossiping scheme. Observe that at some constant number of attributes, as the number of nodes increase in the system, the maximum node stress increases in the gossiping approach, while it decreases in our approach as the load of aggregation is spread across more nodes. Simulations with other session sizes (4 and 16) yield similar results. Administrative Hierarchy and Robustness: Although the routing protocol of ADHT might lead to an increased number of hops to reach the root for a key as compared to original Pastry, the algorithm conforms to the path convergence and locality properties and thus provides administrative isolation property. In Figure 10, we quantify the increased path length by comparisons with unmodified Pastry for different sized networks with different branching factors of the domain hierarchy tree. To quantify the path convergence property, we perform simulations with a large number of probe pairs - each pair probing for a random key starting from two randomly chosen nodes. In Figure 11, we plot the percentage of probe pairs for unmodified pastry that do not conform to the path convergence property. When the branching factor is low, the domain hierarchy tree is deeper resulting in a large difference between 0 1 2 3 4 5 6 7 10 100 1000 10000 100000 PathLength Number of Nodes ADHT bf=4 ADHT bf=16 ADHT bf=64 PASTRY bf=4,16,64 Figure 10: Average path length to root in Pastry versus ADHT for different branching factors. Note that all lines corresponding to Pastry overlap. 0 2 4 6 8 10 12 14 16 10 100 1000 10000 100000 Percentageofviolations Number of Nodes bf=4 bf=16 bf=64 Figure 11: Percentage of probe pairs whose paths to the root did not conform to the path convergence property with Pastry. U pdate-All U pdate-U p U pdate-Local 0 200 400 600 800 Latency(inms) Average Latency U pdate-All U pdate-U p U pdate-Local 0 1000 2000 3000 Latency(inms) Average Latency (a) (b) Figure 12: Latency of probes for aggregate at global root level with three different modes of aggregate propagation on (a) department machines, and (b) PlanetLab machines Pastry and ADHT in the average path length; but it is at these small domain sizes, that the path convergence fails more often with the original Pastry. 7.2 Testbed experiments We run our prototype on 180 department machines (some machines ran multiple node instances, so this configuration has a total of 283 SDIMS nodes) and also on 69 machines of the PlanetLab [27] testbed. We measure the performance of our system with two micro-benchmarks. In the first micro-benchmark, we install three aggregation functions of types Update-Local, Update-Up, and Update-All, perform update operation on all nodes for all three aggregation functions, and measure the latencies incurred by probes for the global aggregate from all nodes in the system. Figure 12 387 0 20 40 60 80 100 120 140 0 5 10 15 20 25 2700 2720 2740 2760 2780 2800 2820 2840 Latency(inms) ValuesObserved Time(in sec) Values latency Node Killed Figure 13: Micro-benchmark on department network showing the behavior of the probes from a single node when failures are happening at some other nodes. All 283 nodes assign a value of 10 to the attribute. 10 100 1000 10000 100000 0 50 100 150 200 250 300 350 400 450 500 500 550 600 650 700 Latency(inms) ValuesObserved Time(in sec) Values latency Node Killed Figure 14: Probe performance during failures on 69 machines of PlanetLab testbed shows the observed latencies for both testbeds. Notice that the latency in Update-Local is high compared to the Update-UP policy. This is because latency in Update-Local is affected by the presence of even a single slow machine or a single machine with a high latency network connection. In the second benchmark, we examine robustness. We install one aggregation function of type Update-Up that performs sum operation on an integer valued attribute. Each node updates the attribute with the value 10. Then we monitor the latencies and results returned on the probe operation for global aggregate on one chosen node, while we kill some nodes after every few probes. Figure 13 shows the results on the departmental testbed. Due to the nature of the testbed (machines in a department), there is little change in the latencies even in the face of reconfigurations. In Figure 14, we present the results of the experiment on PlanetLab testbed. The root node of the aggregation tree is terminated after about 275 seconds. There is a 5X increase in the latencies after the death of the initial root node as a more distant node becomes the root node after repairs. In both experiments, the values returned on probes start reflecting the correct situation within a short time after the failures. From both the testbed benchmark experiments and the simulation experiments on flexibility and scalability, we conclude that (1) the flexibility provided by SDIMS allows applications to tradeoff read-write overheads (Figure 8), read latency, and sensitivity to slow machines (Figure 12), (2) a good default aggregation strategy is Update-Up which has moderate overheads on both reads and writes (Figure 8), has moderate read latencies (Figure 12), and is scalable with respect to both nodes and attributes (Figure 9), and (3) small domain sizes are the cases where DHT algorithms fail to provide path convergence more often and SDIMS ensures path convergence with only a moderate increase in path lengths (Figure 11). 7.3 Applications SDIMS is designed as a general distributed monitoring and control infrastructure for a broad range of applications. Above, we discuss some simple microbenchmarks including a multicast membership service and a calculate-sum function. Van Renesse et al. [38] provide detailed examples of how such a service can be used for a peer-to-peer caching directory, a data-diffusion service, a publishsubscribe system, barrier synchronization, and voting. Additionally, we have initial experience using SDIMS to construct two significant applications: the control plane for a large-scale distributed file system [12] and a network monitor for identifying heavy hitters that consume excess resources. Distributed file system control: The PRACTI (Partial Replication, Arbitrary Consistency, Topology Independence) replication system provides a set of mechanisms for data replication over which arbitrary control policies can be layered. We use SDIMS to provide several key functions in order to create a file system over the lowlevel PRACTI mechanisms. First, nodes use SDIMS as a directory to handle read misses. When a node n receives an object o, it updates the (ReadDir, o) attribute with the value n; when n discards o from its local store, it resets (ReadDir, o) to NULL. At each virtual node, the ReadDir aggregation function simply selects a random non-null child value (if any) and we use the Update-Up policy for propagating updates. Finally, to locate a nearby copy of an object o, a node n1 issues a series of probe requests for the (ReadDir, o) attribute, starting with level = 1 and increasing the level value with each repeated probe request until a non-null node ID n2 is returned. n1 then sends a demand read request to n2, and n2 sends the data if it has it. Conversely, if n2 does not have a copy of o, it sends a nack to n1, and n1 issues a retry probe with the down parameter set to a value larger than used in the previous probe in order to force on-demand re-aggregation, which will yield a fresher value for the retry. Second, nodes subscribe to invalidations and updates to interest sets of files, and nodes use SDIMS to set up and maintain perinterest-set network-topology-sensitive spanning trees for propagating this information. To subscribe to invalidations for interest set i, a node n1 first updates the (Inval, i) attribute with its identity n1, and the aggregation function at each virtual node selects one non-null child value. Finally, n1 probes increasing levels of the the (Inval, i) attribute until it finds the first node n2 = n1; n1 then uses n2 as its parent in the spanning tree. n1 also issues a continuous probe for this attribute at this level so that it is notified of any change to its spanning tree parent. Spanning trees for streams of pushed updates are maintained in a similar manner. In the future, we plan to use SDIMS for at least two additional services within this replication system. First, we plan to use SDIMS to track the read and write rates to different objects; prefetch algorithms will use this information to prioritize replication [40, 41]. Second, we plan to track the ranges of invalidation sequence numbers seen by each node for each interest set in order to augment the spanning trees described above with additional hole filling to allow nodes to locate specific invalidations they have missed. Overall, our initial experience with using SDIMS for the PRACTII replication system suggests that (1) the general aggregation interface provided by SDIMS simplifies the construction of distributed applications-given the low-level PRACTI mechanisms, 388 we were able to construct a basic file system that uses SDIMS for several distinct control tasks in under two weeks and (2) the weak consistency guarantees provided by SDIMS meet the requirements of this application-each node"s controller effectively treats information from SDIMS as hints, and if a contacted node does not have the needed data, the controller retries, using SDIMS on-demand reaggregation to obtain a fresher hint. Distributed heavy hitter problem: The goal of the heavy hitter problem is to identify network sources, destinations, or protocols that account for significant or unusual amounts of traffic. As noted by Estan et al. [13], this information is useful for a variety of applications such as intrusion detection (e.g., port scanning), denial of service detection, worm detection and tracking, fair network allocation, and network maintenance. Significant work has been done on developing high-performance stream-processing algorithms for identifying heavy hitters at one router, but this is just a first step; ideally these applications would like not just one router"s views of the heavy hitters but an aggregate view. We use SDIMS to allow local information about heavy hitters to be pooled into a view of global heavy hitters. For each destination IP address IPx, a node updates the attribute (DestBW,IPx) with the number of bytes sent to IPx in the last time window. The aggregation function for attribute type DestBW is installed with the Update-UP strategy and simply adds the values from child nodes. Nodes perform continuous probe for global aggregate of the attribute and raise an alarm when the global aggregate value goes above a specified limit. Note that only nodes sending data to a particular IP address perform probes for the corresponding attribute. Also note that techniques from [25] can be extended to hierarchical case to tradeoff precision for communication bandwidth. 8. RELATED WORK The aggregation abstraction we use in our work is heavily influenced by the Astrolabe [38] project. Astrolabe adopts a PropagateAll and unstructured gossiping techniques to attain robustness [5]. However, any gossiping scheme requires aggressive replication of the aggregates. While such aggressive replication is efficient for read-dominated attributes, it incurs high message cost for attributes with a small read-to-write ratio. Our approach provides a flexible API for applications to set propagation rules according to their read-to-write ratios. Other closely related projects include Willow [39], Cone [4], DASIS [1], and SOMO [45]. Willow, DASIS and SOMO build a single tree for aggregation. Cone builds a tree per attribute and requires a total order on the attribute values. Several academic [15, 21, 42] and commercial [37] distributed monitoring systems have been designed to monitor the status of large networked systems. Some of them are centralized where all the monitoring data is collected and analyzed at a central host. Ganglia [15, 23] uses a hierarchical system where the attributes are replicated within clusters using multicast and then cluster aggregates are further aggregated along a single tree. Sophia [42] is a distributed monitoring system designed with a declarative logic programming model where the location of query execution is both explicit in the language and can be calculated during evaluation. This research is complementary to our work. TAG [21] collects information from a large number of sensors along a single tree. The observation that DHTs internally provide a scalable forest of reduction trees is not new. Plaxton et al."s [28] original paper describes not a DHT, but a system for hierarchically aggregating and querying object location data in order to route requests to nearby copies of objects. Many systems-building upon both Plaxton"s bit-correcting strategy [32, 46] and upon other strategies [24, 29, 35]-have chosen to hide this power and export a simple and general distributed hash table abstraction as a useful building block for a broad range of distributed applications. Some of these systems internally make use of the reduction forest not only for routing but also for caching [32], but for simplicity, these systems do not generally export this powerful functionality in their external interface. Our goal is to develop and expose the internal reduction forest of DHTs as a similarly general and useful abstraction. Although object location is a predominant target application for DHTs, several other applications like multicast [8, 9, 33, 36] and DNS [11] are also built using DHTs. All these systems implicitly perform aggregation on some attribute, and each one of them must be designed to handle any reconfigurations in the underlying DHT. With the aggregation abstraction provided by our system, designing and building of such applications becomes easier. Internal DHT trees typically do not satisfy domain locality properties required in our system. Castro et al. [7] and Gummadi et al. [17] point out the importance of path convergence from the perspective of achieving efficiency and investigate the performance of Pastry and other DHT algorithms, respectively. SkipNet [18] provides domain restricted routing where a key search is limited to the specified domain. This interface can be used to ensure path convergence by searching in the lowest domain and moving up to the next domain when the search reaches the root in the current domain. Although this strategy guarantees path convergence, it loses the aggregation tree abstraction property of DHTs as the domain constrained routing might touch a node more than once (as it searches forward and then backward to stay within a domain). 9. CONCLUSIONS This paper presents a Scalable Distributed Information Management System (SDIMS) that aggregates information in large-scale networked systems and that can serve as a basic building block for a broad range of applications. For large scale systems, hierarchical aggregation is a fundamental abstraction for scalability. We build our system by extending ideas from Astrolabe and DHTs to achieve (i) scalability with respect to both nodes and attributes through a new aggregation abstraction that helps leverage DHT"s internal trees for aggregation, (ii) flexibility through a simple API that lets applications control propagation of reads and writes, (iii) administrative isolation through simple augmentations of current DHT algorithms, and (iv) robustness to node and network reconfigurations through lazy reaggregation, on-demand reaggregation, and tunable spatial replication. Acknowlegements We are grateful to J.C. Browne, Robert van Renessee, Amin Vahdat, Jay Lepreau, and the anonymous reviewers for their helpful comments on this work. 10. REFERENCES [1] K. Albrecht, R. Arnold, M. Gahwiler, and R. Wattenhofer. Join and Leave in Peer-to-Peer Systems: The DASIS approach. Technical report, CS, ETH Zurich, 2003. [2] G. Back, W. H. Hsieh, and J. Lepreau. Processes in KaffeOS: Isolation, Resource Management, and Sharing in Java. In Proc. OSDI, Oct 2000. [3] G. Banga, P. Druschel, and J. Mogul. Resource Containers: A New Facility for Resource Management in Server Systems. In OSDI99, Feb. 1999. [4] R. Bhagwan, P. Mahadevan, G. Varghese, and G. M. Voelker. Cone: A Distributed Heap-Based Approach to Resource Selection. Technical Report CS2004-0784, UCSD, 2004. 389 [5] K. P. Birman. The Surprising Power of Epidemic Communication. In Proceedings of FuDiCo, 2003. [6] B. Bloom. Space/time tradeoffs in hash coding with allowable errors. Comm. of the ACM, 13(7):422-425, 1970. [7] M. Castro, P. Druschel, Y. C. Hu, and A. Rowstron. Exploiting Network Proximity in Peer-to-Peer Overlay Networks. Technical Report MSR-TR-2002-82, MSR. [8] M. Castro, P. Druschel, A.-M. Kermarrec, A. Nandi, A. Rowstron, and A. Singh. SplitStream: High-bandwidth Multicast in a Cooperative Environment. In SOSP, 2003. [9] M. Castro, P. Druschel, A.-M. Kermarrec, and A. Rowstron. SCRIBE: A Large-scale and Decentralised Application-level Multicast Infrastructure. IEEE JSAC (Special issue on Network Support for Multicast Communications), 2002. [10] J. Challenger, P. Dantzig, and A. Iyengar. A scalable and highly available system for serving dynamic data at frequently accessed web sites. In In Proceedings of ACM/IEEE, Supercomputing "98 (SC98), Nov. 1998. [11] R. Cox, A. Muthitacharoen, and R. T. Morris. Serving DNS using a Peer-to-Peer Lookup Service. In IPTPS, 2002. [12] M. Dahlin, L. Gao, A. Nayate, A. Venkataramani, P. Yalagandula, and J. Zheng. PRACTI replication for large-scale systems. Technical Report TR-04-28, The University of Texas at Austin, 2004. [13] C. Estan, G. Varghese, and M. Fisk. Bitmap algorithms for counting active flows on high speed links. In Internet Measurement Conference 2003, 2003. [14] Y. Fu, J. Chase, B. Chun, S. Schwab, and A. Vahdat. SHARP: An architecture for secure resource peering. In Proc. SOSP, Oct. 2003. [15] Ganglia: Distributed Monitoring and Execution System. http://ganglia.sourceforge.net. [16] S. Gribble, A. Halevy, Z. Ives, M. Rodrig, and D. Suciu. What Can Peer-to-Peer Do for Databases, and Vice Versa? In Proceedings of the WebDB, 2001. [17] K. Gummadi, R. Gummadi, S. D. Gribble, S. Ratnasamy, S. Shenker, and I. Stoica. The Impact of DHT Routing Geometry on Resilience and Proximity. In SIGCOMM, 2003. [18] N. J. A. Harvey, M. B. Jones, S. Saroiu, M. Theimer, and A. Wolman. SkipNet: A Scalable Overlay Network with Practical Locality Properties. In USITS, March 2003. [19] R. Huebsch, J. M. Hellerstein, N. Lanham, B. T. Loo, S. Shenker, and I. Stoica. Querying the Internet with PIER. In Proceedings of the VLDB Conference, May 2003. [20] C. Intanagonwiwat, R. Govindan, and D. Estrin. Directed diffusion: a scalable and robust communication paradigm for sensor networks. In MobiCom, 2000. [21] S. R. Madden, M. J. Franklin, J. M. Hellerstein, and W. Hong. TAG: a Tiny AGgregation Service for ad-hoc Sensor Networks. In OSDI, 2002. [22] D. Malkhi. Dynamic Lookup Networks. In FuDiCo, 2002. [23] M. L. Massie, B. N. Chun, and D. E. Culler. The ganglia distributed monitoring system: Design, implementation, and experience. In submission. [24] P. Maymounkov and D. Mazieres. Kademlia: A Peer-to-peer Information System Based on the XOR Metric. In Proceesings of the IPTPS, March 2002. [25] C. Olston and J. Widom. Offering a precision-performance tradeoff for aggregation queries over replicated data. In VLDB, pages 144-155, Sept. 2000. [26] K. Petersen, M. Spreitzer, D. Terry, M. Theimer, and A. Demers. Flexible Update Propagation for Weakly Consistent Replication. In Proc. SOSP, Oct. 1997. [27] Planetlab. http://www.planet-lab.org. [28] C. G. Plaxton, R. Rajaraman, and A. W. Richa. Accessing Nearby Copies of Replicated Objects in a Distributed Environment. In ACM SPAA, 1997. [29] S. Ratnasamy, P. Francis, M. Handley, R. Karp, and S. Shenker. A Scalable Content Addressable Network. In Proceedings of ACM SIGCOMM, 2001. [30] S. Ratnasamy, S. Shenker, and I. Stoica. Routing Algorithms for DHTs: Some Open Questions. In IPTPS, March 2002. [31] T. Roscoe, R. Mortier, P. Jardetzky, and S. Hand. InfoSpect: Using a Logic Language for System Health Monitoring in Distributed Systems. In Proceedings of the SIGOPS European Workshop, 2002. [32] A. Rowstron and P. Druschel. Pastry: Scalable, Distributed Object Location and Routing for Large-scale Peer-to-peer Systems. In Middleware, 2001. [33] S.Ratnasamy, M.Handley, R.Karp, and S.Shenker. Application-level Multicast using Content-addressable Networks. In Proceedings of the NGC, November 2001. [34] W. Stallings. SNMP, SNMPv2, and CMIP. Addison-Wesley, 1993. [35] I. Stoica, R. Morris, D. Karger, F. Kaashoek, and H. Balakrishnan. Chord: A scalable Peer-To-Peer lookup service for internet applications. In ACM SIGCOMM, 2001. [36] S.Zhuang, B.Zhao, A.Joseph, R.Katz, and J.Kubiatowicz. Bayeux: An Architecture for Scalable and Fault-tolerant Wide-Area Data Dissemination. In NOSSDAV, 2001. [37] IBM Tivoli Monitoring. www.ibm.com/software/tivoli/products/monitor. [38] R. VanRenesse, K. P. Birman, and W. Vogels. Astrolabe: A Robust and Scalable Technology for Distributed System Monitoring, Management, and Data Mining. TOCS, 2003. [39] R. VanRenesse and A. Bozdog. Willow: DHT, Aggregation, and Publish/Subscribe in One Protocol. In IPTPS, 2004. [40] A. Venkataramani, P. Weidmann, and M. Dahlin. Bandwidth constrained placement in a wan. In PODC, Aug. 2001. [41] A. Venkataramani, P. Yalagandula, R. Kokku, S. Sharif, and M. Dahlin. Potential costs and benefits of long-term prefetching for content-distribution. Elsevier Computer Communications, 25(4):367-375, Mar. 2002. [42] M. Wawrzoniak, L. Peterson, and T. Roscoe. Sophia: An Information Plane for Networked Systems. In HotNets-II, 2003. [43] R. Wolski, N. Spring, and J. Hayes. The network weather service: A distributed resource performance forecasting service for metacomputing. Journal of Future Generation Computing Systems, 15(5-6):757-768, Oct 1999. [44] P. Yalagandula and M. Dahlin. SDIMS: A scalable distributed information management system. Technical Report TR-03-47, Dept. of Computer Sciences, UT Austin, Sep 2003. [45] Z. Zhang, S.-M. Shi, and J. Zhu. SOMO: Self-Organized Metadata Overlay for Resource Management in P2P DHT. In IPTPS, 2003. [46] B. Y. Zhao, J. D. Kubiatowicz, and A. D. Joseph. Tapestry: An Infrastructure for Fault-tolerant Wide-area Location and Routing. Technical Report UCB/CSD-01-1141, UC Berkeley, Apr. 2001. 390

autonomous dht;temporal heterogeneity;administrative isolation;distribute hash table;lazy re-aggregation;write-dominated attribute;large-scale networked system;update-upk-downj strategy;information management system;freepastry framework;tunable spatial replication;distributed hash table;aggregation management layer;distributed operating system backbone;availability;read-dominated attribute;network system monitor;virtual node;networked system monitoring;eventual consistency

train_C-61

Authority Assignment in Distributed Multi-Player Proxy-based Games

We present a proxy-based gaming architecture and authority assignment within this architecture that can lead to better game playing experience in Massively Multi-player Online games. The proposed game architecture consists of distributed game clients that connect to game proxies (referred to as communication proxies) which forward game related messages from the clients to one or more game servers. Unlike proxy-based architectures that have been proposed in the literature where the proxies replicate all of the game state, the communication proxies in the proposed architecture support clients that are in proximity to it in the physical network and maintain information about selected portions of the game space that are relevant only to the clients that they support. Using this architecture, we propose an authority assignment mechanism that divides the authority for deciding the outcome of different actions/events that occur within the game between client and servers on a per action/event basis. We show that such division of authority leads to a smoother game playing experience by implementing this mechanism in a massively multi-player online game called RPGQuest. In addition, we argue that cheat detection techniques can be easily implemented at the communication proxies if they are made aware of the game-play mechanics.

1. INTRODUCTION In Massively Multi-player On-line Games (MMOG), game clients who are positioned across the Internet connect to a game server to interact with other clients in order to be part of the game. In current architectures, these interactions are direct in that the game clients and the servers exchange game messages with each other. In addition, current MMOGs delegate all authority to the game server to make decisions about the results pertaining to the actions that game clients take and also to decide upon the result of other game related events. Such centralized authority has been implemented with the claim that this improves the security and consistency required in a gaming environment. A number of works have shown the effect of network latency on distributed multi-player games [1, 2, 3, 4]. It has been shown that network latency has real impact on practical game playing experience [3, 5]. Some types of games can function quite well even in the presence of large delays. For example, [4] shows that in a modern RPG called Everquest 2, the breakpoint of the game when adding artificial latency was 1250ms. This is accounted to the fact that the combat system used in Everquest 2 is queueing based and has very low interaction. For example, a player queues up 4 or 5 spells they wish to cast, each of these spells take 1-2 seconds to actually perform, giving the server plenty of time to validate these actions. But there are other games such as FPS games that break even in the presence of moderate network latencies [3, 5]. Latency compensation techniques have been proposed to alleviate the effect of latency [1, 6, 7] but it is obvious that if MMOGs are to increase in interactivity and speed, more architectures will have to be developed that address responsiveness, accuracy and consistency of the gamestate. In this paper, we propose two important features that would make game playing within MMOGs more responsive for movement and scalable. First, we propose that centralized server-based architectures be made hierarchical through the introduction of communication proxies so that game updates made by clients that are time sensitive, such as movement, can be more efficiently distributed to other players within their game-space. Second, we propose that assignment of authority in terms of who makes the decision on client actions such as object pickups and hits, and collisions between players, be distributed between the clients and the servers in order to distribute the computing load away from the central server. In order to move towards more complex real-time networked games, we believe that definitions of authority must be refined. Most currently implemented MMOGs have game servers that have almost absolute authority. We argue that there is no single consistent view of the virtual game space that can be maintained on any one component within a network that has significant latency, such as the one that many MMOG players would experience. We believe that in most cases, the client with the most accurate view of an entity is the best suited to make decisions for that entity when the causality of that action will not immediately affect any other players. In this paper we define what it means to have authority within the context of events and objects in a virtual game space. We then show the benefits of delegating authority for different actions and game events between the clients and server. In our model, the game space consists of game clients (representing the players) and objects that they control. We divide the client actions and game events (we will collectively refer to these as events) such as collisions, hits etc. into three different categories, a) events for which the game client has absolute authority, b) events for which the game server has absolute authority, and c) events for which the authority changes dynamically from client to the server and vice-versa. Depending on who has the authority, that entity will make decisions on the events that happen within a game space. We propose that authority for all decisions that pertain to a single player or object in the game that neither affects the other players or objects, nor are affected by the actions of other players be delegated to that player"s game client. These type of decisions would include collision detection with static objects within the virtual game space and hit detection with linear path bullets (whose trajectory is fixed and does not change with time) fired by other players. Authority for decisions that could be affected by two or more players should be delegated to the impartial central server, in some cases, to ensure that no conflicts occur and in other cases can be delegated to the clients responsible for those players. For example, collision detection of two players that collide with each other and hit detection of non-linear bullets (that changes trajectory with time) should be delegated to the server. Decision on events such as item pickup (for example, picking up items in a game to accumulate points) should be delegated to a server if there are multiple players within close proximity of an item and any one of the players could succeed in picking the item; for item pick-up contention where the client realizes that no other player, except its own player, is within a certain range of the item, the client could be delegated the responsibility to claim the item. The client"s decision can always be accurately verified by the server. In summary, we argue that while current authority models that only delegate responsibility to the server to make authoritative decisions on events is more secure than allowing the clients to make the decisions, these types of models add undesirable delays to events that could very well be decided by the clients without any inconsistency being introduced into the game. As networked games become more complex, our architecture will become more applicable. This architecture is applicable for massively multiplayer games where the speed and accuracy of game-play are a major concern while consistency between player game-states is still desired. We propose that a mixed authority assignment mechanism such as the one outlined above be implemented in high interaction MMOGs. Our paper has the following contributions. First we propose an architecture that uses communication proxies to enable clients to connect to the game server. A communication proxy in the proposed architecture maintains information only about portions of the game space that are relevant to clients connected to it and is able to process the movement information of objects and players within these portions. In addition, it is capable of multicasting this information only to a relevant subset of other communication proxies. These functionalities of a communication proxy leads to a decrease in latency of event update and subsequently, better game playing experience. Second, we propose a mixed authority assignment mechanism as described above that improves game playing experience. Third, we implement the proposed mixed authority assignment mechanism within a MMOG called RPGQuest [8] to validate its viability within MMOGs. In Section 2, we describe the proxy-based game architecture in more detail and illustrate its advantages. In Section 3, we provide a generic description of the mixed authority assignment mechanism and discuss how it improves game playing experience. In Section 4, we show the feasibility of implementing the proposed mixed authority assignment mechanism within existing MMOGs by describing a proof-of-concept implementation within an existing MMOG called RPGQuest. Section 5 discusses related work. In Section 6, we present our conclusions and discuss future work. 2. PROXY-BASED GAME ARCHITECTURE Massively Multi-player Online Games (MMOGs) usually consist of a large game space in which the players and different game objects reside and move around and interact with each-other. State information about the whole game space could be kept in a single central server which we would refer to as a Central-Server Architecture. But to alleviate the heavy demand on the processing for handling the large player population and the objects in the game in real-time, a MMOG is normally implemented using a distributed server architecture where the game space is further sub-divided into regions so that each region has relatively smaller number of players and objects that can be handled by a single server. In other words, the different game regions are hosted by different servers in a distributed fashion. When a player moves out of one game region to another adjacent one, the player must communicate with a different server (than it was currently communicating with) hosting the new region. The servers communicate with one another to hand off a player or an object from one region to another. In this model, the player on the client machine has to establish multiple gaming sessions with different servers so that it can roam in the entire game space. We propose a communication proxy based architecture where a player connects to a (geographically) nearby proxy instead of connecting to a central server in the case of a centralserver architecture or to one of the servers in case of dis2 The 5th Workshop on Network & System Support for Games 2006 - NETGAMES 2006 tributed server architecture. In the proposed architecture, players who are close by geographically join a particular proxy. The proxy then connects to one or more game servers, as needed by the set of players that connect to it and maintains persistent transport sessions with these server. This alleviates the problem of each player having to connect directly to multiple game servers, which can add extra connection setup delay. Introduction of communication proxies also mitigates the overhead of a large number of transport sessions that must be managed and reduces required network bandwidth [9] and processing at the game servers both with central server and distributed server architectures. With central server architectures, communication proxies reduce the overhead at the server by not requiring the server to terminate persistent transport sessions from every one of the clients. With distributed-server architectures, additionally, communication proxies eliminate the need for the clients to maintain persistent transport sessions to every one of the servers. Figure 1 shows the proposed architecture. Figure 1: Architecture of the gaming environment. Note that the communication proxies need not be cognizant of the game. They host a number of players and inform the servers which players are hosted by the proxy in question. Also note that the players hosted by a proxy may not be in the same game space. That is, a proxy hosts players that are geographically close to it, but the players themselves can reside in different parts of the game space. The proxy communicates with the servers responsible for maintaining the game spaces subscribed by the different players. The proxies communicate with one another in a peer-to-peer to fashion. The responsiveness of the game can be improved for updates that do not need to wait on processing at a central authority. In this way, information about players can be disseminated faster before even the game server gets to know about it. This definitely improves the responsiveness of the game. However, it ignores consistency that is critical in MMORPGs. The notion that an architecture such as this one can still maintain temporal consistency will be discussed in detail in Section 3. Figure 2 shows and example of the working principle of the proposed architecture. Assume that the game space is divided into 9 regions and there are three servers responsible for managing the regions. Server S1 owns regions 1 and 2, S2 manages 4, 5, 7, and 8, and S3 is responsible for 3, 6 and 9. Figure 2: An example. There are four communication proxies placed in geographically distant locations. Players a, b, c join proxy P1, proxy P2 hosts players d, e, f, players g, h are with proxy P3, whereas players i, j, k, l are with proxy P4. Underneath each player, the figure shows which game region the player is located currently. For example, players a, b, c are in regions 1, 2, 6, respectively. Therefore, proxy P1 must communicate with servers S1 and S3. The reader can verify the rest of the links between the proxies and the servers. Players can move within the region and between regions. Player movement within a region will be tracked by the proxy hosting the player and this movement information (for example, the player"s new coordinates) will be multicast to a subset of other relevant communication proxies directly. At the same time, this information will be sent to the server responsible for that region with the indication that this movement has already been communicated to all the other relevant communication proxies (so that the server does not have to relay this information to all the proxies). For example, if player a moves within region 1, this information will be communicated by proxy P1 to server S1 and multicast to proxies P3 and P4. Note that proxies that do not keep state information about this region at this point in time (because they do not have any clients within that region) such as P2 do not have to receive this movement information. If a player is at the boundary of a region and moves into a new region, there are two possibilities. The first possibility is that the proxy hosting the player can identify the region into which the player is moving (based on the trajectory information) because it is also maintaining state information about the new region at that point in time. In this case, the proxy can update movement information directly at the other relevant communication proxies and also send information to the appropriate server informing of the movement (this may require handoff between servers as we will describe). Consider the scenario where player a is at the boundary of region 1 and proxy P1 can identify that the player is moving into region 2. Because proxy P1 is currently keeping state information about region 2, it can inform all The 5th Workshop on Network & System Support for Games 2006 - NETGAMES 2006 3 the other relevant communication proxies (in this example, no other proxy maintains information about region 2 at this point and so no update needs to be sent to any of the other proxies) about this movement and then inform the server independently. In this particular case, server S1 is responsible for region 2 as well and so no handoff between servers would be needed. Now consider another scenario where player j moves from region 9 to region 8 and that proxy P4 is able to identify this movement. Again, because proxy P4 maintains state information about region 8, it can inform any other relevant communication proxies (again, none in this example) about this movement. But now, regions 9 and 8 are managed by different servers (servers S3 and S2 respectively) and thus a hand-off between these servers is needed. We propose that in this particular scenario, the handoff be managed by the proxy P4 itself. When the proxy sends movement update to server S3 (informing the server that the player is moving out of its region), it would also send a message to server S2 informing the server of the presence and location of the player in one of its region. In the intra-region and inter-region scenarios described above, the proxy is able to manage movement related information, update only the relevant communication proxies about the movement, update the servers with the movement and enable handoff of a player between the servers if needed. In this way, the proxy performs movement updates without involving the servers in any way in this time-critical function thereby speeding up the game and improving game playing experience for the players. We consider this the fast path for movement update. We envision the proxies to be just communication proxies in that they do not know about the workings of specific games. They merely process movement information of players and objects and communicate this information to the other proxies and the servers. If the proxies are made more intelligent in that they understand more of the game logic, it is possible for them to quickly check on claims made by the clients and mitigate cheating. The servers could perform the same functionality but with more delay. Even without being aware of game logic, the proxies can provide additional functionalities such as timestamping messages to make the game playing experience more accurate [10] and fair [11]. The second possibility that should be considered is when players move between regions. It is possible that a player moves from one region to another but the proxy that is hosting the player is not able to determine the region into which the player is moving, a) the proxy does not maintain state information about all the regions into which the player could potentially move, or b) the proxy is not able to determine which region the player may move into (even if maintains state information about all these regions). In this case, we propose that the proxy be not responsible for making the movement decision, but instead communicate the movement indication to the server responsible for the region within which the player is currently located. The server will then make the movement decision and then a) inform all the proxies including the proxy hosting the player, and b) initiate handoff with another server if the player moves into a region managed by another server. We consider this the slow path for movement update in that the servers need to be involved in determining the new position of the player. In the example, assume that player a moves from region 1 to region 4. Proxy P1 does not maintain state information about region 4 and thus would pass the movement information to server S1. The server will identify that the player has moved into region 4 and would inform proxy P1 as well as proxy P2 (which is the only other proxy that maintains information about region 4 at this point in time). Server S1 will also initiate a handoff of player a with server S2. Proxy P1 will now start maintaining state information about region 4 because one of its hosted players, player a has moved into this region. It will do so by requesting and receiving the current state information about region 4 from server S2 which is responsible for this region. Thus, a proxy architecture allows us to make use of faster movement updates through the fast path through a proxy if and when possible as opposed to conventional server-based architectures that always have to use the slow path through the server for movement updates. By selectively maintaining relevant regional game state information at the proxies, we are able to achieve this capability in our architecture without the need for maintaining the complete game state at every proxy. 3. ASSIGNMENT OF AUTHORITY As a MMOG is played, the players and the game objects that are part of the game, continually change their state. For example, consider a player who owns a tank in a battlefield game. Based on action of the player, the tank changes its position in the game space, the amount of ammunition the tank contains changes as it fires at other tanks, the tank collects bonus firing power based on successful hits, etc. Similarly objects in the battlefield, such as flags, buildings etc. change their state when a flag is picked up by a player (i.e. tank) or a building is destroyed by firing at it. That is, some decision has to be made on the state of each player and object as the game progresses. Note that the state of a player and/or object can contain several parameters (e.g., position, amount of ammunition, fuel storage, points collected, etc), and if any of the parameters changes, the state of the player/object changes. In a client-server based game, the server controls all the players and the objects. When a player at a client machine makes a move, the move is transmitted to the server over the network. The server then analyzes the move, and if the move is a valid one, changes the state of the player at the server and informs the client of the change. The client subsequently updates the state of the player and renders the player at the new location. In this case the authority to change the state of the player resides with the server entirely and the client simply follows what the server instructs it to do. Most of the current first person shooter (FPS) games and role playing games (RPG) fall under this category. In current FPS games, much like in RPG games, the client is not trusted. All moves and actions that it makes are validated. If a client detects that it has hit another player with a bullet, it proceeds assuming that it is a hit. Meanwhile, an update is sent to the server and the server will send back a message either affirming or denying that the player was hit. If the remote player was not hit, then the client will know that it 4 The 5th Workshop on Network & System Support for Games 2006 - NETGAMES 2006 did not actually make the shot. If it did make the hit, an update will also be sent from the server to the other clients informing them that the other player was hit. A difference that occurs in some RPGs is that they use very dumb client programs. Some RPGs do not maintain state information at the client and therefore, cannot predict anything such as hits at the client. State information is not maintained because the client is not trusted with it. In RPGs, a cheating player with a hacked game client can use state information stored at the client to gain an advantage and find things such as hidden treasure or monsters lurking around the corner. This is a reason why most MMORPGs do not send a lot of state information to the client and causes the game to be less responsive and have lower interaction game-play than FPS games. In a peer-to-peer game, each peer controls the player and object that it owns. When a player makes a move, the peer machine analyzes the move and if it is a valid one, changes the state of the player and places the player in new position. Afterwards, the owner peer informs all other peers about the new state of the player and the rest of the peers update the state of the player. In this scenario, the authority to change the state of the player is given to the owning peer and all other peers simply follow the owner. For example, Battle Zone Flag (BzFlag) [12] is a multiplayer client-server game where the client has all authority for making decisions. It was built primarily with LAN play in mind and cheating as an afterthought. Clients in BzFlag are completely authoritative and when they detect that they were hit by a bullet, they send an update to the server which simply forwards the message along to all other players. The server does no sort of validation. Each of the above two traditional approaches has its own set of advantages and disadvantages. The first approach, which we will refer to as server authoritative henceforth, uses a centralized method to assign authority. While a centralized approach can keep the state of the game (i.e., state of all the players and objects) consistent across any number of client machines, it suffers from delayed response in game-play as any move that a player at the client machine makes must go through one round-trip delay to the server before it can take effect on the client"s screen. In addition to the round-trip delay, there is also queuing delay in processing the state change request at the server. This can result in additional processing delay, and can also bring in severe scalability problems if there are large number of clients playing the game. One definite advantage of the server authoritative approach is that it can easily detect if a client is cheating and can take appropriate action to prevent cheating. The peer-to-peer approach, henceforth referred to as client authoritative, can make games very responsive. However, it can make the game state inconsistent for a few players and tie break (or roll back) has to be performed to bring the game back to a consistent state. Neither tie break nor roll back is a desirable feature of online gaming. For example, assume that for a game, the goal of each player is to collect as many flags as possible from the game space (e.g. BzFlag). When two players in proximity try to collect the same flag at the same time, depending on the algorithm used at the client-side, both clients may determine that it is the winner, although in reality only one player can pick the flag up. Both players will see on their screen that it is the winner. This makes the state of the game inconsistent. Ways to recover from this inconsistency are to give the flag to only one player (using some tie break rule) or roll the game back so that the players can try again. Neither of these two approaches is a pleasing experience for online gaming. Another problem with client authoritative approach is that of cheating by clients as there is no cross checking of the validation of the state changes authorized by the owner client. We propose to use a hybrid approach to assign the authority dynamically between the client and the server. That is, we assign the authority to the client to make the game responsive, and use the server"s authority only when the client"s individual authoritative decisions can make the game state inconsistent. By moving the authority of time critical updates to the client, we avoid the added delay caused by requiring the server to validate these updates. For example, in the flag pickup game, the clients will be given the authority to pickup flags only when other players are not within a range that they could imminently pickup a flag. Only when two or more players are close by so that more than one player may claim to have picked up a flag, the authority for movement and flag pickup would go to the central server so that the game state does not become inconsistent. We believe that in a large game-space where a player is often in a very wide open and sparsely populated area such as those often seen in the game Second Life [13], this hybrid architecture would be very beneficial because of the long periods that the client would have authority to send movement updates for itself. This has two advantages over the centralauthority approach, it distributes the processing load down to the clients for the majority of events and it allows for a more responsive game that does not need to wait on a server for validation. We believe that our notion of authority can be used to develop a globally consistent state model of the evolution of a game. Fundamentally, the consistent state of the system is the one that is defined by the server. However, if local authority is delegated to the client, in this case, the client"s state is superimposed on the server"s state to determine the correct global state. For example, if the client is authoritative with respect to movement of a player, then the trajectory of the player is the true trajectory and must replace the server"s view of the player"s trajectory. Note that this could be problematic and lead to temporal inconsistency only if, for example, two or more entities are moving in the same region and can interact with each other. In this situation, the client authority must revert to the server and the sever would then make decisions. Thus, the client is only authoritative in situations where there is no potential to imminently interact with other players. We believe that in complex MMOGs, when allowing more rapid movement, it will still be the case that local authority is possible for significant spans of game time. Note that it might also be possible to minimize the occurrences of the Dead Man Shooting problem described in [14]. This could be done by allowing the client to be authoritative for more actions such as its player"s own death and disallowing other players from making preemptive decisions based on a remote player. The 5th Workshop on Network & System Support for Games 2006 - NETGAMES 2006 5 One reason why the client-server based architecture has gained popularity is due to belief that the fastest route to the other clients is through the server. While this may be true, we aim to create a new architecture where decisions do not always have to be made at the game server and the fastest route to a client is actually through a communication proxy located close to the client. That is, the shortest distance in our architecture is not through the game server but through the communication proxy. After a client makes an action such as movement, it will simultaneously distribute it directly to the clients and the game server by way of the communications proxy. We note that our architecture however is not practical for a game where game players setup their own servers in an ad-hoc fashion and do not have access to proxies at the various ISPs. This proxy and distributed authority architecture can be used to its full potential only when the proxies can be placed at strategic places within the main ISPs and evenly distributed geographically. Our game architecture does not assume that the client is not to be trusted. We are designing our architecture on the fact that there will be sufficient cheat deterring and detection mechanisms present so that it will be both undesirable and very difficult to cheat [15]. In our proposed approach, we can make the games cheat resilient by using the proxybased architecture when client authoritative decisions take place. In order to achieve this, the proxies have to be game cognizant so that decisions made by a client can be cross checked by a proxy that the client connects to. For example, assume that in a game a plane controlled by a client moves in the game space. It is not possible for the plane to go through a building unharmed. In a client authoritative mode, it is possible for the client to cheat by maneuvering the plane through a building and claiming the plane to be unharmed. However, when such move is published by the client, the proxy, being aware of the game space that the plane is in, can quickly check that the client has misused the authority and then can block such move. This allows us to distribute authority to make decisions about the clients. In the following section we use a multiplayer game called RPGQuest to implement different authoritative schemes and discuss our experience with the implementation. Our implementation shows the viability of our proposed solution. 4. IMPLEMENTATION EXPERIENCE We have experimented with the authority assignment mechanism described in the last section by implementing the mechanisms in a game called RPGQuest. A screen shot from this game is shown in Figure 3. The purpose of the implementation is to test its feasibility in a real game. RPGQuest is a basic first person game where the player can move around a three dimensional environment. Objects are placed within the game world and players gain points for each object that is collected. The game clients connect to a game server which allows many players to coexist in the same game world. The basic functionality of this game is representative of current online first person shooter and role playing games. The game uses the DirectX 8 graphics API and DirectPlay networking API. In this section we will discuss the three different versions of the game that we experimented with. Figure 3: The RPGQuest Game. The first version of the game, which is the original implementation of RPGQuest, was created with a completely authoritative server and a non-authoritative client. Authority given to the server includes decisions of when a player collides with static objects and other players and when a player picks up an object. This version of the game performs well up to 100ms round-trip latency between the client and the server. There is little lag between the time player hits a wall and the time the server corrects the player"s position. However, as more latency is induced between the client and server, the game becomes increasingly difficult to play. With the increased latency, the messages coming from the server correcting the player when it runs into a wall are not received fast enough. This causes the player to pass through the wall for the period that it is waiting for the server to resolve the collision. When studying the source code of the original version of the RPGQuest game, there is a substantial delay that is unavoidable each time an action must be validated by the server. Whenever a movement update is sent to the server, the client must then wait whatever the round trip delay is, plus some processing time at the server in order to receive its validated or corrected position. This is obviously unacceptable in any game where movement or any other rapidly changing state information must be validated and disseminated to the other clients rapidly. In order to get around this problem, we developed a second version of the game, which gives all authority to the client. The client was delegated the authority to validate its own movement and the authority to pick up objects without validation from the server. In this version of the game when a player moves around the game space, the client validates that the player"s new position does not intersect with any walls or static objects. A position update is then sent to the server which then immediately forwards the update to the other clients within the region. The update does not have to go through any extra processing or validation. This game model of complete authority given to the client is beneficial with respect to movement. When latencies of 6 The 5th Workshop on Network & System Support for Games 2006 - NETGAMES 2006 100ms and up are induced into the link between the client and server, the game is still playable since time critical aspects of the game like movement do not have to wait on a reply from the server. When a player hits a wall, the collision is processed locally and does not have to wait on the server to resolve the collision. Although game playing experience with respect to responsiveness is improved when the authority for movement is given to the client, there are still aspects of games that do not benefit from this approach. The most important of these is consistency. Although actions such as movement are time critical, other actions are not as time critical, but instead require consistency among the player states. An example of a game aspect that requires consistency is picking up objects that should only be possessed by a single player. In our client authoritative version of RPGQuest clients send their own updates to all other players whenever they pick up an object. From our tests we have realized this is a problem because when there is a realistic amount of latency between the client and server, it is possible for two players to pick up the same object at the same time. When two players attempt to pick up an object at physical times which are close to each other, the update sent by the player who picked up the object first will not reach the second player in time for it to see that the object has already been claimed. The two players will now both think that they own the object. This is why a server is still needed to be authoritative in this situation and maintain consistency throughout the players. These two versions of the RPGQuest game has showed us why it is necessary to mix the two absolute models of authority. It is better to place authority on the client for quickly changing actions such as movement. It is not desirable to have to wait for server validation on a movement that could change before the reply is even received. It is also sometimes necessary to place consistency over efficiency in aspects of the game that cannot tolerate any inconsistencies such as object ownership. We believe that as the interactivity of games increases, our architecture of mixed authority that does not rely on server validation will be necessary. To test the benefits and show the feasibility of our architecture of mixed authority, we developed a third version of the RPGQuest game that distributed authority for different actions between the client and server. In this version, in the interest of consistency, the server remained authoritative for deciding who picked up an object. The client was given full authority to send positional updates to other clients and verify its own position without the need to verify its updates with the server. When the player tries to move their avatar, the client verifies that the move will not cause it to move through a wall. A positional update is then sent to the server which then simply forwards it to the other clients within the region. This eliminates any extra processing delay that would occur at the server and is also a more accurate means of verification since the client has a more accurate view of its own state than the server. This version of the RPGQuest game where authority is distributed between the client and server is an improvement from the server authoritative version. The client has no delay in waiting for an update for its own position and other clients do not have to wait on the server to verify the update. The inconsistencies where two clients can pick up the same object in the client authoritative architecture are not present in this version of the client. However, the benefits of mixed authority will not truly be seen until an implementation of our communication proxy is integrated into the game. With the addition of the communication proxy, after the client verifies its own positional updates it will be able to send the update to all clients within its region through a low latency link instead of having to first go through the game server which could possibly be in a very remote location. The coding of the different versions of the game was very simple. The complexity of the client increased very slightly in the client authoritative and hybrid models. The original dumb clients of RPGQuest know the position of other players; it is not just sent a screen snapshot from the server. The server updates each client with the position of all nearby clients. The dumb clients use client side prediction to fill in the gaps between the updates they receive. The only extra processing the client has to do in the hybrid architecture is to compare its current position to the positions of all objects (walls, boxes, etc.) in its area. This obviously means that each client will have to already have downloaded the locations of all static objects within its current region. 5. RELATED WORK It has been noted that in addition to latency, bandwidth requirements also dictate the type of gaming architecture to be used. In [16], different types of architectures are studied with respect to bandwidth efficiencies and latency. It is pointed out that Central Server architectures are not scalable because of bandwidth requirements at the server but the overhead for consistency checks are limited as they are performed at the server. A Peer-to-Peer architecture, on the other hand, is scalable but there is a significant overhead for consistency checks as this is required at every player. The paper proposes a hybrid architecture which is Peer-toPeer in terms of message exchange (and thereby is scalable) where a Central Server is used for off-line consistency checks (thereby mitigating consistency check overhead). The paper provides an implementation example of BZFlag which is a peer-to-peer game which is modified to transfer all authority to a central server. In essence, this paper advocates an authority architecture which is server based even for peerto-peer games, but does not consider division of authority between a client and a server to minimize latency which could affect game playing experience even with the type of latency found in server based games (where all authority is with the server). There is also previous work that has suggested that proxy based architectures be used to alleviate the latency problem and in addition use proxies to provide congestion control and cheat-proof mechanisms in distributed multi-player games [17]. In [18], a proxy server-network architecture is presented that is aimed at improving scalability of multiplayer games and lowering latency in server-client data transmission. The main goal of this work is to improve scalability of First-Person Shooter (FPS) and RPG games. The further objective is to improve the responsiveness MMOGs by providing low latency communications between the client and The 5th Workshop on Network & System Support for Games 2006 - NETGAMES 2006 7 server. The architecture uses interconnected proxy servers that each have a full view of the global game state. Proxy servers are located at various different ISPs. It is mentioned in this work that dividing the game space among multiple games servers such as the federated model presented in [19] is inefficient for a relatively fast game flow and that the proposed architecture alleviates this problem because users do not have to connect to a different server whenever they cross the server boundary. This architecture still requires all proxies to be aware of the overall game state over the whole game space unlike our work where we require the proxies to maintain only partial state information about the game space. Fidelity based agent architectures have been proposed in [20, 21]. These works propose a distributed client-server architecture for distributed interactive simulations where different servers are responsible for different portions of the game space. When an object moves from one portion to another, there is a handoff from one server to another. Although these works propose an architecture where different portions of the simulation space are managed by different servers, they do not address the issue of decreasing the bandwidth required through the use of communication proxies. Our work differs from the above discussed previous works by proposing a) a distributed proxy-based architecture to decrease bandwidth requirements at the clients and the servers without requiring the proxies to keep state information about the whole game space, b) a dynamic authority assignment technique to reduce latency (by performing consistency checks locally at the client whenever possible) by splitting the authority between the clients and servers on a per object basis, and c) proposing that cheat detection can be built into the proxies if they are provided more information about the specific game instead of using them purely as communication proxies (although this idea has not been implemented yet and is part of our future work). 6. CONCLUSIONS AND FUTURE WORK In this paper, we first proposed a proxy-based architecture for MMOGs that enables MMOGs to scale to a large number of users by mitigating the need for a large number of transport sessions to be maintained and decreasing both bandwidth overhead and latency of event update. Second, we proposed a mixed authority assignment mechanism that divides authority for making decisions on actions and events within the game between the clients and server and argued how such an authority assignment leads to better game playing experience without sacrificing the consistency of the game. Third, to validate the viability of the mixed authority assignment mechanism, we implemented it within a MMOG called RPGQuest and described our implementation experience. In future work, we propose to implement the communications proxy architecture described in this paper and integrate the mixed authority mechanism within this architecture. We propose to evaluate the benefits of the proxy-based architecture in terms of scalability, accuracy and responsiveness. We also plan to implement a version of the RPGQuest game with dynamic assignment of authority to allow players the authority to pickup objects when no other players are near. As discussed earlier, this will allow for a more efficient and responsive game in certain situations and alleviate some of the processing load from the server. Also, since so much trust is put into the clients of our architecture, it will be necessary to integrate into the architecture many of the cheat detection schemes that have been proposed in the literature. Software such as Punkbuster [22] and a reputation system like those proposed by [23] and [15] would be integral to the operation of an architecture such as ours which has a lot of trust placed on the client. We further propose to make the proxies in our architecture more game cognizant so that cheat detection mechanisms can be built into the proxies themselves. 7. REFERENCES [1] Y. W. Bernier. Latency Compensation Methods in Client/Server In-game Protocol Design and Optimization. In Proc. of Game Developers Conference"01, 2001. [2] Lothar Pantel and Lars C. Wolf. On the impact of delay on real-time multiplayer games. In NOSSDAV "02: Proceedings of the 12th international workshop on Network and operating systems support for digital audio and video, pages 23-29, New York, NY, USA, 2002. ACM Press. [3] G. Armitage. Sensitivity of Quake3 Players to Network Latency. In Proc. of IMW2001, Workshop Poster Session, November 2001. http://www.geocities.com/ gj armitage/q3/quake-results.html. [4] Tobias Fritsch, Hartmut Ritter, and Jochen Schiller. The effect of latency and network limitations on mmorpgs: a field study of everquest2. In NetGames "05: Proceedings of 4th ACM SIGCOMM workshop on Network and system support for games, pages 1-9, New York, NY, USA, 2005. ACM Press. [5] Tom Beigbeder, Rory Coughlan, Corey Lusher, John Plunkett, Emmanuel Agu, and Mark Claypool. The effects of loss and latency on user performance in unreal tournament 2003. In NetGames "04: Proceedings of 3rd ACM SIGCOMM workshop on Network and system support for games, pages 144-151, New York, NY, USA, 2004. ACM Press. [6] Y. Lin, K. Guo, and S. Paul. Sync-MS: Synchronized Messaging Service for Real-Time Multi-Player Distributed Games. In Proc. of 10th IEEE International Conference on Network Protocols (ICNP), Nov 2002. [7] Katherine Guo, Sarit Mukherjee, Sampath Rangarajan, and Sanjoy Paul. A fair message exchange framework for distributed multi-player games. In NetGames "03: Proceedings of the 2nd workshop on Network and system support for games, pages 29-41, New York, NY, USA, 2003. ACM Press. [8] T. Barron. Multiplayer Game Programming, chapter 16-17, pages 672-731. Prima Tech"s Game Development Series. Prima Publishing, 2001. 8 The 5th Workshop on Network & System Support for Games 2006 - NETGAMES 2006 [9] Carsten Griwodz and P˚al Halvorsen. The fun of using tcp for an mmorpg. In NOSSDAV "06: Proceedings of the International Workshop on Network and Operating Systems Support for Digital Audio and VIdeo, New York, NY, USA, 2006. ACM Press. [10] Sudhir Aggarwal, Hemant Banavar, Amit Khandelwal, Sarit Mukherjee, and Sampath Rangarajan. Accuracy in dead-reckoning based distributed multi-player games. In NetGames "04: Proceedings of 3rd ACM SIGCOMM workshop on Network and system support for games, pages 161-165, New York, NY, USA, 2004. ACM Press. [11] Sudhir Aggarwal, Hemant Banavar, Sarit Mukherjee, and Sampath Rangarajan. Fairness in dead-reckoning based distributed multi-player games. In NetGames "05: Proceedings of 4th ACM SIGCOMM workshop on Network and system support for games, pages 1-10, New York, NY, USA, 2005. ACM Press. [12] Riker, T. et al. Bzflag. http://www.bzflag.org, 2000-2006. [13] Linden Lab. Second life. http://secondlife.com, 2003. [14] Martin Mauve. How to keep a dead man from shooting. In IDMS "00: Proceedings of the 7th International Workshop on Interactive Distributed Multimedia Systems and Telecommunication Services, pages 199-204, London, UK, 2000. Springer-Verlag. [15] Max Skibinsky. Massively Multiplayer Game Development 2, chapter The Quest for Holy ScalePart 2: P2P Continuum, pages 355-373. Charles River Media, 2005. [16] Joseph D. Pellegrino and Constantinos Dovrolis. Bandwidth requirement and state consistency in three multiplayer game architectures. In NetGames "03: Proceedings of the 2nd workshop on Network and system support for games, pages 52-59, New York, NY, USA, 2003. ACM Press. [17] M. Mauve J. Widmer and S. Fischer. A Generic Proxy Systems for Networked Computer Games. In Proc. of the Workshop on Network Games, Netgames 2002, April 2002. [18] S. Gorlatch J. Muller, S. Fischer and M.Mauve. A Proxy Server Network Architecture for Real-Time Computer Games. In Euor-Par 2004 Parallel Processing: 10th International EURO-PAR Conference, August-September 2004. [19] H. Hazeyama T. Limura and Y. Kadobayashi. Zoned Federation of Game Servers: A Peer-to-Peer Approach to Scalable Multiplayer On-line Games. In Proc. of ACM Workshop on Network Games, Netgames 2004, August-September 2004. [20] B. Kelly and S. Aggarwal. A Framework for a Fidelity Based Agent Architecture for Distributed Interactive Simulation. In Proc. 14th Workshop on Standards for Distributed Interactive Simulation, pages 541-546, March 1996. [21] S. Aggarwal and B. Kelly. Hierarchical Structuring for Distributed Interactive Simulation. In Proc. 13th Workshop on Standards for Distributed Interactive Simulation, pages 125-132, Sept 1995. [22] Even Balance, Inc. Punkbuster. http://www.evenbalance.com/, 2001-2006. [23] Y. Wang and J. Vassileva. Trust and Reputation Model in Peer-to-Peer Networks. In Third International Conference on Peer-to-Peer Computing, 2003. The 5th Workshop on Network & System Support for Games 2006 - NETGAMES 2006 9

cheat-proof mechanism;latency compensation;role playing game;artificial latency;mmog;central-server architecture;assignment of authority;authority;proxy-based game architecture;first person shooter;communication proxy;distribute multi-player game;client authoritative approach;authority assignment;multi-player online game

train_C-62

Network Monitors and Contracting Systems: Competition and Innovation

Today"s Internet industry suffers from several well-known pathologies, but none is as destructive in the long term as its resistance to evolution. Rather than introducing new services, ISPs are presently moving towards greater commoditization. It is apparent that the network"s primitive system of contracts does not align incentives properly. In this study, we identify the network"s lack of accountability as a fundamental obstacle to correcting this problem: Employing an economic model, we argue that optimal routes and innovation are impossible unless new monitoring capability is introduced and incorporated with the contracting system. Furthermore, we derive the minimum requirements a monitoring system must meet to support first-best routing and innovation characteristics. Our work does not constitute a new protocol; rather, we provide practical and specific guidance for the design of monitoring systems, as well as a theoretical framework to explore the factors that influence innovation.

1. INTRODUCTION Many studies before us have noted the Internet"s resistance to new services and evolution. In recent decades, numerous ideas have been developed in universities, implemented in code, and even written into the routers and end systems of the network, only to languish as network operators fail to turn them on on a large scale. The list includes Multicast, IPv6, IntServ, and DiffServ. Lacking the incentives just to activate services, there seems to be little hope of ISPs devoting adequate resources to developing new ideas. In the long term, this pathology stands out as a critical obstacle to the network"s continued success (Ratnasamy, Shenker, and McCanne provide extensive discussion in [11]). On a smaller time scale, ISPs shun new services in favor of cost cutting measures. Thus, the network has characteristics of a commodity market. Although in theory, ISPs have a plethora of routing policies at their disposal, the prevailing strategy is to route in the cheapest way possible [2]. On one hand, this leads directly to suboptimal routing. More importantly, commoditization in the short term is surely related to the lack of innovation in the long term. When the routing decisions of others ignore quality characteristics, ISPs are motivated only to lower costs. There is simply no reward for introducing new services or investing in quality improvements. In response to these pathologies and others, researchers have put forth various proposals for improving the situation. These can be divided according to three high-level strategies: The first attempts to improve the status quo by empowering end-users. Clark, et al., suggest that giving end-users control over routing would lead to greater service diversity, recognizing that some payment mechanism must also be provided [5]. Ratnasamy, Shenker, and McCanne postulate a link between network evolution and user-directed routing [11]. They propose a system of Anycast to give end-users the ability to tunnel their packets to an ISP that introduces a desirable protocol. The extra traffic to the ISP, the authors suggest, will motivate the initial investment. The second strategy suggests a revision of the contracting system. This is exemplified by MacKie-Mason and Varian, who propose a smart market to control access to network resources [10]. Prices are set to the market-clearing level based on bids that users associate to their traffic. In another direction, Afergan and Wroclawski suggest that prices should be explicitly encoded in the routing protocols [2]. They argue that such a move would improve stability and align incentives. The third high-level strategy calls for greater network accountability. In this vein, Argyraki, et al., propose a system of packet obituaries to provide feedback as to which ISPs drop packets [3]. They argue that such feedback would help reveal which ISPs were adequately meeting their contractual obligations. Unlike the first two strategies, we are not aware of any previous studies that have connected accountability with the pathologies of commoditization or lack of innovation. It is clear that these three strategies are closely linked to each other (for example, [2], [5], and [9] each argue that giving end-users routing control within the current contracting system is problematic). Until today, however, the relationship between them has been poorly understood. There is currently little theoretical foundation to compare the relative merits of each proposal, and a particular lack of evidence linking accountability with innovation and service differentiation. This paper will address both issues. We will begin by introducing an economic network model that relates accountability, contracts, competition, and innovation. Our model is highly stylized and may be considered preliminary: it is based on a single source sending data to a single destination. Nevertheless, the structure is rich enough to expose previously unseen features of network behavior. We will use our model for two main purposes: First, we will use our model to argue that the lack of accountability in today"s network is a fundamental obstacle to overcoming the pathologies of commoditization and lack of innovation. In other words, unless new monitoring capabilities are introduced, and integrated with the system of contracts, the network cannot achieve optimal routing and innovation characteristics. This result provides motivation for the remainder of the paper, in which we explore how accountability can be leveraged to overcome these pathologies and create a sustainable industry. We will approach this problem from a clean-slate perspective, deriving the level of accountability needed to sustain an ideal competitive structure. When we say that today"s Internet has poor accountability, we mean that it reveals little information about the behavior - or misbehavior - of ISPs. This well-known trait is largely rooted in the network"s history. In describing the design philosophy behind the Internet protocols, Clark lists accountability as the least important among seven second level goals. [4] Accordingly, accountability received little attention during the network"s formative years. Clark relates this to the network"s military context, and finds that had the network been designed for commercial development, accountability would have been a top priority. Argyraki, et al., conjecture that applying the principles of layering and transparency may have led to the network"s lack of accountability [3]. According to these principles, end hosts should be informed of network problems only to the extent that they are required to adapt. They notice when packet drops occur so that they can perform congestion control and retransmit packets. Details of where and why drops occur are deliberately concealed. The network"s lack of accountability is highly relevant to a discussion of innovation because it constrains the system of contracts. This is because contracts depend upon external institutions to function - the judge in the language of incomplete contract theory, or simply the legal system. Ultimately, if a judge cannot verify that some condition holds, she cannot enforce a contract based on that condition. Of course, the vast majority of contracts never end up in court. Especially when a judge"s ruling is easily predicted, the parties will typically comply with the contract terms on their own volition. This would not be possible, however, without the judge acting as a last resort. An institution to support contracts is typically complex, but we abstract it as follows: We imagine that a contract is an algorithm that outputs a payment transfer among a set of ISPs (the parties) at every time. This payment is a function of the past and present behaviors of the participants, but only those that are verifiable. Hence, we imagine that a contract only accepts proofs as inputs. We will call any process that generates these proofs a contractible monitor. Such a monitor includes metering or sensing devices on the physical network, but it is a more general concept. Constructing a proof of a particular behavior may require readings from various devices distributed among many ISPs. The contractible monitor includes whatever distributed algorithmic mechanism is used to motivate ISPs to share this private information. Figure 1 demonstrates how our model of contracts fits together. We make the assumption that all payments are mediated by contracts. This means that without contractible monitors that attest to, say, latency, payments cannot be conditioned on latency. Figure 1: Relationship between monitors and contracts With this model, we may conclude that the level of accountability in today"s Internet only permits best effort contracts. Nodes cannot condition payments on either quality or path characteristics. Is there anything wrong with best-effort contracts? The reader might wonder why the Internet needs contracts at all. After all, in non-network industries, traditional firms invest in research and differentiate their products, all in the hopes of keeping their customers and securing new ones. One might believe that such market forces apply to ISPs as well. We may adopt this as our null hypothesis: Null hypothesis: Market forces are sufficient to maintain service diversity and innovation on a network, at least to the same extent as they do in traditional markets. There is a popular intuitive argument that supports this hypothesis, and it may be summarized as follows: Intuitive argument supporting null hypothesis: 1. Access providers try to increase their quality to get more consumers. 2. Access providers are themselves customers for second hop ISPs, and the second hops will therefore try to provide highquality service in order to secure traffic from access providers. Access providers try to select high quality transit because that increases their quality. 3. The process continues through the network, giving every ISP a competitive reason to increase quality. We are careful to model our network in continuous time, in order to capture the essence of this argument. We can, for example, specify equilibria in which nodes switch to a new next hop in the event of a quality drop. Moreover, our model allows us to explore any theoretically possible punishments against cheaters, including those that are costly for end-users to administer. By contrast, customers in the real world rarely respond collectively, and often simply seek the best deal currently offered. These constraints limit their ability to punish cheaters. Even with these liberal assumptions, however, we find that we must reject our null hypothesis. Our model will demonstrate that identifying a cheating ISP is difficult under low accountability, limiting the threat of market driven punishment. We will define an index of commoditization and show that it increases without bound as data paths grow long. Furthermore, we will demonstrate a framework in which an ISP"s maximum research investment decreases hyperbolically with its distance from the end-user. Network Behavior Monitor Contract Proof Payments 184 To summarize, we argue that the Internet"s lack of accountability must be addressed before the pathologies of commoditization and lack of innovation can be resolved. This leads us to our next topic: How can we leverage accountability to overcome these pathologies? We approach this question from a clean-slate perspective. Instead of focusing on incremental improvements, we try to imagine how an ideal industry would behave, then derive the level of accountability needed to meet that objective. According to this approach, we first craft a new equilibrium concept appropriate for network competition. Our concept includes the following requirements: First, we require that punishing ISPs that cheat is done without rerouting the path. Rerouting is likely to prompt end-users to switch providers, punishing access providers who administer punishments correctly. Next, we require that the equilibrium cannot be threatened by a coalition of ISPs that exchanges illicit side payments. Finally, we require that the punishment mechanism that enforces contracts does not punish innocent nodes that are not in the coalition. The last requirement is somewhat unconventional from an economic perspective, but we maintain that it is crucial for any reasonable solution. Although ISPs provide complementary services when they form a data path together, they are likely to be horizontal competitors as well. If innocent nodes may be punished, an ISP may decide to deliberately cheat and draw punishment onto itself and its neighbors. By cheating, the ISP may save resources, thereby ensuring that the punishment is more damaging to the other ISPs, which probably compete with the cheater directly for some customers. In the extreme case, the cheater may force the other ISPs out of business, thereby gaining a monopoly on some routes. Applying this equilibrium concept, we derive the monitors needed to maintain innovation and optimize routes. The solution is surprisingly simple: contractible monitors must report the quality of the rest of the path, from each ISP to the destination. It turns out that this is the correct minimum accountability requirement, as opposed to either end-to-end monitors or hop-by-hop monitors, as one might initially suspect. Rest of path monitors can be implemented in various ways. They may be purely local algorithms that listen for packet echoes. Alternately, they can be distributed in nature. We describe a way to construct a rest of path monitor out of monitors for individual ISP quality and for the data path. This requires a mechanism to motivate ISPs to share their monitor outputs with each other. The rest of path monitor then includes the component monitors and the distributed algorithmic mechanism that ensures that information is shared as required. This example shows that other types of monitors may be useful as building blocks, but must be combined to form rest of path monitors in order to achieve ideal innovation characteristics. Our study has several practical implications for future protocol design. We show that new monitors must be implemented and integrated with the contracting system before the pathologies of commoditization and lack of innovation can be overcome. Moreover, we derive exactly what monitors are needed to optimize routes and support innovation. In addition, our results provide useful input for clean-slate architectural design, and we use several novel techniques that we expect will be applicable to a variety of future research. The rest of this paper is organized as follows: In section 2, we lay out our basic network model. In section 3, we present a lowaccountability network, modeled after today"s Internet. We demonstrate how poor monitoring causes commoditization and a lack of innovation. In section 4, we present verifiable monitors, and show that proofs, even without contracts, can improve the status quo. In section 5, we turn our attention to contractible monitors. We show that rest of path monitors can support competition games with optimal routing and innovation. We further show that rest of path monitors are required to support such competition games. We continue by discussing how such monitors may be constructed using other monitors as building blocks. In section 6, we conclude and present several directions for future research. 2. BASIC NETWORK MODEL A source, S, wants to send data to destination, D. S and D are nodes on a directed, acyclic graph, with a finite set of intermediate nodes, { }NV ,...2,1= , representing ISPs. All paths lead to D, and every node not connected to D has at least two choices for next hop. We will represent quality by a finite dimensional vector space, Q, called the quality space. Each dimension represents a distinct network characteristic that end-users care about. For example, latency, loss probability, jitter, and IP version can each be assigned to a dimension. To each node, i, we associate a vector in the quality space, Qqi ∈ . This corresponds to the quality a user would experience if i were the only ISP on the data path. Let N Q∈q be the vector of all node qualities. Of course, when data passes through multiple nodes, their qualities combine in some way to yield a path quality. We represent this by an associative binary operation, *: QQQ →× . For path ( )nvvv ,...,, 21 , the quality is given by nvvv qqq ∗∗∗ ...21 . The * operation reflects the characteristics of each dimension of quality. For example, * can act as an addition in the case of latency, multiplication in the case of loss probability, or a minimumargument function in the case of security. When data flows along a complete path from S to D, the source and destination, generally regarded as a single player, enjoy utility given by a function of the path quality, →Qu : . Each node along the path, i, experiences some cost of transmission, ci. 2.1 Game Dynamics Ultimately, we are most interested in policies that promote innovation on the network. In this study, we will use innovation in a fairly general sense. Innovation describes any investment by an ISP that alters its quality vector so that at least one potential data path offers higher utility. This includes researching a new routing algorithm that decreases the amount of jitter users experience. It also includes deploying a new protocol that supports quality of service. Even more broadly, buying new equipment to decrease S D 185 latency may also be regarded as innovation. Innovation may be thought of as the micro-level process by which the network evolves. Our analysis is limited in one crucial respect: We focus on inventions that a single ISP can implement to improve the end-user experience. This excludes technologies that require adoption by all ISPs on the network to function. Because such technologies do not create a competitive advantage, rewarding them is difficult and may require intellectual property or some other market distortion. We defer this interesting topic to future work. At first, it may seem unclear how a large-scale distributed process such as innovation can be influenced by mechanical details like networks monitors. Our model must draw this connection in a realistic fashion. The rate of innovation depends on the profits that potential innovators expect in the future. The reward generated by an invention must exceed the total cost to develop it, or the inventor will not rationally invest. This reward, in turn, is governed by the competitive environment in which the firm operates, including the process by which firms select prices, and agree upon contracts with each other. Of course, these decisions depend on how routes are established, and how contracts determine actual monetary exchanges. Any model of network innovation must therefore relate at least three distinct processes: innovation, competition, and routing. We select a game dynamics that makes the relation between these processes as explicit as possible. This is represented schematically in Figure 2. The innovation stage occurs first, at time 2−=t . In this stage, each agent decides whether or not to make research investments. If she chooses not to, her quality remains fixed. If she makes an investment, her quality may change in some way. It is not necessary for us to specify how such changes take place. The agents" choices in this stage determine the vector of qualities, q, common knowledge for the rest of the game. Next, at time 1−=t , agents participate in the competition stage, in which contracts are agreed upon. In today"s industry, these contracts include prices for transit access, and peering agreements. Since access is provided on a best-effort basis, a transit agreement can simply be represented by its price. Other contracting systems we will explore will require more detail. Finally, beginning at 0=t , firms participate in the routing stage. Other research has already employed repeated games to study routing, for example [1], [12]. Repetition reveals interesting effects not visible in a single stage game, such as informal collusion to elevate prices in [12]. We use a game in continuous time in order to study such properties. For example, we will later ask whether a player will maintain higher quality than her contracts require, in the hope of keeping her customer base or attracting future customers. Our dynamics reflect the fact that ISPs make innovation decisions infrequently. Although real firms have multiple opportunities to innovate, each opportunity is followed by a substantial length of time in which qualities are fixed. The decision to invest focuses on how the firm"s new quality will improve the contracts it can enter into. Hence, our model places innovation at the earliest stage, attempting to capture a single investment decision. Contracting decisions are made on an intermediate time scale, thus appearing next in the dynamics. Routing decisions are made very frequently, mainly to maximize immediate profit flows, so they appear in the last stage. Because of this ordering, our model does not allow firms to route strategically to affect future innovation or contracting decisions. In opposition, Afergan and Wroclawski argue that contracts are formed in response to current traffic patterns, in a feedback loop [2]. Although we are sympathetic to their observation, such an addition would make our analysis intractable. Our model is most realistic when contracting decisions are infrequent. Throughout this paper, our solution concept will be a subgame perfect equilibrium (SPE). An SPE is a strategy point that is a Nash equilibrium when restricted to each subgame. Three important subgames have been labeled in Figure 2. The innovation game includes all three stages. The competition game includes only the competition stage and the routing stage. The routing game includes only the routing stage. An SPE guarantees that players are forward-looking. This means, for example, that in the competition stage, firms must act rationally, maximizing their expected profits in the routing stage. They cannot carry out threats they made in the innovation stage if it lowers their expected payoff. Our schematic already suggests that the routing game is crucial for promoting innovation. To support innovation, the competition game must somehow reward ISPs with high quality. But that means that the routing game must tend to route to nodes with high quality. If the routing game always selects the lowest-cost routes, for example, innovation will not be supported. We will support this observation with analysis later. 2.2 The Routing Game The routing game proceeds in continuous time, with all players discounting by a common factor, r. The outputs from previous stages, q and the set of contracts, are treated as exogenous parameters for this game. For each time 0≥t , each node must select a next hop to route data to. Data flows across the resultant path, causing utility flow to S and D, and a flow cost to the nodes on the path, as described above. Payment flows are also created, based on the contracts in place. Relating our game to the familiar repeated prisoners" dilemma, imagine that we are trying to impose a high quality, but costly path. As we argued loosely above, such paths must be sustainable in order to support innovation. Each ISP on the path tries to maximize her own payment, net of costs, so she may not want to cooperate with our plan. Rather, if she can find a way to save on costs, at the expense of the high quality we desire, she will be tempted to do so. Innovation Game Competition Game Routing Game Innovation stage Competition stage Routing stageQualities (q) Contracts (prices) Profits t = -2 t = -1 t ∈ [ 0 , ) Figure 2: Game Dynamics 186 Analogously to the prisoners" dilemma, we will call such a decision cheating. A little more formally, Cheating refers to any action that an ISP can take, contrary to some target strategy point that we are trying to impose, that enhances her immediate payoff, but compromises the quality of the data path. One type of cheating relates to the data path. Each node on the path has to pay the next node to deliver its traffic. If the next node offers high quality transit, we may expect that a lower quality node will offer a lower price. Each node on the path will be tempted to route to a cheaper next hop, increasing her immediate profits, but lowering the path quality. We will call this type of action cheating in route. Another possibility we can model, is that a node finds a way to save on its internal forwarding costs, at the expense of its own quality. We will call this cheating internally to distinguish it from cheating in route. For example, a node might drop packets beyond the rate required for congestion control, in order to throttle back TCP flows and thus save on forwarding costs [3]. Alternately, a node employing quality of service could give high priority packets a lower class of service, thus saving on resources and perhaps allowing itself to sell more high priority service. If either cheating in route or cheating internally is profitable, the specified path will not be an equilibrium. We assume that cheating can never be caught instantaneously. Rather, a cheater can always enjoy the payoff from cheating for some positive time, which we label 0t . This includes the time for other players to detect and react to the cheating. If the cheater has a contract which includes a customer lock-in period, 0t also includes the time until customers are allowed to switch to a new ISP. As we will see later, it is socially beneficial to decrease 0t , so such lock-in is detrimental to welfare. 3. PATHOLOGIES OF A LOWACCOUNTABILITY NETWORK In order to motivate an exploration of monitoring systems, we begin in this section by considering a network with a poor degree of accountability, modeled after today"s Internet. We will show how the lack of monitoring necessarily leads to poor routing and diminishes the rate of innovation. Thus, the network"s lack of accountability is a fundamental obstacle to resolving these pathologies. 3.1 Accountability in the Current Internet First, we reflect on what accountability characteristics the present Internet has. Argyraki, et al., point out that end hosts are given minimal information about packet drops [3]. Users know when drops occur, but not where they occur, nor why. Dropped packets may represent the innocent signaling of congestion, or, as we mentioned above, they may be a form of cheating internally. The problem is similar for other dimensions of quality, or in fact more acute. Finding an ISP that gives high priority packets a lower class of service, for example, is further complicated by the lack of even basic diagnostic tools. In fact, it is similarly difficult to identify an ISP that cheats in route. Huston notes that Internet traffic flows do not always correspond to routing information [8]. An ISP may hand a packet off to a neighbor regardless of what routes that neighbor has advertised. Furthermore, blocks of addresses are summarized together for distant hosts, so a destination may not even be resolvable until packets are forwarded closer. One might argue that diagnostic tools like ping and traceroute can identify cheaters. Unfortunately, Argyraki, et al., explain that these tools only reveal whether probe packets are echoed, not the fate of past packets [3]. Thus, for example, they are ineffective in detecting low-frequency packet drops. Even more fundamentally, a sophisticated cheater can always spot diagnostic packets and give them special treatment. As a further complication, a cheater may assume different aliases for diagnostic packets arriving over different routes. As we will see below, this gives the cheater a significant advantage in escaping punishment for bad behavior, even if the data path is otherwise observable. 3.2 Modeling Low-Accountability As the above evidence suggests, the current industry allows for very little insight into the behavior of the network. In this section, we attempt to capture this lack of accountability in our model. We begin by defining a monitor, our model of the way that players receive external information about network behavior, A monitor is any distributed algorithmic mechanism that runs on the network graph, and outputs, to specific nodes, informational statements about current or past network behavior. We assume that all external information about network behavior is mediated in this way. The accountability properties of the Internet can be represented by the following monitors: E2E (End to End): A monitor that informs S/D about what the total path quality is at any time (this is the quality they experience). ROP (Rest of Path): A monitor that informs each node along the data path what the quality is for the rest of the path to the destination. PRc (Packets Received): A monitor that tells nodes how much data they accept from each other, so that they can charge by volume. It is important to note, however, that this information is aggregated over many source-destination pairs. Hence, for the sake of realism, it cannot be used to monitor what the data path is. Players cannot measure the qualities of other, single nodes, just the rest of the path. Nodes cannot see the path past the next hop. This last assumption is stricter than needed for our results. The critical ingredient is that nodes cannot verify that the path avoids a specific hop. This holds, for example, if the path is generally visible, except nodes can use different aliases for different parents. Similar results also hold if alternate paths always converge after some integer number, m, of hops. It is important to stress that E2E and ROP are not the contractible monitors we described in the introduction - they do not generate proofs. Thus, even though a player observes certain information, she generally cannot credibly share it with another player. For example, if a node after the first hop starts cheating, the first hop will detect the sudden drop in quality for the rest of the path, but the first hop cannot make the source believe this observation - the 187 source will suspect that the first hop was the cheater, and fabricated the claim against the rest of the path. Typically, E2E and ROP are envisioned as algorithms that run on a single node, and listen for packet echoes. This is not the only way that they could be implemented, however; an alternate strategy is to aggregate quality measurements from multiple points in the network. These measurements can originate in other monitors, located at various ISPs. The monitor then includes the component monitors as well as whatever mechanisms are in place to motivate nodes to share information honestly as needed. For example, if the source has monitors that reveal the qualities of individual nodes, they could be combined with path information to create an ROP monitor. Since we know that contracts only accept proofs as input, we can infer that payments in this environment can only depend on the number of packets exchanged between players. In other words, contracts are best-effort. For the remainder of this section, we will assume that contracts are also linear - there is a constant payment flow so long as a node accepts data, and all conditions of the contract are met. Other, more complicated tariffs are also possible, and are typically used to generate lock-in. We believe that our parameter t0 is sufficient to describe lock-in effects, and we believe that the insights in this section apply equally to any tariffs that are bounded so that the routing game remains continuous at infinity. Restricting attention to linear contracts allows us to represent some node i"s contract by its price, pi. Because we further know that nodes cannot observe the path after the next hop, we can infer that contracts exist only between neighboring nodes on the graph. We will call this arrangement of contracts bilateral. When a competition game exclusively uses bilateral contracts, we will call it a bilateral contract competition game. We first focus on the routing game and ask whether a high quality route can be maintained, even when a low quality route is cheaper. Recall that this is a requirement in order for nodes to have any incentive to innovate. If nodes tend to route to low price next hops, regardless of quality, we say that the network is commoditized. To measure this tendency, we define an index of commoditization as follows: For a node on the data path, i, define its quality premium, minppd ji −= , where pj is the flow payment to the next hop in equilibrium, and pmin is the price of the lowest cost next hop. Definition: The index of commoditization, CI , is the average, over each node on the data path, i, of i"s flow profit as a fraction of i"s quality premium, ( ) ijii dpcp /−− . CI ranges from 0, when each node spends all of its potential profit on its quality premium, to infinite, when a node absorbs positive profit, but uses the lowest price next hop. A high value for CI implies that nodes are spending little of their money inflow on purchasing high quality for the rest of the path. As the next claim shows, this is exactly what happens as the path grows long: Claim 1. If the only monitors are E2E, ROP, and PRc, ∞→CI as ∞→n , where n is the number of nodes on the data path. To show that this is true, we first need the following lemma, which will establish the difficulty of punishing nodes in the network. First a bit of notation: Recall that a cheater can benefit from its actions for 00 >t before other players can react. When a node cheats, it can expect a higher profit flow, at least until it is caught and other players react, perhaps by diverting traffic. Let node i"s normal profit flow be iπ , and her profit flow during cheating be some greater value, yi. We will call the ratio, iiy π/ , the temptation to cheat. Lemma 1. If the only monitors are E2E, ROP, and PRc, the discounted time, −nt rt e 0 , needed to punish a cheater increases at least as fast as the product of the temptations to cheat along the data path, ∏ −− ≥ 0 0 pathdataon 0 t rt i i i t rt e y e n π (1) Corollary. If nodes share a minimum temptation to cheat, π/y , the discounted time needed to punish cheating increases at least exponentially in the length of the data path, n, −− ≥ 0 00 t rt nt rt e y e n π (2) Since it is the discounted time that increases exponentially, the actual time increases faster than exponentially. If n is so large that tn is undefined, the given path cannot be maintained in equilibrium. Proof. The proof proceeds by induction on the number of nodes on the equilibrium data path, n. For 1=n , there is a single node, say i. By cheating, the node earns extra profit ( ) − − 0 0 t rt ii ey π . If node i is then punished until time 1t , the extra profit must be cancelled out by the lost profit between time 0t and 1t , −1 0 t t rt i eπ . A little manipulation gives −− = 01 00 t rt i i t rt e y e π , as required. For 1>n , assume for induction that the claim holds for 1−n . The source does not know whether the cheater is the first hop, or after the first hop. Because the source does not know the data path after the first hop, it is unable to punish nodes beyond it. If it chooses a new first hop, it might not affect the rest of the data path. Because of this, the source must rely on the first hop to punish cheating nodes farther along the path. The first hop needs discounted time, ∏ −0 0 hopfirstafter t rt i i i e y π , to accomplish this by assumption. So the source must give the first hop this much discounted time in order to punish defectors further down the line (and the source will expect poor quality during this period). Next, the source must be protected against a first hop that cheats, and pretends that the problem is later in the path. The first hop can 188 do this for the full discounted time, ∏ −0 0 hopfirstafter t rt i i i e y π , so the source must punish the first hop long enough to remove the extra profit it can make. Following the same argument as for 1=n , we can show that the full discounted time is ∏ −0 0 pathdataon t rt i i i e y π , which completes the proof. The above lemma and its corollary show that punishing cheaters becomes more and more difficult as the data path grows long, until doing so is impossible. To capture some intuition behind this result, imagine that you are an end user, and you notice a sudden drop in service quality. If your data only travels through your access provider, you know it is that provider"s fault. You can therefore take your business elsewhere, at least for some time. This threat should motivate your provider to maintain high quality. Suppose, on the other hand, that your data traverses two providers. When you complain to your ISP, he responds, yes, we know your quality went down, but it"s not our fault, it"s the next ISP. Give us some time to punish them and then normal quality will resume. If your access provider is telling the truth, you will want to listen, since switching access providers may not even route around the actual offender. Thus, you will have to accept lower quality service for some longer time. On the other hand, you may want to punish your access provider as well, in case he is lying. This means you have to wait longer to resume normal service. As more ISPs are added to the path, the time increases in a recursive fashion. With this lemma in hand, we can return to prove Claim 1. Proof of Claim 1. Fix an equilibrium data path of length n. Label the path nodes 1,2,…,n. For each node i, let i"s quality premium be '11 ++ −= iii ppd . Then we have, [ ] = − = − + + = − + ++ = + −=− −− −− = −− − = −− = n i i n i iii iii n i iii ii n i i iii C g npcp pcp n pcp pp nd pcp n I 1 1 1 1 1 1 1 1 1 11 1 1 1 1 1 '1 '11 , (3) where gi is node i"s temptation to cheat by routing to the lowest price next hop. Lemma 1 tells us that Tg n i i <∏ =1 , where ( )01 rt eT − −= . It requires a bit of calculus to show that IC is minimized by setting each gi equal to n T /1 . However, as ∞→n , we have 1/1 →n T , which shows that ∞→CI . According to the claim, as the data path grows long, it increasingly resembles a lowest-price path. Since lowest-price routing does not support innovation, we may speculate that innovation degrades with the length of the data path. Though we suspect stronger claims are possible, we can demonstrate one such result by including an extra assumption: Available Bargain Path: A competitive market exists for lowcost transit, such that every node can route to the destination for no more than flow payment, lp . Claim 2. Under the available bargain path assumption, if node i , a distance n from S, can invest to alter its quality, and the source will spend no more than sP for a route including node i"s new quality, then the payment to node i, p, decreases hyperbolically with n, ( ) ( ) s n l P n T pp 1 1/1 − +≤ − , (4) where ( )01 rt eT − −= is the bound on the product of temptations from the previous claim. Thus, i will spend no more than ( ) ( )− + − s n l P n T p r 1 1 1/1 on this quality improvement, which approaches the bargain path"s payment, r pl , as ∞→n . The proof is given in the appendix. As a node gets farther from the source, its maximum payment approaches the bargain price, pl. Hence, the reward for innovation is bounded by the same amount. Large innovations, meaning substantially more expensive than rpl / , will not be pursued deep into the network. Claim 2 can alternately be viewed as a lower bound on how much it costs to elicit innovation in a network. If the source S wants node i to innovate, it needs to get a motivating payment, p, to i during the routing stage. However, it must also pay the nodes on the way to i a premium in order to motivate them to route properly. The claim shows that this premium increases with the distance to i, until it dwarfs the original payment, p. Our claims stand in sharp contrast to our null hypothesis from the introduction. Comparing the intuitive argument that supported our hypothesis with these claims, we can see that we implicitly used an oversimplified model of market pressure (as either present or not). As is now clear, market pressure relies on the decisions of customers, but these are limited by the lack of information. Hence, competitive forces degrade as the network deepens. 4. VERIFIABLE MONITORS In this section, we begin to introduce more accountability into the network. Recall that in the previous section, we assumed that players couldn"t convince each other of their private information. What would happen if they could? If a monitor"s informational signal can be credibly conveyed to others, we will call it a verifiable monitor. The monitor"s output in this case can be thought of as a statement accompanied by a proof, a string that can be processed by any player to determine that the statement is true. A verifiable monitor is a distributed algorithmic mechanism that runs on the network graph, and outputs, to specific nodes, proofs about current or past network behavior. Along these lines, we can imagine verifiable counterparts to E2E and ROP. We will label these E2Ev and ROPv. With these monitors, each node observes the quality of the rest of the path and can also convince other players of these observations by giving them a proof. 189 By adding verifiability to our monitors, identifying a single cheater is straightforward. The cheater is the node that cannot produce proof that the rest of path quality decreased. This means that the negative results of the previous section no longer hold. For example, the following lemma stands in contrast to Lemma 1. Lemma 2. With monitors E2Ev, ROPv, and PRc, and provided that the node before each potential cheater has an alternate next hop that isn"t more expensive, it is possible to enforce any data path in SPE so long as the maximum temptation is less than what can be deterred in finite time, − ≤ 0 0 max 1 t rt er y π (5) Proof. This lemma follows because nodes can share proofs to identify who the cheater is. Only that node must be punished in equilibrium, and the preceding node does not lose any payoff in administering the punishment. With this lemma in mind, it is easy to construct counterexamples to Claim 1 and Claim 2 in this new environment. Unfortunately, there are at least four reasons not to be satisfied with this improved monitoring system. The first, and weakest reason is that the maximum temptation remains finite, causing some distortion in routes or payments. Each node along a route must extract some positive profit unless the next hop is also the cheapest. Of course, if t0 is small, this effect is minimal. The second, and more serious reason is that we have always given our source the ability to commit to any punishment. Real world users are less likely to act collectively, and may simply search for the best service currently offered. Since punishment phases are generally characterized by a drop in quality, real world end-users may take this opportunity to shop for a new access provider. This will make nodes less motivated to administer punishments. The third reason is that Lemma 2 does not apply to cheating by coalitions. A coalition node may pretend to punish its successor, but instead enjoy a secret payment from the cheating node. Alternately, a node may bribe its successor to cheat, if the punishment phase is profitable, and so forth. The required discounted time for punishment may increase exponentially in the number of coalition members, just as in the previous section! The final reason not to accept this monitoring system is that when a cheater is punished, the path will often be routed around not just the offender, but around other nodes as well. Effectively, innocent nodes will be punished along with the guilty. In our abstract model, this doesn"t cause trouble since the punishment falls off the equilibrium path. The effects are not so benign in the real world. When ISPs lie in sequence along a data path, they contribute complementary services, and their relationship is vertical. From the perspective of other source-destination pairs, however, these same firms are likely to be horizontal competitors. Because of this, a node might deliberately cheat, in order to trigger punishment for itself and its neighbors. By cheating, the node will save money to some extent, so the cheater is likely to emerge from the punishment phase better off than the innocent nodes. This may give the cheater a strategic advantage against its competitors. In the extreme, the cheater may use such a strategy to drive neighbors out of business, and thereby gain a monopoly on some routes. 5. CONTRACTIBLE MONITORS At the end of the last section, we identified several drawbacks that persist in an environment with E2Ev, ROPv, and PRc. In this section, we will show how all of these drawbacks can be overcome. To do this, we will require our third and final category of monitor: A contractible monitor is simply a verifiable monitor that generates proofs that can serve as input to a contract. Thus, contractible is jointly a property of the monitor and the institutions that must verify its proofs. Contractibility requires that a court, 1. Can verify the monitor"s proofs. 2. Can understand what the proofs and contracts represent to the extent required to police illegal activity. 3. Can enforce payments among contracting parties. Understanding the agreements between companies has traditionally been a matter of reading contracts on paper. This may prove to be a harder task in a future network setting. Contracts may plausibly be negotiated by machine, be numerous, even per-flow, and be further complicated by the many dimensions of quality. When a monitor (together with institutional infrastructure) meets these criteria, we will label it with a subscript c, for contractible. The reader may recall that this is how we labeled the packets received monitor, PRc, which allows ISPs to form contracts with per-packet payments. Similarly, E2Ec and ROPc are contractible versions of the monitors we are now familiar with. At the end of the previous section, we argued for some desirable properties that we"d like our solution to have. Briefly, we would like to enforce optimal data paths with an equilibrium concept that doesn"t rely on re-routing for punishment, is coalition proof, and doesn"t punish innocent nodes when a coalition cheats. We will call such an equilibrium a fixed-route coalition-proof protect-theinnocent equilibrium. As the next claim shows, ROPc allows us to create a system of linear (price, quality) contracts under just such an equilibrium. Claim 3. With ROPc, for any feasible and consistent assignment of rest of path qualities to nodes, and any corresponding payment schedule that yields non-negative payoffs, these qualities can be maintained with bilateral contracts in a fixed-route coalition-proof protect-the-innocent equilibrium. Proof: Fix any data path consistent with the given rest of path qualities. Select some monetary punishment, P, large enough to prevent any cheating for time t0 (the discounted total payment from the source will work). Let each node on the path enter into a contract with its parent, which fixes an arbitrary payment schedule so long as the rest of path quality is as prescribed. When the parent node, which has ROPc, submits a proof that the rest of path quality is less than expected, the contract awards her an instantaneous transfer, P, from the downstream node. Such proofs can be submitted every 0t for the previous interval. Suppose now that a coalition, C, decides to cheat. The source measures a decrease in quality, and according to her contract, is awarded P from the first hop. This means that there is a net outflow of P from the ISPs as a whole. Suppose that node i is not in C. In order for the parent node to claim P from i, it must submit proof that the quality of the path starting at i is not as prescribed. This means 190 that there is a cheater after i. Hence, i would also have detected a change in quality, so i can claim P from the next node on the path. Thus, innocent nodes are not punished. The sequence of payments must end by the destination, so the net outflow of P must come from the nodes in C. This establishes all necessary conditions of the equilibrium. Essentially, ROPc allows for an implementation of (price, quality) contracts. Building upon this result, we can construct competition games in which nodes offer various qualities to each other at specified prices, and can credibly commit to meet these performance targets, even allowing for coalitions and a desire to damage other ISPs. Example 1. Define a Stackelberg price-quality competition game as follows: Extend the partial order of nodes induced by the graph to any complete ordering, such that downstream nodes appear before their parents. In this order, each node selects a contract to offer to its parents, consisting of a rest of path quality, and a linear price. In the routing game, each node selects a next hop at every time, consistent with its advertised rest of path quality. The Stackelberg price-quality competition game can be implemented in our model with ROPc monitors, by using the strategy in the proof, above. It has the following useful property: Claim 4. The Stackelberg price-quality competition game yields optimal routes in SPE. The proof is given in the appendix. This property is favorable from an innovation perspective, since firms that invest in high quality will tend to fall on the optimal path, gaining positive payoff. In general, however, investments may be over or under rewarded. Extra conditions may be given under which innovation decisions approach perfect efficiency for large innovations. We omit the full analysis here. Example 2. Alternately, we can imagine that players report their private information to a central authority, which then assigns all contracts. For example, contracts could be computed to implement the cost-minimizing VCG mechanism proposed by Feigenbaum, et al. in [7]. With ROPc monitors, we can adapt this mechanism to maximize welfare. For node, i, on the optimal path, L, the net payment must equal, essentially, its contribution to the welfare of S, D, and the other nodes. If L" is an optimal path in the graph with i removed, the profit flow to i is, ( ) ( ) ∈≠∈ +−− ', ' Lj j ijLj jLL ccququ , (6) where Lq and 'Lq are the qualities of the two paths. Here, (price, quality) contracts ensure that nodes report their qualities honestly. The incentive structure of the VCG mechanism is what motivates nodes to report their costs accurately. A nice feature of this game is that individual innovation decisions are efficient, meaning that a node will invest in an innovation whenever the investment cost is less than the increased welfare of the optimal data path. Unfortunately, the source may end up paying more than the utility of the path. Notice that with just E2Ec, a weaker version of Claim 3 holds. Bilateral (price, quality) contracts can be maintained in an equilibrium that is fixed-route and coalition-proof, but not protectthe-innocent. This is done by writing contracts to punish everyone on the path when the end to end quality drops. If the path length is n, the first hop pays nP to the source, the second hop pays ( )Pn 1− to the first, and so forth. This ensures that every node is punished sufficiently to make cheating unprofitable. For the reasons we gave previously, we believe that this solution concept is less than ideal, since it allows for malicious nodes to deliberately trigger punishments for potential competitors. Up to this point, we have adopted fixed-route coalition-proof protect-the-innocent equilibrium as our desired solution concept, and shown that ROPc monitors are sufficient to create some competition games that are desirable in terms of service diversity and innovation. As the next claim will show, rest of path monitoring is also necessary to construct such games under our solution concept. Before we proceed, what does it mean for a game to be desirable from the perspective of service diversity and innovation? We will use a very weak assumption, essentially, that the game is not fully commoditized for any node. The claim will hold for this entire class of games. Definition: A competition game is nowhere-commoditized if for each node, i, not adjacent to D, there is some assignment of qualities and marginal costs to nodes, such that the optimal data path includes i, and i has a positive temptation to cheat. In the case of linear contracts, it is sufficient to require that ∞<CI , and that every node make positive profit under some assignment of qualities and marginal costs. Strictly speaking, ROPc monitors are not the only way to construct these desirable games. To prove the next claim, we must broaden our notion of rest of path monitoring to include the similar ROPc" monitor, which attests to the quality starting at its own node, through the end of the path. Compare the two monitors below: ROPc: gives a node proof that the path quality from the next node to the destination is not correct. ROPc": gives a node proof that the path quality from that node to the destination is correct. We present a simplified version of this claim, by including an assumption that only one node on the path can cheat at a time (though conspirators can still exchange side payments). We will discuss the full version after the proof. Claim 5. Assume a set of monitors, and a nowhere-commoditized bilateral contract competition game that always maintains the optimal quality in fixed-route coalition-proof protect-the-innocent equilibrium, with only one node allowed to cheat at a time. Then for each node, i, not adjacent to D, either i has an ROPc monitor, or i"s children each have an ROPc" monitor. Proof: First, because of the fixed-route assumption, punishments must be purely monetary. Next, when cheating occurs, if the payment does not go to the source or destination, it may go to another coalition member, rendering it ineffective. Thus, the source must accept some monetary compensation, net of its normal flow payment, when cheating occurs. Since the source only contracts with the first hop, it must accept this money from the first hop. The source"s contract must therefore distinguish when the path quality is normal from when it is lowered by cheating. To do so, it can either accept proofs 191 from the source, that the quality is lower than required, or it can accept proofs from the first hop, that the quality is correct. These nodes will not rationally offer the opposing type of proof. By definition, any monitor that gives the source proof that the path quality is wrong is an ROPc monitor. Any monitor that gives the first hop proof that the quality is correct is a ROPc" monitor. Thus, at least one of these monitors must exist. By the protect-the-innocent assumption, if cheating occurs, but the first hop is not a cheater, she must be able to claim the same size reward from the next ISP on the path, and thus pass on the punishment. The first hop"s contract with the second must then distinguish when cheating occurs after the first hop. By argument similar to that for the source, either the first hop has a ROPc monitor, or the second has a ROPc" monitor. This argument can be iterated along the entire path to the penultimate node before D. Since the marginal costs and qualities can be arranged to make any path the optimal path, these statements must hold for all nodes and their children, which completes the proof. The two possibilities for monitor correspond to which node has the burden of proof. In one case, the prior node must prove the suboptimal quality to claim its reward. In the other, the subsequent node must prove that the quality was correct to avoid penalty. Because the two monitors are similar, it seems likely that they require comparable costs to implement. If submitting the proofs is costly, it seems natural that nodes would prefer to use the ROPc monitor, placing the burden of proof on the upstream node. Finally, we note that it is straightforward to derive the full version of the claim, which allows for multiple cheaters. The only complication is that cheaters can exchange side payments, which makes any money transfers between them redundant. Because of this, we have to further generalize our rest of path monitors, so they are less constrained in the case that there are cheaters on either side. 5.1 Implementing Monitors Claim 5 should not be interpreted as a statement that each node must compute the rest of path quality locally, without input from other nodes. Other monitors, besides ROPc and ROPc" can still be used, loosely speaking, as building blocks. For instance, network tomography is concerned with measuring properties of the network interior with tools located at the edge. Using such techniques, our source might learn both individual node qualities and the data path. This is represented by the following two monitors: SHOPc i : (source-based hop quality) A monitor that gives the source proof of what the quality of node i is. SPATHc: (source-based path) A monitor that gives the source proof of what the data path is at any time, at least as far as it matches the equilibrium path. With these monitors, a punishment mechanism can be designed to fulfill the conditions of Claim 5. It involves the source sharing the proofs it generates with nodes further down the path, which use them to determine bilateral payments. Ultimately however, the proof of Claim 5 shows us that each node i"s bilateral contracts require proof of the rest of path quality. This means that node i (or possibly its children) will have to combine the proofs that they receive to generate a proof of the rest of path quality. Thus, the combined process is itself a rest of path monitor. What we have done, all in all, is constructed a rest of path monitor using SPATHc and SHOPc i as building blocks. Our new monitor includes both the component monitors and whatever distributed algorithmic mechanism exists to make sure nodes share their proofs correctly. This mechanism can potentially involve external institutions. For a concrete example, suppose that when node i suspects it is getting poor rest of path quality from its successor, it takes the downstream node to court. During the discovery process, the court subpoenas proofs of the path and of node qualities from the source (ultimately, there must be some threat to ensure the source complies). Finally, for the court to issue a judgment, one party or the other must compile a proof of what the rest of path quality was. Hence, the entire discovery process acts as a rest of path monitor, albeit a rather costly monitor in this case. Of course, mechanisms can be designed to combine these monitors at much lower cost. Typically, such mechanisms would call for automatic sharing of proofs, with court intervention only as a last resort. We defer these interesting mechanisms to future work. As an aside, intuition might dictate that SHOPc i generates more information than ROPc; after all, inferring individual node qualities seems a much harder problem. Yet, without path information, SHOPc i is not sufficient for our first-best innovation result. The proof of this demonstrates a useful technique: Claim 6. With monitors E2E, ROP, SHOPc i and PRc, and a nowhere-commoditized bilateral contract competition game, the optimal quality cannot be maintained for all assignments of quality and marginal cost, in fixed-route coalition-proof protect-theinnocent equilibrium. Proof: Because nodes cannot verify the data path, they cannot form a proof of what the rest of path quality is. Hence, ROPc monitors do not exist, and therefore the requirements of Claim 5 cannot hold. 6. CONCLUSIONS AND FUTURE WORK It is our hope that this study will have a positive impact in at least three different ways. The first is practical: we believe our analysis has implications for the design of future monitoring protocols and for public policy. For protocol designers, we first provide fresh motivation to create monitoring systems. We have argued that the poor accountability of the Internet is a fundamental obstacle to alleviating the pathologies of commoditization and lack of innovation. Unless accountability improves, these pathologies are guaranteed to remain. Secondly, we suggest directions for future advances in monitoring. We have shown that adding verifiability to monitors allows for some improvements in the characteristics of competition. At the same time, this does not present a fully satisfying solution. This paper has suggested a novel standard for monitors to aspire to - one of supporting optimal routes in innovative competition games under fixed-route coalition-proof protect-the-innocent equilibrium. We have shown that under bilateral contracts, this specifically requires contractible rest of path monitors. This is not to say that other types of monitors are unimportant. We included an example in which individual hop quality monitors and a path monitor can also meet our standard for sustaining competition. However, in order for this to happen, a mechanism must be included 192 to combine proofs from these monitors to form a proof of rest of path quality. In other words, the monitors must ultimately be combined to form contractible rest of path monitors. To support service differentiation and innovation, it may be easier to design rest of path monitors directly, thereby avoiding the task of designing mechanisms for combining component monitors. As far as policy implications, our analysis points to the need for legal institutions to enforce contracts based on quality. These institutions must be equipped to verify proofs of quality, and police illegal contracting behavior. As quality-based contracts become numerous and complicated, and possibly negotiated by machine, this may become a challenging task, and new standards and regulations may have to emerge in response. This remains an interesting and unexplored area for research. The second area we hope our study will benefit is that of clean-slate architectural design. Traditionally, clean-slate design tends to focus on creating effective and elegant networks for a static set of requirements. Thus, the approach is often one of engineering, which tends to neglect competitive effects. We agree with Ratnasamy, Shenker, and McCanne, that designing for evolution should be a top priority [11]. We have demonstrated that the network"s monitoring ability is critical to supporting innovation, as are the institutions that support contracting. These elements should feature prominently in new designs. Our analysis specifically suggests that architectures based on bilateral contracts should include contractible rest of path monitoring. From a clean-slate perspective, these monitors can be transparently and fully integrated with the routing and contracting systems. Finally, the last contribution our study makes is methodological. We believe that the mathematical formalization we present is applicable to a variety of future research questions. While a significant literature addresses innovation in the presence of network effects, to the best of our knowledge, ours is the first model of innovation in a network industry that successfully incorporates the actual topological structure as input. This allows the discovery of new properties, such as the weakening of market forces with the number of ISPs on a data path that we observe with lowaccountability. Our method also stands in contrast to the typical approach of distributed algorithmic mechanism design. Because this field is based on a principle-agent framework, contracts are usually proposed by the source, who is allowed to make a take it or leave it offer to network nodes. Our technique allows contracts to emerge from a competitive framework, so the source is limited to selecting the most desirable contract. We believe this is a closer reflection of the industry. Based on the insights in this study, the possible directions for future research are numerous and exciting. To some degree, contracting based on quality opens a Pandora"s Box of pressing questions: Do quality-based contracts stand counter to the principle of network neutrality? Should ISPs be allowed to offer a choice of contracts at different quality levels? What anti-competitive behaviors are enabled by quality-based contracts? Can a contracting system support optimal multicast trees? In this study, we have focused on bilateral contracts. This system has seemed natural, especially since it is the prevalent system on the current network. Perhaps its most important benefit is that each contract is local in nature, so both parties share a common, familiar legal jurisdiction. There is no need to worry about who will enforce a punishment against another ISP on the opposite side of the planet, nor is there a dispute over whose legal rules to apply in interpreting a contract. Although this benefit is compelling, it is worth considering other systems. The clearest alternative is to form a contract between the source and every node on the path. We may call these source contracts. Source contracting may present surprising advantages. For instance, since ISPs do not exchange money with each other, an ISP cannot save money by selecting a cheaper next hop. Additionally, if the source only has contracts with nodes on the intended path, other nodes won"t even be willing to accept packets from this source since they won"t receive compensation for carrying them. This combination seems to eliminate all temptation for a single cheater to cheat in route. Because of this and other encouraging features, we believe source contracts are a fertile topic for further study. Another important research task is to relax our assumption that quality can be measured fully and precisely. One possibility is to assume that monitoring is only probabilistic or suffers from noise. Even more relevant is the possibility that quality monitors are fundamentally incomplete. A quality monitor can never anticipate every dimension of quality that future applications will care about, nor can it anticipate a new and valuable protocol that an ISP introduces. We may define a monitor space as a subspace of the quality space that a monitor can measure, QM ⊂ , and a corresponding monitoring function that simply projects the full range of qualities onto the monitor space, MQm →: . Clearly, innovations that leave quality invariant under m are not easy to support - they are invisible to the monitoring system. In this environment, we expect that path monitoring becomes more important, since it is the only way to ensure data reaches certain innovator ISPs. Further research is needed to understand this process. 7. ACKNOWLEDGEMENTS We would like to thank the anonymous reviewers, Jens Grossklags, Moshe Babaioff, Scott Shenker, Sylvia Ratnasamy, and Hal Varian for their comments. This work is supported in part by the National Science Foundation under ITR award ANI-0331659. 8. REFERENCES [1] Afergan, M. Using Repeated Games to Design IncentiveBased Routing Systems. In Proceedings of IEEE INFOCOM (April 2006). [2] Afergan, M. and Wroclawski, J. On the Benefits and Feasibility of Incentive Based Routing Infrastructure. In ACM SIGCOMM'04 Workshop on Practice and Theory of Incentives in Networked Systems (PINS) (August 2004). [3] Argyraki, K., Maniatis, P., Cheriton, D., and Shenker, S. Providing Packet Obituaries. In Third Workshop on Hot Topics in Networks (HotNets) (November 2004). [4] Clark, D. D. The Design Philosophy of the DARPA Internet Protocols. In Proceedings of ACM SIGCOMM (1988). 193 [5] Clark, D. D., Wroclawski, J., Sollins, K. R., and Braden, R. Tussle in cyberspace: Defining tomorrow's internet. In Proceedings of ACM SIGCOMM (August 2002). [6] Dang-Nguyen, G. and Pénard, T. Interconnection Agreements: Strategic Behaviour and Property Rights. In Brousseau, E. and Glachant, J.M. Eds. The Economics of Contracts: Theories and Applications, Cambridge University Press, 2002. [7] Feigenbaum, J., Papadimitriou, C., Sami, R., and Shenker, S. A BGP-based Mechanism for Lowest-Cost Routing. Distributed Computing 18 (2005), pp. 61-72. [8] Huston, G. Interconnection, Peering, and Settlements. Telstra, Australia. [9] Liu, Y., Zhang, H., Gong, W., and Towsley, D. On the Interaction Between Overlay Routing and Traffic Engineering. In Proceedings of IEEE INFOCOM (2005). [10] MacKie-Mason, J. and Varian, H. Pricing the Internet. In Kahin, B. and Keller, J. Eds. Public access to the Internet. Englewood Cliffs, NJ; Prentice-Hall, 1995. [11] Ratnasamy, S., Shenker, S., and McCanne, S. Towards an Evolvable Internet Architecture. In Proceeding of ACM SIGCOMM (2005). [12] Shakkottai, S., and Srikant, R. Economics of Network Pricing with Multiple ISPs. In Proceedings of IEEE INFOCOM (2005). 9. APPENDIX Proof of Claim 2. Node i must fall on the equilibrium data path to receive any payment. Let the prices along the data path be ppppP nS == ,..., 21 , with marginal costs, ncc ,...,1 . We may assume the prices on the path are greater than lp or the claim follows trivially. Each node along the data path can cheat in route by giving data to the bargain path at price no more than lp . So node j"s temptation to cheat is at least 11 ++ − − ≥ −− −− jj l jjj ljj pp pp pcp pcp . Then Lemma 1 gives, 1 13221 1... − − − −> − − ⋅⋅ − − − − ≥ n S l n lll P pp n pp pp pp pp pp pp T (7) This can be rearranged to give ( ) ( ) s n l P n T pp 1 1/1 − +≤ − , as required. The rest of the claim simply recognizes that rp / is the greatest reward node i can receive for its investment, so it will not invest sums greater than this. Proof of Claim 4. Label the nodes 1,2,.. N in the order in which they select contracts. Let subgame n be the game that begins with n choosing its contract. Let Ln be the set of possible paths restricted to nodes n,…,N. That is, Ln is the set of possible routes from S to reach some node that has already moved. For subgame n, define the local welfare over paths nLl ∈ , and their possible next hops, nj < as follows, ( ) ( ) j li ipathjl pcqqujlV −−= ∈ *, , (8) where ql is the quality of path l in the set {n,…,N}, and pathjq and pj are the quality and price of the contract j has offered. For induction, assume that subgame n + 1 maximizes local welfare. We show that subgame n does as well. If node n selects next hop k, we can write the following relation, ( ) ( )( ) ( )( ) nknn knlVpcpknlVnlV π−=++−= ,,,,, , (9) where n is node n"s profit if the path to n is chosen. This path is chosen whenever ( )nlV , is maximal over Ln+1 and possible next hops. If ( )( )knlV ,, is maximal over Ln, it is also maximal over the paths in Ln+1 that don"t lead to n. This means that node n can choose some n small enough so that ( )nlV , is maximal over Ln+1, so the route will lead to k. Conversely, if ( )( )knlV ,, is not maximal over Ln, either V is greater for another of n"s next hops, in which case n will select that one in order to increase n, or V is greater for some path in Ln+1 that don"t lead to n, in which case ( )nlV , cannot be maximal for any nonnegative n. Thus, we conclude that subgame n maximizes local welfare. For the initial case, observe that this assumption holds for the source. Finally, we deduce that subgame 1, which is the entire game, maximizes local welfare, which is equivalent to actual welfare. Hence, the Stackelberg price-quality game yields an optimal route. 194

monitor;clean-slate architectural design;contracting system;verifiable monitor;innovation;contract;routing policy;network monitor;routing stagequality;contractible monitor;commoditization;smart market;incentive

train_C-65

Shooter Localization and Weapon Classification with Soldier-Wearable Networked Sensors

The paper presents a wireless sensor network-based mobile countersniper system. A sensor node consists of a helmetmounted microphone array, a COTS MICAz mote for internode communication and a custom sensorboard that implements the acoustic detection and Time of Arrival (ToA) estimation algorithms on an FPGA. A 3-axis compass provides self orientation and Bluetooth is used for communication with the soldier"s PDA running the data fusion and the user interface. The heterogeneous sensor fusion algorithm can work with data from a single sensor or it can fuse ToA or Angle of Arrival (AoA) observations of muzzle blasts and ballistic shockwaves from multiple sensors. The system estimates the trajectory, the range, the caliber and the weapon type. The paper presents the system design and the results from an independent evaluation at the US Army Aberdeen Test Center. The system performance is characterized by 1degree trajectory precision and over 95% caliber estimation accuracy for all shots, and close to 100% weapon estimation accuracy for 4 out of 6 guns tested.

1. INTRODUCTION The importance of countersniper systems is underscored by the constant stream of news reports coming from the Middle East. In October 2006 CNN reported on a new tactic employed by insurgents. A mobile sniper team moves around busy city streets in a car, positions itself at a good standoff distance from dismounted US military personnel, takes a single well-aimed shot and immediately melts in the city traffic. By the time the soldiers can react, they are gone. A countersniper system that provides almost immediate shooter location to every soldier in the vicinity would provide clear benefits to the warfigthers. Our team introduced PinPtr, the first sensor networkbased countersniper system [17, 8] in 2003. The system is based on potentially hundreds of inexpensive sensor nodes deployed in the area of interest forming an ad-hoc multihop network. The acoustic sensors measure the Time of Arrival (ToA) of muzzle blasts and ballistic shockwaves, pressure waves induced by the supersonic projectile, send the data to a base station where a sensor fusion algorithm determines the origin of the shot. PinPtr is characterized by high precision: 1m average 3D accuracy for shots originating within or near the sensor network and 1 degree bearing precision for both azimuth and elevation and 10% accuracy in range estimation for longer range shots. The truly unique characteristic of the system is that it works in such reverberant environments as cluttered urban terrain and that it can resolve multiple simultaneous shots at the same time. This capability is due to the widely distributed sensing and the unique sensor fusion approach [8]. The system has been tested several times in US Army MOUT (Military Operations in Urban Terrain) facilities. The obvious disadvantage of such a system is its static nature. Once the sensors are distributed, they cover a certain area. Depending on the operation, the deployment may be needed for an hour or a month, but eventually the area looses its importance. It is not practical to gather and reuse the sensors, especially under combat conditions. Even if the sensors are cheap, it is still a waste and a logistical problem to provide a continuous stream of sensors as the operations move from place to place. As it is primarily the soldiers that the system protects, a natural extension is to mount the sensors on the soldiers themselves. While there are vehiclemounted countersniper systems [1] available commercially, we are not aware of a deployed system that protects dismounted soldiers. A helmet-mounted system was developed in the mid 90s by BBN [3], but it was not continued beyond the Darpa program that funded it. 113 To move from a static sensor network-based solution to a highly mobile one presents significant challenges. The sensor positions and orientation need to be constantly monitored. As soldiers may work in groups of as little as four people, the number of sensors measuring the acoustic phenomena may be an order of magnitude smaller than before. Moreover, the system should be useful to even a single soldier. Finally, additional requirements called for caliber estimation and weapon classification in addition to source localization. The paper presents the design and evaluation of our soldierwearable mobile countersniper system. It describes the hardware and software architecture including the custom sensor board equipped with a small microphone array and connected to a COTS MICAz mote [12]. Special emphasis is paid to the sensor fusion technique that estimates the trajectory, range, caliber and weapon type simultaneously. The results and analysis of an independent evaluation of the system at the US Army Aberdeen Test Center are also presented. 2. APPROACH The firing of a typical military rifle, such as the AK47 or M16, produces two distinct acoustic phenomena. The muzzle blast is generated at the muzzle of the gun and travels at the speed sound. The supersonic projectile generates an acoustic shockwave, a kind of sonic boom. The wavefront has a conical shape, the angle of which depends on the Mach number, the speed of the bullet relative to the speed of sound. The shockwave has a characteristic shape resembling a capital N. The rise time at both the start and end of the signal is very fast, under 1 μsec. The length is determined by the caliber and the miss distance, the distance between the trajectory and the sensor. It is typically a few hundred μsec. Once a trajectory estimate is available, the shockwave length can be used for caliber estimation. Our system is based on four microphones connected to a sensorboard. The board detects shockwaves and muzzle blasts and measures their ToA. If at least three acoustic channels detect the same event, its AoA is also computed. If both the shockwave and muzzle blast AoA are available, a simple analytical solution gives the shooter location as shown in Section 6. As the microphones are close to each other, typically 2-4, we cannot expect very high precision. Also, this method does not estimate a trajectory. In fact, an infinite number of trajectory-bullet speed pairs satisfy the observations. However, the sensorboards are also connected to COTS MICAz motes and they share their AoA and ToA measurements, as well as their own location and orientation, with each other using a multihop routing service [9]. A hybrid sensor fusion algorithm then estimates the trajectory, the range, the caliber and the weapon type based on all available observations. The sensorboard is also Bluetooth capable for communication with the soldier"s PDA or laptop computer. A wired USB connection is also available. The sensorfusion algorithm and the user interface get their data through one of these channels. The orientation of the microphone array at the time of detection is provided by a 3-axis digital compass. Currently the system assumes that the soldier"s PDA is GPS-capable and it does not provide self localization service itself. However, the accuracy of GPS is a few meters degrading the Figure 1: Acoustic sensorboard/mote assembly . overall accuracy of the system. Refer to Section 7 for an analysis. The latest generation sensorboard features a Texas Instruments CC-1000 radio enabling the high-precision radio interferometric self localization approach we have developed separately [7]. However, we leave the integration of the two technologies for future work. 3. HARDWARE Since the first static version of our system in 2003, the sensor nodes have been built upon the UC Berkeley/Crossbow MICA product line [11]. Although rudimentary acoustic signal processing can be done on these microcontroller-based boards, they do not provide the required computational performance for shockwave detection and angle of arrival measurements, where multiple signals from different microphones need to be processed in parallel at a high sampling rate. Our 3rd generation sensorboard is designed to be used with MICAz motes-in fact it has almost the same size as the mote itself (see Figure 1). The board utilizes a powerful Xilinx XC3S1000 FPGA chip with various standard peripheral IP cores, multiple soft processor cores and custom logic for the acoustic detectors (Figure 2). The onboard Flash (4MB) and PSRAM (8MB) modules allow storing raw samples of several acoustic events, which can be used to build libraries of various acoustic signatures and for refining the detection cores off-line. Also, the external memory blocks can store program code and data used by the soft processor cores on the FPGA. The board supports four independent analog channels sampled at up to 1 MS/s (million samples per seconds). These channels, featuring an electret microphone (Panasonic WM64PNT), amplifiers with controllable gain (30-60 dB) and a 12-bit serial ADC (Analog Devices AD7476), reside on separate tiny boards which are connected to the main sensorboard with ribbon cables. This partitioning enables the use of truly different audio channels (eg.: slower sampling frequency, different gain or dynamic range) and also results in less noisy measurements by avoiding long analog signal paths. The sensor platform offers a rich set of interfaces and can be integrated with existing systems in diverse ways. An RS232 port and a Bluetooth (BlueGiga WT12) wireless link with virtual UART emulation are directly available on the board and provide simple means to connect the sensor to PCs and PDAs. The mote interface consists of an I2 C bus along with an interrupt and GPIO line (the latter one is used 114 Figure 2: Block diagram of the sensorboard. for precise time synchronization between the board and the mote). The motes are equipped with IEEE 802.15.4 compliant radio transceivers and support ad-hoc wireless networking among the nodes and to/from the base station. The sensorboard also supports full-speed USB transfers (with custom USB dongles) for uploading recorded audio samples to the PC. The on-board JTAG chain-directly accessible through a dedicated connector-contains the FPGA part and configuration memory and provides in-system programming and debugging facilities. The integrated Honeywell HMR3300 digital compass module provides heading, pitch and roll information with 1◦ accuracy, which is essential for calculating and combining directional estimates of the detected events. Due to the complex voltage requirements of the FPGA, the power supply circuitry is implemented on the sensorboard and provides power both locally and to the mote. We used a quad pack of rechargeable AA batteries as the power source (although any other configuration is viable that meets the voltage requirements). The FPGA core (1.2 V) and I/O (3.3 V) voltages are generated by a highly efficient buck switching regulator. The FPGA configuration (2.5 V) and a separate 3.3 V power net are fed by low current LDOs, the latter one is used to provide independent power to the mote and to the Bluetooth radio. The regulators-except the last one-can be turned on/off from the mote or through the Bluetooth radio (via GPIO lines) to save power. The first prototype of our system employed 10 sensor nodes. Some of these nodes were mounted on military kevlar helmets with the microphones directly attached to the surface at about 20 cm separation as shown in Figure 3(a). The rest of the nodes were mounted in plastic enclosures (Figure 3(b)) with the microphones placed near the corners of the boxes to form approximately 5 cm×10 cm rectangles. 4. SOFTWARE ARCHITECTURE The sensor application relies on three subsystems exploiting three different computing paradigms as they are shown in Figure 4. Although each of these execution models suit their domain specific tasks extremely well, this diversity (a) (b) Figure 3: Sensor prototypes mounted on a kevlar helmet (a) and in a plastic box on a tripod (b). presents a challenge for software development and system integration. The sensor fusion and user interface subsystem is running on PDAs and were implemented in Java. The sensing and signal processing tasks are executed by an FPGA, which also acts as a bridge between various wired and wireless communication channels. The ad-hoc internode communication, time synchronization and data sharing are the responsibilities of a microcontroller based radio module. Similarly, the application employs a wide variety of communication protocols such as Bluetooth and IEEE 802.14.5 wireless links, as well as optional UARTs, I2 C and/or USB buses. Soldier Operated Device (PDA/Laptop) FPGA Sensor Board Mica Radio Module 2.4 GHz Wireless Link Radio Control Message Routing Acoustic Event Encoder Sensor Time Synch. Network Time Synch.Remote Control Time stamping Interrupts Virtual Register Interface C O O R D I N A T O R A n a l o g c h a n n e l s Compass PicoBlaze Comm. Interface PicoBlaze WT12 Bluetooth Radio MOTE IF:I2C,Interrupts USB PSRAM U A R T U A R T MB det SW det REC Bluetooth Link User Interface Sensor Fusion Location Engine GPS Message (Dis-)AssemblerSensor Control Figure 4: Software architecture diagram. The sensor fusion module receives and unpacks raw measurements (time stamps and feature vectors) from the sensorboard through the Bluetooth link. Also, it fine tunes the execution of the signal processing cores by setting parameters through the same link. Note that measurements from other nodes along with their location and orientation information also arrive from the sensorboard which acts as a gateway between the PDA and the sensor network. The handheld device obtains its own GPS location data and di115 rectly receives orientation information through the sensorboard. The results of the sensor fusion are displayed on the PDA screen with low latency. Since, the application is implemented in pure Java, it is portable across different PDA platforms. The border between software and hardware is considerably blurred on the sensor board. The IP cores-implemented in hardware description languages (HDL) on the reconfigurable FPGA fabric-closely resemble hardware building blocks. However, some of them-most notably the soft processor cores-execute true software programs. The primary tasks of the sensor board software are 1) acquiring data samples from the analog channels, 2) processing acoustic data (detection), and 3) providing access to the results and run-time parameters through different interfaces. As it is shown in Figure 4, a centralized virtual register file contains the address decoding logic, the registers for storing parameter values and results and the point to point data buses to and from the peripherals. Thus, it effectively integrates the building blocks within the sensorboard and decouples the various communication interfaces. This architecture enabled us to deploy the same set of sensors in a centralized scenario, where the ad-hoc mote network (using the I2 C interface) collected and forwarded the results to a base station or to build a decentralized system where the local PDAs execute the sensor fusion on the data obtained through the Bluetooth interface (and optionally from other sensors through the mote interface). The same set of registers are also accessible through a UART link with a terminal emulation program. Also, because the low-level interfaces are hidden by the register file, one can easily add/replace these with new ones (eg.: the first generation of motes supported a standard μP interface bus on the sensor connector, which was dropped in later designs). The most important results are the time stamps of the detected events. These time stamps and all other timing information (parameters, acoustic event features) are based on a 1 MHz clock and an internal timer on the FPGA. The time conversion and synchronization between the sensor network and the board is done by the mote by periodically requesting the capture of the current timer value through a dedicated GPIO line and reading the captured value from the register file through the I2 C interface. Based on the the current and previous readings and the corresponding mote local time stamps, the mote can calculate and maintain the scaling factor and offset between the two time domains. The mote interface is implemented by the I2 C slave IP core and a thin adaptation layer which provides a data and address bus abstraction on top of it. The maximum effective bandwidth is 100 Kbps through this interface. The FPGA contains several UART cores as well: for communicating with the on-board Bluetooth module, for controlling the digital compass and for providing a wired RS232 link through a dedicated connector. The control, status and data registers of the UART modules are available through the register file. The higher level protocols on these lines are implemented by Xilinx PicoBlaze microcontroller cores [13] and corresponding software programs. One of them provides a command line interface for test and debug purposes, while the other is responsible for parsing compass readings. By default, they are connected to the RS232 port and to the on-board digital compass line respectively, however, they can be rewired to any communication interface by changing the register file base address in the programs (e.g. the command line interface can be provided through the Bluetooth channel). Two of the external interfaces are not accessible through the register file: a high speed USB link and the SRAM interface are tied to the recorder block. The USB module implements a simple FIFO with parallel data lines connected to an external FT245R USB device controller. The RAM driver implements data read/write cycles with correct timing and is connected to the on-board pseudo SRAM. These interfaces provide 1 MB/s effective bandwidth for downloading recorded audio samples, for example. The data acquisition and signal processing paths exhibit clear symmetry: the same set of IP cores are instantiated four times (i.e. the number of acoustic channels) and run independently. The signal paths meet only just before the register file. Each of the analog channels is driven by a serial A/D core for providing a 20 MHz serial clock and shifting in 8-bit data samples at 1 MS/s and a digital potentiometer driver for setting the required gain. Each channel has its own shockwave and muzzle blast detector, which are described in Section 5. The detectors fetch run-time parameter values from the register file and store their results there as well. The coordinator core constantly monitors the detection results and generates a mote interrupt promptly upon full detection or after a reasonable timeout after partial detection. The recorder component is not used in the final deployment, however, it is essential for development purposes for refining parameter values for new types of weapons or for other acoustic sources. This component receives the samples from all channels and stores them in circular buffers in the PSRAM device. If the signal amplitude on one of the channels crosses a predefined threshold, the recorder component suspends the sample collection with a predefined delay and dumps the contents of the buffers through the USB link. The length of these buffers and delays, the sampling rate, the threshold level and the set of recorded channels can be (re)configured run-time through the register file. Note that the core operates independently from the other signal processing modules, therefore, it can be used to validate the detection results off-line. The FPGA cores are implemented in VHDL, the PicoBlaze programs are written in assembly. The complete configuration occupies 40% of the resources (slices) of the FPGA and the maximum clock speed is 30 MHz, which is safely higher than the speed used with the actual device (20MHz). The MICAz motes are responsible for distributing measurement data across the network, which drastically improves the localization and classification results at each node. Besides a robust radio (MAC) layer, the motes require two essential middleware services to achieve this goal. The messages need to be propagated in the ad-hoc multihop network using a routing service. We successfully integrated the Directed Flood-Routing Framework (DFRF) [9] in our application. Apart from automatic message aggregation and efficient buffer management, the most unique feature of DFRF is its plug-in architecture, which accepts custom routing policies. Routing policies are state machines that govern how received messages are stored, resent or discarded. Example policies include spanning tree routing, broadcast, geographic routing, etc. Different policies can be used for different messages concurrently, and the application is able to 116 change the underlying policies at run-time (eg.: because of the changing RF environment or power budget). In fact, we switched several times between a simple but lavish broadcast policy and a more efficient gradient routing on the field. Correlating ToA measurements requires a common time base and precise time synchronization in the sensor network. The Routing Integrated Time Synchronization (RITS) [15] protocol relies on very accurate MAC-layer time-stamping to embed the cumulative delay that a data message accrued since the time of the detection in the message itself. That is, at every node it measures the time the message spent there and adds this to the number in the time delay slot of the message, right before it leaves the current node. Every receiving node can subtract the delay from its current time to obtain the detection time in its local time reference. The service provides very accurate time conversion (few μs per hop error), which is more than adequate for this application. Note, that the motes also need to convert the sensorboard time stamps to mote time as it is described earlier. The mote application is implemented in nesC [5] and is running on top of TinyOS [6]. With its 3 KB RAM and 28 KB program space (ROM) requirement, it easily fits on the MICAz motes. 5. DETECTION ALGORITHM There are several characteristics of acoustic shockwaves and muzzle blasts which distinguish their detection and signal processing algorithms from regular audio applications. Both events are transient by their nature and present very intense stimuli to the microphones. This is increasingly problematic with low cost electret microphones-designed for picking up regular speech or music. Although mechanical damping of the microphone membranes can mitigate the problem, this approach is not without side effects. The detection algorithms have to be robust enough to handle severe nonlinear distortion and transitory oscillations. Since the muzzle blast signature closely follows the shockwave signal and because of potential automatic weapon bursts, it is extremely important to settle the audio channels and the detection logic as soon as possible after an event. Also, precise angle of arrival estimation necessitates high sampling frequency (in the MHz range) and accurate event detection. Moreover, the detection logic needs to process multiple channels in parallel (4 channels on our existing hardware). These requirements dictated simple and robust algorithms both for muzzle blast and shockwave detections. Instead of using mundane energy detectors-which might not be able to distinguish the two different events-the applied detectors strive to find the most important characteristics of the two signals in the time-domain using simple state machine logic. The detectors are implemented as independent IP cores within the FPGA-one pair for each channel. The cores are run-time configurable and provide detection event signals with high precision time stamps and event specific feature vectors. Although the cores are running independently and in parallel, a crude local fusion module integrates them by shutting down those cores which missed their events after a reasonable timeout and by generating a single detection message towards the mote. At this point, the mote can read and forward the detection times and features and is responsible to restart the cores afterwards. The most conspicuous characteristics of an acoustic shockwave (see Figure 5(a)) are the steep rising edges at the be0 200 400 600 800 1000 1200 1400 1600 -1 -0.8 -0.6 -0.4 -0.2 0 0.2 0.4 0.6 0.8 1 Shockwave (M16) Time (µs) Amplitude 1 3 5 2 4 len (a) s[t] - s[t-D] > E tstart := t s[t] - s[t-D] < E s[t] - s[t-D] > E & t - t_start > Lmin s[t] - s[t-D] < E len := t - tstart IDLE 1 FIRST EDGE DONE 3 SECOND EDGE 4 FIRST EDGE 2 FOUND 5 t - tstart ≥ Lmax t - tstart ≥ Lmax (b) Figure 5: Shockwave signal generated by a 5.56 × 45 mm NATO projectile (a) and the state machine of the detection algorithm (b). ginning and end of the signal. Also, the length of the N-wave is fairly predictable-as it is described in Section 6.5-and is relatively short (200-300 μs). The shockwave detection core is continuously looking for two rising edges within a given interval. The state machine of the algorithm is shown in Figure 5(b). The input parameters are the minimum steepness of the edges (D, E), and the bounds on the length of the wave (Lmin, Lmax). The only feature calculated by the core is the length of the observed shockwave signal. In contrast to shockwaves, the muzzle blast signatures are characterized by a long initial period (1-5 ms) where the first half period is significantly shorter than the second half [4]. Due to the physical limitations of the analog circuitry described at the beginning of this section, irregular oscillations and glitches might show up within this longer time window as they can be clearly seen in Figure 6(a). Therefore, the real challenge for the matching detection core is to identify the first and second half periods properly. The state machine (Figure 6(b)) does not work on the raw samples directly but is fed by a zero crossing (ZC) encoder. After the initial triggering, the detector attempts to collect those ZC segments which belong to the first period (positive amplitude) while discarding too short (in our terminology: garbage) segments-effectively implementing a rudimentary low-pass filter in the ZC domain. After it encounters a sufficiently long negative segment, it runs the same collection logic for the second half period. If too much garbage is discarded in the collection phases, the core resets itself to prevent the (false) detection of the halves from completely different periods separated by rapid oscillation or noise. Finally, if the constraints on the total length and on the length ratio hold, the core generates a detection event along with the actual length, amplitude and energy of the period calculated concurrently. The initial triggering mechanism is based on two amplitude thresholds: one static (but configurable) amplitude level and a dynamically computed one. The latter one is essential to adapt the sensor to different ambient noise environments and to temporarily suspend the muzzle blast detector after a shock wave event (oscillations in the analog section or reverberations in the sensor enclosure might otherwise trigger false muzzle blast detections). The dynamic noise level is estimated by a single pole recursive low-pass filter (cutoff @ 0.5 kHz ) on the FPGA. 117 0 1000 2000 3000 4000 5000 6000 7000 8000 9000 10000 -1 -0.8 -0.6 -0.4 -0.2 0 0.2 0.4 0.6 0.8 1 Time (µs) Amplitude Muzzle blast (M16) 1 2 3 4 5 len2 + len1 (a) IDLE 1 SECOND ZC 3 PENDING ZC 4 FIRST ZC 2 FOUND 5 amplitude threshold long positive ZC long negative ZC valid full period max garbage wrong sign garbage collect first period garbage collect first period garbage (b) Figure 6: Muzzle blast signature (a) produced by an M16 assault rifle and the corresponding detection logic (b). The detection cores were originally implemented in Java and evaluated on pre-recorded signals because of much faster test runs and more convenient debugging facilities. Later on, they were ported to VHDL and synthesized using the Xilinx ISE tool suite. The functional equivalence between the two implementations were tested by VHDL test benches and Python scripts which provided an automated way to exercise the detection cores on the same set of pre-recorded signals and to compare the results. 6. SENSOR FUSION The sensor fusion algorithm receives detection messages from the sensor network and estimates the bullet trajectory, the shooter position, the caliber of the projectile and the type of the weapon. The algorithm consists of well separated computational tasks outlined below: 1. Compute muzzle blast and shockwave directions of arrivals for each individual sensor (see 6.1). 2. Compute range estimates. This algorithm can analytically fuse a pair of shockwave and muzzle blast AoA estimates. (see 6.2). 3. Compute a single trajectory from all shockwave measurements (see 6.3). 4. If trajectory available compute range (see 6.4). else compute shooter position first and then trajectory based on it. (see 6.4) 5. If trajectory available compute caliber (see 6.5). 6. If caliber available compute weapon type (see 6.6). We describe each step in the following sections in detail. 6.1 Direction of arrival The first step of the sensor fusion is to calculate the muzzle blast and shockwave AoA-s for each sensorboard. Each sensorboard has four microphones that measure the ToA-s. Since the microphone spacing is orders of magnitude smaller than the distance to the sound source, we can approximate the approaching sound wave front with a plane (far field assumption). Let us formalize the problem for 3 microphones first. Let P1, P2 and P3 be the position of the microphones ordered by time of arrival t1 < t2 < t3. First we apply a simple geometry validation step. The measured time difference between two microphones cannot be larger than the sound propagation time between the two microphones: |ti − tj| <= |Pi − Pj |/c + ε Where c is the speed of sound and ε is the maximum measurement error. If this condition does not hold, the corresponding detections are discarded. Let v(x, y, z) be the normal vector of the unknown direction of arrival. We also use r1(x1, y1, z1), the vector from P1 to P2 and r2(x2, y2, z2), the vector from P1 to P3. Let"s consider the projection of the direction of the motion of the wave front (v) to r1 divided by the speed of sound (c). This gives us how long it takes the wave front to propagate form P1 to P2: vr1 = c(t2 − t1) The same relationship holds for r2 and v: vr2 = c(t3 − t1) We also know that v is a normal vector: vv = 1 Moving from vectors to coordinates using the dot product definition leads to a quadratic system: xx1 + yy1 + zz1 = c(t2 − t1) xx2 + yy2 + zz2 = c(t3 − t1) x2 + y2 + z2 = 1 We omit the solution steps here, as they are straightforward, but long. There are two solutions (if the source is on the P1P2P3 plane the two solutions coincide). We use the fourth microphone"s measurement-if there is one-to eliminate one of them. Otherwise, both solutions are considered for further processing. 6.2 Muzzle-shock fusion u v 11,tP 22,tP tP, 2P′ Bullet trajectory Figure 7: Section plane of a shot (at P) and two sensors (at P1 and at P2). One sensor detects the muzzle blast"s, the other the shockwave"s time and direction of arrivals. Consider the situation in Figure 7. A shot was fired from P at time t. Both P and t are unknown. We have one muzzle blast and one shockwave detections by two different sensors 118 with AoA and hence, ToA information available. The muzzle blast detection is at position P1 with time t1 and AoA u. The shockwave detection is at P2 with time t2 and AoA v. u and v are normal vectors. It is shown below that these measurements are sufficient to compute the position of the shooter (P). Let P2 be the point on the extended shockwave cone surface where PP2 is perpendicular to the surface. Note that PP2 is parallel with v. Since P2 is on the cone surface which hits P2, a sensor at P2 would detect the same shockwave time of arrival (t2). The cone surface travels at the speed of sound (c), so we can express P using P2: P = P2 + cv(t2 − t). P can also be expressed from P1: P = P1 + cu(t1 − t) yielding P1 + cu(t1 − t) = P2 + cv(t2 − t). P2P2 is perpendicular to v: (P2 − P2)v = 0 yielding (P1 + cu(t1 − t) − cv(t2 − t) − P2)v = 0 containing only one unknown t. One obtains: t = (P1−P2)v c +uvt1−t2 uv−1 . From here we can calculate the shoter position P. Let"s consider the special single sensor case where P1 = P2 (one sensor detects both shockwave and muzzle blast AoA). In this case: t = uvt1−t2 uv−1 . Since u and v are not used separately only uv, the absolute orientation of the sensor can be arbitrary, we still get t which gives us the range. Here we assumed that the shockwave is a cone which is only true for constant projectile speeds. In reality, the angle of the cone slowly grows; the surface resembles one half of an American football. The decelerating bullet results in a smaller time difference between the shockwave and the muzzle blast detections because the shockwave generation slows down with the bullet. A smaller time difference results in a smaller range, so the above formula underestimates the true range. However, it can still be used with a proper deceleration correction function. We leave this for future work. 6.3 Trajectory estimation Danicki showed that the bullet trajectory and speed can be computed analytically from two independent shockwave measurements where both ToA and AoA are measured [2]. The method gets more sensitive to measurement errors as the two shockwave directions get closer to each other. In the special case when both directions are the same, the trajectory cannot be computed. In a real world application, the sensors are typically deployed on a plane approximately. In this case, all sensors located on one side of the trajectory measure almost the same shockwave AoA. To avoid this error sensitivity problem, we consider shockwave measurement pairs only if the direction of arrival difference is larger than a certain threshold. We have multiple sensors and one sensor can report two different directions (when only three microphones detect the shockwave). Hence, we typically have several trajectory candidates, i.e. one for each AoA pair over the threshold. We applied an outlier filtering and averaging method to fuse together the shockwave direction and time information and come up with a single trajectory. Assume that we have N individual shockwave AoA measurements. Let"s take all possible unordered pairs where the direction difference is above the mentioned threshold and compute the trajectory for each. This gives us at most N(N−1) 2 trajectories. A trajectory is represented by one point pi and the normal vector vi (where i is the trajectory index). We define the distance of two trajectories as the dot product of their normal vectors: D(i, j) = vivj For each trajectory a neighbor set is defined: N(i) := {j|D(i, j) < R} where R is a radius parameter. The largest neighbor set is considered to be the core set C, all other trajectories are outliers. The core set can be found in O(N2 ) time. The trajectories in the core set are then averaged to get the final trajectory. It can happen that we cannot form any sensor pairs because of the direction difference threshold. It means all sensors are on the same side of the trajectory. In this case, we first compute the shooter position (described in the next section) that fixes p making v the only unknown. To find v in this case, we use a simple high resolution grid search and minimize an error function based on the shockwave directions. We have made experiments to utilize the measured shockwave length in the trajectory estimation. There are some promising results, but it needs further research. 6.4 Shooter position estimation The shooter position estimation algorithm aggregates the following heterogenous information generated by earlier computational steps: 1. trajectory, 2. muzzle blast ToA at a sensor, 3. muzzle blast AoA at a sensor, which is effectively a bearing estimate to the shooter, and 4. range estimate at a sensor (when both shockwave and muzzle blast AoA are available). Some sensors report only ToA, some has bearing estimate(s) also and some has range estimate(s) as well, depending on the number of successful muzzle blast and shockwave detections by the sensor. For an example, refer to Figure 8. Note that a sensor may have two different bearing and range estimates. 3 detections gives two possible AoA-s for muzzle blast (i.e. bearing) and/or shockwave. Furthermore, the combination of two different muzzle blast and shockwave AoA-s may result in two different ranges. 119 11111 ,,,, rrvvt ′′ 22 ,vt 333 ,, vvt ′ 4t 5t 6t bullet trajectory shooter position Figure 8: Example of heterogenous input data for the shooter position estimation algorithm. All sensors have ToA measurements (t1, t2, t3, t4, t5), one sensor has a single bearing estimate (v2), one sensor has two possible bearings (v3, v3) and one sensor has two bearing and two range estimates (v1, v1,r1, r1) In a multipath environment, these detections will not only contain gaussian noise, but also possibly large errors due to echoes. It has been showed in our earlier work that a similar problem can be solved efficiently with an interval arithmetic based bisection search algorithm [8]. The basic idea is to define a discrete consistency function over the area of interest and subdivide the space into 3D boxes. For any given 3D box, this function gives the number of measurements supporting the hypothesis that the shooter was within that box. The search starts with a box large enough to contain the whole area of interest, then zooms in by dividing and evaluating boxes. The box with the maximum consistency is divided until the desired precision is reached. Backtracking is possible to avoid getting stuck in a local maximum. This approach has been shown to be fast enough for online processing. Note, however, that when the trajectory has already been calculated in previous steps, the search needs to be done only on the trajectory making it orders of magnitude faster. Next let us describe how the consistency function is calculated in detail. Consider B, a three dimensional box, we would like to compute the consistency value of. First we consider only the ToA information. If one sensor has multiple ToA detections, we use the average of those times, so one sensor supplies at most one ToA estimate. For each ToA, we can calculate the corresponding time of the shot, since the origin is assumed to be in box B. Since it is a box and not a single point, this gives us an interval for the shot time. The maximum number of overlapping time intervals gives us the value of the consistency function for B. For a detailed description of the consistency function and search algorithm, refer to [8]. Here we extend the approach the following way. We modify the consistency function based on the bearing and range data from individual sensors. A bearing estimate supports B if the line segment starting from the sensor with the measured direction intersects the B box. A range supports B, if the sphere with the radius of the range and origin of the sensor intersects B. Instead of simply checking whether the position specified by the corresponding bearing-range pairs falls within B, this eliminates the sensor"s possible orientation error. The value of the consistency function is incremented by one for each bearing and range estimate that is consistent with B. 6.5 Caliber estimation The shockwave signal characteristics has been studied before by Whitham [20]. He showed that the shockwave period T is related to the projectile diameter d, the length l, the perpendicular miss distance b from the bullet trajectory to the sensor, the Mach number M and the speed of sound c. T = 1.82Mb1/4 c(M2−1)3/8 d l1/4 ≈ 1.82d c (Mb l )1/4 0 100 200 300 400 500 600 0 10 20 30 miss distance (m)shockwavelength(microseconds) .50 cal 5.56 mm 7.62 mm Figure 9: Shockwave length and miss distance relationship. Each data point represents one sensorboard after an aggregation of the individual measurements of the four acoustic channels. Three different caliber projectiles have been tested (196 shots, 10 sensors). To illustrate the relationship between miss distance and shockwave length, here we use all 196 shots with three different caliber projectiles fired during the evaluation. (During the evaluation we used data obtained previously using a few practice shots per weapon.) 10 sensors (4 microphones by sensor) measured the shockwave length. For each sensor, we considered the shockwave length estimation valid if at least three out of four microphones agreed on a value with at most 5 microsecond variance. This filtering leads to a 86% report rate per sensor and gets rid of large measurement errors. The experimental data is shown in Figure 9. Whitham"s formula suggests that the shockwave length for a given caliber can be approximated with a power function of the miss distance (with a 1/4 exponent). Best fit functions on our data are: .50 cal: T = 237.75b0.2059 7.62 mm: T = 178.11b0.1996 5.56 mm: T = 144.39b0.1757 To evaluate a shot, we take the caliber whose approximation function results in the smallest RMS error of the filtered sensor readings. This method has less than 1% caliber estimation error when an accurate trajectory estimate is available. In other words, caliber estimation only works if enough shockwave detections are made by the system to compute a trajectory. 120 6.6 Weapon estimation We analyzed all measured signal characteristics to find weapon specific information. Unfortunately, we concluded that the observed muzzle blast signature is not characteristic enough of the weapon for classification purposes. The reflections of the high energy muzzle blast from the environment have much higher impact on the muzzle blast signal shape than the weapon itself. Shooting the same weapon from different places caused larger differences on the recorded signal than shooting different weapons from the same place. 0 100 200 300 400 500 600 700 800 900 0 100 200 300 400 range (m) speed(m/s) AK-47 M240 Figure 10: AK47 and M240 bullet deceleration measurements. Both weapons have the same caliber. Data is approximated using simple linear regression. 0 100 200 300 400 500 600 700 800 900 1000 0 50 100 150 200 250 300 350 range (m) speed(m/s) M16 M249 M4 Figure 11: M16, M249 and M4 bullet deceleration measurements. All weapons have the same caliber. Data is approximated using simple linear regression. However, the measured speed of the projectile and its caliber showed good correlation with the weapon type. This is because for a given weapon type and ammunition pair, the muzzle velocity is nearly constant. In Figures 10 and 11 we can see the relationship between the range and the measured bullet speed for different calibers and weapons. In the supersonic speed range, the bullet deceleration can be approximated with a linear function. In case of the 7.62 mm caliber, the two tested weapons (AK47, M240) can be clearly separated (Figure 10). Unfortunately, this is not necessarily true for the 5.56 mm caliber. The M16 with its higher muzzle speed can still be well classified, but the M4 and M249 weapons seem practically undistinguishable (Figure 11). However, this may be partially due to the limited number of practice shots we were able to take before the actual testing began. More training data may reveal better separation between the two weapons since their published muzzle velocities do differ somewhat. The system carries out weapon classification in the following manner. Once the trajectory is known, the speed can be calculated for each sensor based on the shockwave geometry. To evaluate a shot, we choose the weapon type whose deceleration function results in the smallest RMS error of the estimated range-speed pairs for the estimated caliber class. 7. RESULTS An independent evaluation of the system was carried out by a team from NIST at the US Army Aberdeen Test Center in April 2006 [19]. The experiment was setup on a shooting range with mock-up wooden buildings and walls for supporting elevated shooter positions and generating multipath effects. Figure 12 shows the user interface with an aerial photograph of the site. 10 sensor nodes were deployed on surveyed points in an approximately 30×30 m area. There were five fixed targets behind the sensor network. Several firing positions were located at each of the firing lines at 50, 100, 200 and 300 meters. These positions were known to the evaluators, but not to the operators of the system. Six different weapons were utilized: AK47 and M240 firing 7.62 mm projectiles, M16, M4 and M249 with 5.56mm ammunition and the .50 caliber M107. Note that the sensors remained static during the test. The primary reason for this is that nobody is allowed downrange during live fire tests. Utilizing some kind of remote control platform would have been too involved for the limited time the range was available for the test. The experiment, therefore, did not test the mobility aspect of the system. During the one day test, there were 196 shots fired. The results are summarized in Table 1. The system detected all shots successfully. Since a ballistic shockwave is a unique acoustic phenomenon, it makes the detection very robust. There were no false positives for shockwaves, but there were a handful of false muzzle blast detections due to parallel tests of artillery at a nearby range. Shooter Local- Caliber Trajectory Trajectory Distance No. Range ization Accu- Azimuth Distance Error of (m) Rate racy Error (deg) Error (m) (m) Shots 50 93% 100% 0.86 0.91 2.2 54 100 100% 100% 0.66 1.34 8.7 54 200 96% 100% 0.74 2.71 32.8 54 300 97% 97% 1.49 6.29 70.6 34 All 96% 99.5% 0.88 2.47 23.0 196 Table 1: Summary of results fusing all available sensor observations. All shots were successfully detected, so the detection rate is omitted. Localization rate means the percentage of shots that the sensor fusion was able to estimate the trajectory of. The caliber accuracy rate is relative to the shots localized and not all the shots because caliber estimation requires the trajectory. The trajectory error is broken down to azimuth in degrees and the actual distance of the shooter from the trajectory. The distance error shows the distance between the real shooter position and the estimated shooter position. As such, it includes the error caused by both the trajectory and that of the range estimation. Note that the traditional bearing and range measures are not good ones for a distributed system such as ours because of the lack of a single reference point. 121 Figure 12: The user interface of the system showing the experimental setup. The 10 sensor nodes are labeled by their ID and marked by dark circles. The targets are black squares marked T-1 through T-5. The long white arrows point to the shooter position estimated by each sensor. Where it is missing, the corresponding sensor did not have enough detections to measure the AoA of either the muzzle blast, the shockwave or both. The thick black line and large circle indicate the estimated trajectory and the shooter position as estimated by fusing all available detections from the network. This shot from the 100-meter line at target T-3 was localized almost perfectly by the sensor network. The caliber and weapon were also identified correctly. 6 out of 10 nodes were able to estimate the location alone. Their bearing accuracy is within a degree, while the range is off by less than 10% in the worst case. The localization rate characterizes the system"s ability to successfully estimate the trajectory of shots. Since caliber estimation and weapon classification relies on the trajectory, non-localized shots are not classified either. There were 7 shots out of 196 that were not localized. The reason for missed shots is the trajectory ambiguity problem that occurs when the projectile passes on one side of all the sensors. In this case, two significantly different trajectories can generate the same set of observations (see [8] and also Section 6.3). Instead of estimating which one is more likely or displaying both possibilities, we decided not to provide a trajectory at all. It is better not to give an answer other than a shot alarm than misleading the soldier. Localization accuracy is broken down to trajectory accuracy and range estimation precision. The angle of the estimated trajectory was better than 1 degree except for the 300 m range. Since the range should not affect trajectory estimation as long as the projectile passes over the network, we suspect that the slightly worse angle precision for 300 m is due to the hurried shots we witnessed the soldiers took near the end of the day. This is also indicated by another datapoint: the estimated trajectory distance from the actual targets has an average error of 1.3 m for 300 m shots, 0.75 m for 200 m shots and 0.6 m for all but 300 m shots. As the distance between the targets and the sensor network was fixed, this number should not show a 2× improvement just because the shooter is closer. Since the angle of the trajectory itself does not characterize the overall error-there can be a translation alsoTable 1 also gives the distance of the shooter from the estimated trajectory. These indicate an error which is about 1-2% of the range. To put this into perspective, a trajectory estimate for a 100 m shot will very likely go through or very near the window the shooter is located at. Again, we believe that the disproportionally larger errors at 300 m are due to human errors in aiming. As the ground truth was obtained by knowing the precise location of the shooter and the target, any inaccuracy in the actual trajectory directly adds to the perceived error of the system. We call the estimation of the shooter"s position on the calculated trajectory range estimation due to the lack of a better term. The range estimates are better than 5% accurate from 50 m and 10% for 100 m. However, this goes to 20% or worse for longer distances. We did not have a facility to test system before the evaluation for ranges beyond 100 m. During the evaluation, we ran into the problem of mistaking shockwave echoes for muzzle blasts. These echoes reached the sensors before the real muzzle blast for long range shots only, since the projectile travels 2-3× faster than the speed of sound, so the time between the shockwave (and its possible echo from nearby objects) and the muzzle blast increases with increasing ranges. This resulted in underestimating the range, since the system measured shorter times than the real ones. Since the evaluation we finetuned the muzzle blast detection algorithm to avoid this problem. Distance M16 AK47 M240 M107 M4 M249 M4-M249 50m 100% 100% 100% 100% 11% 25% 94% 100m 100% 100% 100% 100% 22% 33% 100% 200m 100% 100% 100% 100% 50% 22% 100% 300m 67% 100% 83% 100% 33% 0% 57% All 96% 100% 97% 100% 23% 23% 93% Table 2: Weapon classification results. The percentages are relative to the number of shots localized and not all shots, as the classification algorithm needs to know the trajectory and the range. Note that the difference is small; there were 189 shots localized out of the total 196. The caliber and weapon estimation accuracy rates are based on the 189 shots that were successfully localized. Note that there was a single shot that was falsely classified by the caliber estimator. The 73% overall weapon classification accuracy does not seem impressive. But if we break it down to the six different weapons tested, the picture changes dramatically as shown in Table 2. For four of the weapons (AK14, M16, M240 and M107), the classification rate is almost 100%. There were only two shots out of approximately 140 that were missed. The M4 and M249 proved to be too similar and they were mistaken for each other most of the time. One possible explanation is that we had only a limited number of test shots taken with these weapons right before the evaluation and used the wrong deceleration approximation function. Either this or a similar mistake was made 122 since if we simply used the opposite of the system"s answer where one of these weapons were indicated, the accuracy would have improved 3x. If we consider these two weapons a single weapon class, then the classification accuracy for it becomes 93%. Note that the AK47 and M240 have the same caliber (7.62 mm), just as the M16, M4 and M249 do (5.56 mm). That is, the system is able to differentiate between weapons of the same caliber. We are not aware of any system that classifies weapons this accurately. 7.1 Single sensor performance As was shown previously, a single sensor alone is able to localize the shooter if it can determine both the muzzle blast and the shockwave AoA, that is, it needs to measure the ToA of both on at least three acoustic channels. While shockwave detection is independent of the range-unless the projectile becomes subsonic-, the likelihood of muzzle blast detection beyond 150 meters is not enough for consistently getting at least three per sensor node for AoA estimation. Hence, we only evaluate the single sensor performance for the 104 shots that were taken from 50 and 100 m. Note that we use the same test data as in the previous section, but we evaluate individually for each sensor. Table 3 summarizes the results broken down by the ten sensors utilized. Since this is now not a distributed system, the results are given relative to the position of the given sensor, that is, a bearing and range estimate is provided. Note that many of the common error sources of the networked system do not play a role here. Time synchronization is not applicable. The sensor"s absolute location is irrelevant (just as the relative location of multiple sensors). The sensor"s orientation is still important though. There are several disadvantages of the single sensor case compared to the networked system: there is no redundancy to compensate for other errors and to perform outlier rejection, the localization rate is markedly lower, and a single sensor alone is not able to estimate the caliber or classify the weapon. Sensor id 1 2 3 5 7 8 9 10 11 12 Loc. rate 44% 37% 53% 52% 19% 63% 51% 31% 23% 44% Bearing (deg) 0.80 1.25 0.60 0.85 1.02 0.92 0.73 0.71 1.28 1.44 Range (m) 3.2 6.1 4.4 4.7 4.6 4.6 4.1 5.2 4.8 8.2 Table 3: Single sensor accuracy for 108 shots fired from 50 and 100 meters. Localization rate refers to the percentage of shots the given sensor alone was able to localize. The bearing and range values are average errors. They characterize the accuracy of localization from the given sensor"s perspective. The data indicates that the performance of the sensors varied significantly especially considering the localization rate. One factor has to be the location of the given sensor including how far it was from the firing lines and how obstructed its view was. Also, the sensors were hand-built prototypes utilizing nowhere near production quality packaging/mounting. In light of these factors, the overall average bearing error of 0.9 degrees and range error of 5 m with a microphone spacing of less than 10 cm are excellent. We believe that professional manufacturing and better microphones could easily achieve better performance than the best sensor in our experiment (>60% localization rate and 3 m range error). Interestingly, the largest error in range was a huge 90 m clearly due to some erroneous detection, yet the largest bearing error was less than 12 degrees which is still a good indication for the soldier where to look. The overall localization rate over all single sensors was 42%, while for 50 m shots only, this jumped to 61%. Note that the firing range was prepared to simulate an urban area to some extent: there were a few single- and two-storey wooden structures built both in and around the sensor deployment area and the firing lines. Hence, not all sensors had line-of-sight to all shooting positions. We estimate that 10% of the sensors had obstructed view to the shooter on average. Hence, we can claim that a given sensor had about 50% chance of localizing a shot within 130 m. (Since the sensor deployment area was 30 m deep, 100 m shots correspond to actual distances between 100 and 130 m.) Again, we emphasize that localization needs at least three muzzle blast and three shockwave detections out of a possible four for each per sensor. The detection rate for single sensors-corresponding to at least one shockwave detection per sensor-was practically 100%. 0% 10% 20% 30% 40% 50% 60% 70% 80% 90% 100% 0 1 2 3 4 5 6 7 8 9 10 number of sensors percentageofshots Figure 13: Histogram showing what fraction of the 104 shots taken from 50 and 100 meters were localized by at most how many individual sensors alone. 13% of the shots were missed by every single sensor, i.e., none of them had both muzzle blast and shockwave AoA detections. Note that almost all of these shots were still accurately localized by the networked system, i.e. the sensor fusion using all available observations in the sensor network. It would be misleading to interpret these results as the system missing half the shots. As soldiers never work alone and the sensor node is relatively cheap to afford having every soldier equipped with one, we also need to look at the overall detection rates for every shot. Figure 13 shows the histogram of the percentage of shots vs. the number of individual sensors that localized it. 13% of shots were not localized by any sensor alone, but 87% was localized by at least one sensor out of ten. 7.2 Error sources In this section, we analyze the most significant sources of error that affect the performance of the networked shooter localization and weapon classification system. In order to correlate the distributed observations of the acoustic events, the nodes need to have a common time and space reference. Hence, errors in the time synchronization, node localization and node orientation all degrade the overall accuracy of the system. 123 Our time synchronization approach yields errors significantly less than 100 microseconds. As the sound travels about 3 cm in that time, time synchronization errors have a negligible effect on the system. On the other hand, node location and orientation can have a direct effect on the overall system performance. Notice that to analyze this, we do not have to resort to simulation, instead we can utilize the real test data gathered at Aberdeen. But instead of using the real sensor locations known very accurately and the measured and calibrated almost perfect node orientations, we can add error terms to them and run the sensor fusion. This exactly replicates how the system would have performed during the test using the imprecisely known locations and orientations. Another aspect of the system performance that can be evaluated this way is the effect of the number of available sensors. Instead of using all ten sensors in the data fusion, we can pick any subset of the nodes to see how the accuracy degrades as we decrease the number of nodes. The following experiment was carried out. The number of sensors were varied from 2 to 10 in increments of 2. Each run picked the sensors randomly using a uniform distribution. At each run each node was randomly moved to a new location within a circle around its true position with a radius determined by a zero-mean Gaussian distribution. Finally, the node orientations were perturbed using a zeromean Gaussian distribution. Each combination of parameters were generated 100 times and utilized for all 196 shots. The results are summarized in Figure 14. There is one 3D barchart for each of the experiment sets with the given fixed number of sensors. The x-axis shows the node location error, that is, the standard deviation of the corresponding Gaussian distribution that was varied between 0 and 6 meters. The y-axis shows the standard deviation of the node orientation error that was varied between 0 and 6 degrees. The z-axis is the resulting trajectory azimuth error. Note that the elevation angles showed somewhat larger errors than the azimuth. Since all the sensors were in approximately a horizontal plane and only a few shooter positions were out of the same plane and only by 2 m or so, the test was not sufficient to evaluate this aspect of the system. There are many interesting observation one can make by analyzing these charts. Node location errors in this range have a small effect on accuracy. Node orientation errors, on the other hand, noticeably degrade the performance. Still the largest errors in this experiment of 3.5 degrees for 6 sensors and 5 degrees for 2 sensors are still very good. Note that as the location and orientation errors increase and the number of sensors decrease, the most significantly affected performance metric is the localization rate. See Table 4 for a summary. Successful localization goes down from almost 100% to 50% when we go from 10 sensors to 2 even without additional errors. This is primarily caused by geometry: for a successful localization, the bullet needs to pass over the sensor network, that is, at least one sensor should be on the side of the trajectory other than the rest of the nodes. (This is a simplification for illustrative purposes. If all the sensors and the trajectory are not coplanar, localization may be successful even if the projectile passes on one side of the network. See Section 6.3.) As the numbers of sensors decreased in the experiment by randomly selecting a subset, the probability of trajectories abiding by this rule decreased. This also means that even if there are 0 2 4 6 0 2 4 6 0 1 2 3 4 5 6 azimutherror(degree) position error (m) orientation error (degree) 2 sensors 0 2 4 6 0 2 4 6 0 1 2 3 4 5 6 azimutherror(degree) position error (m) orientation error (degree) 4 sensors 0 2 4 6 0 2 4 6 0 1 2 3 4 5 6 azimutherror(degree) position error (m) orientation error (degree) 6 sensors 0 2 4 6 0 2 4 6 0 1 2 3 4 5 6 azimutherror(degree) position error (m) orientation error (degree) 8 sensors Figure 14: The effect of node localization and orientation errors on azimuth accuracy with 2, 4, 6 and 8 nodes. Note that the chart for 10 nodes is almost identical for the 8-node case, hence, it is omitted. 124 many sensors (i.e. soldiers), but all of them are right next to each other, the localization rate will suffer. However, when the sensor fusion does provide a result, it is still accurate even with few available sensors and relatively large individual errors. A very few consistent observation lead to good accuracy as the inconsistent ones are discarded by the algorithm. This is also supported by the observation that for the cases with the higher number of sensors (8 or 10), the localization rate is hardly affected by even large errors. Errors/Sensors 2 4 6 8 10 0 m, 0 deg 54% 87% 94% 95% 96% 2 m, 2 deg 53% 80% 91% 96% 96% 6 m, 0 deg 43% 79% 88% 94% 94% 0 m, 6 deg 44% 78% 90% 93% 94% 6 m, 6 deg 41% 73% 85% 89% 92% Table 4: Localization rate as a function of the number of sensors used, the sensor node location and orientation errors. One of the most significant observations on Figure 14 and Table 4 is that there is hardly any difference in the data for 6, 8 and 10 sensors. This means that there is little advantage of adding more nodes beyond 6 sensors as far as the accuracy is concerned. The speed of sound depends on the ambient temperature. The current prototype considers it constant that is typically set before a test. It would be straightforward to employ a temperature sensor to update the value of the speed of sound periodically during operation. Note also that wind may adversely affect the accuracy of the system. The sensor fusion, however, could incorporate wind speed into its calculations. It would be more complicated than temperature compensation, but could be done. Other practical issues also need to be looked at before a real world deployment. Silencers reduce the muzzle blast energy and hence, the effective range the system can detect it at. However, silencers do not effect the shockwave and the system would still detect the trajectory and caliber accurately. The range and weapon type could not be estimated without muzzle blast detections. Subsonic weapons do not produce a shockwave. However, this is not of great significance, since they have shorter range, lower accuracy and much less lethality. Hence, their use is not widespread and they pose less danger in any case. Another issue is the type of ammunition used. Irregular armies may use substandard, even hand manufactured bullets. This effects the muzzle velocity of the weapon. For weapon classification to work accurately, the system would need to be calibrated with the typical ammunition used by the given adversary. 8. RELATED WORK Acoustic detection and recognition has been under research since the early fifties. The area has a close relevance to the topic of supersonic flow mechanics [20]. Fansler analyzed the complex near-field pressure waves that occur within a foot of the muzzle blast. Fansler"s work gives a good idea of the ideal muzzle blast pressure wave without contamination from echoes or propagation effects [4]. Experiments with greater distances from the muzzle were conducted by Stoughton [18]. The measurements of the ballistic shockwaves using calibrated pressure transducers at known locations, measured bullet speeds, and miss distances of 355 meters for 5.56 mm and 7.62 mm projectiles were made. Results indicate that ground interaction becomes a problem for miss distances of 30 meters or larger. Another area of research is the signal processing of gunfire acoustics. The focus is on the robust detection and length estimation of small caliber acoustic shockwaves and muzzle blasts. Possible techniques for classifying signals as either shockwaves or muzzle blasts includes short-time Fourier Transform (STFT), the Smoothed Pseudo Wigner-Ville distribution (SPWVD), and a discrete wavelet transformation (DWT). Joint time-frequency (JTF) spectrograms are used to analyze the typical separation of the shockwave and muzzle blast transients in both time and frequency. Mays concludes that the DWT is the best method for classifying signals as either shockwaves or muzzle blasts because it works well and is less expensive to compute than the SPWVD [10]. The edges of the shockwave are typically well defined and the shockwave length is directly related to the bullet characteristics. A paper by Sadler [14] compares two shockwave edge detection methods: a simple gradient-based detector, and a multi-scale wavelet detector. It also demonstrates how the length of the shockwave, as determined by the edge detectors, can be used along with Whithams equations [20] to estimate the caliber of a projectile. Note that the available computational performance on the sensor nodes, the limited wireless bandwidth and real-time requirements render these approaches infeasible on our platform. A related topic is the research and development of experimental and prototype shooter location systems. Researchers at BBN have developed the Bullet Ears system [3] which has the capability to be installed in a fixed position or worn by soldiers. The fixed system has tetrahedron shaped microphone arrays with 1.5 meter spacing. The overall system consists of two to three of these arrays spaced 20 to 100 meters from each other. The soldier-worn system has 12 microphones as well as a GPS antenna and orientation sensors mounted on a helmet. There is a low speed RF connection from the helmet to the processing body. An extensive test has been conducted to measure the performance of both type of systems. The fixed systems performance was one order of magnitude better in the angle calculations while their range performance where matched. The angle accuracy of the fixed system was dominantly less than one degree while it was around five degrees for the helmet mounted one. The range accuracy was around 5 percent for both of the systems. The problem with this and similar centralized systems is the need of the one or handful of microphone arrays to be in line-of-sight of the shooter. A sensor networked based solution has the advantage of widely distributed sensing for better coverage, multipath effect compensation and multiple simultaneous shot resolution [8]. This is especially important for operation in acoustically reverberant urban areas. Note that BBN"s current vehicle-mounted system called BOOMERANG, a modified version of Bullet Ears, is currently used in Iraq [1]. The company ShotSpotter specializes in law enforcement systems that report the location of gunfire to police within seconds. The goal of the system is significantly different than that of military systems. Shotspotter reports 25 m typical accuracy which is more than enough for police to 125 respond. They are also manufacturing experimental soldier wearable and UAV mounted systems for military use [16], but no specifications or evaluation results are publicly available. 9. CONCLUSIONS The main contribution of this work is twofold. First, the performance of the overall distributed networked system is excellent. Most noteworthy are the trajectory accuracy of one degree, the correct caliber estimation rate of well over 90% and the close to 100% weapon classification rate for 4 of the 6 weapons tested. The system proved to be very robust when increasing the node location and orientation errors and decreasing the number of available sensors all the way down to a couple. The key factor behind this is the sensor fusion algorithm"s ability to reject erroneous measurements. It is also worth mentioning that the results presented here correspond to the first and only test of the system beyond 100 m and with six different weapons. We believe that with the lessons learned in the test, a consecutive field experiment could have showed significantly improved results especially in range estimation beyond 100 m and weapon classification for the remaining two weapons that were mistaken for each other the majority of the times during the test. Second, the performance of the system when used in standalone mode, that is, when single sensors alone provided localization, was also very good. While the overall localization rate of 42% per sensor for shots up to 130 m could be improved, the bearing accuracy of less than a degree and the average 5% range error are remarkable using the handmade prototypes of the low-cost nodes. Note that 87% of the shots were successfully localized by at least one of the ten sensors utilized in standalone mode. We believe that the technology is mature enough that a next revision of the system could be a commercial one. However, important aspects of the system would still need to be worked on. We have not addresses power management yet. A current node runs on 4 AA batteries for about 12 hours of continuous operation. A deployable version of the sensor node would need to be asleep during normal operation and only wake up when an interesting event occurs. An analog trigger circuit could solve this problem, however, the system would miss the first shot. Instead, the acoustic channels would need to be sampled and stored in a circular buffer. The rest of the board could be turned off. When a trigger wakes up the board, the acoustic data would be immediately available. Experiments with a previous generation sensor board indicated that this could provide a 10x increase in battery life. Other outstanding issues include weatherproof packaging and ruggedization, as well as integration with current military infrastructure. 10. REFERENCES [1] BBN technologies website. http://www.bbn.com. [2] E. Danicki. Acoustic sniper localization. Archives of Acoustics, 30(2):233-245, 2005. [3] G. L. Duckworth et al. Fixed and wearable acoustic counter-sniper systems for law enforcement. In E. M. Carapezza and D. B. Law, editors, Proc. SPIE Vol. 3577, p. 210-230, pages 210-230, Jan. 1999. [4] K. Fansler. Description of muzzle blast by modified scaling models. Shock and Vibration, 5(1):1-12, 1998. [5] D. Gay, P. Levis, R. von Behren, M. Welsh, E. Brewer, and D. Culler. The nesC language: a holistic approach to networked embedded systems. Proceedings of Programming Language Design and Implementation (PLDI), June 2003. [6] J. Hill, R. Szewczyk, A. Woo, S. Hollar, D. Culler, and K. Pister. System architecture directions for networked sensors. in Proc. of ASPLOS 2000, Nov. 2000. [7] B. Kus´y, G. Balogh, P. V¨olgyesi, J. Sallai, A. N´adas, A. L´edeczi, M. Mar´oti, and L. Meertens. Node-density independent localization. Information Processing in Sensor Networks (IPSN 06) SPOTS Track, Apr. 2006. [8] A. L´edeczi, A. N´adas, P. V¨olgyesi, G. Balogh, B. Kus´y, J. Sallai, G. Pap, S. D´ora, K. Moln´ar, M. Mar´oti, and G. Simon. Countersniper system for urban warfare. ACM Transactions on Sensor Networks, 1(1):153-177, Nov. 2005. [9] M. Mar´oti. Directed flood-routing framework for wireless sensor networks. In Proceedings of the 5th ACM/IFIP/USENIX International Conference on Middleware, pages 99-114, New York, NY, USA, 2004. Springer-Verlag New York, Inc. [10] B. Mays. Shockwave and muzzle blast classification via joint time frequency and wavelet analysis. Technical report, Army Research Lab Adelphi MD 20783-1197, Sept. 2001. [11] TinyOS Hardware Platforms. http://tinyos.net/scoop/special/hardware. [12] Crossbow MICAz (MPR2400) Radio Module. http://www.xbow.com/Products/productsdetails. aspx?sid=101. [13] PicoBlaze User Resources. http://www.xilinx.com/ipcenter/processor_ central/picoblaze/picoblaze_user_resources.htm. [14] B. M. Sadler, T. Pham, and L. C. Sadler. Optimal and wavelet-based shock wave detection and estimation. Acoustical Society of America Journal, 104:955-963, Aug. 1998. [15] J. Sallai, B. Kus´y, A. L´edeczi, and P. Dutta. On the scalability of routing-integrated time synchronization. 3rd European Workshop on Wireless Sensor Networks (EWSN 2006), Feb. 2006. [16] ShotSpotter website. http: //www.shotspotter.com/products/military.html. [17] G. Simon, M. Mar´oti, A. L´edeczi, G. Balogh, B. Kus´y, A. N´adas, G. Pap, J. Sallai, and K. Frampton. Sensor network-based countersniper system. In SenSys "04: Proceedings of the 2nd international conference on Embedded networked sensor systems, pages 1-12, New York, NY, USA, 2004. ACM Press. [18] R. Stoughton. Measurements of small-caliber ballistic shock waves in air. Acoustical Society of America Journal, 102:781-787, Aug. 1997. [19] B. A. Weiss, C. Schlenoff, M. Shneier, and A. Virts. Technology evaluations and performance metrics for soldier-worn sensors for assist. In Performance Metrics for Intelligent Systems Workshop, Aug. 2006. [20] G. Whitham. Flow pattern of a supersonic projectile. Communications on pure and applied mathematics, 5(3):301, 1952. 126

1degree trajectory precision;weapon type;weapon classification;range;sensorboard;trajectory;internode communication;datum fusion;acoustic source localization;sensor network;caliber estimation accuracy;helmetmounted microphone array;caliber;caliber estimation;wireless sensor network-based mobile countersniper system;self orientation

train_C-66

Heuristics-Based Scheduling of Composite Web Service Workloads

Web services can be aggregated to create composite workflows that provide streamlined functionality for human users or other systems. Although industry standards and recent research have sought to define best practices and to improve end-to-end workflow composition, one area that has not fully been explored is the scheduling of a workflow"s web service requests to actual service provisioning in a multi-tiered, multi-organisation environment. This issue is relevant to modern business scenarios where business processes within a workflow must complete within QoS-defined limits. Because these business processes are web service consumers, service requests must be mapped and scheduled across multiple web service providers, each with its own negotiated service level agreement. In this paper we provide heuristics for scheduling service requests from multiple business process workflows to web service providers such that a business value metric across all workflows is maximised. We show that a genetic search algorithm is appropriate to perform this scheduling, and through experimentation we show that our algorithm scales well up to a thousand workflows and produces better mappings than traditional approaches.

1. INTRODUCTION Web services can be composed into workflows to provide streamlined end-to-end functionality for human users or other systems. Although previous research efforts have looked at ways to intelligently automate the composition of web services into workflows (e.g. [1, 9]), an important remaining problem is the assignment of web service requests to the underlying web service providers in a multi-tiered runtime scenario within constraints. In this paper we address this scheduling problem and examine means to manage a large number of business process workflows in a scalable manner. The problem of scheduling web service requests to providers is relevant to modern business domains that depend on multi-tiered service provisioning. Consider the example shown in Figure 1 that illustrates our problem space. Workflows comprise multiple related business processes that are web service consumers; here we assume that the workflows represent requested service from customers or automated systems and that the workflow has already been composed with an existing choreography toolkit. These workflows are then submitted to a portal (not shown) that acts as a scheduling agent between the web service consumers and the web service providers. In this example, a workflow could represent the actions needed to instantiate a vacation itinerary, where one business process requests booking an airline ticket, another business process requests a hotel room, and so forth. Each of these requests target a particular service type (e.g. airline reservations, hotel reservations, car reservations, etc.), and for each service type, there are multiple instances of service providers that publish a web service interface. An important challenge is that the workflows must meet some quality-of-service (QoS) metric, such as end-to-end completion time of all its business processes, and that meeting or failing this goal results in the assignment of a quantitative business value metric for the workflow; intuitively, it is desired that all workflows meet their respective QoS goals. We further leverage the notion that QoS service agreements are generally agreed-upon between the web service providers and the scheduling agent such that the providers advertise some level of guaranteed QoS to the scheduler based upon runtime conditions such as turnaround time and maximum available concurrency. The resulting problem is then to schedule and assign the business processes" requests for service types to one of the service providers for that type. The scheduling must be done such that the aggregate business value across all the workflows is maximised. In Section 3 we state the scenario as a combinatorial problem and utilise a genetic search algorithm [5] to find the best assignment of web service requests to providers. This approach converges towards an assignment that maximises the overall business value for all the workflows. In Section 4 we show through experimentation that this search heuristic finds better assignments than other algorithms (greedy, round-robin, and proportional). Further, this approach allows us to scale the number of simultaneous workflows (up to one thousand workflows in our experiments) and yet still find effective schedules. 2. RELATED WORK In the context of service assignment and scheduling, [11] maps web service calls to potential servers using linear programming, but their work is concerned with mapping only single workflows; our principal focus is on scalably scheduling multiple workflows (up 30 Service Type SuperHotels.com Business Process Business Process Workflow ... Business Process Business Process ... HostileHostels.com IncredibleInns.com Business Process Business Process Business Process ... Business Process Service Provider SkyHighAirlines.com SuperCrazyFlights.com Business Process . . . . . . Advertised QoS Service Agreement CarRentalService.com Figure 1: An example scenario demonstrating the interaction between business processes in workflows and web service providers. Each business process accesses a service type and is then mapped to a service provider for that type. to one thousand as we show later) using different business metrics and a search heuristic. [10] presents a dynamic provisioning approach that uses both predictive and reactive techniques for multi-tiered Internet application delivery. However, the provisioning techniques do not consider the challenges faced when there are alternative query execution plans and replicated data sources. [8] presents a feedback-based scheduling mechanism for multi-tiered systems with back-end databases, but unlike our work, it assumes a tighter coupling between the various components of the system. Our work also builds upon prior scheduling research. The classic job-shop scheduling problem, shown to be NP-complete [4] [3], is similar to ours in that tasks within a job must be scheduled onto machinery (c.f. our scenario is that business processes within a workflow must be scheduled onto web service providers). The salient differences are that the machines can process only one job at a time (we assume servers can multi-task but with degraded performance and a maximum concurrency level), tasks within a job cannot simultaneously run on different machines (we assume business processes can be assigned to any available server), and the principal metric of performance is the makespan, which is the time for the last task among all the jobs to complete (and as we show later, optimising on the makespan is insufficient for scheduling the business processes, necessitating different metrics). 3. DESIGN In this section we describe our model and discuss how we can find scheduling assignments using a genetic search algorithm. 3.1 Model We base our model on the simplified scenario shown in Figure 1. Specifically, we assume that users or automated systems request the execution of a workflow. The workflows comprise business processes, each of which makes one web service invocation to a service type. Further, business processes have an ordering in the workflow. The arrangement and execution of the business processes and the data flow between them are all managed by a composition or choreography tool (e.g. [1, 9]). Although composition languages can use sophisticated flow-control mechanisms such as conditional branches, for simplicity we assume the processes execute sequentially in a given order. This scenario can be naturally extended to more complex relationships that can be expressed in BPEL [7], which defines how business processes interact, messages are exchanged, activities are ordered, and exceptions are handled. Due to space constraints, we focus on the problem space presented here and will extend our model to more advanced deployment scenarios in the future. Each workflow has a QoS requirement to complete within a specified number of time units (e.g. on the order of seconds, as detailed in the Experiments section). Upon completion (or failure), the workflow is assigned a business value. We extended this approach further and considered different types of workflow completion in order to model differentiated QoS levels that can be applied by businesses (for example, to provide tiered customer service). We say that a workflow is successful if it completes within its QoS requirement, acceptable if it completes within a constant factor κ 31 of its QoS bound (in our experiments we chose κ=3), or failing if it finishes beyond κ times its QoS bound. For each category, a business value score is assigned to the workflow, with the successful category assigned the highest positive score, followed by acceptable and then failing. The business value point distribution is non-uniform across workflows, further modelling cases where some workflows are of higher priority than others. Each service type is implemented by a number of different service providers. We assume that the providers make service level agreements (SLAs) to guarantee a level of performance defined by the completion time for completing a web service invocation. Although SLAs can be complex, in this paper we assume for simplicity that the guarantees can take the form of a linear performance degradation under load. This guarantee is defined by several parameters: α is the expected completion time (for example, on the order of seconds) if the assigned workload of web service requests is less than or equal to β, the maximum concurrency, and if the workload is higher than β, the expected completion for a workload of size ω is α+ γ(ω − β) where γ is a fractional coefficient. In our experiments we vary α, β, and γ with different distributions. Ideally, all workflows would be able to finish within their QoS limits and thus maximise the aggregate business value across all workflows. However, because we model service providers with degrading performance under load, not all workflows will achieve their QoS limit: it may easily be the case that business processes are assigned to providers who are overloaded and cannot complete within the respective workflow"s QoS limit. The key research problem, then, is to assign the business processes to the web service providers with the goal of optimising on the aggregate business value of all workflows. Given that the scope of the optimisation is the entire set of workflows, it may be that the best scheduling assignments may result in some workflows having to fail in order for more workflows to succeed. This intuitive observation suggests that traditional scheduling approaches such as round-robin or proportional assignments will not fare well, which is what we observe and discuss in Section 4. On the other hand, an exhaustive search of all the possible assignments will find the best schedule, but the computational complexity is prohibitively high. Suppose there are W workflows with an average of B business processes per workflow. Further, in the worst case each business process requests one service type, for which there are P providers. There are thus W · PB combinations to explore to find the optimal assignments of business processes to providers. Even for small configurations (e.g. W =10, B=5, P=10), the computational time for exhaustive search is significant, and in our work we look to scale these parameters. In the next subsection, discuss how a genetic search algorithm can be used to converge toward the optimum scheduling assignments. 3.2 Genetic algorithm Given an exponential search space of business process assignments to web service providers, the problem is to find the optimal assignment that produces the overall highest aggregate business value across all workflows. To explore the solution space, we use a genetic algorithm (GA) search heuristic that simulates Darwinian natural selection by having members of a population compete to survive in order to pass their genetic chromosomes onto the next generation; after successive generations, there is a tendency for the chromosomes to converge toward the best combination [5] [6]. Although other search heuristics exist that can solve optimization problems (e.g. simulated annealing or steepest-ascent hillclimbing), the business process scheduling problem fits well with a GA because potential solutions can be represented in a matrix form and allows us to use prior research in effective GA chromosome recombination to form new members of the population (e.g. [2]). 0 1 2 3 4 0 1 2 0 2 1 1 0 1 0 1 0 2 1 2 0 0 1 Figure 2: An example chromosome representing a scheduling assignment of (workflow,service type) → service provider. Each row represents a workflow, and each column represents a service type. For example, here there are 3 workflows (0 to 2) and 5 service types (0 to 4). In workflow 0, any request for service type 3 goes to provider 2. Note that the service provider identifier is within a range limited to its service type (i.e. its column), so the 2 listed for service type 3 is a different server from server 2 in other columns. Chromosome representation of a solution. In Figure 2 we show an example chromosome that encodes one scheduling assignment. The representation is a 2-dimensional matrix that maps {workflow, service type} to a service provider. For a business process in workflow i and utilising service type j, the (i, j)th entry in the table is the identifier for the service provider to which the business process is assigned. Note that the service provider identifier is within a range limited to its service type. GA execution. A GA proceeds as follows. Initially a random set of chromosomes is created for the population. The chromosomes are evaluated (hashed) to some metric, and the best ones are chosen to be parents. In our problem, the evaluation produces the net business value across all workflows after executing all business processes once they are assigned to their respective service providers according to the mapping in the chromosome. The parents recombine to produce children, simulating sexual crossover, and occasionally a mutation may arise which produces new characteristics that were not available in either parent. The principal idea is that we would like the children to be different from the parents (in order to explore more of the solution space) yet not too different (in order to contain the portions of the chromosome that result in good scheduling assignments). Note that finding the global optimum is not guaranteed because the recombination and mutation are stochastic. GA recombination and mutation. As mentioned, the chromosomes are 2-dimensional matrices that represent scheduling assignments. To simulate sexual recombination of two chromosomes to produce a new child chromosome, we applied a one-point crossover scheme twice (once along each dimension). The crossover is best explained by analogy to Cartesian space as follows. A random point is chosen in the matrix to be coordinate (0, 0). Matrix elements from quadrants II and IV from the first parent and elements from quadrants I and III from the second parent are used to create the new child. This approach follows GA best practices by keeping contiguous chromosome segments together as they are transmitted from parent to child. The uni-chromosome mutation scheme randomly changes one of the service provider assignments to another provider within the available range. Other recombination and mutation schemes are an area of research in the GA community, and we look to explore new operators in future work. GA evaluation function. An important GA component is the evaluation function. Given a particular chromosome representing one scheduling mapping, the function deterministically calculates the net business value across all workloads. The business processes in each workload are assigned to service providers, and each provider"s completion time is calculated based on the service agreement guarantee using the parameters mentioned in Section 3.1, namely the unloaded completion time α, the maximum concur32 rency β, and a coefficient γ that controls the linear performance degradation under heavy load. Note that the evaluation function can be easily replaced if desired; for example, other evaluation functions can model different service provider guarantees or parallel workflows. 4. EXPERIMENTS AND RESULTS In this section we show the benefit of using our GA-based scheduler. Because we wanted to scale the scenarios up to a large number of workflows (up to 1000 in our experiments), we implemented a simulation program that allowed us to vary parameters and to measure the results with different metrics. The simulator was written in standard C++ and was run on a Linux (Fedora Core) desktop computer running at 2.8 GHz with 1GB of RAM. We compared our algorithm against alternative candidates: • A well-known round-robin algorithm that assigns each business process in circular fashion to the service providers for a particular service type. This approach provides the simplest scheme for load-balancing. • A random-proportional algorithm that proportionally assigns business processes to the service providers; that is, for a given service type, the service providers are ranked by their guaranteed completion time, and business processes are assigned proportionally to the providers based on their completion time. (We also tried a proportionality scheme based on both the completion times and maximum concurrency but attained the same results, so only the former scheme"s results are shown here.) • A strawman greedy algorithm that always assigns business processes to the service provider that has the fastest guaranteed completion time. This algorithm represents a naive approach based on greedy, local observations of each workflow without taking into consideration all workflows. In the experiments that follow, all results were averaged across 20 trials, and to help normalise the effects of randomisation used during the GA, each trial started by reading in pre-initialised data from disk. In Table 1 we list our experimental parameters. In Figure 3 we show the results of running our GA against the three candidate alternatives. The x-axis shows the number for workflows scaled up to 1000, and the y-axis shows the aggregate business value for all workflows. As can be seen, the GA consistently produces the highest business value even as the number of workflows grows; at 1000 workflows, the GA produces a 115% improvement over the next-best alternative. (Note that although we are optimising against the business value metric we defined earlier, genetic algorithms are able to converge towards the optimal value of any metric, as long as the evaluation function can consistently measure a chromosome"s value with that metric.) As expected, the greedy algorithm performs very poorly because it does the worst job at balancing load: all business processes for a given service type are assigned to only one server (the one advertised to have the fastest completion time), and as more business processes arrive, the provider"s performance degrades linearly. The round-robin scheme is initially outperformed by the randomproportional scheme up to around 120 workflows (as shown in the magnified graph of Figure 4), but as the number of workflows increases, the round-robin scheme consistently wins over randomproportional. The reason is that although the random-proportional scheme assigns business processes to providers proportionally according to the advertised completion times (which is a measure of the power of the service provider), even the best providers will eventually reach a real-world maximum concurrency for the large -2000 -1000 0 1000 2000 3000 4000 5000 6000 7000 0 200 400 600 800 1000 Aggregatebusinessvalueacrossallworkflows Total number of workflows Business value scores of scheduling algorithms Genetic algorithm Round robin Random proportional Greedy Figure 3: Net business value scores of different scheduling algorithms. -500 0 500 1000 1500 2000 2500 3000 3500 4000 0 50 100 150 200Aggregatebusinessvalueacrossallworkflows Total number of workflows Business value scores of scheduling algorithms Genetic algorithm Round robin Random proportional Greedy Figure 4: Magnification of the left-most region in Figure 3. number of workflows that we are considering. For a very large number of workflows, the round-robin scheme is able to better balance the load across all service providers. To better understand the behaviour resulting from the scheduling assignments, we show the workflow completion results in Figures 5, 6, and 7 for 100, 500, and 900 workflows, respectively. These figures show the percentage of workflows that are successful (can complete with their QoS limit), acceptable (can complete within κ=3 times their QoS limit), and failed (cannot complete within κ=3 times their QoS limit). The GA consistently produces the highest percentage of successful workflows (resulting in higher business values for the aggregate set of workflows). Further, the round-robin scheme produces better results than the random-proportional for a large number of workflows but does not perform as well as the GA. In Figure 8 we graph the makespan resulting from the same experiments above. Makespan is a traditional metric from the job scheduling community measuring elapsed time for the last job to complete. While useful, it does not capture the high-level business value metric that we are optimising against. Indeed, the makespan is oblivious to the fact that we provide multiple levels of completion (successful, acceptable, and failed) and assign business value scores accordingly. For completeness, we note that the GA provides the fastest makespan, but it is matched by the round robin algorithm. The GA produces better business values (as shown in Figure 3) because it is able to search the solution space to find better mappings that produce more successful workflows (as shown in Figures 5 to 7). We also looked at the effect of the scheduling algorithms on balancing the load. Figure 9 shows the percentage of services providers that were accessed while the workflows ran. As expected, the greedy algorithm always hits one service provider; on the other hand, the round-robin algorithm is the fastest to spread the business 33 Experimental parameter Comment Workflows 5 to 1000 Business processes per workflow uniform random: 1 - 10 Service types 10 Service providers per service type uniform random: 1 - 10 Workflow QoS goal uniform random: 10-30 seconds Service provider completion time (α) uniform random: 1 - 12 seconds Service provider maximum concurrency (β) uniform random: 1 - 12 Service provider degradation coefficient (γ) uniform random: 0.1 - 0.9 Business value for successful workflows uniform random: 10 - 50 points Business value for acceptable workflows uniform random: 0 - 10 points Business value for failed workflows uniform random: -10 - 0 points GA: number of parents 20 GA: number of children 80 GA: number of generations 1000 Table 1: Experimental parameters Failed Acceptable (completed but not within QoS) Successful (completed within QoS) 0% 20% 40% 60% 80% 100% RoundRobinRandProportionalGreedyGeneticAlg Percentageofallworkflows Workflow behaviour, 100 workflows Figure 5: Workflow behaviour for 100 workflows. Failed Acceptable (completed but not within QoS) Successful (completed within QoS) 0% 20% 40% 60% 80% 100% RoundRobinRandProportionalGreedyGeneticAlg Percentageofallworkflows Workflow behaviour, 500 workflows Figure 6: Workflow behaviour for 500 workflows. Failed Acceptable (completed but not within QoS) Successful (completed within QoS) 0% 20% 40% 60% 80% 100% RoundRobinRandProportionalGreedyGeneticAlg Percentageofallworkflows Workflow behaviour, 500 workflows Figure 7: Workflow behaviour for 900 workflows. 0 50 100 150 200 250 300 0 200 400 600 800 1000 Makespan[seconds] Number of workflows Maximum completion time for all workflows Genetic algorithm Round robin Random proportional Greedy Figure 8: Maximum completion time for all workflows. This value is the makespan metric used in traditional scheduling research. Although useful, the makespan does not take into consideration the business value scoring in our problem domain. processes. Figure 10 is the percentage of accessed service providers (that is, the percentage of service providers represented in Figure 9) that had more assigned business processes than their advertised maximum concurrency. For example, in the greedy algorithm only one service provider is utilised, and this one provider quickly becomes saturated. On the other hand, the random-proportional algorithm uses many service providers, but because business processes are proportionally assigned with more assignments going to the better providers, there is a tendency for a smaller percentage of providers to become saturated. For completeness, we show the performance of the genetic algorithm itself in Figure 11. The algorithm scales linearly with an increasing number of workflows. We note that the round-robin, random-proportional, and greedy algorithms all finished within 1 second even for the largest workflow configuration. However, we feel that the benefit of finding much higher business value scores justifies the running time of the GA; further we would expect that the running time will improve with both software tuning as well as with a computer faster than our off-the-shelf PC. 5. CONCLUSION Business processes within workflows can be orchestrated to access web services. In this paper we looked at multi-tiered service provisioning where web service requests to service types can be mapped to different service providers. The resulting problem is that in order to support a very large number of workflows, the assignment of business process to web service provider must be intelligent. We used a business value metric to measure the be34 0 0.2 0.4 0.6 0.8 1 0 200 400 600 800 1000 Percentageofallserviceproviders Number of workflows Service providers utilised Genetic algorithm Round robin Random proportional Greedy Figure 9: The percentage of service providers utilized during workload executions. The Greedy algorithm always hits the one service provider, while the Round Robin algorithm spreads requests evenly across the providers. 0 0.2 0.4 0.6 0.8 1 0 200 400 600 800 1000 Percentageofallserviceproviders Number of workflows Service providers saturated Genetic algorithm Round robin Random proportional Greedy Figure 10: The percentage of service providers that are saturated among those providers who were utilized (that is, percentage of the service providers represented in Figure 9). A saturated service provider is one whose workload is greater that its advertised maximum concurrency. 0 5 10 15 20 25 0 200 400 600 800 1000 Runningtimeinseconds Total number of workflows Running time of genetic algorithm GA running time Figure 11: Running time of the genetic algorithm. haviour of workflows meeting or failing QoS values, and we optimised our scheduling to maximise the aggregate business value across all workflows. Since the solution space of scheduler mappings is exponential, we used a genetic search algorithm to search the space and converge toward the best schedule. With a default configuration for all parameters and using our business value scoring, the GA produced up to 115% business value improvement over the next best algorithm. Finally, because a genetic algorithm will converge towards the optimal value using any metric (even other than the business value metric we used), we believe our approach has strong potential for continuing work. In future work, we look to acquire real-world traces of web service instances in order to get better estimates of service agreement guarantees, although we expect that such guarantees between the providers and their consumers are not generally available to the public. We will also look at other QoS metrics such as CPU and I/O usage. For example, we can analyse transfer costs with varying bandwidth, latency, data size, and data distribution. Further, we hope to improve our genetic algorithm and compare it to more scheduler alternatives. Finally, since our work is complementary to existing work in web services choreography (because we rely on pre-configured workflows), we look to integrate our approach with available web service workflow systems expressed in BPEL. 6. REFERENCES [1] A. Ankolekar, et al. DAML-S: Semantic Markup For Web Services, In Proc. of the Int"l Semantic Web Working Symposium, 2001. [2] L. Davis. Job Shop Scheduling with Genetic Algorithms, In Proc. of the Int"l Conference on Genetic Algorithms, 1985. [3] H.-L. Fang, P. Ross, and D. Corne. A Promising Genetic Algorithm Approach to Job-Shop Scheduling, Rescheduling, and Open-Shop Scheduling Problems , In Proc. on the 5th Int"l Conference on Genetic Algorithms, 1993. [4] M. Gary and D. Johnson. Computers and Intractability: A Guide to the Theory of NP-Completeness, Freeman, 1979. [5] J. Holland. Adaptation in Natural and Artificial Systems: An Introductory Analysis with Applications to Biology, Control, and Artificial Intelligence, MIT Press, 1992. [6] D. Goldberg. Genetic Algorithms in Search, Optimization and Machine Learning, Kluwer Academic Publishers, 1989. [7] Business Processes in a Web Services World, www-128.ibm.com/developerworks/ webservices/library/ws-bpelwp/. [8] G. Soundararajan, K. Manassiev, J. Chen, A. Goel, and C. Amza. Back-end Databases in Shared Dynamic Content Server Clusters, In Proc. of the IEEE Int"l Conference on Autonomic Computing, 2005. [9] B. Srivastava and J. Koehler. Web Service Composition Current Solutions and Open Problems, ICAP, 2003. [10] B. Urgaonkar, P. Shenoy, A. Chandra, and P. Goyal. Dynamic Provisioning of Multi-Tier Internet Applications, In Proc. of the IEEE Int"l Conference on Autonomic Computing, 2005. [11] L. Zeng, B. Benatallah, M. Dumas, J. Kalagnanam, and Q. Sheng. Quality Driven Web Services Composition, In Proc. of the WWW Conference, 2003. 35

qo;service request;scheduling service;qos-defined limit;multi-tiered system;end-to-end workflow composition;workflows;multi-organisation environment;business process workflow;web service;streamlined functionality;business value metric;scheduling agent;heuristic;schedule

train_C-67

A Holistic Approach to High-Performance Computing: Xgrid Experience

The Ringling School of Art and Design is a fully accredited fouryear college of visual arts and design. With a student to computer ratio of better than 2-to-1, the Ringling School has achieved national recognition for its large-scale integration of technology into collegiate visual art and design education. We have found that Mac OS X is the best operating system to train future artists and designers. Moreover, we can now buy Macs to run high-end graphics, nonlinear video editing, animation, multimedia, web production, and digital video applications rather than expensive UNIX workstations. As visual artists cross from paint on canvas to creating in the digital realm, the demand for a highperformance computing environment grows. In our public computer laboratories, students use the computers most often during the workday; at night and on weekends the computers see only light use. In order to harness the lost processing time for tasks such as video rendering, we are testing Xgrid, a suite of Mac OS X applications recently developed by Apple for parallel and distributed high-performance computing. As with any new technology deployment, IT managers need to consider a number of factors as they assess, plan, and implement Xgrid. Therefore, we would like to share valuable information we learned from our implementation of an Xgrid environment with our colleagues. In our report, we will address issues such as assessing the needs for grid computing, potential applications, management tools, security, authentication, integration into existing infrastructure, application support, user training, and user support. Furthermore, we will discuss the issues that arose and the lessons learned during and after the implementation process.

1. INTRODUCTION Grid computing does not have a single, universally accepted definition. The technology behind grid computing model is not new. Its roots lie in early distributed computing models that date back to early 1980s, where scientists harnessed the computing power of idle workstations to let compute intensive applications to run on multiple workstations, which dramatically shortening processing times. Although numerous distributed computing models were available for discipline-specific scientific applications, only recently have the tools became available to use general-purpose applications on a grid. Consequently, the grid computing model is gaining popularity and has become a show piece of "utility computing". Since in the IT industry, various computing models are used interchangeably with grid computing, we first sort out the similarities and difference between these computing models so that grid computing can be placed in perspective. 1.1 Clustering A cluster is a group of machines in a fixed configuration united to operate and be managed as a single entity to increase robustness and performance. The cluster appears as a single high-speed system or a single highly available system. In this model, resources can not enter and leave the group as necessary. There are at least two types of clusters: parallel clusters and highavailability clusters. Clustered machines are generally in spatial proximity, such as in the same server room, and dedicated solely to their task. In a high-availability cluster, each machine provides the same service. If one machine fails, another seamlessly takes over its workload. For example, each computer could be a web server for a web site. Should one web server "die," another provides the service, so that the web site rarely, if ever, goes down. A parallel cluster is a type of supercomputer. Problems are split into many parts, and individual cluster members are given part of the problem to solve. An example of a parallel cluster is composed of Apple Power Mac G5 computers at Virginia Tech University [1]. 1.2 Distributed Computing Distributed computing spatially expands network services so that the components providing the services are separated. The major objective of this computing model is to consolidate processing power over a network. A simple example is spreading services such as file and print serving, web serving, and data storage across multiple machines rather than a single machine handling all the tasks. Distributed computing can also be more fine-grained, where even a single application is broken into parts and each part located on different machines: a word processor on one server, a spell checker on a second server, etc. 1.3 Utility Computing Literally, utility computing resembles common utilities such as telephone or electric service. A service provider makes computing resources and infrastructure management available to a customer as needed, and charges for usage rather than a flat rate. The important thing to note is that resources are only used as needed, and not dedicated to a single customer. 1.4 Grid Computing Grid computing contains aspects of clusters, distributed computing, and utility computing. In the most basic sense, grid turns a group of heterogeneous systems into a centrally managed but flexible computing environment that can work on tasks too time intensive for the individual systems. The grid members are not necessarily in proximity, but must merely be accessible over a network; the grid can access computers on a LAN, WAN, or anywhere in the world via the Internet. In addition, the computers comprising the grid need not be dedicated to the grid; rather, they can function as normal workstations, and then advertise their availability to the grid when not in use. The last characteristic is the most fundamental to the grid described in this paper. A well-known example of such an ad hoc grid is the SETI@home project [2] of the University of California at Berkeley, which allows any person in the world with a computer and an Internet connection to donate unused processor time for analyzing radio telescope data. 1.5 Comparing the Grid and Cluster A computer grid expands the capabilities of the cluster by loosing its spatial bounds, so that any computer accessible through the network gains the potential to augment the grid. A fundamental grid feature is that it scales well. The processing power of any machine added to the grid is immediately availably for solving problems. In addition, the machines on the grid can be generalpurpose workstations, which keep down the cost of expanding the grid. 2. ASSESSING THE NEED FOR GRID COMPUTING Effective use of a grid requires a computation that can be divided into independent (i.e., parallel) tasks. The results of each task cannot depend on the results of any other task, and so the members of the grid can solve the tasks in parallel. Once the tasks have been completed, the results can be assembled into the solution. Examples of parallelizable computations are the Mandelbrot set of fractals, the Monte Carlo calculations used in disciplines such as Solid State Physics, and the individual frames of a rendered animation. This paper is concerned with the last example. 2.1 Applications Appropriate for Grid Computing The applications used in grid computing must either be specifically designed for grid use, or scriptable in such a way that they can receive data from the grid, process the data, and then return results. In other words, the best candidates for grid computing are applications that run the same or very similar computations on a large number of pieces of data without any dependencies on the previous calculated results. Applications heavily dependent on data handling rather than processing power are generally more suitable to run on a traditional environment than on a grid platform. Of course, the applications must also run on the computing platform that hosts the grid. Our interest is in using the Alias Maya application [3] with Apple"s Xgrid [4] on Mac OS X. Commercial applications usually have strict license requirements. This is an important concern if we install a commercial application such as Maya on all members of our grid. By its nature, the size of the grid may change as the number of idle computers changes. How many licenses will be required? Our resolution of this issue will be discussed in a later section. 2.2 Integration into the Existing Infrastructure The grid requires a controller that recognizes when grid members are available, and parses out job to available members. The controller must be able to see members on the network. This does not require that members be on the same subnet as the controller, but if they are not, any intervening firewalls and routers must be configured to allow grid traffic. 3. XGRID Xgrid is Apple"s grid implementation. It was inspired by Zilla, a desktop clustering application developed by NeXT and acquired by Apple. In this report we describe the Xgrid Technology Preview 2, a free download that requires Mac OS X 10.2.8 or later and a minimum 128 MB RAM [5]. Xgrid, leverages Apple"s traditional ease of use and configuration. If the grid members are on the same subnet, by default Xgrid automatically discovers available resources through Rendezvous [6]. Tasks are submitted to the grid through a GUI interface or by the command line. A System Preference Pane controls when each computer is available to the grid. It may be best to view Xgrid as a facilitator. The Xgrid architecture handles software and data distribution, job execution, and result aggregation. However, Xgrid does not perform the actual calculations. 3.1 Xgrid Components Xgrid has three major components: the client, controller, and the agent. Each component is included in the default installation, and any computer can easily be configured to assume any role. In 120 fact, for testing purposes, a computer can simultaneously assume all roles in local mode. The more typical production use is called cluster mode. The client submits jobs to the controller through the Xgrid GUI or command line. The client defines how the job will be broken into tasks for the grid. If any files or executables must be sent as part of a job, they must reside on the client or at a location accessible to the client. When a job is complete, the client can retrieve the results from the controller. A client can only connect to a single controller at a time. The controller runs the GridServer process. It queues tasks received from clients, distributes those tasks to the agents, and handles failover if an agent cannot complete a task. In Xgrid Technology Preview 2, a controller can handle a maximum of 10,000 agent connections. Only one controller can exist per logical grid. The agents run the GridAgent process. When the GridAgent process starts, it registers with a controller; an agent can only be connected to one controller at a time. Agents receive tasks from their controller, perform the specified computations, and then send the results back to the controller. An agent can be configured to always accept tasks, or to just accept them when the computer is not otherwise busy. 3.2 Security and Authentication By default, Xgrid requires two passwords. First, a client needs a password to access a controller. Second, the controller needs a password to access an agent. Either password requirement can be disabled. Xgrid uses two-way-random mutual authentication protocol with MD5 hashes. At this time, data encryption is only used for passwords. As mentioned earlier, an agent registers with a controller when the GridAgent process starts. There is no native method for the controller to reject agents, and so it must accept any agent that registers. This means that any agent could submit a job that consumes excessive processor and disk space on the agents. Of course, since Mac OS X is a BSD-based operating system, the controller could employ Unix methods of restricting network connections from agents. The Xgrid daemons run as the user nobody, which means the daemons can read, write, or execute any file according to world permissions. Thus, Xgrid jobs can execute many commands and write to /tmp and /Volumes. In general, this is not a major security risk, but is does require a level of trust between all members of the grid. 3.3 Using Xgrid 3.3.1 Installation Basic Xgrid installation and configuration is described both in Apple documentation [5] and online at the University of Utah web site [8]. The installation is straightforward and offers no options for customization. This means that every computer on which Xgrid is installed has the potential to be a client, controller, or agent. 3.3.2 Agent and Controller Configuration The agents and controllers can be configured through the Xgrid Preference Pane in the System Preferences or XML files in /Library/Preferences. Here the GridServer and GridAgent processes are started, passwords set, and the controller discovery method used by agents is selected. By default, agents use Rendezvous to find a controller, although the agents can also be configured to look for a specific host. The Xgrid Preference Pane also sets whether the Agents will always accept jobs, or only accept jobs when idle. In Xgrid terms, idle either means that the Xgrid screen saver has activated, or the mouse and keyboard have not been used for more than 15 minutes. Even if the agent is configured to always accept tasks, if the computer is being used these tasks will run in the background at a low priority. However, if an agent only accepts jobs when idle, any unfinished task being performed when the computer ceases being idle are immediately stopped and any intermediary results lost. Then the controller assigns the task to another available member of the grid. Advertising the controller via Rendezvous can be disabled by editing /Library/Preferences/com.apple.xgrid.controller.plist. This, however, will not prevent an agent from connecting to the controller by hostname. 3.3.3 Sending Jobs from an Xgrid Client The client sends jobs to the controller either through the Xgrid GUI or the command line. The Xgrid GUI submits jobs via small applications called plug-ins. Sample plug-ins are provided by Apple, but they are only useful as simple testing or as examples of how to create a custom plug-in. If we are to employ Xgrid for useful work, we will require a custom plug-in. James Reynolds details the creation of custom plug-ins on the University of Utah Mac OS web site [8]. Xgrid stores plug-ins in /Library/Xgrid/Plug-ins or ~/Library/Xgrid/Plug-ins, depending on whether the plug-in was installed with Xgrid or created by a user. The core plug-in parameter is the command, which includes the executable the agents will run. Another important parameter is the working directory. This directory contains necessary files that are not installed on the agents or available to them over a network. The working directory will always be copied to each agent, so it is best to keep this directory small. If the files are installed on the agents or available over a network, the working directory parameter is not needed. The command line allows the options available with the GUI plug-in, but it can be slightly more cumbersome. However, the command line probably will be the method of choice for serious work. The command arguments must be included in a script unless they are very basic. This can be a shell, perl, or python script, as long as the agent can interpret it. 3.3.4 Running the Xgrid Job When the Xgrid job is started, the command tells the controller how to break the job into tasks for the agents. Then the command is tarred and gzipped and sent to each agent; if there is a working directory, this is also tarred and gzipped and sent to the agents. 121 The agents extract these files into /tmp and run the task. Recall that since the GridAgent process runs as the user nobody, everything associated with the command must be available to nobody. Executables called by the command should be installed on the agents unless they are very simple. If the executable depends on libraries or other files, it may not function properly if transferred, even if the dependent files are referenced in the working directory. When the task is complete, the results are available to the client. In principle, the results are sent to the client, but whether this actually happens depends on the command. If the results are not sent to the client, they will be in /tmp on each agent. When available, a better solution is to direct the results to a network volume accessible to the client. 3.4 Limitations and Idiosyncrasies Since Xgrid is only in its second preview release, there are some rough edges and limitations. Apple acknowledges some limitations [7]. For example, the controller cannot determine whether an agent is trustworthy and the controller always copies the command and working directory to the agent without checking to see if these exist on the agent. Other limitations are likely just a by-product of an unfinished work. Neither the client nor controller can specify which agents will receive the tasks, which is particularly important if the agents contain a variety of processor types and speeds and the user wants to optimize the calculations. At this time, the best solution to this problem may be to divide the computers into multiple logical grids. There is also no standard way to monitor the progress of a running job on each agent. The Xgrid GUI and command line indicate which agents are working on tasks, but gives no indication of progress. Finally, at this time only Mac OS X clients can submit jobs to the grid. The framework exists to allow third parties to write plug-ins for other Unix flavors, but Apple has not created them. 4. XGRID IMPLEMENTATION Our goal is an Xgrid render farm for Alias Maya. The Ringling School has about 400 Apple Power Mac G4"s and G5"s in 13 computer labs. The computers range from 733 MHz singleprocessor G4"s and 500 MHz and 1 GHz dual-processor G4"s to 1.8 GHz dual-processor G5"s. All of these computers are lightly used in the evening and on weekends and represent an enormous processing resource for our student rendering projects. 4.1 Software Installation During our Xgrid testing, we loaded software on each computer multiple times, including the operating systems. We saved time by facilitating our installations with the remote administration daemon (radmind) software developed at the University of Michigan [9], [10]. Everything we installed for testing was first created as a radmind base load or overload. Thus, Mac OS X, Mac OS X Developer Tools, Xgrid, POV-Ray [11], and Alias Maya were stored on a radmind server and then installed on our test computers when needed. 4.2 Initial Testing We used six 1.8 GHz dual-processor Apple Power Mac G5"s for our Xgrid tests. Each computer ran Mac OS X 10.3.3 and contained 1 GB RAM. As shown in Figure 1, one computer served as both client and controller, while the other five acted as agents. Before attempting Maya rendering with Xgrid, we performed basic calculations to cement our understanding of Xgrid. Apple"s Xgrid documentation is sparse, so finding helpful web sites facilitated our learning. We first ran the Mandelbrot set plug-in provided by Apple, which allowed us to test the basic functionality of our grid. Then we performed benchmark rendering with the Open Source Application POV-Ray, as described by Daniel Côté [12] and James Reynolds [8]. Our results showed that one dual-processor G5 rendering the benchmark POV-Ray image took 104 minutes. Breaking the image into three equal parts and using Xgrid to send the parts to three agents required 47 minutes. However, two agents finished their rendering in 30 minutes, while the third agent used 47 minutes; the entire render was only as fast as the slowest agent. These results gave us two important pieces of information. First, the much longer rendering time for one of the tasks indicated that we should be careful how we split jobs into tasks for the agents. All portions of the rendering will not take equal amounts of time, even if the pixel size is the same. Second, since POV-Ray cannot take advantage of both processors in a G5, neither can an Xgrid task running POV-Ray. Alias Maya does not have this limitation. 4.3 Rendering with Alias Maya 6 We first installed Alias Maya 6 for Mac OS X on the client/controller and each agent. Maya 6 requires licenses for use as a workstation application. However, if it is just used for rendering from the command line or a script, no license is needed. We thus created a minimal installation of Maya as a radmind overload. The application was installed in a hidden directory inside /Applications. This was done so that normal users of the workstations would not find and attempt to run Maya, which would fail because these installations are not licensed for such use. In addition, Maya requires the existence of a directory ending in the path /maya. The directory must be readable and writable by the Maya user. For a user running Maya on a Mac OS X workstation, the path would usually be ~/Documents/maya. Unless otherwise specified, this directory will be the default location for Maya data and output files. If the directory does not Figure 1. Xgrid test grid. Client/ Controller Agent 1 Agent 2 Agent 3 Agent 4 Agent 5 Network Volume Jobs Data Data 122 exist, Maya will try to create it, even if the user specifies that the data and output files exist in other locations. However, Xgrid runs as the user nobody, which does not have a home directory. Maya is unable to create the needed directory, and looks instead for /Alias/maya. This directory also does not exist, and the user nobody has insufficient rights to create it. Our solution was to manually create /Alias/maya and give the user nobody read and write permissions. We also created a network volume for storage of both the rendering data and the resulting rendered frames. This avoided sending the Maya files and associated textures to each agent as part of a working directory. Such a solution worked well for us because our computers are geographically close on a LAN; if greater distance had separated the agents from the client/controller, specifying a working directory may have been a better solution. Finally, we created a custom GUI plug-in for Xgrid. The plug-in command calls a Perl script with three arguments. Two arguments specify the beginning and end frames of the render and the third argument the number of frames in each job (which we call the cluster size). The script then calculates the total number of jobs and parses them out to the agents. For example, if we begin at frame 201 and end at frame 225, with 5 frames for each job, the plug-in will create 5 jobs and send them out to the agents. Once the jobs are sent to the agents, the script executes the /usr/sbin/Render command on each agent with the parameters appropriate for the particular job. The results are sent to the network volume. With the setup described, we were able to render with Alias Maya 6 on our test grid. Rendering speed was not important at this time; our first goal was to implement the grid, and in that we succeeded. 4.3.1 Pseudo Code for Perl Script in Custom Xgrid Plug-in In this section we summarize in simplified pseudo code format the Perl script used in our Xgrig plug-in. agent_jobs{ • Read beginning frame, end frame, and cluster size of render. • Check whether the render can be divided into an integer number of jobs based on the cluster size. • If there are not an integer number of jobs, reduce the cluster size of the last job and set its last frame to the end frame of the render. • Determine the start frame and end frame for each job. • Execute the Render command. } 4.4 Lessons Learned Rendering with Maya from the Xgrid GUI was not trivial. The lack of Xgrid documentation and the requirements of Maya combined into a confusing picture, where it was difficult to decide the true cause of the problems we encountered. Trial and error was required to determine the best way to set up our grid. The first hurdle was creating the directory /Alias/maya with read and write permissions for the user nobody. The second hurdle was learning that we got the best performance by storing the rendering data on a network volume. The last major hurdle was retrieving our results from the agents. Unlike the POV-Ray rendering tests, our initial Maya results were never returned to the client; instead, Maya stored the results in /tmp on each agent. Specifying in the plug-in where to send the results would not change this behavior. We decided this was likely a Maya issue rather than an Xgrid issue, and the solution was to send the results to the network volume via the Perl script. 5. FUTURE PLANS Maya on Xgrid is not yet ready to be used by the students of Ringling School. In order to do this, we must address at least the following concerns. • Continue our rendering tests through the command line rather than the GUI plug-in. This will be essential for the following step. • Develop an appropriate interface for users to send jobs to the Xgrid controller. This will probably be an extension to the web interface of our existing render farm, where the student specifies parameters that are placed in a script that issues the Render command. • Perform timed Maya rendering tests with Xgrid. Part of this should compare the rendering times for Power Mac G4"s and G5"s. 6. CONCLUSION Grid computing continues to advance. Recently, the IT industry has witnessed the emergence of numerous types of contemporary grid applications in addition to the traditional grid framework for compute intensive applications. For instance, peer-to-peer applications such as Kazaa, are based on storage grids that do not share processing power but instead an elegant protocol to swap files between systems. Although in our campuses we discourage students from utilizing peer-to-peer applications from music sharing, the same protocol can be utilized on applications such as decision support and data mining. The National Virtual Collaboratory grid project [13] will link earthquake researchers across the U.S. with computing resources, allowing them to share extremely large data sets, research equipment, and work together as virtual teams over the Internet. There is an assortment of new grid players in the IT world expanding the grid computing model and advancing the grid technology to the next level. SAP [14] is piloting a project to grid-enable SAP ERP applications, Dell [15] has partnered with Platform Computing to consolidate computing resources and provide grid-enabled systems for compute intensive applications, Oracle has integrated support for grid computing in their 10g release [16], United Devices [17] offers hosting service for gridon-demand, and Sun Microsystems continues their research and development of Sun"s N1 Grid engine [18] which combines grid and clustering platforms. Simply, the grid computing is up and coming. The potential benefits of grid computing are colossal in higher education learning while the implementation costs are low. Today, it would be difficult to identify an application with as high a return on investment as grid computing in information technology divisions in higher education institutions. It is a mistake to overlook this technology with such a high payback. 123 7. ACKNOWLEDGMENTS The authors would like to thank Scott Hanselman of the IT team at the Ringling School of Art and Design for providing valuable input in the planning of our Xgrid testing. We would also like to thank the posters of the Xgrid Mailing List [13] for providing insight into many areas of Xgrid. 8. REFERENCES [1] Apple Academic Research, http://www.apple.com/education/science/profiles/vatech/. [2] SETI@home: Search for Extraterrestrial Intelligence at home. http://setiathome.ssl.berkeley.edu/. [3] Alias, http://www.alias.com/. [4] Apple Computer, Xgrid, http://www.apple.com/acg/xgrid/. [5] Xgrid Guide, http://www.apple.com/acg/xgrid/, 2004. [6] Apple Mac OS X Features, http://www.apple.com/macosx/features/rendezvous/. [7] Xgrid Manual Page, 2004. [8] James Reynolds, Xgrid Presentation, University of Utah, http://www.macos.utah.edu:16080/xgrid/, 2004. [9] Research Systems Unix Group, Radmind, University of Michigan, http://rsug.itd.umich.edu/software/radmind. [10]Using the Radmind Command Line Tools to Maintain Multiple Mac OS X Machines, http://rsug.itd.umich.edu/software/radmind/files/radmindtutorial-0.8.1.pdf. [11]POV-Ray, http://www.povray.org/. [12]Daniel Côté, Xgrid example: Parallel graphics rendering in POVray, http://unu.novajo.ca/simple/, 2004. [13]NEESgrid, http://www.neesgrid.org/. [14]SAP, http://www.sap.com/. [15]Platform Computing, http://platform.com/. [16]Grid, http://www.oracle.com/technologies/grid/. [17]United Devices, Inc., http://ud.com/. [18]N1 Grid Engine 6, http://www.sun.com/ software/gridware/index.html/. [19]Xgrig Users Mailing List, http://www.lists.apple.com/mailman/listinfo/xgridusers/. 124

render;multimedia;xgrid environment;xgrid;nonlinear video editing;macintosh os x;grid computing;large-scale integration of technology;web production;animation;digital video application;rendezvous;design;visual art;highperformance computing;design education;high-end graphic;mac os x;operating system;cluster

train_C-68

An Evaluation of Availability Latency in Carrier-based Vehicular ad-hoc Networks

On-demand delivery of audio and video clips in peer-to-peer vehicular ad-hoc networks is an emerging area of research. Our target environment uses data carriers, termed zebroids, where a mobile device carries a data item on behalf of a server to a client thereby minimizing its availability latency. In this study, we quantify the variation in availability latency with zebroids as a function of a rich set of parameters such as car density, storage per device, repository size, and replacement policies employed by zebroids. Using analysis and extensive simulations, we gain novel insights into the design of carrier-based systems. Significant improvements in latency can be obtained with zebroids at the cost of a minimal overhead. These improvements occur even in scenarios with lower accuracy in the predictions of the car routes. Two particularly surprising findings are: (1) a naive random replacement policy employed by the zebroids shows competitive performance, and (2) latency improvements obtained with a simplified instantiation of zebroids are found to be robust to changes in the popularity distribution of the data items.

1. INTRODUCTION Technological advances in areas of storage and wireless communications have now made it feasible to envision on-demand delivery of data items, for e.g., video and audio clips, in vehicular peer-topeer networks. In prior work, Ghandeharizadeh et al. [10] introduce the concept of vehicles equipped with a Car-to-Car-Peer-toPeer device, termed AutoMata, for in-vehicle entertainment. The notable features of an AutoMata include a mass storage device offering hundreds of gigabytes (GB) of storage, a fast processor, and several types of networking cards. Even with today"s 500 GB disk drives, a repository of diverse entertainment content may exceed the storage capacity of a single AutoMata. Such repositories constitute the focus of this study. To exchange data, we assume each AutoMata is configured with two types of networking cards: 1) a low-bandwidth networking card with a long radio-range in the order of miles that enables an AutoMata device to communicate with a nearby cellular or WiMax station, 2) a high-bandwidth networking card with a limited radio-range in the order of hundreds of feet. The high bandwidth connection supports data rates in the order of tens to hundreds of Megabits per second and represents the ad-hoc peer to peer network between the vehicles. This is labelled as the data plane and is employed to exchange data items between devices. The low-bandwidth connection serves as the control plane, enabling AutoMata devices to exchange meta-data with one or more centralized servers. This connection offers bandwidths in the order of tens to hundreds of Kilobits per second. The centralized servers, termed dispatchers, compute schedules of data delivery along the data plane using the provided meta-data. These schedules are transmitted to the participating vehicles using the control plane. The technical feasibility of such a two-tier architecture is presented in [7], with preliminary results to demonstrate the bandwidth of the control plane is sufficient for exchange of control information needed for realizing such an application. In a typical scenario, an AutoMata device presents a passenger with a list of data items1 , showing both the name of each data item and its availability latency. The latter, denoted as δ, is defined as the earliest time at which the client encounters a copy of its requested data item. A data item is available immediately when it resides in the local storage of the AutoMata device serving the request. Due to storage constraints, an AutoMata may not store the entire repository. In this case, availability latency is the time from when the user issues a request until when the AutoMata encounters another car containing the referenced data item. (The terms car and AutoMata are used interchangeably in this study.) The availability latency for an item is a function of the current location of the client, its destination and travel path, the mobility model of the AutoMata equipped vehicles, the number of replicas constructed for the different data items, and the placement of data item replicas across the vehicles. A method to improve the availability latency is to employ data carriers which transport a replica of the requested data item from a server car containing it to a client that requested it. These data carriers are termed ‘zebroids". Selection of zebroids is facilitated by the two-tiered architecture. The control plane enables centralized information gathering at a dispatcher present at a base-station.2 Some examples of control in1 Without loss of generality, the term data item might be either traditional media such as text or continuous media such as an audio or video clip. 2 There may be dispatchers deployed at a subset of the base-stations for fault-tolerance and robustness. Dispatchers between basestations may communicate via the wired infrastructure. 75 formation are currently active requests, travel path of the clients and their destinations, and paths of the other cars. For each client request, the dispatcher may choose a set of z carriers that collaborate to transfer a data item from a server to a client (z-relay zebroids). Here, z is the number of zebroids such that 0 ≤ z < N, where N is the total number of cars. When z = 0 there are no carriers, requiring a server to deliver the data item directly to the client. Otherwise, the chosen relay team of z zebroids hand over the data item transitively to one another to arrive at the client, thereby reducing availability latency (see Section 3.1 for details). To increase robustness, the dispatcher may employ multiple relay teams of z-carriers for every request. This may be useful in scenarios where the dispatcher has lower prediction accuracy in the information about the routes of the cars. Finally, storage constraints may require a zebroid to evict existing data items from its local storage to accommodate the client requested item. In this study, we quantify the following main factors that affect availability latency in the presence of zebroids: (i) data item repository size, (ii) car density, (iii) storage capacity per car, (iv) client trip duration, (v) replacement scheme employed by the zebroids, and (vi) accuracy of the car route predictions. For a significant subset of these factors, we address some key questions pertaining to use of zebroids both via analysis and extensive simulations. Our main findings are as follows. A naive random replacement policy employed by the zebroids shows competitive performance in terms of availability latency. With such a policy, substantial improvements in latency can be obtained with zebroids at a minimal replacement overhead. In more practical scenarios, where the dispatcher has inaccurate information about the car routes, zebroids continue to provide latency improvements. A surprising result is that changes in popularity of the data items do not impact the latency gains obtained with a simple instantiation of z-relay zebroids called one-instantaneous zebroids (see Section 3.1). This study suggests a number of interesting directions to be pursued to gain better understanding of design of carrier-based systems that improve availability latency. Related Work: Replication in mobile ad-hoc networks has been a widely studied topic [11, 12, 15]. However, none of these studies employ zebroids as data carriers to reduce the latency of the client"s requests. Several novel and important studies such as ZebraNet [13], DakNet [14], Data Mules [16], Message Ferries [20], and Seek and Focus [17] have analyzed factors impacting intermittently connected networks consisting of data carriers similar in spirit to zebroids. Factors considered by each study are dictated by their assumed environment and target application. A novel characteristic of our study is the impact on availability latency for a given database repository of items. A detailed description of related works can be obtained in [9]. The rest of this paper is organized as follows. Section 2 provides an overview of the terminology along with the factors that impact availability latency in the presence of zebroids. Section 3 describes how the zebroids may be employed. Section 4 provides details of the analysis methodology employed to capture the performance with zebroids. Section 5 describes the details of the simulation environment used for evaluation. Section 6 enlists the key questions examined in this study and answers them via analysis and simulations. Finally, Section 7 presents brief conclusions and future research directions. 2. OVERVIEW AND TERMINOLOGY Table 1 summarizes the notation of the parameters used in the paper. Below we introduce some terminology used in the paper. Assume a network of N AutoMata-equipped cars, each with storage capacity of α bytes. The total storage capacity of the system is ST =N ·α. There are T data items in the database, each with Database Parameters T Number of data items. Si Size of data item i fi Frequency of access to data item i. Replication Parameters Ri Normalized frequency of access to data item i ri Number of replicas for data item i n Characterizes a particular replication scheme. δi Average availability latency of data item i δagg Aggregate availability latency, δagg = T j=1 δj · fj AutoMata System Parameters G Number of cells in the map (2D-torus). N Number of AutoMata devices in the system. α Storage capacity per AutoMata. γ Trip duration of the client AutoMata. ST Total storage capacity of the AutoMata system, ST = N · α. Table 1: Terms and their definitions size Si. The frequency of access to data item i is denoted as fi, with T j=1 fj = 1. Let the trip duration of the client AutoMata under consideration be γ. We now define the normalized frequency of access to the data item i, denoted by Ri, as: Ri = (fi)n T j=1(fj)n ; 0 ≤ n ≤ ∞ (1) The exponent n characterizes a particular replication technique. A square-root replication scheme is realized when n = 0.5 [5]. This serves as the base-line for comparison with the case when zebroids are deployed. Ri is normalized to a value between 0 and 1. The number of replicas for data item i, denoted as ri, is: ri = min (N, max (1, Ri·N·α Si )). This captures the case when at least one copy of every data item must be present in the ad-hoc network at all times. In cases where a data item may be lost from the ad-hoc network, this equation becomes ri = min (N, max (0, Ri·N·α Si )). In this case, a request for the lost data item may need to be satisfied by fetching the item from a remote server. The availability latency for a data item i, denoted as δi, is defined as the earliest time at which a client AutoMata will find the first replica of the item accessible to it. If this condition is not satisfied, then we set δi to γ. This indicates that data item i was not available to the client during its journey. Note that since there is at least one replica in the system for every data item i, by setting γ to a large value we ensure that the client"s request for any data item i will be satisfied. However, in most practical circumstances γ may not be so large as to find every data item. We are interested in the availability latency observed across all data items. Hence, we augment the average availability latency for every data item i with its fi to obtain the following weighted availability latency (δagg) metric: δagg = T i=1 δi · fi Next, we present our solution approach describing how zebroids are selected. 3. SOLUTION APPROACH 3.1 Zebroids When a client references a data item missing from its local storage, the dispatcher identifies all cars with a copy of the data item as servers. Next, the dispatcher obtains the future routes of all cars for a finite time duration equivalent to the maximum time the client is willing to wait for its request to be serviced. Using this information, the dispatcher schedules the quickest delivery path from any of the servers to the client using any other cars as intermediate carriers. Hence, it determines the optimal set of forwarding decisions 76 that will enable the data item to be delivered to the client in the minimum amount of time. Note that the latency along the quickest delivery path that employs a relay team of z zebroids is similar to that obtained with epidemic routing [19] under the assumptions of infinite storage and no interference. A simple instantiation of z-relay zebroids occurs when z = 1 and the client"s request triggers a transfer of a copy of the requested data item from a server to a zebroid in its vicinity. Such a zebroid is termed one-instantaneous zebroid. In some cases, the dispatcher might have inaccurate information about the routes of the cars. Hence, a zebroid scheduled on the basis of this inaccurate information may not rendezvous with its target client. To minimize the likelihood of such scenarios, the dispatcher may schedule multiple zebroids. This may incur additional overhead due to redundant resource utilization to obtain the same latency improvements. The time required to transfer a data item from a server to a zebroid depends on its size and the available link bandwidth. With small data items, it is reasonable to assume that this transfer time is small, especially in the presence of the high bandwidth data plane. Large data items may be divided into smaller chunks enabling the dispatcher to schedule one or more zebroids to deliver each chunk to a client in a timely manner. This remains a future research direction. Initially, number of replicas for each data item replicas might be computed using Equation 1. This scheme computes the number of data item replicas as a function of their popularity. It is static because number of replicas in the system do not change and no replacements are performed. Hence, this is referred to as the ‘nozebroids" environment. We quantify the performance of the various replacement policies with reference to this base-line that does not employ zebroids. One may assume a cold start phase, where initially only one or few copies of every data item exist in the system. Many storage slots of the cars may be unoccupied. When the cars encounter one another they construct new replicas of some selected data items to occupy the empty slots. The selection procedure may be to choose the data items uniformly at random. New replicas are created as long as a car has a certain threshold of its storage unoccupied. Eventually, majority of the storage capacity of a car will be exhausted. 3.2 Carrier-based Replacement policies The replacement policies considered in this paper are reactive since a replacement occurs only in response to a request issued for a certain data item. When the local storage of a zebroid is completely occupied, it needs to replace one of its existing items to carry the client requested data item. For this purpose, the zebroid must select an appropriate candidate for eviction. This decision process is analogous to that encountered in operating system paging where the goal is to maximize the cache hit ratio to prevent disk access delay [18]. The carrier-based replacement policies employed in our study are Least Recently Used (LRU), Least Frequently Used (LFU) and Random (where a eviction candidate is chosen uniformly at random). We have considered local and global variants of the LRU/LFU policies which determine whether local or global knowledge of contents of the cars known at the dispatcher is used for the eviction decision at a zebroid (see [9] for more details). The replacement policies incur the following overheads. First, the complexity associated with the implementation of a policy. Second, the bandwidth used to transfer a copy of a data item from a server to the zebroid. Third, the average number of replacements incurred by the zebroids. Note that in the no-zebroids case neither overhead is incurred. The metrics considered in this study are aggregate availability latency, δagg, percentage improvement in δagg with zebroids as compared to the no-zebroids case and average number of replacements incurred per client request which is an indicator of the overhead incurred by zebroids. Note that the dispatchers with the help of the control plane may ensure that no data item is lost from the system. In other words, at least one replica of every data item is maintained in the ad-hoc network at all times. In such cases, even though a car may meet a requesting client earlier than other servers, if its local storage contains data items with only a single copy in the system, then such a car is not chosen as a zebroid. 4. ANALYSIS METHODOLOGY Here, we present the analytical evaluation methodology and some approximations as closed-form equations that capture the improvements in availability latency that can be obtained with both oneinstantaneous and z-relay zebroids. First, we present some preliminaries of our analysis methodology. • Let N be the number of cars in the network performing a 2D random walk on a √ G× √ G torus. An additional car serves as a client yielding a total of N + 1 cars. Such a mobility model has been used widely in the literature [17, 16] chiefly because it is amenable to analysis and provides a baseline against which performance of other mobility models can be compared. Moreover, this class of Markovian mobility models has been used to model the movements of vehicles [3, 21]. • We assume that all cars start from the stationary distribution and perform independent random walks. Although for sparse density scenarios, the independence assumption does hold, it is no longer valid when N approaches G. • Let the size of data item repository of interest be T. Also, data item i has ri replicas. This implies ri cars, identified as servers, have a copy of this data item when the client requests item i. All analysis results presented in this section are obtained assuming that the client is willing to wait as long as it takes for its request to be satisfied (unbounded trip duration γ = ∞). With the random walk mobility model on a 2D-torus, there is a guarantee that as long as there is at least one replica of the requested data item in the network, the client will eventually encounter this replica [2]. Extensions to the analysis that also consider finite trip durations can be obtained in [9]. Consider a scenario where no zebroids are employed. In this case, the expected availability latency for the data item is the expected meeting time of the random walk undertaken by the client with any of the random walks performed by the servers. Aldous et al. [2] show that the the meeting time of two random walks in such a setting can be modelled as an exponential distribution with the mean C = c · G · log G, where the constant c 0.17 for G ≥ 25. The meeting time, or equivalently the availability latency δi, for the client requesting data item i is the time till it encounters any of these ri replicas for the first time. This is also an exponential distribution with the following expected value (note that this formulation is valid only for sparse cases when G >> ri): δi = cGlogG ri The aggregate availability latency without employing zebroids is then this expression averaged over all data items, weighted by their frequency of access: δagg(no − zeb) = T i=1 fi · c · G · log G ri = T i=1 fi · C ri (2) 77 4.1 One-instantaneous zebroids Recall that with one-instantaneous zebroids, for a given request, a new replica is created on a car in the vicinity of the server, provided this car meets the client earlier than any of the ri servers. Moreover, this replica is spawned at the time step when the client issues the request. Let Nc i be the expected total number of nodes that are in the same cell as any of the ri servers. Then, we have Nc i = (N − ri) · (1 − (1 − 1 G )ri ) (3) In the analytical model, we assume that Nc i new replicas are created, so that the total number of replicas is increased to ri +Nc i . The availability latency is reduced since the client is more likely to meet a replica earlier. The aggregated expected availability latency in the case of one-instantaneous zebroids is then given by, δagg(zeb) = T i=1 fi · c · G · log G ri + Nc i = T i=1 fi · C ri + Nc i (4) Note that in obtaining this expression, for ease of analysis, we have assumed that the new replicas start from random locations in the torus (not necessarily from the same cell as the original ri servers). It thus treats all the Nc i carriers independently, just like the ri original servers. As we shall show below by comparison with simulations, this approximation provides an upper-bound on the improvements that can be obtained because it results in a lower expected latency at the client. It should be noted that the procedure listed above will yield a similar latency to that employed by a dispatcher employing oneinstantaneous zebroids (see Section 3.1). Since the dispatcher is aware of all future car movements it would only transfer the requested data item on a single zebroid, if it determines that the zebroid will meet the client earlier than any other server. This selected zebroid is included in the Nc i new replicas. 4.2 z-relay zebroids To calculate the expected availability latency with z-relay zebroids, we use a coloring problem analog similar to an approach used by Spyropoulos et al. [17]. Details of the procedure to obtain a closed-form expression are given in [9]. The aggregate availability latency (δagg) with z-relay zebroids is given by, δagg(zeb) = T i=1 [fi · C N + 1 · 1 N + 1 − ri · (N · log N ri − log (N + 1 − ri))] (5) 5. SIMULATION METHODOLOGY The simulation environment considered in this study comprises of vehicles such as cars that carry a fraction of the data item repository. A prediction accuracy parameter inherently provides a certain probabilistic guarantee on the confidence of the car route predictions known at the dispatcher. A value of 100% implies that the exact routes of all cars are known at all times. A 70% value for this parameter indicates that the routes predicted for the cars will match the actual ones with probability 0.7. Note that this probability is spread across the car routes for the entire trip duration. We now provide the preliminaries of the simulation study and then describe the parameter settings used in our experiments. • Similar to the analysis methodology, the map used is a 2D torus. A Markov mobility model representing a unbiased 2D random walk on the surface of the torus describes the movement of the cars across this torus. • Each grid/cell is a unique state of this Markov chain. In each time slot, every car makes a transition from a cell to any of its neighboring 8 cells. The transition is a function of the current location of the car and a probability transition matrix Q = [qij] where qij is the probability of transition from state i to state j. Only AutoMata equipped cars within the same cell may communicate with each other. • The parameters γ, δ have been discretized and expressed in terms of the number of time slots. • An AutoMata device does not maintain more than one replica of a data item. This is because additional replicas occupy storage without providing benefits. • Either one-instantaneous or z-relay zebroids may be employed per client request for latency improvement. • Unless otherwise mentioned, the prediction accuracy parameter is assumed to be 100%. This is because this study aims to quantify the effect of a large number of parameters individually on availability latency. Here, we set the size of every data item, Si, to be 1. α represents the number of storage slots per AutoMata. Each storage slot stores one data item. γ represents the duration of the client"s journey in terms of the number of time slots. Hence the possible values of availability latency are between 0 and γ. δ is defined as the number of time slots after which a client AutoMata device will encounter a replica of the data item for the first time. If a replica for the data item requested was encountered by the client in the first cell then we set δ = 0. If δ > γ then we set δ = γ indicating that no copy of the requested data item was encountered by the client during its entire journey. In all our simulations, for illustration we consider a 5 × 5 2D-torus with γ set to 10. Our experiments indicate that the trends in the results scale to maps of larger size. We simulated a skewed distribution of access to the T data items that obeys Zipf"s law with a mean of 0.27. This distribution is shown to correspond to sale of movie theater tickets in the United States [6]. We employ a replication scheme that allocates replicas for a data item as a function of the square-root of the frequency of access of that item. The square-root replication scheme is shown to have competitive latency performance over a large parameter space [8]. The data item replicas are distributed uniformly across the AutoMata devices. This serves as the base-line no-zebroids case. The square-root scheme also provides the initial replica distribution when zebroids are employed. Note that the replacements performed by the zebroids will cause changes to the data item replica distribution. Requests generated as per the Zipf distribution are issued one at a time. The client car that issues the request is chosen in a round-robin manner. After a maximum period of γ, the latency encountered by this request is recorded. In all the simulation results, each point is an average of 200,000 requests. Moreover, the 95% confidence intervals determined for the results are quite tight for the metrics of latency and replacement overhead. Hence, we only present them for the metric that captures the percentage improvement in latency with respect to the no-zebroids case. 6. RESULTS In this section, we describe our evaluation results where the following key questions are addressed. With a wide choice of replacement schemes available for a zebroid, what is their effect on availability latency? A more central question is: Do zebroids provide 78 0 20 40 60 80 100 1.5 2 2.5 3 3.5 Number of cars Aggregate availability latency (δ agg ) lru_global lfu_global lru_local lfu_local random Figure 1: Figure 1 shows the availability latency when employing one-instantaneous zebroids as a function of (N,α) values, when the total storage in the system is kept fixed, ST = 200. significant improvements in availability latency? What is the associated overhead incurred in employing these zebroids? What happens to these improvements in scenarios where a dispatcher may have imperfect information about the car routes? What inherent trade-offs exist between car density and storage per car with regards to their combined as well as individual effect on availability latency in the presence of zebroids? We present both simple analysis and detailed simulations to provide answers to these questions as well as gain insights into design of carrier-based systems. 6.1 How does a replacement scheme employed by a zebroid impact availability latency? For illustration, we present ‘scale-up" experiments where oneinstantaneous zebroids are employed (see Figure 1). By scale-up, we mean that α and N are changed proportionally to keep the total system storage, ST , constant. Here, T = 50 and ST = 200. We choose the following values of (N,α) = {(20,10), (25,8), (50,4), (100,2)}. The figure indicates that a random replacement scheme shows competitive performance. This is because of several reasons. Recall that the initial replica distribution is set as per the squareroot replication scheme. The random replacement scheme does not alter this distribution since it makes replacements blind to the popularity of a data item. However, the replacements cause dynamic data re-organization so as to better serve the currently active request. Moreover, the mobility pattern of the cars is random, hence, the locations from which the requests are issued by clients are also random and not known a priori at the dispatcher. These findings are significant because a random replacement policy can be implemented in a simple decentralized manner. The lru-global and lfu-global schemes provide a latency performance that is worse than random. This is because these policies rapidly develop a preference for the more popular data items thereby creating a larger number of replicas for them. During eviction, the more popular data items are almost never selected as a replacement candidate. Consequently, there are fewer replicas for the less popular items. Hence, the initial distribution of the data item replicas changes from square-root to that resembling linear replication. The higher number of replicas for the popular data items provide marginal additional benefits, while the lower number of replicas for the other data items hurts the latency performance of these global policies. The lfu-local and lru-local schemes have similar performance to random since they do not have enough history of local data item requests. We speculate that the performance of these local policies will approach that of their global variants for a large enough history of data item requests. On account of the competitive performance shown by a random policy, for the remainder of the paper, we present the performance of zebroids that employ a random replacement policy. 6.2 Do zebroids provide significant improvements in availability latency? We find that in many scenarios employing zebroids provides substantial improvements in availability latency. 6.2.1 Analysis We first consider the case of one-instantaneous zebroids. Figure 2.a shows the variation in δagg as a function of N for T = 10 and α = 1 with a 10 × 10 torus using Equation 4. Both the x and y axes are drawn to a log-scale. Figure 2.b show the % improvement in δagg obtained with one-instantaneous zebroids. In this case, only the x-axis is drawn to a log-scale. For illustration, we assume that the T data items are requested uniformly. Initially, when the network is sparse the analytical approximation for improvements in latency with zebroids, obtained from Equations 2 and 4, closely matches the simulation results. However, as N increases, the sparseness assumption for which the analysis is valid, namely N << G, is no longer true. Hence, the two curves rapidly diverge. The point at which the two curves move away from each other corresponds to a value of δagg ≤ 1. Moreover, as mentioned earlier, the analysis provides an upper bound on the latency improvements, as it treats the newly created replicas given by Nc i independently. However, these Nc i replicas start from the same cell as one of the server replicas ri. Finally, the analysis captures a oneshot scenario where given an initial data item replica distribution, the availability latency is computed. The new replicas created do not affect future requests from the client. On account of space limitations, here, we summarize the observations in the case when z-relay zebroids are employed. The interested reader can obtain further details in [9]. Similar observations, like the one-instantaneous zebroid case, apply since the simulation and analysis curves again start diverging when the analysis assumptions are no longer valid. However, the key observation here is that the latency improvement with z-relay zebroids is significantly better than the one-instantaneous zebroids case, especially for lower storage scenarios. This is because in sparse scenarios, the transitive hand-offs between the zebroids creates higher number of replicas for the requested data item, yielding lower availability latency. Moreover, it is also seen that the simulation validation curve for the improvements in δagg with z-relay zebroids approaches that of the one-instantaneous zebroid case for higher storage (higher N values). This is because one-instantaneous zebroids are a special case of z-relay zebroids. 6.2.2 Simulation We conduct simulations to examine the entire storage spectrum obtained by changing car density N or storage per car α to also capture scenarios where the sparseness assumptions for which the analysis is valid do not hold. We separate the effect of N and α by capturing the variation of N while keeping α constant (case 1) and vice-versa (case 2) both with z-relay and one-instantaneous zebroids. Here, we set the repository size as T = 25. Figure 3 captures case 1 mentioned above. Similar trends are observed with case 2, a complete description of those results are available in [9]. With Figure 3.b, keeping α constant, initially increasing car density has higher latency benefits because increasing N introduces more zebroids in the system. As N is further increased, ω reduces because the total storage in the system goes up. Consequently, the number of replicas per data item goes up thereby increasing the 79 number of servers. Hence, the replacement policy cannot find a zebroid as often to transport the requested data item to the client earlier than any of the servers. On the other hand, the increased number of servers benefits the no-zebroids case in bringing δagg down. The net effect results in reduction in ω for larger values of N. 10 1 10 2 10 3 10 −1 10 0 10 1 10 2 Number of cars no−zebroidsanal no−zebroids sim one−instantaneous anal one−instantaneoussim Aggregate Availability latency (δagg ) 2.a) δagg 10 1 10 2 10 3 0 10 20 30 40 50 60 70 80 90 100 Number of cars % Improvement in δagg wrt no−zebroids (ω) analytical upper−bound simulation 2.b) ω Figure 2: Figure 2 shows the latency performance with oneinstantaneous zebroids via simulations along with the analytical approximation for a 10 × 10 torus with T = 10. The trends mentioned above are similar to that obtained from the analysis. However, somewhat counter-intuitively with relatively higher system storage, z-relay zebroids provide slightly lower improvements in latency as compared to one-instantaneous zebroids. We speculate that this is due to the different data item replica distributions enforced by them. Note that replacements performed by the zebroids cause fluctuations in these replica distributions which may effect future client requests. We are currently exploring suitable choices of parameters that can capture these changing replica distributions. 6.3 What is the overhead incurred with improvements in latency with zebroids? We find that the improvements in latency with zebroids are obtained at a minimal replacement overhead (< 1 per client request). 6.3.1 Analysis With one-instantaneous zebroids, for each client request a maximum of one zebroid is employed for latency improvement. Hence, the replacement overhead per client request can amount to a maximum of one. Recall that to calculate the latency with one-instantaneous 0 50 100 150 200 250 300 350 400 0 1 2 3 4 5 6 Number of cars Aggregate availability latency (δagg ) no−zebroids one−instantaneous z−relays 3.a 0 50 100 150 200 250 300 350 400 0 10 20 30 40 50 60 Number of cars % Improvement in δagg wrt no−zebroids (ω) one−instantaneous z−relays 3.b Figure 3: Figure 3 depicts the latency performance with both one-instantaneous and z-relay zebroids as a function of the car density when α = 2 and T = 25. zebroids, Nc i new replicas are created in the same cell as the servers. Now a replacement is only incurred if one of these Nc i newly created replicas meets the client earlier than any of the ri servers. Let Xri and XNc i respectively be random variables that capture the minimum time till any of the ri and Nc i replicas meet the client. Since Xri and XNc i are assumed to be independent, by the property of exponentially distributed random variables we have, Overhead/request = 1 · P(XNc i < Xri ) + 0 · P(Xri ≤ XNc i ) (6) Overhead/request = ri C ri C + Nc i C = ri ri + Nc i (7) Recall that the number of replicas for data item i, ri, is a function of the total storage in the system i.e., ri = k·N ·α where k satisfies the constraint 1 ≤ ri ≤ N. Using this along with Equation 2, we get Overhead/request = 1 − G G + N · (1 − k · α) (8) Now if we keep the total system storage N · α constant since G and T are also constant, increasing N increases the replacement overhead. However, if N ·α is constant then increasing N causes α 80 0 20 40 60 80 100 0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5 Number of cars one−instantaneous zebroids Average number of replacements per request (N=20,α=10) (N=25,α=8) (N=50,α=4) (N=100,α=2) Figure 4: Figure 4 captures replacement overhead when employing one-instantaneous zebroids as a function of (N,α) values, when the total storage in the system is kept fixed, ST = 200. to go down. This implies that a higher replacement overhead is incurred for higher N and lower α values. Moreover, when ri = N, this means that every car has a replica of data item i. Hence, no zebroids are employed when this item is requested, yielding an overhead/request for this item as zero. Next, we present simulation results that validate our analysis hypothesis for the overhead associated with deployment of one-instantaneous zebroids. 6.3.2 Simulation Figure 4 shows the replacement overhead with one-instantaneous zebroids when (N,α) are varied while keeping the total system storage constant. The trends shown by the simulation are in agreement with those predicted by the analysis above. However, the total system storage can be changed either by varying car density (N) or storage per car (α). On account of similar trends, here we present the case when α is kept constant and N is varied (Figure 5). We refer the reader to [9] for the case when α is varied and N is held constant. We present an intuitive argument for the behavior of the perrequest replacement overhead curves. When the storage is extremely scarce so that only one replica per data item exists in the AutoMata network, the number of replacements performed by the zebroids is zero since any replacement will cause a data item to be lost from the system. The dispatcher ensures that no data item is lost from the system. At the other end of the spectrum, if storage becomes so abundant that α = T then the entire data item repository can be replicated on every car. The number of replacements is again zero since each request can be satisfied locally. A similar scenario occurs if N is increased to such a large value that another car with the requested data item is always available in the vicinity of the client. However, there is a storage spectrum in the middle where replacements by the scheduled zebroids result in improvements in δagg (see Figure 3). Moreover, we observe that for sparse storage scenarios, the higher improvements with z-relay zebroids are obtained at the cost of a higher replacement overhead when compared to the one-instantaneous zebroids case. This is because in the former case, each of these z zebroids selected along the lowest latency path to the client needs to perform a replacement. However, the replacement overhead is still less than 1 indicating that on an average less than one replacement per client request is needed even when z-relay zebroids are employed. 0 50 100 150 200 250 300 350 400 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 Number of cars z−relays one−instantaneous Average number of replacements per request Figure 5: Figure 5 shows the replacement overhead with zebroids for the cases when N is varied keeping α = 2. 10 20 30 40 50 60 70 80 90 100 0 0.5 1 1.5 2 2.5 3 3.5 4 Prediction percentage no−zebroids (N=50) one−instantaneous (N=50) z−relays (N=50) no−zebroids (N=200) one−instantaneous (N=200) z−relays (N=200) Aggregate Availability Latency (δagg ) Figure 6: Figure 6 shows δagg for different car densities as a function of the prediction accuracy metric with α = 2 and T = 25. 6.4 What happens to the availability latency with zebroids in scenarios with inaccuracies in the car route predictions? We find that zebroids continue to provide improvements in availability latency even with lower accuracy in the car route predictions. We use a single parameter p to quantify the accuracy of the car route predictions. 6.4.1 Analysis Since p represents the probability that a car route predicted by the dispatcher matches the actual one, hence, the latency with zebroids can be approximated by, δerr agg = p · δagg(zeb) + (1 − p) · δagg(no − zeb) (9) δerr agg = p · δagg(zeb) + (1 − p) · C ri (10) Expressions for δagg(zeb) can be obtained from Equations 4 (one-instantaneous) or 5 (z-relay zebroids). 6.4.2 Simulation Figure 6 shows the variation in δagg as a function of this route prediction accuracy metric. We observe a smooth reduction in the 81 improvement in δagg as the prediction accuracy metric reduces. For zebroids that are scheduled but fail to rendezvous with the client due to the prediction error, we tag any such replacements made by the zebroids as failed. It is seen that failed replacements gradually increase as the prediction accuracy reduces. 6.5 Under what conditions are the improvements in availability latency with zebroids maximized? Surprisingly, we find that the improvements in latency obtained with one-instantaneous zebroids are independent of the input distribution of the popularity of the data items. 6.5.1 Analysis The fractional difference (labelled ω) in the latency between the no-zebroids and one-instantaneous zebroids is obtained from equations 2, 3, and 4 as ω = T i=1 fi·C ri − T i=1 fi·C ri+(N−ri)·(1−(1− 1 G )ri ) T i=1 fi·C ri (11) Here C = c·G·log G. This captures the fractional improvement in the availability latency obtained by employing one-instantaneous zebroids. Let α = 1, making the total storage in the system ST = N. Assuming the initial replica distribution is as per the squareroot replication scheme, we have, ri = √ fi·N T j=1 √ fj . Hence, we get fi = K2 ·r2 i N2 , where K = T j=1 fj. Using this, along with the approximation (1 − x)n 1 − n · x for small x, we simplify the above equation to get, ω = 1 − T i=1 ri 1+ N−ri G T i=1 ri In order to determine when the gains with one-instantaneous zebroids are maximized, we can frame an optimization problem as follows: Maximize ω, subject to T i=1 ri = ST THEOREM 1. With a square-root replication scheme, improvements obtained with one-instantaneous zebroids are independent of the input popularity distribution of the data items. (See [9] for proof) 6.5.2 Simulation We perform simulations with two different frequency distribution of data items: Uniform and Zipfian (with mean= 0.27). Similar latency improvements with one-instantaneous zebroids are obtained in both cases. This result has important implications. In cases with biased popularity toward certain data items, the aggregate improvements in latency across all data item requests still remain the same. Even in scenarios where the frequency of access to the data items changes dynamically, zebroids will continue to provide similar latency improvements. 7. CONCLUSIONS AND FUTURE RESEARCH DIRECTIONS In this study, we examined the improvements in latency that can be obtained in the presence of carriers that deliver a data item from a server to a client. We quantified the variation in availability latency as a function of a rich set of parameters such as car density, storage per car, title database size, and replacement policies employed by zebroids. Below we summarize some key future research directions we intend to pursue. To better reflect reality we would like to validate the observations obtained from this study with some real world simulation traces of vehicular movements (for example using CORSIM [1]). This will also serve as a validation for the utility of the Markov mobility model used in this study. We are currently analyzing the performance of zebroids on a real world data set comprising of an ad-hoc network of buses moving around a small neighborhood in Amherst [4]. Zebroids may also be used for delivery of data items that carry delay sensitive information with a certain expiry. Extensions to zebroids that satisfy such application requirements presents an interesting future research direction. 8. ACKNOWLEDGMENTS This research was supported in part by an Annenberg fellowship and NSF grants numbered CNS-0435505 (NeTS NOSS), CNS-0347621 (CAREER), and IIS-0307908. 9. REFERENCES [1] Federal Highway Administration. Corridor simulation. Version 5.1, http://www.ops.fhwa.dot.gov/trafficanalysistools/cors im.htm. [2] D. Aldous and J. Fill. Reversible markov chains and random walks on graphs. Under preparation. [3] A. Bar-Noy, I. Kessler, and M. Sidi. Mobile Users: To Update or Not to Update. In IEEE Infocom, 1994. [4] J. Burgess, B. Gallagher, D. Jensen, and B. Levine. MaxProp: Routing for Vehicle-Based Disruption-Tolerant Networking. In IEEE Infocom, April 2006. [5] E. Cohen and S. Shenker. Replication Strategies in Unstructured Peer-to-Peer Networks. In SIGCOMM, 2002. [6] A. Dan, D. Dias, R. Mukherjee, D. Sitaram, and R. Tewari. Buffering and Caching in Large-Scale Video Servers. In COMPCON, 1995. [7] S. Ghandeharizadeh, S. Kapadia, and B. Krishnamachari. PAVAN: A Policy Framework for Content Availabilty in Vehicular ad-hoc Networks. In VANET, New York, NY, USA, 2004. ACM Press. [8] S. Ghandeharizadeh, S. Kapadia, and B. Krishnamachari. Comparison of Replication Strategies for Content Availability in C2P2 networks. In MDM, May 2005. [9] S. Ghandeharizadeh, S. Kapadia, and B. Krishnamachari. An Evaluation of Availability Latency in Carrier-based Vehicular ad-hoc Networks. Technical report, Department of Computer Science, University of Southern California,CENG-2006-1, 2006. [10] S. Ghandeharizadeh and B. Krishnamachari. C2p2: A peer-to-peer network for on-demand automobile information services. In Globe. IEEE, 2004. [11] T. Hara. Effective Replica Allocation in ad-hoc Networks for Improving Data Accessibility. In IEEE Infocom, 2001. [12] H. Hayashi, T. Hara, and S. Nishio. A Replica Allocation Method Adapting to Topology Changes in ad-hoc Networks. In DEXA, 2005. [13] P. Juang, H. Oki, Y. Wang, M. Martonosi, L. Peh, and D. Rubenstein. Energy-efficient computing for wildlife tracking: design tradeoffs and early experiences with ZebraNet. SIGARCH Comput. Archit. News, 2002. [14] A. Pentland, R. Fletcher, and A. Hasson. DakNet: Rethinking Connectivity in Developing Nations. Computer, 37(1):78-83, 2004. [15] F. Sailhan and V. Issarny. Cooperative Caching in ad-hoc Networks. In MDM, 2003. [16] R. Shah, S. Roy, S. Jain, and W. Brunette. Data mules: Modeling and analysis of a three-tier architecture for sparse sensor networks. Elsevier ad-hoc Networks Journal, 1, September 2003. [17] T. Spyropoulos, K. Psounis, and C. Raghavendra. Single-Copy Routing in Intermittently Connected Mobile Networks. In SECON, April 2004. [18] A. Tanenbaum. Modern Operating Systems, 2nd Edition, Chapter 4, Section 4.4 . Prentice Hall, 2001. [19] A. Vahdat and D. Becker. Epidemic routing for partially-connected ad-hoc networks. Technical report, Department of Computer Science, Duke University, 2000. [20] W. Zhao, M. Ammar, and E. Zegura. A message ferrying approach for data delivery in sparse mobile ad-hoc networks. In MobiHoc, pages 187-198, New York, NY, USA, 2004. ACM Press. [21] M. Zonoozi and P. Dassanayake. User Mobility Modeling and Characterization of Mobility Pattern. IEEE Journal on Selected Areas in Communications, 15:1239-1252, September 1997. 82

naive random replacement policy;zebroid;vehicular network;termed zebroid;audio and video clip;mobility;car density;storage per device;datum carrier;latency;simplified instantiation of zebroid;automaton;availability latency;repository size;markov model;zebroid simplified instantiation;mobile device;peer-to-peer vehicular ad-hoc network;replacement policy

train_C-69

pTHINC: A Thin-Client Architecture for Mobile Wireless Web

Although web applications are gaining popularity on mobile wireless PDAs, web browsers on these systems can be quite slow and often lack adequate functionality to access many web sites. We have developed pTHINC, a PDA thinclient solution that leverages more powerful servers to run full-function web browsers and other application logic, then sends simple screen updates to the PDA for display. pTHINC uses server-side screen scaling to provide high-fidelity display and seamless mobility across a broad range of different clients and screen sizes, including both portrait and landscape viewing modes. pTHINC also leverages existing PDA control buttons to improve system usability and maximize available screen resolution for application display. We have implemented pTHINC on Windows Mobile and evaluated its performance on mobile wireless devices. Our results compared to local PDA web browsers and other thin-client approaches demonstrate that pTHINC provides superior web browsing performance and is the only PDA thin client that effectively supports crucial browser helper applications such as video playback.

1. INTRODUCTION The increasing ubiquity of wireless networks and decreasing cost of hardware is fueling a proliferation of mobile wireless handheld devices, both as standalone wireless Personal Digital Assistants (PDA) and popular integrated PDA/cell phone devices. These devices are enabling new forms of mobile computing and communication. Service providers are leveraging these devices to deliver pervasive web access, and mobile web users already often use these devices to access web-enabled information such as news, email, and localized travel guides and maps. It is likely that within a few years, most of the devices accessing the web will be mobile. Users typically access web content by running a web browser and associated helper applications locally on the PDA. Although native web browsers exist for PDAs, they deliver subpar performance and have a much smaller feature set and more limited functionality than their desktop computing counterparts [10]. As a result, PDA web browsers are often not able to display web content from web sites that leverage more advanced web technologies to deliver a richer web experience. This fundamental problem arises for two reasons. First, because PDAs have a completely different hardware/software environment from traditional desktop computers, web applications need to be rewritten and customized for PDAs if at all possible, duplicating development costs. Because the desktop application market is larger and more mature, most development effort generally ends up being spent on desktop applications, resulting in greater functionality and performance than their PDA counterparts. Second, PDAs have a more resource constrained environment than traditional desktop computers to provide a smaller form factor and longer battery life. Desktop web browsers are large, complex applications that are unable to run on a PDA. Instead, developers are forced to significantly strip down these web browsers to provide a usable PDA web browser, thereby crippling PDA browser functionality. Thin-client computing provides an alternative approach for enabling pervasive web access from handheld devices. A thin-client computing system consists of a server and a client that communicate over a network using a remote display protocol. The protocol enables graphical displays to be virtualized and served across a network to a client device, while application logic is executed on the server. Using the remote display protocol, the client transmits user input to the server, and the server returns screen updates of the applications from the server to the client. Using a thin-client model for mobile handheld devices, PDAs can become simple stateless clients that leverage the remote server capabilities to execute web browsers and other helper applications. The thin-client model provides several important benefits for mobile wireless web. First, standard desktop web applications can be used to deliver web content to PDAs without rewriting or adapting applications to execute on a PDA, reducing development costs and leveraging existing software investments. Second, complex web applications can be executed on powerful servers instead of running stripped down versions on more resource constrained PDAs, providing greater functionality and better performance [10]. Third, web applications can take advantage of servers with faster networks and better connectivity, further boosting application performance. Fourth, PDAs can be even simpler devices since they do not need to perform complex application logic, potentially reducing energy consumption and extend143 ing battery life. Finally, PDA thin clients can be essentially stateless appliances that do not need to be backed up or restored, require almost no maintenance or upgrades, and do not store any sensitive data that can be lost or stolen. This model provides a viable avenue for medical organizations to comply with HIPAA regulations [6] while embracing mobile handhelds in their day to day operations. Despite these potential advantages, thin clients have been unable to provide the full range of these benefits in delivering web applications to mobile handheld devices. Existing thin clients were not designed for PDAs and do not account for important usability issues in the context of small form factor devices, resulting in difficulty in navigating displayed web content. Furthermore, existing thin clients are ineffective at providing seamless mobility across the heterogeneous mix of device display sizes and resolutions. While existing thin clients can already provide faster performance than native PDA web browsers in delivering HTML web content, they do not effectively support more display-intensive web helper applications such as multimedia video, which is increasingly an integral part of available web content. To harness the full potential of thin-client computing in providing mobile wireless web on PDAs, we have developed pTHINC (PDA THin-client InterNet Computing). pTHINC builds on our previous work on THINC [1] to provide a thinclient architecture for mobile handheld devices. pTHINC virtualizes and resizes the display on the server to efficiently deliver high-fidelity screen updates to a broad range of different clients, screen sizes, and screen orientations, including both portrait and landscape viewing modes. This enables pTHINC to provide the same persistent web session across different client devices. For example, pTHINC can provide the same web browsing session appropriately scaled for display on a desktop computer and a PDA so that the same cookies, bookmarks, and other meta-data are continuously available on both machines simultaneously. pTHINC"s virtual display approach leverages semantic information available in display commands, and client-side video hardware to provide more efficient remote display mechanisms that are crucial for supporting more display-intensive web applications. Given limited display resolution on PDAs, pTHINC maximizes the use of screen real estate for remote display by moving control functionality from the screen to readily available PDA control buttons, improving system usability. We have implemented pTHINC on Windows Mobile and demonstrated that it works transparently with existing applications, window systems, and operating systems, and does not require modifying, recompiling, or relinking existing software. We have quantitatively evaluated pTHINC against local PDA web browsers and other thin-client approaches on Pocket PC devices. Our experimental results demonstrate that pTHINC provides superior web browsing performance and is the only PDA thin client that effectively supports crucial browser helper applications such as video playback. This paper presents the design and implementation of pTHINC. Section 2 describes the overall usage model and usability characteristics of pTHINC. Section 3 presents the design and system architecture of pTHINC. Section 4 presents experimental results measuring the performance of pTHINC on web applications and comparing it against native PDA browsers and other popular PDA thin-client systems. Section 5 discusses related work. Finally, we present some concluding remarks. 2. PTHINC USAGE MODEL pTHINC is a thin-client system that consists of a simple client viewer application that runs on the PDA and a server that runs on a commodity PC. The server leverages more powerful PCs to to run web browsers and other application logic. The client takes user input from the PDA stylus and virtual keyboard and sends them to the server to pass to the applications. Screen updates are then sent back from the server to the client for display to the user. When the pTHINC PDA client is started, the user is presented with a simple graphical interface where information such as server address and port, user authentication information, and session settings can be provided. pTHINC first attempts to connect to the server and perform the necessary handshaking. Once this process has been completed, pTHINC presents the user with the most recent display of his session. If the session does not exist, a new session is created. Existing sessions can be seamlessly continued without changes in the session setting or server configuration. Unlike other thin-client systems, pTHINC provides a user with a persistent web session model in which a user can launch a session running a web browser and associated applications at the server, then disconnect from that session and reconnect to it again anytime. When a user reconnects to the session, all of the applications continue running where the user left off, so that the user can continue working as though he or she never disconnected. The ability to disconnect and reconnect to a session at anytime is an important benefit for mobile wireless PDA users which may have intermittent network connectivity. pTHINC"s persistent web session model enables a user to reconnect to a web session from devices other than the one on which the web session was originally initiated. This provides users with seamless mobility across different devices. If a user loses his PDA, he can easily use another PDA to access his web session. Furthermore, pTHINC allows users to use non-PDA devices to access web sessions as well. A user can access the same persistent web session on a desktop PC as on a PDA, enabling a user to use the same web session from any computer. pTHINC"s persistent web session model addresses a key problem encountered by mobile web users, the lack of a common web environment across computers. Web browsers often store important information such as bookmarks, cookies, and history, which enable them to function in a much more useful manner. The problem that occurs when a user moves between computers is that this data, which is specific to a web browser installation, cannot move with the user. Furthermore, web browsers often need helper applications to process different media content, and those applications may not be consistently available across all computers. pTHINC addresses this problem by enabling a user to remotely use the exact same web browser environment and helper applications from any computer. As a result, pTHINC can provide a common, consistent web browsing environment for mobile users across different devices without requiring them to attempt to repeatedly synchronize different web browsing environments across multiple machines. To enable a user to access the same web session on different devices, pTHINC must provide mechanisms to support different display sizes and resolutions. Toward this end, pTHINC provides a zoom feature that enables a user to zoom in and out of a display and allows the display of a web 144 Figure 1: pTHINC shortcut keys session to be resized to fit the screen of the device being used. For example, if the server is running a web session at 1024×768 but the client is a PDA with a display resolution of 640×480, pTHINC will resize the desktop display to fit the full display in the smaller screen of the PDA. pTHINC provides the PDA user with the option to increase the size of the display by zooming in to different parts of the display. Users are often familiar with the general layout of commonly visited websites, and are able to leverage this resizing feature to better navigate through web pages. For example, a user can zoom out of the display to view the entire page content and navigate hyperlinks, then zoom in to a region of interest for a better view. To enable a user to access the same web session on different devices, pTHINC must also provide mechanisms to support different display orientations. In a desktop environment, users are typically accustomed to having displays presented in landscape mode where the screen width is larger than its height. However, in a PDA environment, the choice is not always obvious. Some users may prefer having the display in portrait mode, as it is easier to hold the device in their hands, while others may prefer landscape mode in order to minimize the amount of side-scrolling necessary to view a web page. To accommodate PDA user preferences, pTHINC provides an orientation feature that enables it to seamless rotate the display between landscape and portrait mode. The landscape mode is particularly useful for pTHINC users who frequently access their web sessions on both desktop and PDA devices, providing those users with the same familiar landscape setting across different devices. Because screen space is a relatively scarce resource on PDAs, pTHINC runs in fullscreen mode to maximize the screen area available to display the web session. To be able to use all of the screen on the PDA and still allow the user to control and interact with it, pTHINC reuses the typical shortcut buttons found on PDAs to perform all the control functions available to the user. The buttons used by pTHINC do not require any OS environment changes; they are simply intercepted by the pTHINC client application when they are pressed. Figure 1 shows how pTHINC utilizes the shortcut buttons to provide easy navigation and improve the overall user experience. These buttons are not device specific, and the layout shown is common to widelyused PocketPC devices. pTHINC provides six shortcuts to support its usage model: • Rotate Screen: The record button on the left edge is used to rotate the screen between portrait and landscape mode each time the button is pressed. • Zoom Out: The leftmost button on the bottom front is used to zoom out the display of the web session providing a bird"s eye view of the web session. • Zoom In: The second leftmost button on the bottom front is used to zoom in the display of the web session to more clearly view content of interest. • Directional Scroll: The middle button on the bottom front is used to scroll around the display using a single control button in a way that is already familiar to PDA users. This feature is particularly useful when the user has zoomed in to a region of the display such that only part of the display is visible on the screen. • Show/Hide Keyboard: The second rightmost button on the bottom front is used to bring up a virtual keyboard drawn on the screen for devices which have no physical keyboard. The virtual keyboard uses standard PDA OS mechanisms, providing portability across different PDA environments. • Close Session: The rightmost button on the bottom front is used to disconnect from the pTHINC session. pTHINC uses the PDA touch screen, stylus, and standard user interface mechanisms to provide a user interface pointand-click metaphor similar to that provided by the mouse in a traditional desktop computing environment. pTHINC does not use a cursor since PDA environments do not provide one. Instead, a user can use the stylus to tap on different sections of the touch screen to indicate input focus. A single tap on the touch screen generates a corresponding single click mouse event. A double tap on the touch screen generates a corresponding double click mouse event. pTHINC provides two-button mouse emulation by using the stylus to press down on the screen for one second to generate a right mouse click. All of these actions are identical to the way users already interact with PDA applications in the common PocketPC environment. In web browsing, users can click on hyperlinks and focus on input boxes by simply tapping on the desired screen area of interest. Unlike local PDA web browsers and other PDA applications, pTHINC leverages more powerful desktop user interface metaphors to enable users to manipulate multiple open application windows instead of being limited to a single application window at any given moment. This provides increased browsing flexibility beyond what is currently available on PDA devices. Similar to a desktop environment, browser windows and other application windows can be moved around by pressing down and dragging the stylus similar to a mouse. 3. PTHINC SYSTEM ARCHITECTURE pTHINC builds on the THINC [1] remote display architecture to provide a thin-client system for PDAs. pTHINC virtualizes the display at the server by leveraging the video device abstraction layer, which sits below the window server and above the framebuffer. This is a well-defined, low-level, device-dependent layer that exposes the video hardware to the display system. pTHINC accomplishes this through a simple virtual display driver that intercepts drawing commands, packetizes, and sends them over the network. 145 While other thin-client approaches intercept display commands at other layers of the display subsystem, pTHINC"s display virtualization approach provides some key benefits in efficiently supporting PDA clients. For example, intercepting display commands at a higher layer between applications and the window system as is done by X [17] requires replicating and running a great deal of functionality on the PDA that is traditionally provided by the desktop window system. Given both the size and complexity of traditional window systems, attempting to replicate this functionality in the restricted PDA environment would have proven to be a daunting, and perhaps unfeasible task. Furthermore, applications and the window system often require tight synchronization in their operation and imposing a wireless network between them by running the applications on the server and the window system on the client would significantly degrade performance. On the other hand, intercepting at a lower layer by extracting pixels out of the framebuffer as they are rendered provides a simple solution that requires very little functionality on the PDA client, but can also result in degraded performance. The reason is that by the time the remote display server attempts to send screen updates, it has lost all semantic information that may have helped it encode efficiently, and it must resort to using a generic and expensive encoding mechanism on the server, as well as a potentially expensive decoding mechanism on the limited PDA client. In contrast to both the high and low level interception approaches, pTHINC"s approach of intercepting at the device driver provides an effective balance between client and server simplicity, and the ability to efficiently encode and decode screen updates. By using a low-level virtual display approach, pTHINC can efficiently encode application display commands using only a small set of low-level commands. In a PDA environment, this set of commands provides a crucial component in maintaining the simplicity of the client in the resourceconstrained PDA environment. The display commands are shown in Table 1, and work as follows. COPY instructs the client to copy a region of the screen from its local framebuffer to another location. This command improves the user experience by accelerating scrolling and opaque window movement without having to resend screen data from the server. SFILL, PFILL, and BITMAP are commands that paint a fixed-size region on the screen. They are useful for accelerating the display of solid window backgrounds, desktop patterns, backgrounds of web pages, text drawing, and certain operations in graphics manipulation programs. SFILL fills a sizable region on the screen with a single color. PFILL replicates a tile over a screen region. BITMAP performs a fill using a bitmap of ones and zeros as a stipple to apply a foreground and background color. Finally, RAW is used to transmit unencoded pixel data to be displayed verbatim on a region of the screen. This command is invoked as a last resort if the server is unable to employ any other command, and it is the only command that may be compressed to mitigate its impact on network bandwidth. pTHINC delivers its commands using a non-blocking, serverpush update mechanism, where as soon as display updates are generated on the server, they are sent to the client. Clients are not required to explicitly request display updates, thus minimizing the impact that the typical varying network latency of wireless links may have on the responsiveness of the system. Keeping in mind that resource Command Description COPY Copy a frame buffer area to specified coordinates SFILL Fill an area with a given pixel color value PFILL Tile an area with a given pixel pattern BITMAP Fill a region using a bit pattern RAW Display raw pixel data at a given location Table 1: pTHINC Protocol Display Commands constrained PDAs and wireless networks may not be able to keep up with a fast server generating a large number of updates, pTHINC is able to coalesce, clip, and discard updates automatically if network loss or congestion occurs, or the client cannot keep up with the rate of updates. This type of behavior proves crucial in a web browsing environment, where for example, a page may be redrawn multiple times as it is rendered on the fly by the browser. In this case, the PDA will only receive and render the final result, which clearly is all the user is interesting in seeing. pTHINC prioritizes the delivery of updates to the PDA using a Shortest-Remaining-Size-First (SRSF) preemptive update scheduler. SRSF is analogous to Shortest-RemainingProcessing-Time scheduling, which is known to be optimal for minimizing mean response time in an interactive system. In a web browsing environment, short jobs are associated with text and basic page layout components such as the page"s background, which are critical web content for the user. On the other hand, large jobs are often lower priority beautifying elements, or, even worse, web page banners and advertisements, which are of questionable value to the user as he or she is browsing the page. Using SRSF, pTHINC is able to maximize the utilization of the relatively scarce bandwidth available on the wireless connection between the PDA and the server. 3.1 Display Management To enable users to just as easily access their web browser and helper applications from a desktop computer at home as from a PDA while on the road, pTHINC provides a resize mechanism to zoom in and out of the display of a web session. pTHINC resizing is completely supported by the server, not the client. The server resamples updates to fit within the PDAs viewport before they are transmitted over the network. pTHINC uses Fant"s resampling algorithm to resize pixel updates. This provides smooth, visually pleasing updates with properly antialiasing and has only modest computational requirements. pTHINC"s resizing approach has a number of advantages. First, it allows the PDA to leverage the vastly superior computational power of the server to use high quality resampling algorithms and produce higher quality updates for the PDA to display. Second, resizing the screen does not translate into additional resource requirements for the PDA, since it does not need to perform any additional work. Finally, better utilization of the wireless network is attained since rescaling the updates reduces their bandwidth requirements. To enable users to orient their displays on a PDA to provide a viewing experience that best accommodates user preferences and the layout of web pages or applications, pTHINC provides a display rotation mechanism to switch between landscape and portrait viewing modes. pTHINC display rotation is completely supported by the client, not the server. pTHINC does not explicitly recalculate the ge146 ometry of display updates to perform rotation, which would be computationally expensive. Instead, pTHINC simply changes the way data is copied into the framebuffer to switch between display modes. When in portrait mode, data is copied along the rows of the framebuffer from left to right. When in landscape mode, data is copied along the columns of the framebuffer from top to bottom. These very fast and simple techniques replace one set of copy operations with another and impose no performance overhead. pTHINC provides its own rotation mechanism to support a wide range of devices without imposing additional feature requirements on the PDA. Although some newer PDA devices provide native support for different orientations, this mechanism is not dynamic and requires the user to rotate the PDA"s entire user interface before starting the pTHINC client. Windows Mobile provides native API mechanisms for PDA applications to rotate their UI on the fly, but these mechanisms deliver poor performance and display quality as the rotation is performed naively and is not completely accurate. 3.2 Video Playback Video has gradually become an integral part of the World Wide Web, and its presence will only continue to increase. Web sites today not only use animated graphics and flash to deliver web content in an attractive manner, but also utilize streaming video to enrich the web interface. Users are able to view pre-recorded and live newscasts on CNN, watch sports highlights on ESPN, and even search through large collection of videos on Google Video. To allow applications to provide efficient video playback, interfaces have been created in display systems that allow video device drivers to expose their hardware capabilities back to the applications. pTHINC takes advantage of these interfaces and its virtual device driver approach to provide a virtual bridge between the remote client and its hardware and the applications, and transparently support video playback. On top of this architecture, pTHINC uses the YUV colorspace to encode the video content, which provides a number of benefits. First, it has become increasingly common for PDA video hardware to natively support YUV and be able to perform the colorspace conversion and scaling automatically. As a result, pTHINC is able to provide fullscreen video playback without any performance hits. Second, the use of YUV allows for a more efficient representation of RGB data without loss of quality, by taking advantage of the human eye"s ability to better distinguish differences in brightness than in color. In particular, pTHINC uses the YV12 format, which allows full color RGB data to be encoded using just 12 bits per pixel. Third, YUV data is produced as one of the last steps of the decoding process of most video codecs, allowing pTHINC to provide video playback in a manner that is format independent. Finally, even if the PDA"s video hardware is unable to accelerate playback, the colorspace conversion process is simple enough that it does not impose unreasonable requirements on the PDA. A more concrete example of how pTHINC leverages the PDA video hardware to support video playback can be seen in our prototype implementation on the popular Dell Axim X51v PDA, which is equipped with the Intel 2700G multimedia accelerator. In this case, pTHINC creates an offscreen buffer in video memory and writes and reads from this memory region data on the YV12 format. When a new video frame arrives, video data is copied from the buffer to Figure 2: Experimental Testbed an overlay surface in video memory, which is independent of the normal surface used for traditional drawing. As the YV12 data is put onto the overlay, the Intel accelerator automatically performs both colorspace conversion and scaling. By using the overlay surface, pTHINC has no need to redraw the screen once video playback is over since the overlapped surface is unaffected. In addition, specific overlay regions can be manipulated by leveraging the video hardware, for example to perform hardware linear interpolation to smooth out the frame and display it fullscreen, and to do automatic rotation when the client runs in landscape mode. 4. EXPERIMENTAL RESULTS We have implemented a pTHINC prototype that runs the client on widely-used Windows Mobile-based Pocket PC devices and the server on both Windows and Linux operating systems. To demonstrate its effectiveness in supporting mobile wireless web applications, we have measured its performance on web applications. We present experimental results on different PDA devices for two popular web applications, browsing web pages and playing video content from the web. We compared pTHINC against native web applications running locally on the PDA to demonstrate the improvement that pTHINC can provide over the traditional fat-client approach. We also compared pTHINC against three of the most widely used thin clients that can run on PDAs, Citrix Meta-FrameXP [2], Microsoft Remote Desktop [3] and VNC (Virtual Network Computing) [16]. We follow common practice and refer to Citrix MetaFrameXP and Microsoft Remote Desktop by their respective remote display protocols, ICA (Independent Computing Architecture) and RDP (Remote Desktop Protocol). 4.1 Experimental Testbed We conducted our web experiments using two different wireless Pocket PC PDAs in an isolated Wi-Fi network testbed, as shown in Figure 2. The testbed consisted of two PDA client devices, a packet monitor, a thin-client server, and a web server. Except for the PDAs, all of the other machines were IBM Netfinity 4500R servers with dual 933 MHz Intel PIII CPUs and 512 MB RAM and were connected on a switched 100 Mbps FastEthernet network. The web server used was Apache 1.3.27, the network emulator was NISTNet 2.0.12, and the packet monitor was Ethereal 0.10.9. The PDA clients connected to the testbed through a 802.11b Lucent Orinoco AP-2000 wireless access point. All experiments using the wireless network were conducted within ten feet of the access point, so we considered the amount of packet loss to be negligible in our experiments. Two Pocket PC PDAs were used to provide results across both older, less powerful models and newer higher performance models. The older model was a Dell Axim X5 with 147 Client 1024×768 640×480 Depth Resize Clip RDP no yes 8-bit no yes VNC yes yes 16-bit no no ICA yes yes 16-bit yes no pTHINC yes yes 24-bit yes no Table 2: Thin-client Testbed Configuration Setting a 400 MHz Intel XScale PXA255 CPU and 64 MB RAM running Windows Mobile 2003 and a Dell TrueMobile 1180 2.4Ghz CompactFlash card for wireless networking. The newer model was a Dell Axim X51v with a 624 MHz Intel XScale XPA270 CPU and 64 MB RAM running Windows Mobile 5.0 and integrated 802.11b wireless networking. The X51v has an Intel 2700G multimedia accelerator with 16MB video memory. Both PDAs are capable of 16-bit color but have different screen sizes and display resolutions. The X5 has a 3.5 inch diagonal screen with 240×320 resolution. The X51v has a 3.7 inch diagonal screen with 480×640. The four thin clients that we used support different levels of display quality as summarized in Table 2. The RDP client only supports a fixed 640×480 display resolution on the server with 8-bit color depth, while other platforms provide higher levels of display quality. To provide a fair comparison across all platforms, we conducted our experiments with thin-client sessions configured for two possible resolutions, 1024×768 and 640×480. Both ICA and VNC were configured to use the native PDA resolution of 16-bit color depth. The current pTHINC prototype uses 24-bit color directly and the client downsamples updates to the 16-bit color depth available on the PDA. RDP was configured using only 8-bit color depth since it does not support any better color depth. Since both pTHINC and ICA provide the ability to view the display resized to fit the screen, we measured both clients with and without the display resized to fit the PDA screen. Each thin client was tested using landscape rather than portrait mode when available. All systems run on the X51v could run in landscape mode because the hardware provides a landscape mode feature. However, the X5 does not provide this functionality. Only pTHINC directly supports landscape mode, so it was the only system that could run in landscape mode on both the X5 and X51v. To provide a fair comparison, we also standardized on common hardware and operating systems whenever possible. All of the systems used the Netfinity server as the thin-client server. For the two systems designed for Windows servers, ICA and RDP, we ran Windows 2003 Server on the server. For the other systems which support X-based servers, VNC and pTHINC, we ran the Debian Linux Unstable distribution with the Linux 2.6.10 kernel on the server. We used the latest thin-client server versions available on each platform at the time of our experiments, namely Citrix MetaFrame XP Server for Windows Feature Release 3, Microsoft Remote Desktop built into Windows XP and Windows 2003 using RDP 5.2, and VNC 4.0. 4.2 Application Benchmarks We used two web application benchmarks for our experiments based on two common application scenarios, browsing web pages and playing video content from the web. Since many thin-client systems including two of the ones tested are closed and proprietary, we measured their performance in a noninvasive manner by capturing network traffic with a packet monitor and using a variant of slow-motion benchmarking [13] previously developed to measure thin-client performance in PDA environments [10]. This measurement methodology accounts for both the display decoupling that can occur between client and server in thin-client systems as well as client processing time, which may be significant in the case of PDAs. To measure web browsing performance, we used a web browsing benchmark based on the Web Text Page Load Test from the Ziff-Davis i-Bench benchmark suite [7]. The benchmark consists of JavaScript controlled load of 55 pages from the web server. The pages contain both text and graphics with pages varying in size. The graphics are embedded images in GIF and JPEG formats. The original i-Bench benchmark was modified for slow-motion benchmarking by introducing delays of several seconds between the pages using JavaScript. Then two tests were run, one where delays where added between each page, and one where pages where loaded continuously without waiting for them to be displayed on the client. In the first test, delays were sufficiently adjusted in each case to ensure that each page could be received and displayed on the client completely without temporal overlap in transferring the data belonging to two consecutive pages. We used the packet monitor to record the packet traffic for each run of the benchmark, then used the timestamps of the first and last packet in the trace to obtain our latency measures [10]. The packet monitor also recorded the amount of data transmitted between the client and the server. The ratio between the data traffic in the two tests yields a scale factor. This scale factor shows the loss of data between the server and the client due to inability of the client to process the data quickly enough. The product of the scale factor with the latency measurement produces the true latency accounting for client processing time. To run the web browsing benchmark, we used Mozilla Firefox 1.0.4 running on the thin-client server for the thin clients, and Windows Internet Explorer (IE) Mobile for 2003 and Mobile for 5.0 for the native browsers on the X5 and X51v PDAs, respectively. In all cases, the web browser used was sized to fill the entire display region available. To measure video playback performance, we used a video benchmark that consisted of playing a 34.75s MPEG-1 video clip containing a mix of news and entertainment programming at full-screen resolution. The video clip is 5.11 MB and consists of 834 352x240 pixel frames with an ideal frame rate of 24 frames/sec. We measured video performance using slow-motion benchmarking by monitoring resulting packet traffic at two playback rates, 1 frames/second (fps) and 24 fps, and comparing the results to determine playback delays and frame drops that occur at 24 fps to measure overall video quality [13]. For example, 100% quality means that all video frames were played at real-time speed. On the other hand, 50% quality could mean that half the video data was dropped, or that the clip took twice as long to play even though all of the video data was displayed. To run the video benchmark, we used Windows Media Player 9 for Windows-based thin-client servers, MPlayer 1.0 pre 6 for X-based thin-client servers, and Windows Media Player 9 Mobile and 10 Mobile for the native video players running locally on the X5 and X51v PDAs, respectively. In all cases, the video player used was sized to fill the entire display region available. 4.3 Measurements Figures 3 and 4 show the results of running the web brows148 0 1 10 100 pTHINC Resized pTHINCICA Resized ICAVNCRDPLOCAL Latency(s) Platform Axim X5 (640x480 or less) Axim X51v (640x480) Axim X5 (1024x768) Axim X51v (1024x768) Figure 3: Browsing Benchmark: Average Page Latency ing benchmark. For each platform, we show results for up to four different configurations, two on the X5 and two on the X51v, depending on whether each configuration was supported. However, not all platforms could support all configurations. The local browser only runs at the display resolution of the PDA, 480×680 or less for the X51v and the X5. RDP only runs at 640×480. Neither platform could support 1024×768 display resolution. ICA only ran on the X5 and could not run on the X51v because it did not work on Windows Mobile 5. Figure 3 shows the average latency per web page for each platform. pTHINC provides the lowest average web browsing latency on both PDAs. On the X5, pTHINC performs up to 70 times better than other thin-client systems and 8 times better than the local browser. On the X51v, pTHINC performs up to 80 times better than other thin-client systems and 7 times better than the native browser. In fact, all of the thin clients except VNC outperform the local PDA browser, demonstrating the performance benefits of the thin-client approach. Usability studies have shown that web pages should take less than one second to download for the user to experience an uninterrupted web browsing experience [14]. The measurements show that only the thin clients deliver subsecond web page latencies. In contrast, the local browser requires more than 3 seconds on average per web page. The local browser performs worse since it needs to run a more limited web browser to process the HTML, JavaScript, and do all the rendering using the limited capabilities of the PDA. The thin clients can take advantage of faster server hardware and a highly tuned web browser to process the web content much faster. Figure 3 shows that RDP is the next fastest platform after pTHINC. However, RDP is only able to run at a fixed resolution of 640×480 and 8-bit color depth. Furthermore, RDP also clips the display to the size of the PDA screen so that it does not need to send updates that are not visible on the PDA screen. This provides a performance benefit assuming the remaining web content is not viewed, but degrades performance when a user scrolls around the display to view other web content. RDP achieves its performance with significantly lower display quality compared to the other thin clients and with additional display clipping not used by other systems. As a result, RDP performance alone does not provide a complete comparison with the other platforms. In contrast, pTHINC provides the fastest performance while at the same time providing equal or better display quality than the other systems. 0 1 10 100 1000 pTHINC Resized pTHINCICA Resized ICAVNCRDPLOCAL DataSize(KB) Platform Axim X5 (640x480 or less) Axim X51v (640x480) Axim X5 (1024x768) Axim X51v (1024x768) Figure 4: Browsing Benchmark: Average Page Data Transferred Since VNC and ICA provide similar display quality to pTHINC, these systems provide a more fair comparison of different thin-client approaches. ICA performs worse in part because it uses higher-level display primitives that require additional client processing costs. VNC performs worse in part because it loses display data due to its client-pull delivery mechanism and because of the client processing costs in decompressing raw pixel primitives. In both cases, their performance was limited in part because their PDA clients were unable to keep up with the rate at which web pages were being displayed. Figure 3 also shows measurements for those thin clients that support resizing the display to fit the PDA screen, namely ICA and pTHINC. Resizing requires additional processing, which results in slower average web page latencies. The measurements show that the additional delay incurred by ICA when resizing versus not resizing is much more substantial than for pTHINC. ICA performs resizing on the slower PDA client. In contrast, pTHINC leverage the more powerful server to do resizing, reducing the performance difference between resizing and not resizing. Unlike ICA, pTHINC is able to provide subsecond web page download latencies in both cases. Figure 4 shows the data transferred in KB per page when running the slow-motion version of the tests. All of the platforms have modest data transfer requirements of roughly 100 KB per page or less. This is well within the bandwidth capacity of Wi-Fi networks. The measurements show that the local browser does not transfer the least amount of data. This is surprising as HTML is often considered to be a very compact representation of content. Instead, RDP is the most bandwidth efficient platform, largely as a result of using only 8-bit color depth and screen clipping so that it does not transfer the entire web page to the client. pTHINC overall has the largest data requirements, slightly more than VNC. This is largely a result of the current pTHINC prototype"s lack of native support for 16-bit color data in the wire protocol. However, this result also highlights pTHINC"s performance as it is faster than all other systems even while transferring more data. Furthermore, as newer PDA models support full 24-bit color, these results indicate that pTHINC will continue to provide good web browsing performance. Since display usability and quality are as important as performance, Figures 5 to 8 compare screenshots of the different thin clients when displaying a web page, in this case from the popular BBC news website. Except for ICA, all of the screenshots were taken on the X51v in landscape mode 149 Figure 5: Browser Screenshot: RDP 640x480 Figure 6: Browser Screenshot: VNC 1024x768 Figure 7: Browser Screenshot: ICA Resized 1024x768 Figure 8: Browser Screenshot: pTHINC Resized 1024x768 using the maximum display resolution settings for each platform given in Table 2. The ICA screenshot was taken on the X5 since ICA does not run on the X51v. While the screenshots lack the visual fidelity of the actual device display, several observations can be made. Figure 5 shows that RDP does not support fullscreen mode and wastes lots of screen space for controls and UI elements, requiring the user to scroll around in order to access the full contents of the web browsing session. Figure 6 shows that VNC makes better use of the screen space and provides better display quality, but still forces the user to scroll around to view the web page due to its lack of resizing support. Figure 7 shows ICA"s ability to display the full web page given its resizing support, but that its lack of landscape capability and poorer resize algorithm significantly compromise display quality. In contrast, Figure 8 shows pTHINC using resizing to provide a high quality fullscreen display of the full width of the web page. pTHINC maximizes the entire viewing region by moving all controls to the PDA buttons. In addition, pTHINC leverages the server computational power to use a high quality resizing algorithm to resize the display to fit the PDA screen without significant overhead. Figures 9 and 10 show the results of running the video playback benchmark. For each platform except ICA, we show results for an X5 and X51v configuration. ICA could not run on the X51v as noted earlier. The measurements were done using settings that reflected the environment a user would have to access a web session from both a desktop computer and a PDA. As such, a 1024×768 server display resolution was used whenever possible and the video was shown at fullscreen. RDP was limited to 640×480 display resolution as noted earlier. Since viewing the entire video display is the only really usable option, we resized the display to fit the PDA screen for those platforms that supported this feature, namely ICA and pTHINC. Figure 9 shows the video quality for each platform. pTHINC is the only thin client able to provide perfect video playback quality, similar to the native PDA video player. All of the other thin clients deliver very poor video quality. With the exception of RDP on the X51v which provided unacceptable 35% video quality, none of the other systems were even able to achieve 10% video quality. VNC and ICA have the worst quality at 8% on the X5 device. pTHINC"s native video support enables superior video performance, while other thin clients suffer from their inability to distinguish video from normal display updates. They attempt to apply ineffective and expensive compression algorithms on the video data and are unable to keep up with the stream of updates generated, resulting in dropped frames or long playback times. VNC suffers further from its client-pull update model because video frames are generated faster than the rate at which the client can process and send requests to the server to obtain the next display update. Figure 10 shows the total data transferred during 150 0% 20% 40% 60% 80% 100% pTHINCICAVNCRDPLOCAL VideoQuality Platform Axim X5 Axim X51v Figure 9: Video Benchmark: Fullscreen Video Quality 0 1 10 100 pTHINCICAVNCRDPLOCAL VideoDataSize(MB) Platform Axim X5 Axim X51v Figure 10: Video Benchmark: Fullscreen Video Data video playback for each system. The native player is the most bandwidth efficient platform, sending less than 6 MB of data, which corresponds to about 1.2 Mbps of bandwidth. pTHINC"s 100% video quality requires about 25 MB of data which corresponds to a bandwidth usage of less than 6 Mbps. While the other thin clients send less data than THINC, they do so because they are dropping video data, resulting in degraded video quality. Figures 11 to 14 compare screenshots of the different thin clients when displaying the video clip. Except for ICA, all of the screenshots were taken on the X51v in landscape mode using the maximum display resolution settings for each platform given in Table 2. The ICA screenshot was taken on the X5 since ICA does not run on the X51v. Figures 11 and 12 show that RDP and VNC are unable to display the entire video frame on the PDA screen. RDP wastes screen space for UI elements and VNC only shows the top corner of the video frame on the screen. Figure 13 shows that ICA provides resizing to display the entire video frame, but did not proportionally resize the video data, resulting in strange display artifacts. In contrast, Figure 14 shows pTHINC using resizing to provide a high quality fullscreen display of the entire video frame. pTHINC provides visually more appealing video display than RDP, VNC, or ICA. 5. RELATED WORK Several studies have examined the web browsing performance of thin-client computing [13, 19, 10]. The ability for thin clients to improve web browsing performance on wireless PDAs was first quantitatively demonstrated in a previous study by one of the authors [10]. This study demonstrated that thin clients can provide both faster web browsing performance and greater web browsing functionality. The study considered a wide range of web content including content from medical information systems. Our work builds on this previous study and consider important issues such as how usable existing thin clients are in PDA environments, the trade-offs between thin-client usability and performance, performance across different PDA devices, and the performance of thin clients on common web-related applications such as video. Many thin clients have been developed and some have PDA clients, including Microsoft"s Remote Desktop [3], Citrix MetraFrame XP [2], Virtual Network Computing [16, 12], GoToMyPC [5], and Tarantella [18]. These systems were first designed for desktop computing and retrofitted for PDAs. Unlike pTHINC, they do not address key system architecture and usability issues important for PDAs. This limits their display quality, system performance, available screen space, and overall usability on PDAs. pTHINC builds on previous work by two of the authors on THINC [1], extending the server architecture and introducing a client interface and usage model to efficiently support PDA devices for mobile web applications. Other approaches to improve the performance of mobile wireless web browsing have focused on using transcoding and caching proxies in conjunction with the fat client model [11, 9, 4, 8]. They work by pushing functionality to external proxies, and using specialized browsing applications on the PDA device that communicate with the proxy. Our thinclient approach differs fundamentally from these fat-client approaches by pushing all web browser logic to the server, leveraging existing investments in desktop web browsers and helper applications to work seamlessly with production systems without any additional proxy configuration or web browser modifications. With the emergence of web browsing on small display devices, web sites have been redesigned using mechanisms like WAP and specialized native web browsers have been developed to tailor the needs of these devices. Recently, Opera has developed the Opera Mini [15] web browser, which uses an approach similar to the thin-client model to provide access across a number of mobile devices that would normally be incapable of running a web browser. Instead of requiring the device to process web pages, it uses a remote server to pre-process the page before sending it to the phone. 6. CONCLUSIONS We have introduced pTHINC, a thin-client architecture for wireless PDAs. pTHINC provides key architectural and usability mechanisms such as server-side screen resizing, clientside screen rotation using simple copy techniques, YUV video support, and maximizing screen space for display updates and leveraging existing PDA control buttons for UI elements. pTHINC transparently supports traditional desktop browsers and their helper applications on PDA devices and desktop machines, providing mobile users with ubiquitous access to a consistent, personalized, and full-featured web environment across heterogeneous devices. We have implemented pTHINC and measured its performance on web applications compared to existing thin-client systems and native web applications. Our results on multiple mobile wireless devices demonstrate that pTHINC delivers web browsing performance up to 80 times better than existing thin-client systems, and 8 times better than a native PDA browser. In addition, pTHINC is the only PDA thin client 151 Figure 11: Video Screenshot: RDP 640x480 Figure 12: Video Screenshot: VNC 1024x768 Figure 13: Video Screenshot: ICA Resized 1024x768 Figure 14: Video Screenshot: pTHINC Resized 1024x768 that transparently provides full-screen, full frame rate video playback, making web sites with multimedia content accessible to mobile web users. 7. ACKNOWLEDGEMENTS This work was supported in part by NSF ITR grants CCR0219943 and CNS-0426623, and an IBM SUR Award. 8. REFERENCES [1] R. Baratto, L. Kim, and J. Nieh. THINC: A Virtual Display Architecture for Thin-Client Computing. In Proceedings of the 20th ACM Symposium on Operating Systems Principles (SOSP), Oct. 2005. [2] Citrix Metaframe. http://www.citrix.com. [3] B. C. Cumberland, G. Carius, and A. Muir. Microsoft Windows NT Server 4.0, Terminal Server Edition: Technical Reference. Microsoft Press, Redmond, WA, 1999. [4] A. Fox, I. Goldberg, S. D. Gribble, and D. C. Lee. Experience With Top Gun Wingman: A Proxy-Based Graphical Web Browser for the 3Com PalmPilot. In Proceedings of Middleware "98, Lake District, England, September 1998, 1998. [5] GoToMyPC. http://www.gotomypc.com/. [6] Health Insurance Portability and Accountability Act. http://www.hhs.gov/ocr/hipaa/. [7] i-Bench version 1.5. http: //etestinglabs.com/benchmarks/i-bench/i-bench.asp. [8] A. Joshi. On proxy agents, mobility, and web access. Mobile Networks and Applications, 5(4):233-241, 2000. [9] J. Kangasharju, Y. G. Kwon, and A. Ortega. Design and Implementation of a Soft Caching Proxy. Computer Networks and ISDN Systems, 30(22-23):2113-2121, 1998. [10] A. Lai, J. Nieh, B. Bohra, V. Nandikonda, A. P. Surana, and S. Varshneya. Improving Web Browsing on Wireless PDAs Using Thin-Client Computing. In Proceedings of the 13th International World Wide Web Conference (WWW), May 2004. [11] A. Maheshwari, A. Sharma, K. Ramamritham, and P. Shenoy. TranSquid: Transcoding and caching proxy for heterogenous ecommerce environments. In Proceedings of the 12th IEEE Workshop on Research Issues in Data Engineering (RIDE "02), Feb. 2002. [12] .NET VNC Viewer for PocketPC. http://dotnetvnc.sourceforge.net/. [13] J. Nieh, S. J. Yang, and N. Novik. Measuring Thin-Client Performance Using Slow-Motion Benchmarking. ACM Trans. Computer Systems, 21(1):87-115, Feb. 2003. [14] J. Nielsen. Designing Web Usability. New Riders Publishing, Indianapolis, IN, 2000. [15] Opera Mini Browser. http://www.opera.com/products/mobile/operamini/. [16] T. Richardson, Q. Stafford-Fraser, K. R. Wood, and A. Hopper. Virtual Network Computing. IEEE Internet Computing, 2(1), Jan./Feb. 1998. [17] R. W. Scheifler and J. Gettys. The X Window System. ACM Trans. Gr., 5(2):79-106, Apr. 1986. [18] Sun Secure Global Desktop. http://www.sun.com/software/products/sgd/. [19] S. J. Yang, J. Nieh, S. Krishnappa, A. Mohla, and M. Sajjadpour. Web Browsing Performance of Wireless Thin-Client Computing. In Proceedings of the 12th International World Wide Web Conference (WWW), May 2003. 152

pervasive web;remote display;pda thinclient solution;system usability;web browser;thin-client computing;mobility;full-function web browser;pthinc;seamless mobility;functionality;thin-client;local pda web browser;web application;video playback;screen resolution;web browsing performance;high-fidelity display;mobile wireless pda;crucial browser helper application

train_C-71

A Point-Distribution Index and Its Application to Sensor-Grouping in Wireless Sensor Networks

We propose ι, a novel index for evaluation of point-distribution. ι is the minimum distance between each pair of points normalized by the average distance between each pair of points. We find that a set of points that achieve a maximum value of ι result in a honeycomb structure. We propose that ι can serve as a good index to evaluate the distribution of the points, which can be employed in coverage-related problems in wireless sensor networks (WSNs). To validate this idea, we formulate a general sensorgrouping problem for WSNs and provide a general sensing model. We show that locally maximizing ι at sensor nodes is a good approach to solve this problem with an algorithm called Maximizingι Node-Deduction (MIND). Simulation results verify that MIND outperforms a greedy algorithm that exploits sensor-redundancy we design. This demonstrates a good application of employing ι in coverage-related problems for WSNs.

1. INTRODUCTION A wireless sensor network (WSN) consists of a large number of in-situ battery-powered sensor nodes. A WSN can collect the data about physical phenomena of interest [1]. There are many potential applications of WSNs, including environmental monitoring and surveillance, etc. [1][11]. In many application scenarios, WSNs are employed to conduct surveillance tasks in adverse, or even worse, in hostile working environments. One major problem caused is that sensor nodes are subjected to failures. Therefore, fault tolerance of a WSN is critical. One way to achieve fault tolerance is that a WSN should contain a large number of redundant nodes in order to tolerate node failures. It is vital to provide a mechanism that redundant nodes can be working in sleeping mode (i.e., major power-consuming units such as the transceiver of a redundant sensor node can be shut off) to save energy, and thus to prolong the network lifetime. Redundancy should be exploited as much as possible for the set of sensors that are currently taking charge in the surveillance work of the network area [6]. We find that the minimum distance between each pair of points normalized by the average distance between each pair of points serves as a good index to evaluate the distribution of the points. We call this index, denoted by ι, the normalized minimum distance. If points are moveable, we find that maximizing ι results in a honeycomb structure. The honeycomb structure poses that the coverage efficiency is the best if each point represents a sensor node that is providing surveillance work. Employing ι in coverage-related problems is thus deemed promising. This enlightens us that maximizing ι is a good approach to select a set of sensors that are currently taking charge in the surveillance work of the network area. To explore the effectiveness of employing ι in coverage-related problems, we formulate a sensorgrouping problem for high-redundancy WSNs. An algorithm called Maximizing-ι Node-Deduction (MIND) is proposed in which redundant sensor nodes are removed to obtain a large ι for each set of sensors that are currently taking charge in the surveillance work of the network area. We also introduce another greedy solution called Incremental Coverage Quality Algorithm (ICQA) for this problem, which serves as a benchmark to evaluate MIND. The main contribution of this paper is twofold. First, we introduce a novel index ι for evaluation of point-distribution. We show that maximizing ι of a WSN results in low redundancy of the network. Second, we formulate a general sensor-grouping problem for WSNs and provide a general sensing model. With the MIND algorithm we show that locally maximizing ι among each sensor node and its neighbors is a good approach to solve this problem. This demonstrates a good application of employing ι in coveragerelated problems. The rest of the paper is organized as follows. In Section 2, we introduce our point-distribution index ι. We survey related work and formulate a sensor-grouping problem together with a general sensing model in Section 3. Section 4 investigates the application of ι in this grouping problem. We propose MIND for this problem 1171 and introduce ICQA as a benchmark. In Section 5, we present our simulation results in which MIND and ICQA are compared. Section 6 provides conclusion remarks. 2. THE NORMALIZED MINIMUM DISTANCE ι: A POINT-DISTRIBUTION INDEX Suppose there are n points in a Euclidean space Ω. The coordinates of these points are denoted by xi (i = 1, ..., n). It may be necessary to evaluate how the distribution of these points is. There are many metrics to achieve this goal. For example, the Mean Square Error from these points to their mean value can be employed to calculate how these points deviate from their mean (i.e., their central). In resource-sharing evaluation, the Global Fairness Index (GFI) is often employed to measure how even the resource distributes among these points [8], when xi represents the amount of resource that belong to point i. In WSNs, GFI is usually used to calculate how even the remaining energy of sensor nodes is. When n is larger than 2 and the points do not all overlap (That points all overlap means xi = xj, ∀ i, j = 1, 2, ..., n). We propose a novel index called the normalized minimum distance, namely ι, to evaluate the distribution of the points. ι is the minimum distance between each pair of points normalized by the average distance between each pair of points. It is calculated by: ι = min(||xi − xj||) µ (∀ i, j = 1, 2, ..., n; and i = j) (1) where ||xi − xj|| denotes the Euclidean distance between point i and point j in Ω, the min(·) function calculates the minimum distance between each pair of points, and µ is the average distance between each pair of points, which is: µ = ( Pn i=1 Pn j=1,j=i ||xi − xj||) n(n − 1) (2) ι measures how well the points separate from one another. Obviously, ι is in interval [0, 1]. ι is equal to 1 if and only if n is equal to 3 and these three points forms an equilateral triangle. ι is equal to zero if any two points overlap. ι is a very interesting value of a set of points. If we consider each xi (∀i = 1, ..., n) is a variable in Ω, how these n points would look like if ι is maximized? An algorithm is implemented to generate the topology in which ι is locally maximized (The algorithm can be found in [19]). We consider a 2-dimensional space. We select n = 10, 20, 30, ..., 100 and perform this algorithm. In order to avoid that the algorithm converge to local optimum, we select different random seeds to generate the initial points for 1000 time and obtain the best one that results in the largest ι when the algorithm converges. Figure 1 demonstrates what the resulting topology looks like when n = 20 as an example. Suppose each point represents a sensor node. If the sensor coverage model is the Boolean coverage model [15][17][18][14] and the coverage radius of each node is the same. It is exciting to see that this topology results in lowest redundancy because the Vonoroi diagram [2] formed by these nodes (A Vonoroi diagram formed by a set of nodes partitions a space into a set of convex polygons such that points inside a polygon are closest to only one particular node) is a honeycomb-like structure1 . This enlightens us that ι may be employed to solve problems related to sensor-coverage of an area. In WSNs, it is desirable 1 This is how base stations of a wireless cellular network are deployed and why such a network is called a cellular one. 0 20 40 60 80 100 120 140 160 0 20 40 60 80 100 120 140 160 X Y Figure 1: Node Number = 20, ι = 0.435376 that the active sensor nodes that are performing surveillance task should separate from one another. Under the constraint that the sensing area should be covered, the more each node separates from the others, the less the redundancy of the coverage is. ι indicates the quality of such separation. It should be useful for approaches on sensor-coverage related problems. In our following discussions, we will show the effectiveness of employing ι in sensor-grouping problem. 3. THE SENSOR-GROUPING PROBLEM In many application scenarios, to achieve fault tolerance, a WSN contains a large number of redundant nodes in order to tolerate node failures. A node sleeping-working schedule scheme is therefore highly desired to exploit the redundancy of working sensors and let as many nodes as possible sleep. Much work in the literature is on this issue [6]. Yan et al introduced a differentiated service in which a sensor node finds out its responsible working duration with cooperation of its neighbors to ensure the coverage of sampled points [17]. Ye et al developed PEAS in which sensor nodes wake up randomly over time, probe their neighboring nodes, and decide whether they should begin to take charge of surveillance work [18]. Xing et al exploited a probabilistic distributed detection model with a protocol called Coordinating Grid (Co-Grid) [16]. Wang et al designed an approach called Coverage Configuration Protocol (CCP) which introduced the notion that the coverage degree of intersection-points of the neighboring nodes" sensing-perimeters indicates the coverage of a convex region [15]. In our recent work [7], we also provided a sleeping configuration protocol, namely SSCP, in which sleeping eligibility of a sensor node is determined by a local Voronoi diagram. SSCP can provide different levels of redundancy to maintain different requirements of fault tolerance. The major feature of the aforementioned protocols is that they employ online distributed and localized algorithms in which a sensor node determines its sleeping eligibility and/or sleeping time based on the coverage requirement of its sensing area with some information provided by its neighbors. Another major approach for sensor node sleeping-working scheduling issue is to group sensor nodes. Sensor nodes in a network are divided into several disjoint sets. Each set of sensor nodes are able to maintain the required area surveillance work. The sensor nodes are scheduled according to which set they belong to. These sets work successively. Only one set of sensor nodes work at any time. We call the issue sensor-grouping problem. The major advantage of this approach is that it avoids the overhead caused by the processes of coordination of sensor nodes to make decision on whether a sensor node is a candidate to sleep or 1172 work and how long it should sleep or work. Such processes should be performed from time to time during the lifetime of a network in many online distributed and localized algorithms. The large overhead caused by such processes is the main drawback of the online distributed and localized algorithms. On the contrary, roughly speaking, this approach groups sensor nodes in one time and schedules when each set of sensor nodes should be on duty. It does not require frequent decision-making on working/sleeping eligibility2 . In [13] by Slijepcevic et al, the sensing area is divided into regions. Sensor nodes are grouped with the most-constrained leastconstraining algorithm. It is a greedy algorithm in which the priority of selecting a given sensor is determined by how many uncovered regions this sensor covers and the redundancy caused by this sensor. In [5] by Cardei et al, disjoint sensor sets are modeled as disjoint dominating sets. Although maximum dominating sets computation is NP-complete, the authors proposed a graphcoloring based algorithm. Cardei et al also studied similar problem in the domain of covering target points in [4]. The NP-completeness of the problem is proved and a heuristic that computes the sets are proposed. These algorithms are centralized solutions of sensorgrouping problem. However, global information (e.g., the location of each in-network sensor node) of a large scale WSN is also very expensive to obtained online. Also it is usually infeasible to obtain such information before sensor nodes are deployed. For example, sensor nodes are usually deployed in a random manner and the location of each in-network sensor node is determined only after a node is deployed. The solution of sensor-grouping problem should only base on locally obtainable information of a sensor node. That is to say, nodes should determine which group they should join in a fully distributed way. Here locally obtainable information refers to a node"s local information and the information that can be directly obtained from its adjacent nodes, i.e., nodes within its communication range. In Subsection 3.1, we provide a general problem formulation of the sensor-grouping problem. Distributed-solution requirement is formulated in this problem. It is followed by discussion in Subsection 3.2 on a general sensing model, which serves as a given condition of the sensor-grouping problem formulation. To facilitate our discussions, the notations in our following discussions are described as follows. • n: The number in-network sensor nodes. • S(j) (j = 1, 2, ..., m): The jth set of sensor nodes where m is the number of sets. • L(i) (i = 1, 2, ..., n): The physical location of node i. • φ: The area monitored by the network: i.e., the sensing area of the network. • R: The sensing radius of a sensor node. We assume that a sensor node can only be responsible to monitor a circular area centered at the node with a radius equal to R. This is a usual assumption in work that addresses sensor-coverage related problems. We call this circular area the sensing area of a node. 3.1 Problem Formulation We assume that each sensor node can know its approximate physical location. The approximate location information is obtainable if each sensor node carries a GPS receiver or if some localization algorithms are employed (e.g., [3]). 2 Note that if some nodes die, a re-grouping process might also be performed to exploit the remaining nodes in a set of sensor nodes. How to provide this mechanism is beyond the scope of this paper and yet to be explored. Problem 1. Given: • The set of each sensor node i"s sensing neighbors N(i) and the location of each member in N(i); • A sensing model which quantitatively describes how a point P in area φ is covered by sensor nodes that are responsible to monitor this point. We call this quantity the coverage quality of P. • The coverage quality requirement in φ, denoted by s. When the coverage of a point is larger than this threshold, we say this point is covered. For each sensor node i, make a decision on which group S(j) it should join so that: • Area φ can be covered by sensor nodes in each set S(j) • m, the number of sets S(j) is maximized. In this formulation, we call sensor nodes within a circular area centered at a sensor node i with a radius equal to 2 · R the sensing neighbors of node i. This is because sensors nodes in this area, together with node i, may be cooperative to ensure the coverage of a point inside node i"s sensing area. We assume that the communication range of a sensor node is larger than 2 · R, which is also a general assumption in work that addresses sensor-coverage related problems. That is to say, the first given condition in Problem 1 is the information that can be obtained directly from a node"s adjacent nodes. It is therefore locally obtainable information. The last two given conditions in this problem formulation can be programmed into a node before it is deployed or by a node-programming protocol (e.g., [9]) during network runtime. Therefore, the given conditions can all be easily obtained by a sensor-grouping scheme with fully distributed implementation. We reify this problem with a realistic sensing model in next subsection. 3.2 A General Sensing Model As WSNs are usually employed to monitor possible events in a given area, it is therefore a design requirement that an event occurring in the network area must/may be successfully detected by sensors. This issue is usually formulated as how to ensure that an event signal omitted in an arbitrary point in the network area can be detected by sensor nodes. Obviously, a sensing model is required to address this problem so that how a point in the network area is covered can be modeled and quantified. Thus the coverage quality of a WSN can be evaluated. Different applications of WSNs employ different types of sensors, which surely have widely different theoretical and physical characteristics. Therefore, to fulfill different application requirements, different sensing models should be constructed based on the characteristics of the sensors employed. A simple theoretical sensing model is the Boolean sensing model [15][18][17][14]. Boolean sensing model assumes that a sensor node can always detect an event occurring in its responsible sensing area. But most sensors detect events according to the signal strength sensed. Event signals usually fade in relation to the physical distance between an event and the sensor. The larger the distance, the weaker the event signals that can be sensed by the sensor, which results in a reduction of the probability that the event can be successfully detected by the sensor. As in WSNs, event signals are usually electromagnetic, acoustic, or photic signals, they fade exponentially with the increasing of 1173 their transmit distance. Specifically, the signal strength E(d) of an event that is received by a sensor node satisfies: E(d) = α dβ (3) where d is the physical distance from the event to the sensor node; α is related to the signal strength omitted by the event; and β is signal fading factor which is typically a positive number larger than or equal to 2. Usually, α and β are considered as constants. Based on this notion, to be more reasonable, researchers propose collaborative sensing model to capture application requirements: Area coverage can be maintained by a set of collaborative sensor nodes: For a point with physical location L, the point is considered covered by the collaboration of i sensors (denoted by k1, ..., ki) if and only if the following two equations holds [7][10][12]. ∀j = 1, ..., i; L(kj) − L < R. (4) C(L) = iX j=1 (E( L(kj) − L ) > s. (5) C(L) is regarded as the coverage quality of location L in the network area [7][10][12]. However, we notice that defining the sensibility as the sum of the sensed signal strength by each collaborative sensor implies a very special application: Applications must employ the sum of the signal strength to achieve decision-making. To capture generally realistic application requirement, we modify the definition described in Equation (5). The model we adopt in this paper is described in details as follows. We consider the probability P(L, kj ) that an event located at L can be detected by sensor kj is related to the signal strength sensed by kj. Formally, P(L, kj) = γE(d) = δ ( L(kj) − L / + 1)β , (6) where γ is a constant and δ = γα is a constant too. normalizes the distance to a proper scale and the +1 item is to avoid infinite value of P(L, kj). The probability that an event located at L can be detected by any collaborative sensors that satisfied Equation (4) is: P (L) = 1 − iY j=1 (1 − P(L, kj )). (7) As the detection probability P (L) reasonably determines how an event occurring at location L can be detected by the networks, it is a good measure of the coverage quality of location L in a WSN. Specifically, Equation (5) is modified to: C(L) = P (L) = 1 − iY j=1 [1 − δ ( L(kj) − L / + 1)β ] > s. (8) To sum it up, we consider a point at location L is covered if Equation (4) and (8) hold. 4. MAXIMIZING-ι NODE-DEDUCTION ALGORITHM FOR SENSOR-GROUPING PROBLEM Before we process to introduce algorithms to solve the sensor grouping problem, let us define the margin (denoted by θ) of an area φ monitored by the network as the band-like marginal area of φ and all the points on the outer perimeter of θ is ρ distance away from all the points on the inner perimeter of θ. ρ is called the margin length. In a practical network, sensor nodes are usually evenly deployed in the network area. Obviously, the number of sensor nodes that can sense an event occurring in the margin of the network is smaller than the number of sensor nodes that can sense an event occurring in other area of the network. Based on this consideration, in our algorithm design, we ensure the coverage quality of the network area except the margin. The information on φ and ρ is networkbased. Each in-network sensor node can be pre-programmed or on-line informed about φ and ρ, and thus calculate whether a point in its sensing area is in the margin or not. 4.1 Maximizing-ι Node-Deduction Algorithm The node-deduction process of our Maximizing-ι Node-Deduction Algorithm (MIND) is simple. A node i greedily maximizes ι of the sub-network composed by itself, its ungrouped sensing neighbors, and the neighbors that are in the same group of itself. Under the constraint that the coverage quality of its sensing area should be ensured, node i deletes nodes in this sub-network one by one. The candidate to be pruned satisfies that: • It is an ungrouped node. • The deletion of the node will not result in uncovered-points inside the sensing area of i. A candidate is deleted if the deletion of the candidate results in largest ι of the sub-network compared to the deletion of other candidates. This node-deduction process continues until no candidate can be found. Then all the ungrouped sensing neighbors that are not deleted are grouped into the same group of node i. We call the sensing neighbors that are in the same group of node i the group sensing neighbors of node i. We then call node i a finished node, meaning that it has finished the above procedure and the sensing area of the node is covered. Those nodes that have not yet finished this procedure are called unfinished nodes. The above procedure initiates at a random-selected node that is not in the margin. The node is grouped to the first group. It calculates the resulting group sensing neighbors of it based on the above procedure. It informs these group sensing neighbors that they are selected in the group. Then it hands over the above procedure to an unfinished group sensing neighbors that is the farthest from itself. This group sensing neighbor continues this procedure until no unfinished neighbor can be found. Then the first group is formed (Algorithmic description of this procedure can be found at [19]). After a group is formed, another random-selected ungrouped node begins to group itself to the second group and initiates the above procedure. In this way, groups are formed one by one. When a node that involves in this algorithm found out that the coverage quality if its sensing area, except what overlaps the network margin, cannot be ensured even if all its ungrouped sensing neighbors are grouped into the same group as itself, the algorithm stops. MIND is based on locally obtainable information of sensor nodes. It is a distributed algorithm that serves as an approximate solution of Problem 1. 4.2 Incremental Coverage Quality Algorithm: A Benchmark for MIND To evaluate the effectiveness of introducing ι in the sensor-group problem, another algorithm for sensor-group problem called Incremental Coverage Quality Algorithm (ICQA) is designed. Our aim 1174 is to evaluate how an idea, i.e., MIND, based on locally maximize ι performs. In ICQA, a node-selecting process is as follows. A node i greedily selects an ungrouped sensing neighbor in the same group as itself one by one, and informs the neighbor it is selected in the group. The criterion is: • The selected neighbor is responsible to provide surveillance work for some uncovered parts of node i"s sensing area. (i.e., the coverage quality requirement of the parts is not fulfilled if this neighbor is not selected.) • The selected neighbor results in highest improvement of the coverage quality of the neighbor"s sensing area. The improvement of the coverage quality, mathematically, should be the integral of the the improvements of all points inside the neighbor"s sensing area. A numerical approximation is employed to calculate this improvement. Details are presented in our simulation study. This node-selecting process continues until the sensing area of node i is entirely covered. In this way, node i"s group sensing neighbors are found. The above procedure is handed over as what MIND does and new groups are thus formed one by one. And the condition that ICQA stops is the same as MIND. ICQA is also based on locally obtainable information of sensor nodes. ICQA is also a distributed algorithm that serves as an approximate solution of Problem 1. 5. SIMULATION RESULTS To evaluate the effectiveness of employing ι in sensor-grouping problem, we build simulation surveillance networks. We employ MIND and ICQA to group the in-network sensor nodes. We compare the grouping results with respect to how many groups both algorithms find and how the performance of the resulting groups are. Detailed settings of the simulation networks are shown in Table 1. In simulation networks, sensor nodes are randomly deployed in a uniform manner in the network area. Table 1: The settings of the simulation networks Area of sensor field 400m*400m ρ 20m R 80m α, β, γ and 1.0, 2.0, 1.0 and 100.0 s 0.6 For evaluating the coverage quality of the sensing area of a node, we divide the sensing area of a node into several regions and regard the coverage quality of the central point in each region as a representative of the coverage quality of the region. This is a numerical approximation. Larger number of such regions results in better approximation. As sensor nodes are with low computational capacity, there is a tradeoff between the number of such regions and the precision of the resulting coverage quality of the sensing area of a node. In our simulation study, we set this number 12. For evaluating the improvement of coverage quality in ICQA, we sum up all the improvements at each region-center as the total improvement. 5.1 Number of Groups Formed by MIND and ICQA We set the total in-network node number to different values and let the networks perform MIND and ICQA. For each n, simulations run with several random seeds to generate different networks. Results are averaged. Figure 2 shows the group numbers found in networks with different n"s. 500 1000 1500 2000 0 5 10 15 20 25 30 35 40 45 50 Total in−network node number Totalnumberofgroupsfound ICQA MMNP Figure 2: The number of groups found by MIND and ICQA We can see that MIND always outperforms ICQA in terms of the number of groups formed. Obviously, the larger the number of groups can be formed, the more the redundancy of each group is exploited. This output shows that an approach like MIND that aim to maximize ι of the resulting topology can exploits redundancy well. As an example, in case that n = 1500, the results of five networks are listed in Table 2. Table 2: The grouping results of five networks with n = 1500 Net MIND ICQA MIND ICQA Group Number Group Number Average ι Average ι 1 34 31 0.145514 0.031702 2 33 30 0.145036 0.036649 3 33 31 0.156483 0.033578 4 32 31 0.152671 0.029030 5 33 32 0.146560 0.033109 The difference between the average ι of the groups in each network shows that groups formed by MIND result in topologies with larger ι"s. It demonstrates that ι is good indicator of redundancy in different networks. 5.2 The Performance of the Resulting Groups Although MIND forms more groups than ICQA does, which implies longer lifetime of the networks, another importance consideration is how these groups formed by MIND and ICQA perform. We let 10000 events randomly occur in the network area except the margin. We compare how many events happen at the locations where the quality is less than the requirement s = 0.6 when each resulting group is conducting surveillance work (We call the number of such events the failure number of group). Figure 3 shows the average failure numbers of the resulting groups when different node numbers are set. We can see that the groups formed by MIND outperform those formed by ICQA because the groups formed by MIND result in lower failure numbers. This further demonstrates that MIND is a good approach for sensor-grouping problem. 1175 500 1000 1500 2000 0 10 20 30 40 50 60 Total in−network node number averagefailurenumbers ICQA MMNP Figure 3: The failure numbers of MIND and ICQA 6. CONCLUSION This paper proposes ι, a novel index for evaluation of pointdistribution. ι is the minimum distance between each pair of points normalized by the average distance between each pair of points. We find that a set of points that achieve a maximum value of ι result in a honeycomb structure. We propose that ι can serve as a good index to evaluate the distribution of the points, which can be employed in coverage-related problems in wireless sensor networks (WSNs). We set out to validate this idea by employing ι to a sensorgrouping problem. We formulate a general sensor-grouping problem for WSNs and provide a general sensing model. With an algorithm called Maximizing-ι Node-Deduction (MIND), we show that maximizing ι at sensor nodes is a good approach to solve this problem. Simulation results verify that MIND outperforms a greedy algorithm that exploits sensor-redundancy we design in terms of the number and the performance of the groups formed. This demonstrates a good application of employing ι in coverage-related problems. 7. ACKNOWLEDGEMENT The work described in this paper was substantially supported by two grants, RGC Project No. CUHK4205/04E and UGC Project No. AoE/E-01/99, of the Hong Kong Special Administrative Region, China. 8. REFERENCES [1] I. Akyildiz, W. Su, Y. Sankarasubramaniam, and E. Cayirci. A survey on wireless sensor networks. IEEE Communications Magazine, 40(8):102-114, 2002. [2] F. Aurenhammer. Vononoi diagram - a survey of a fundamental geometric data structure. ACM Computing Surveys, 23(2):345-405, September 1991. [3] N. Bulusu, J. Heidemann, and D. Estrin. GPS-less low-cost outdoor localization for very small devices. IEEE Personal Communication, October 2000. [4] M. Cardei and D.-Z. Du. Improving wireless sensor network lifetime through power aware organization. ACM Wireless Networks, 11(3), May 2005. [5] M. Cardei, D. MacCallum, X. Cheng, M. Min, X. Jia, D. Li, and D.-Z. Du. Wireless sensor networks with energy efficient organization. Journal of Interconnection Networks, 3(3-4), December 2002. [6] M. Cardei and J. Wu. Coverage in wireless sensor networks. In Handbook of Sensor Networks, (eds. M. Ilyas and I. Magboub), CRC Press, 2004. [7] X. Chen and M. R. Lyu. A sensibility-based sleeping configuration protocol for dependable wireless sensor networks. CSE Technical Report, The Chinese University of Hong Kong, 2005. [8] R. Jain, W. Hawe, and D. Chiu. A quantitative measure of fairness and discrimination for resource allocation in shared computer systems. Technical Report DEC-TR-301, September 1984. [9] S. S. Kulkarni and L. Wang. MNP: Multihop network reprogramming service for sensor networks. In Proc. of the 25th International Conference on Distributed Computing Systems (ICDCS), June 2005. [10] B. Liu and D. Towsley. A study on the coverage of large-scale sensor networks. In Proc. of the 1st IEEE International Conference on Mobile ad-hoc and Sensor Systems, Fort Lauderdale, FL, October 2004. [11] A. Mainwaring, J. Polastre, R. Szewczyk, D. Culler, and J. Anderson. Wireless sensor networks for habitat monitoring. In Proc. of the ACM International Workshop on Wireless Sensor Networks and Applications, 2002. [12] S. Megerian, F. Koushanfar, G. Qu, G. Veltri, and M. Potkonjak. Explosure in wirless sensor networks: Theory and pratical solutions. Wireless Networks, 8, 2002. [13] S. Slijepcevic and M. Potkonjak. Power efficient organization of wireless sensor networks. In Proc. of the IEEE International Conference on Communications (ICC), volume 2, Helsinki, Finland, June 2001. [14] D. Tian and N. D. Georganas. A node scheduling scheme for energy conservation in large wireless sensor networks. Wireless Communications and Mobile Computing, 3:272-290, May 2003. [15] X. Wang, G. Xing, Y. Zhang, C. Lu, R. Pless, and C. Gill. Integrated coverage and connectivity configuration in wireless sensor networks. In Proc. of the 1st ACM International Conference on Embedded Networked Sensor Systems (SenSys), Los Angeles, CA, November 2003. [16] G. Xing, C. Lu, R. Pless, and J. A. O´ Sullivan. Co-Grid: an efficient converage maintenance protocol for distributed sensor networks. In Proc. of the 3rd International Symposium on Information Processing in Sensor Networks (IPSN), Berkeley, CA, April 2004. [17] T. Yan, T. He, and J. A. Stankovic. Differentiated surveillance for sensor networks. In Proc. of the 1st ACM International Conference on Embedded Networked Sensor Systems (SenSys), Los Angeles, CA, November 2003. [18] F. Ye, G. Zhong, J. Cheng, S. Lu, and L. Zhang. PEAS: A robust energy conserving protocol for long-lived sensor networks. In Proc. of the 23rd International Conference on Distributed Computing Systems (ICDCS), Providence, Rhode Island, May 2003. [19] Y. Zhou, H. Yang, and M. R. Lyu. A point-distribution index and its application in coverage-related problems. CSE Technical Report, The Chinese University of Hong Kong, 2006. 1176

sensor-grouping;sensor group;fault tolerance;sleeping configuration protocol;surveillance;redundancy;sensor coverage;incremental coverage quality algorithm;node-deduction process;point-distribution index;wireless sensor network;honeycomb structure

train_C-72

GUESS: Gossiping Updates for Efficient Spectrum Sensing

Wireless radios of the future will likely be frequency-agile, that is, supporting opportunistic and adaptive use of the RF spectrum. Such radios must coordinate with each other to build an accurate and consistent map of spectral utilization in their surroundings. We focus on the problem of sharing RF spectrum data among a collection of wireless devices. The inherent requirements of such data and the time-granularity at which it must be collected makes this problem both interesting and technically challenging. We propose GUESS, a novel incremental gossiping approach to coordinated spectral sensing. It (1) reduces protocol overhead by limiting the amount of information exchanged between participating nodes, (2) is resilient to network alterations, due to node movement or node failures, and (3) allows exponentially-fast information convergence. We outline an initial solution incorporating these ideas and also show how our approach reduces network overhead by up to a factor of 2.4 and results in up to 2.7 times faster information convergence than alternative approaches.

1. INTRODUCTION There has recently been a huge surge in the growth of wireless technology, driven primarily by the availability of unlicensed spectrum. However, this has come at the cost of increased RF interference, which has caused the Federal Communications Commission (FCC) in the United States to re-evaluate its strategy on spectrum allocation. Currently, the FCC has licensed RF spectrum to a variety of public and private institutions, termed primary users. New spectrum allocation regimes implemented by the FCC use dynamic spectrum access schemes to either negotiate or opportunistically allocate RF spectrum to unlicensed secondary users Permission to make digital or hard copies of all or part of this work for personal or classroom use is granted without fee provided that copies are not made or distributed for profit or commercial advantage and that copies bear this notice and the full citation on the first page. To copy otherwise, to republish, to post on servers or to redistribute to lists, requires prior specific D1 D2 D5 D3 D4 Primary User Shadowed Secondary Users Secondary Users detect Primary's Signal Shadowed Secondary User Figure 1: Without cooperation, shadowed users are not able to detect the presence of the primary user. that can use it when the primary user is absent. The second type of allocation scheme is termed opportunistic spectrum sharing. The FCC has already legislated this access method for the 5 GHz band and is also considering the same for TV broadcast bands [1]. As a result, a new wave of intelligent radios, termed cognitive radios (or software defined radios), is emerging that can dynamically re-tune their radio parameters based on interactions with their surrounding environment. Under the new opportunistic allocation strategy, secondary users are obligated not to interfere with primary users (senders or receivers). This can be done by sensing the environment to detect the presence of primary users. However, local sensing is not always adequate, especially in cases where a secondary user is shadowed from a primary user, as illustrated in Figure 1. Here, coordination between secondary users is the only way for shadowed users to detect the primary. In general, cooperation improves sensing accuracy by an order of magnitude when compared to not cooperating at all [5]. To realize this vision of dynamic spectrum access, two fundamental problems must be solved: (1) Efficient and coordinated spectrum sensing and (2) Distributed spectrum allocation. In this paper, we propose strategies for coordinated spectrum sensing that are low cost, operate on timescales comparable to the agility of the RF environment, and are resilient to network failures and alterations. We defer the problem of spectrum allocation to future work. Spectrum sensing techniques for cognitive radio networks [4, 17] are broadly classified into three regimes; (1) centralized coordinated techniques, (2) decentralized coordinated techniques, and (3) decentralized uncoordinated techniques. We advocate a decentralized coordinated approach, similar in spirit to OSPF link-state routing used in the Internet. This is more effective than uncoordinated approaches because making decisions based only on local information is fallible (as shown in Figure 1). Moreover, compared to cen12 tralized approaches, decentralized techniques are more scalable, robust, and resistant to network failures and security attacks (e.g. jamming). Coordinating sensory data between cognitive radio devices is technically challenging because accurately assessing spectrum usage requires exchanging potentially large amounts of data with many radios at very short time scales. Data size grows rapidly due to the large number (i.e. thousands) of spectrum bands that must be scanned. This data must also be exchanged between potentially hundreds of neighboring secondary users at short time scales, to account for rapid changes in the RF environment. This paper presents GUESS, a novel approach to coordinated spectrum sensing for cognitive radio networks. Our approach is motivated by the following key observations: 1. Low-cost sensors collect approximate data: Most devices have limited sensing resolution because they are low-cost and low duty-cycle devices and thus cannot perform complex RF signal processing (e.g. matched filtering). Many are typically equipped with simple energy detectors that gather only approximate information. 2. Approximate summaries are sufficient for coordination: Approximate statistical summaries of sensed data are sufficient for correlating sensed information between radios, as relative usage information is more important than absolute usage data. Thus, exchanging exact RF information may not be necessary, and more importantly, too costly for the purposes of spectrum sensing. 3. RF spectrum changes incrementally: On most bands, RF spectrum utilization changes infrequently. Moreover, utilization of a specific RF band affects only that band and not the entire spectrum. Therefore, if the usage pattern of a particular band changes substantially, nodes detecting that change can initiate an update protocol to update the information for that band alone, leaving in place information already collected for other bands. This allows rapid detection of change while saving the overhead of exchanging unnecessary information. Based on these observations, GUESS makes the following contributions: 1. A novel approach that applies randomized gossiping algorithms to the problem of coordinated spectrum sensing. These algorithms are well suited to coordinated spectrum sensing due to the unique characteristics of the problem: i.e. radios are power-limited, mobile and have limited bandwidth to support spectrum sensing capabilities. 2. An application of in-network aggregation for dissemination of spectrum summaries. We argue that approximate summaries are adequate for performing accurate radio parameter tuning. 3. An extension of in-network aggregation and randomized gossiping to support incremental maintenance of spectrum summaries. Compared to standard gossiping approaches, incremental techniques can further reduce overhead and protocol execution time by requiring fewer radio resources. The rest of the paper is organized as follows. Section 2 motivates the need for a low cost and efficient approach to coordinated spectrum sensing. Section 3 discusses related work in the area, while Section 4 provides a background on in-network aggregation and randomized gossiping. Sections 5 and 6 discuss extensions and protocol details of these techniques for coordinated spectrum sensing. Section 7 presents simulation results showcasing the benefits of GUESS, and Section 8 presents a discussion and some directions for future work. 2. MOTIVATION To estimate the scale of the problem, In-stat predicts that the number of WiFi-enabled devices sold annually alone will grow to 430 million by 2009 [2]. Therefore, it would be reasonable to assume that a typical dense urban environment will contain several thousand cognitive radio devices in range of each other. As a result, distributed spectrum sensing and allocation would become both important and fundamental. Coordinated sensing among secondary radios is essential due to limited device sensing resolution and physical RF effects such as shadowing. Cabric et al. [5] illustrate the gains from cooperation and show an order of magnitude reduction in the probability of interference with the primary user when only a small fraction of secondary users cooperate. However, such coordination is non-trivial due to: (1) the limited bandwidth available for coordination, (2) the need to communicate this information on short timescales, and (3) the large amount of sensory data that needs to be exchanged. Limited Bandwidth: Due to restrictions of cost and power, most devices will likely not have dedicated hardware for supporting coordination. This implies that both data and sensory traffic will need to be time-multiplexed onto a single radio interface. Therefore, any time spent communicating sensory information takes away from the device"s ability to perform its intended function. Thus, any such coordination must incur minimal network overhead. Short Timescales: Further compounding the problem is the need to immediately propagate updated RF sensory data, in order to allow devices to react to it in a timely fashion. This is especially true due to mobility, as rapid changes of the RF environment can occur due to device and obstacle movements. Here, fading and multi-path interference heavily impact sensing abilities. Signal level can drop to a deep null with just a λ/4 movement in receiver position (3.7 cm at 2 GHz), where λ is the wavelength [14]. Coordination which does not support rapid dissemination of information will not be able to account for such RF variations. Large Sensory Data: Because cognitive radios can potentially use any part of the RF spectrum, there will be numerous channels that they need to scan. Suppose we wish to compute the average signal energy in each of 100 discretized frequency bands, and each signal can have up to 128 discrete energy levels. Exchanging complete sensory information between nodes would require 700 bits per transmission (for 100 channels, each requiring seven bits of information). Exchanging this information among even a small group of 50 devices each second would require (50 time-steps × 50 devices × 700 bits per transmission) = 1.67 Mbps of aggregate network bandwidth. Contrast this to the use of a randomized gossip protocol to disseminate such information, and the use of FM bit vectors to perform in-network aggregation. By applying gossip and FM aggregation, aggregate bandwidth requirements drop to (c·logN time-steps × 50 devices × 700 bits per transmission) = 0.40 Mbps, since 12 time-steps are needed to propagate the data (with c = 2, for illustrative purpoes1 ). This is explained further in Section 4. Based on these insights, we propose GUESS, a low-overhead approach which uses incremental extensions to FM aggregation and randomized gossiping for efficient coordination within a cognitive radio network. As we show in Section 7, 1 Convergence time is correlated with the connectivity topology of the devices, which in turn depends on the environment. 13 X A A X B B X Figure 2: Using FM aggregation to compute average signal level measured by a group of devices. these incremental extensions can further reduce bandwidth requirements by up to a factor of 2.4 over the standard approaches discussed above. 3. RELATED WORK Research in cognitive radio has increased rapidly [4, 17] over the years, and it is being projected as one of the leading enabling technologies for wireless networks of the future [9]. As mentioned earlier, the FCC has already identified new regimes for spectrum sharing between primary users and secondary users and a variety of systems have been proposed in the literature to support such sharing [4, 17]. Detecting the presence of a primary user is non-trivial, especially a legacy primary user that is not cognitive radio aware. Secondary users must be able to detect the primary even if they cannot properly decode its signals. This has been shown by Sahai et al. [16] to be extremely difficult even if the modulation scheme is known. Sophisticated and costly hardware, beyond a simple energy detector, is required to improve signal detection accuracy [16]. Moreover, a shadowed secondary user may not even be able to detect signals from the primary. As a result, simple local sensing approaches have not gained much momentum. This has motivated the need for cooperation among cognitive radios [16]. More recently, some researchers have proposed approaches for radio coordination. Liu et al. [11] consider a centralized access point (or base station) architecture in which sensing information is forwarded to APs for spectrum allocation purposes. APs direct mobile clients to collect such sensing information on their behalf. However, due to the need of a fixed AP infrastructure, such a centralized approach is clearly not scalable. In other work, Zhao et al. [17] propose a distributed coordination approach for spectrum sensing and allocation. Cognitive radios organize into clusters and coordination occurs within clusters. The CORVUS [4] architecture proposes a similar clustering method that can use either a centralized or decentralized approach to manage clusters. Although an improvement over purely centralized approaches, these techniques still require a setup phase to generate the clusters, which not only adds additional delay, but also requires many of the secondary users to be static or quasi-static. In contrast, GUESS does not place such restrictions on secondary users, and can even function in highly mobile environments. 4. BACKGROUND This section provides the background for our approach. We present the FM aggregation scheme that we use to generate spectrum summaries and perform in-network aggregation. We also discuss randomized gossiping techniques for disseminating aggregates in a cognitive radio network. 4.1 FM Aggregation Aggregation is the process where nodes in a distributed network combine data received from neighboring nodes with their local value to generate a combined aggregate. This aggregate is then communicated to other nodes in the network and this process repeats until the aggregate at all nodes has converged to the same value, i.e. the global aggregate. Double-counting is a well known problem in this process, where nodes may contribute more than once to the aggregate, causing inaccuracy in the final result. Intuitively, nodes can tag the aggregate value they transmit with information about which nodes have contributed to it. However, this approach is not scalable. Order and Duplicate Insensitive (ODI) techniques have been proposed in the literature [10, 15]. We adopt the ODI approach pioneered by Flajolet and Martin (FM) for the purposes of aggregation. Next we outline the FM approach; for full details, see [7]. Suppose we want to compute the number of nodes in the network, i.e. the COUNT query. To do so, each node performs a coin toss experiment as follows: toss an unbiased coin, stopping after the first head is seen. The node then sets the ith bit in a bit vector (initially filled with zeros), where i is the number of coin tosses it performed. The intuition is that as the number of nodes doing coin toss experiments increases, the probability of a more significant bit being set in one of the nodes" bit vectors increases. These bit vectors are then exchanged among nodes. When a node receives a bit vector, it updates its local bit vector by bitwise OR-ing it with the received vector (as shown in Figure 2 which computes AVERAGE). At the end of the aggregation process, every node, with high probability, has the same bit vector. The actual value of the count aggregate is then computed using the following formula, AGGF M = 2j−1 /0.77351, where j represents the bit position of the least significant zero in the aggregate bit vector [7]. Although such aggregates are very compact in nature, requiring only O(logN) state space (where N is the number of nodes), they may not be very accurate as they can only approximate values to the closest power of 2, potentially causing errors of up to 50%. More accurate aggregates can be computed by maintaining multiple bit vectors at each node, as explained in [7]. This decreases the error to within O(1/ √ m), where m is the number of such bit vectors. Queries other than count can also be computed using variants of this basic counting algorithm, as discussed in [3] (and shown in Figure 2). Transmitting FM bit vectors between nodes is done using randomized gossiping, discussed next. 4.2 Gossip Protocols Gossip-based protocols operate in discrete time-steps; a time-step is the required amount of time for all transmissions in that time-step to complete. At every time-step, each node having something to send randomly selects one or more neighboring nodes and transmits its data to them. The randomized propagation of information provides fault-tolerance and resilience to network failures and outages. We emphasize that this characteristic of the protocol also allows it to operate without relying on any underlying network structure. Gossip protocols have been shown to provide exponentially fast convergence2 , on the order of O(log N) [10], where N is the number of nodes (or radios). These protocols can therefore easily scale to very dense environments. 2 Convergence refers to the state in which all nodes have the most up-to-date view of the network. 14 Two types of gossip protocols are: • Uniform Gossip: In uniform gossip, at each timestep, each node chooses a random neighbor and sends its data to it. This process repeats for O(log(N)) steps (where N is the number of nodes in the network). Uniform gossip provides exponentially fast convergence, with low network overhead [10]. • Random Walk: In random walk, only a subset of the nodes (termed designated nodes) communicate in a particular time-step. At startup, k nodes are randomly elected as designated nodes. In each time-step, each designated node sends its data to a random neighbor, which becomes designated for the subsequent timestep (much like passing a token). This process repeats until the aggregate has converged in the network. Random walk has been shown to provide similar convergence bounds as uniform gossip in problems of similar context [8, 12]. 5. INCREMENTAL PROTOCOLS 5.1 Incremental FM Aggregates One limitation of FM aggregation is that it does not support updates. Due to the probabilistic nature of FM, once bit vectors have been ORed together, information cannot simply be removed from them as each node"s contribution has not been recorded. We propose the use of delete vectors, an extension of FM to support updates. We maintain a separate aggregate delete vector whose value is subtracted from the original aggregate vector"s value to obtain the resulting value as follows. AGGINC = (2a−1 /0.77351) − (2b−1 /0.77351) (1) Here, a and b represent the bit positions of the least significant zero in the original and delete bit vectors respectively. Suppose we wish to compute the average signal level detected in a particular frequency. To compute this, we compute the SUM of all signal level measurements and divide that by the COUNT of the number of measurements. A SUM aggregate is computed similar to COUNT (explained in Section 4.1), except that each node performs s coin toss experiments, where s is the locally measured signal level. Figure 2 illustrates the sequence by which the average signal energy is computed in a particular band using FM aggregation. Now suppose that the measured signal at a node changes from s to s . The vectors are updated as follows. • s > s: We simply perform (s − s) more coin toss experiments and bitwise OR the result with the original bit vector. • s < s: We increase the value of the delete vector by performing (s − s ) coin toss experiments and bitwise OR the result with the current delete vector. Using delete vectors, we can now support updates to the measured signal level. With the original implementation of FM, the aggregate would need to be discarded and a new one recomputed every time an update occurred. Thus, delete vectors provide a low overhead alternative for applications whose data changes incrementally, such as signal level measurements in a coordinated spectrum sensing environment. Next we discuss how these aggregates can be communicated between devices using incremental routing protocols. 5.2 Incremental Routing Protocol We use the following incremental variants of the routing protocols presented in Section 4.2 to support incremental updates to previously computed aggregates. Update Received OR Local Update Occurs Recovered Susceptible Time-stamp Expires Initial State Additional Update Received Infectious Clean Up Figure 3: State diagram each device passes through as updates proceed in the system • Incremental Gossip Protocol (IGP): When an update occurs, the updated node initiates the gossiping procedure. Other nodes only begin gossiping once they receive the update. Therefore, nodes receiving the update become active and continue communicating with their neighbors until the update protocol terminates, after O(log(N)) time steps. • Incremental Random Walk Protocol (IRWP): When an update (or updates) occur in the system, instead of starting random walks at k random nodes in the network, all k random walks are initiated from the updated node(s). The rest of the protocol proceeds in the same fashion as the standard random walk protocol. The allocation of walks to updates is discussed in more detail in [3], where the authors show that the number of walks has an almost negligible impact on network overhead. 6. PROTOCOL DETAILS Using incremental routing protocols to disseminate incremental FM aggregates is a natural fit for the problem of coordinated spectrum sensing. Here we outline the implementation of such techniques for a cognitive radio network. We continue with the example from Section 5.1, where we wish to perform coordination between a group of wireless devices to compute the average signal level in a particular frequency band. Using either incremental random walk or incremental gossip, each device proceeds through three phases, in order to determine the global average signal level for a particular frequency band. Figure 3 shows a state diagram of these phases. Susceptible: Each device starts in the susceptible state and becomes infectious only when its locally measured signal level changes, or if it receives an update message from a neighboring device. If a local change is observed, the device updates either the original or delete bit vector, as described in Section 5.1, and moves into the infectious state. If it receives an update message, it ORs the received original and delete bit vectors with its local bit vectors and moves into the infectious state. Note, because signal level measurements may change sporadically over time, a smoothing function, such as an exponentially weighted moving average, should be applied to these measurements. Infectious: Once a device is infectious it continues to send its up-to-date bit vectors, using either incremental random walk or incremental gossip, to neighboring nodes. Due to FM"s order and duplicate insensitive (ODI) properties, simultaneously occurring updates are handled seamlessly by the protocol. Update messages contain a time stamp indicating when the update was generated, and each device maintains a lo15 0 200 400 600 800 1000 1 10 100 Number of Measured Signal Changes Executiontime(ms) Incremental Gossip Uniform Gossip (a) Incremental Gossip and Uniform Gossip on Clique 0 200 400 600 800 1000 1 10 100 Number of Measured Signal Changes ExecutionTime(ms). Incremental Random Walk Random Walk (b) Incremental Random Walk and Random Walk on Clique 0 400 800 1200 1600 2000 1 10 100 Number of Measured Signal Changes ExecutionTime(ms). Random Walk Incremental Random Walk (c) Incremental Random Walk and Random Walk on Power-Law Random Graph Figure 4: Execution times of Incremental Protocols 0.9 1.4 1.9 2.4 2.9 1 10 100 Number of Measured Signal Changes OverheadImprovementRatio. (NormalizedtoUniformGossip) Incremental Gossip Uniform Gossip (a) Incremental Gossip and Uniform Gossip on Clique 0.9 1.4 1.9 2.4 2.9 1 10 100 Number of Measured Signal Changes OverheadImprovementRatio. (NormalizedtoRandomWalk) Incremental Random Walk Random Walk (b) Incremental Random Walk and Random Walk on Clique 0.9 1.1 1.3 1.5 1.7 1.9 1 10 100 Number of Measured Signal Changes OverheadImprovementRatio. (NormalizedtoRandomWalk) Random Walk Incremental Random Walk (c) Incremental Random Walk and Random Walk on Power-Law Random Graph Figure 5: Network overhead of Incremental Protocols cal time stamp of when it received the most recent update. Using this information, a device moves into the recovered state once enough time has passed for the most recent update to have converged. As discussed in Section 4.2, this happens after O(log(N)) time steps. Recovered: A recovered device ceases to propagate any update information. At this point, it performs clean-up and prepares for the next infection by entering the susceptible state. Once all devices have entered the recovered state, the system will have converged, and with high probability, all devices will have the up-to-date average signal level. Due to the cumulative nature of FM, even if all devices have not converged, the next update will include all previous updates. Nevertheless, the probability that gossip fails to converge is small, and has been shown to be O(1/N) [10]. For coordinated spectrum sensing, non-incremental routing protocols can be implemented in a similar fashion. Random walk would operate by having devices periodically drop the aggregate and re-run the protocol. Each device would perform a coin toss (biased on the number of walks) to determine whether or not it is a designated node. This is different from the protocol discussed above where only updated nodes initiate random walks. Similar techniques can be used to implement standard gossip. 7. EVALUATION We now provide a preliminary evaluation of GUESS in simulation. A more detailed evaluation of this approach can be found in [3]. Here we focus on how incremental extensions to gossip protocols can lead to further improvements over standard gossiping techniques, for the problem of coordinated spectrum sensing. Simulation Setup: We implemented a custom simulator in C++. We study the improvements of our incremental gossip protocols over standard gossiping in two dimensions: execution time and network overhead. We use two topologies to represent device connectivity: a clique, to eliminate the effects of the underlying topology on protocol performance, and a BRITE-generated [13] power-law random graph (PLRG), to illustrate how our results extend to more realistic scenarios. We simulate a large deployment of 1,000 devices to analyze protocol scalability. In our simulations, we compute the average signal level in a particular band by disseminating FM bit vectors. In each run of the simulation, we induce a change in the measured signal at one or more devices. A run ends when the new average signal level has converged in the network. For each data point, we ran 100 simulations and 95% confidence intervals (error bars) are shown. Simulation Parameters: Each transmission involves sending 70 bits of information to a neighboring node. To compute the AVERAGE aggregate, four bit vectors need to be transmitted: the original SUM vector, the SUM delete vector, the original COUNT vector, and the COUNT delete vector. Non-incremental protocols do not transmit the delete vectors. Each transmission also includes a time stamp of when the update was generated. We assume nodes communicate on a common control channel at 2 Mbps. Therefore, one time-step of protocol execution corresponds to the time required for 1,000 nodes to sequentially send 70 bits at 2 Mbps. Sequential use of the control channel is a worst case for our protocols; in practice, multiple control channels could be used in parallel to reduce execution time. We also assume nodes are loosely time synchronized, the implications of which are discussed further in [3]. Finally, in order to isolate the effect of protocol operation on performance, we do not model the complexities of the wireless channel in our simulations. Incremental Protocols Reduce Execution Time: Figure 4(a) compares the performance of incremental gossip (IGP) with uniform gossip on a clique topology. We observe that both protocols have almost identical execution times. This is expected as IGP operates in a similar fashion to 16 uniform gossip, taking O(log(N)) time-steps to converge. Figure 4(b) compares the execution times of incremental random walk (IRWP) and standard random walk on a clique. IRWP reduces execution time by a factor of 2.7 for a small number of measured signal changes. Although random walk and IRWP both use k random walks (in our simulations k = number of nodes), IRWP initiates walks only from updated nodes (as explained in Section 5.2), resulting in faster information convergence. These improvements carry over to a PLRG topology as well (as shown in Figure 4(c)), where IRWP is 1.33 times faster than random walk. Incremental Protocols Reduce Network Overhead: Figure 5(a) shows the ratio of data transmitted using uniform gossip relative to incremental gossip on a clique. For a small number of signal changes, incremental gossip incurs 2.4 times less overhead than uniform gossip. This is because in the early steps of protocol execution, only devices which detect signal changes communicate. As more signal changes are introduced into the system, gossip and incremental gossip incur approximately the same overhead. Similarly, incremental random walk (IRWP) incurs much less overhead than standard random walk. Figure 5(b) shows a 2.7 fold reduction in overhead for small numbers of signal changes on a clique. Although each protocol uses the same number of random walks, IRWP uses fewer network resources than random walk because it takes less time to converge. This improvement also holds true on more complex PLRG topologies (as shown in Figure 5(c)), where we observe a 33% reduction in network overhead. From these results it is clear that incremental techniques yield significant improvements over standard approaches to gossip, even on complex topologies. Because spectrum utilization is characterized by incremental changes to usage, incremental protocols are ideally suited to solve this problem in an efficient and cost effective manner. 8. DISCUSSION AND FUTURE WORK We have only just scratched the surface in addressing the problem of coordinated spectrum sensing using incremental gossiping. Next, we outline some open areas of research. Spatial Decay: Devices performing coordinated sensing are primarily interested in the spectrum usage of their local neighborhood. Therefore, we recommend the use of spatially decaying aggregates [6], which limits the impact of an update on more distant nodes. Spatially decaying aggregates work by successively reducing (by means of a decay function) the value of the update as it propagates further from its origin. One challenge with this approach is that propagation distance cannot be determined ahead of time and more importantly, exhibits spatio-temporal variations. Therefore, finding the optimal decay function is non-trivial, and an interesting subject of future work. Significance Threshold: RF spectrum bands continually experience small-scale changes which may not necessarily be significant. Deciding if a change is significant can be done using a significance threshold β, below which any observed change is not propagated by the node. Choosing an appropriate operating value for β is application dependent, and explored further in [3]. Weighted Readings: Although we argued that most devices will likely be equipped with low-cost sensing equipment, there may be situations where there are some special infrastructure nodes that have better sensing abilities than others. Weighting their measurements more heavily could be used to maintain a higher degree of accuracy. Determining how to assign such weights is an open area of research. Implementation Specifics: Finally, implementing gossip for coordinated spectrum sensing is also open. If implemented at the MAC layer, it may be feasible to piggy-back gossip messages over existing management frames (e.g. networking advertisement messages). As well, we also require the use of a control channel to disseminate sensing information. There are a variety of alternatives for implementing such a channel, some of which are outlined in [4]. The trade-offs of different approaches to implementing GUESS is a subject of future work. 9. CONCLUSION Spectrum sensing is a key requirement for dynamic spectrum allocation in cognitive radio networks. The nature of the RF environment necessitates coordination between cognitive radio devices. We propose GUESS, an approximate yet low overhead approach to perform efficient coordination between cognitive radios. The fundamental contributions of GUESS are: (1) an FM aggregation scheme for efficient innetwork aggregation, (2) a randomized gossiping approach which provides exponentially fast convergence and robustness to network alterations, and (3) incremental variations of FM and gossip which we show can reduce the communication time by up to a factor of 2.7 and reduce network overhead by up to a factor of 2.4. Our preliminary simulation results showcase the benefits of this approach and we also outline a set of open problems that make this a new and exciting area of research. 10. REFERENCES [1] Unlicensed Operation in the TV Broadcast Bands and Additional Spectrum for Unlicensed Devices Below 900 MHz in the 3 GHz band, May 2004. Notice of Proposed Rule-Making 04-186, Federal Communications Commission. [2] In-Stat: Covering the Full Spectrum of Digital Communications Market Research, from Vendor to End-user, December 2005. http://www.in-stat.com/catalog/scatalogue.asp?id=28. [3] N. Ahmed, D. Hadaller, and S. Keshav. Incremental Maintenance of Global Aggregates. UW. Technical Report CS-2006-19, University of Waterloo, ON, Canada, 2006. [4] R. W. Brodersen, A. Wolisz, D. Cabric, S. M. Mishra, and D. Willkomm. CORVUS: A Cognitive Radio Approach for Usage of Virtual Unlicensed Spectrum. Technical report, July 2004. [5] D. Cabric, S. M. Mishra, and R. W. Brodersen. Implementation Issues in Spectrum Sensing for Cognitive Radios. In Asilomar Conference, 2004. [6] E. Cohen and H. Kaplan. Spatially-Decaying Aggregation Over a Network: Model and Algorithms. In Proceedings of SIGMOD 2004, pages 707-718, New York, NY, USA, 2004. ACM Press. [7] P. Flajolet and G. N. Martin. Probabilistic Counting Algorithms for Data Base Applications. J. Comput. Syst. Sci., 31(2):182-209, 1985. [8] C. Gkantsidis, M. Mihail, and A. Saberi. Random Walks in Peer-to-Peer Networks. In Proceedings of INFOCOM 2004, pages 1229-1240, 2004. [9] E. Griffith. Previewing Intel"s Cognitive Radio Chip, June 2005. http://www.internetnews.com/wireless/article.php/3513721. [10] D. Kempe, A. Dobra, and J. Gehrke. Gossip-Based Computation of Aggregate Information. In FOCS 2003, page 482, Washington, DC, USA, 2003. IEEE Computer Society. [11] X. Liu and S. Shankar. Sensing-based Opportunistic Channel Access. In ACM Mobile Networks and Applications (MONET) Journal, March 2005. [12] Q. Lv, P. Cao, E. Cohen, K. Li, and S. Shenker. Search and Replication in Unstructured Peer-to-Peer Networks. In Proceedings of ICS, 2002. [13] A. Medina, A. Lakhina, I. Matta, and J. Byers. BRITE: an Approach to Universal Topology Generation. In Proceedings of MASCOTS conference, Aug. 2001. [14] S. M. Mishra, A. Sahai, and R. W. Brodersen. Cooperative Sensing among Cognitive Radios. In ICC 2006, June 2006. [15] S. Nath, P. B. Gibbons, S. Seshan, and Z. R. Anderson. Synopsis Diffusion for Robust Aggregation in Sensor Networks. In Proceedings of SenSys 2004, pages 250-262, 2004. [16] A. Sahai, N. Hoven, S. M. Mishra, and R. Tandra. Fundamental Tradeoffs in Robust Spectrum Sensing for Opportunistic Frequency Reuse. Technical Report UC Berkeley, 2006. [17] J. Zhao, H. Zheng, and G.-H. Yang. Distributed Coordination in Dynamic Spectrum Allocation Networks. In Proceedings of DySPAN 2005, Baltimore (MD), Nov. 2005. 17

coordinate spectrum sense;coordinated sensing;cognitive radio;spatially decaying aggregate;rf interference;spectrum allocation;opportunistic spectrum sharing;innetwork aggregation;incremental algorithm;rf spectrum;spectrum sensing;incremental gossip protocol;fm aggregation;gossip protocol

train_C-74

Adapting Asynchronous Messaging Middleware to ad-hoc Networking

The characteristics of mobile environments, with the possibility of frequent disconnections and fluctuating bandwidth, have forced a rethink of traditional middleware. In particular, the synchronous communication paradigms often employed in standard middleware do not appear to be particularly suited to ad-hoc environments, in which not even the intermittent availability of a backbone network can be assumed. Instead, asynchronous communication seems to be a generally more suitable paradigm for such environments. Message oriented middleware for traditional systems has been developed and used to provide an asynchronous paradigm of communication for distributed systems, and, recently, also for some specific mobile computing systems. In this paper, we present our experience in designing, implementing and evaluating EMMA (Epidemic Messaging Middleware for ad-hoc networks), an adaptation of Java Message Service (JMS) for mobile ad-hoc environments. We discuss in detail the design challenges and some possible solutions, showing a concrete example of the feasibility and suitability of the application of the asynchronous paradigm in this setting and outlining a research roadmap for the coming years.

1. INTRODUCTION With the increasing popularity of mobile devices and their widespread adoption, there is a clear need to allow the development of a broad spectrum of applications that operate effectively over such an environment. Unfortunately, this is far from simple: mobile devices are increasingly heterogeneous in terms of processing capabilities, memory size, battery capacity, and network interfaces. Each such configuration has substantially different characteristics that are both statically different - for example, there is a major difference in capability between a Berkeley mote and an 802.11g-equipped laptop - and that vary dynamically, as in situations of fluctuating bandwidth and intermittent connectivity. Mobile ad hoc environments have an additional element of complexity in that they are entirely decentralised. In order to craft applications for such complex environments, an appropriate form of middleware is essential if cost effective development is to be achieved. In this paper, we examine one of the foundational aspects of middleware for mobile ad-hoc environments: that of the communication primitives. Traditionally, the most frequently used middleware primitives for communication assume the simultaneous presence of both end points on a network, since the stability and pervasiveness of the networking infrastructure is not an unreasonable assumption for most wired environments. In other words, most communication paradigms are synchronous: object oriented middleware such as CORBA and Java RMI are typical examples of middleware based on synchronous communication. In recent years, there has been growing interest in platforms based on asynchronous communication paradigms, such as publish-subscribe systems [6]: these have been exploited very successfully where there is application level asynchronicity. From a Gartner Market Report [7]: Given messageoriented-middleware"s (MOM) popularity, scalability, flexibility, and affinity with mobile and wireless architectures, by 2004, MOM will emerge as the dominant form of communication middleware for linking mobile and enterprise applications (0.7 probability).... Moreover, in mobile ad-hoc systems, the likelihood of network fragmentation means that synchronous communication may in any case be impracticable, giving situations in which delay tolerant asynchronous traffic is the only form of traffic that could be supported. 121 Middleware 2004 Companion Middleware for mobile ad-hoc environments must therefore support semi-synchronous or completely asynchronous communication primitives if it is to avoid substantial limitations to its utility. Aside from the intellectual challenge in supporting this model, this work is also interesting because there are a number of practical application domains in allowing inter-community communication in undeveloped areas of the globe. Thus, for example, projects that have been carried out to help populations that live in remote places of the globe such as Lapland [3] or in poor areas that lack fixed connectivity infrastructure [9]. There have been attempts to provide mobile middleware with these properties, including STEAM, LIME, XMIDDLE, Bayou (see [11] for a more complete review of mobile middleware). These models differ quite considerably from the existing traditional middleware in terms of primitives provided. Furthermore, some of them fail in providing a solution for the true ad-hoc scenarios. If the projected success of MOM becomes anything like a reality, there will be many programmers with experience of it. The ideal solution to the problem of middleware for ad-hoc systems is, then, to allow programmers to utilise the same paradigms and models presented by common forms of MOM and to ensure that these paradigms are supportable within the mobile environment. This approach has clear advantages in allowing applications developed on standard middleware platforms to be easily deployed on mobile devices. Indeed, some research has already led to the adaptation of traditional middleware platforms to mobile settings, mainly to provide integration between mobile devices and existing fixed networks in a nomadic (i.e., mixed) environment [4]. With respect to message oriented middleware, the current implementations, however, either assume the existence of a backbone network to which the mobile hosts connect from time to time while roaming [10], or assume that nodes are always somehow reachable through a path [18]. No adaptation to heterogeneous or completely ad-hoc scenarios, with frequent disconnection and periodically isolated clouds of hosts, has been attempted. In the remainder of this paper we describe an initial attempt to adapt message oriented middleware to suit mobile and, more specifically, mobile ad-hoc networks. In our case, we elected to examine JMS, as one of the most widely known MOM systems. In the latter part of this paper, we explore the limitations of our results and describe the plans we have to take the work further. 2. MESSAGE ORIENTED MIDDLEWARE AND JAVA MESSAGE SERVICE (JMS) Message-oriented middleware systems support communication between distributed components via message-passing: the sender sends a message to identified queues, which usually reside on a server. A receiver retrieves the message from the queue at a different time and may acknowledge the reply using the same asynchronous mechanism. Message-oriented middleware thus supports asynchronous communication in a very natural way, achieving de-coupling of senders and receivers. A sender is able to continue processing as soon as the middleware has accepted the message; eventually, the receiver will send an acknowledgment message and the sender will be able to collect it at a convenient time. However, given the way they are implemented, these middleware systems usually require resource-rich devices, especially in terms of memory and disk space, where persistent queues of messages that have been received but not yet processed, are stored. Sun Java Message Service [5], IBM WebSphere MQ [6], Microsoft MSMQ [12] are examples of very successful message-oriented middleware for traditional distributed systems. The Java Messaging Service (JMS) is a collection of interfaces for asynchronous communication between distributed components. It provides a common way for Java programs to create, send and receive messages. JMS users are usually referred to as clients. The JMS specification further defines providers as the components in charge of implementing the messaging system and providing the administrative and control functionality (i.e., persistence and reliability) required by the system. Clients can send and receive messages, asynchronously, through the JMS provider, which is in charge of the delivery and, possibly, of the persistence of the messages. There are two types of communication supported: point to point and publish-subscribe models. In the point to point model, hosts send messages to queues. Receivers can be registered with some specific queues, and can asynchronously retrieve the messages and then acknowledge them. The publish-subscribe model is based on the use of topics that can be subscribed to by clients. Messages are sent to topics by other clients and are then received in an asynchronous mode by all the subscribed clients. Clients learn about the available topics and queues through Java Naming and Directory Interface (JNDI) [14]. Queues and topics are created by an administrator on the provider and are registered with the JNDI interface for look-up. In the next section, we introduce the challenges of mobile networks, and show how JMS can be adapted to cope with these requirements. 3. JMS FOR MOBILE COMPUTING Mobile networks vary very widely in their characteristics, from nomadic networks in which modes relocate whilst offline through to ad-hoc networks in which modes move freely and in which there is no infrastructure. Mobile ad-hoc networks are most generally applicable in situations where survivability and instant deployability are key: most notably in military applications and disaster relief. In between these two types of "mobile" networks, there are, however, a number of possible heterogeneous combinations, where nomadic and ad-hoc paradigms are used to interconnect totally unwired areas to more structured networks (such as a LAN or the Internet). Whilst the JMS specification has been extensively implemented and used in traditional distributed systems, adaptations for mobile environments have been proposed only recently. The challenges of porting JMS to mobile settings are considerable; however, in view of its widespread acceptance and use, there are considerable advantages in allowing the adaptation of existing applications to mobile environments and in allowing the interoperation of applications in the wired and wireless regions of a network. In [10], JMS was adapted to a nomadic mobile setting, where mobile hosts can be JMS clients and communicate through the JMS provider that, however, sits on a backbone network, providing reliability and persistence. The client prototype presented in [10] is very lightweight, due to the delegation of all the heavyweight functionality to the Middleware for Pervasive and ad-hoc Computing 122 provider on the wired network. However, this approach is somewhat limited in terms of widespread applicability and scalability as a consequence of the concentration of functionality in the wired portion of the network. If JMS is to be adapted to completely ad-hoc environments, where no fixed infrastructure is available, and where nodes change location and status very dynamically, more issues must be taken into consideration. Firstly, discovery needs to use a resilient but distributed model: in this extremely dynamic environment, static solutions are unacceptable. As discussed in Section 2, a JMS administrator defines queues and topics on the provider. Clients can then learn about them using the Java Naming and Directory Interface (JNDI). However, due to the way JNDI is designed, a JNDI node (or more than one) needs to be in reach in order to obtain a binding of a name to an address (i.e., knowing where a specific queue/topic is). In mobile ad-hoc environments, the discovery process cannot assume the existence of a fixed set of discovery servers that are always reachable, as this would not match the dynamicity of ad-hoc networks. Secondly, a JMS Provider, as suggested by the JMS specification, also needs to be reachable by each node in the network, in order to communicate. This assumes a very centralised architecture, which again does not match the requirements of a mobile ad-hoc setting, in which nodes may be moving and sparse: a more distributed and dynamic solution is needed. Persistence is, however, essential functionality in asynchronous communication environments as hosts are, by definition, connected at different times. In the following section, we will discuss our experience in designing and implementing JMS for mobile ad-hoc networks. 4. JMSFOR MOBILE ad-hoc NETWORKS 4.1 Adaptation of JMS for Mobile ad-hoc Networks Developing applications for mobile networks is yet more challenging: in addition to the same considerations as for infrastructured wireless environments, such as the limited device capabilities and power constraints, there are issues of rate of change of network connectivity, and the lack of a static routing infrastructure. Consequently, we now describe an initial attempt to adapt the JMS specification to target the particular requirements related to ad-hoc scenarios. As discussed in Section 3, a JMS application can use either the point to point and the publish-subscribe styles of messaging. Point to Point Model The point to point model is based on the concept of queues, that are used to enable asynchronous communication between the producer of a message and possible different consumers. In our solution, the location of queues is determined by a negotiation process that is application dependent. For example, let us suppose that it is possible to know a priori, or it is possible to determine dynamically, that a certain host is the receiver of the most part of messages sent to a particular queue. In this case, the optimum location of the queue may well be on this particular host. In general, it is worth noting that, according to the JMS specification and suggested design patterns, it is common and preferable for a client to have all of its messages delivered to a single queue. Queues are advertised periodically to the hosts that are within transmission range or that are reachable by means of the underlying synchronous communication protocol, if provided. It is important to note that, at the middleware level, it is logically irrelevant whether or not the network layer implements some form of ad-hoc routing (though considerably more efficient if it does); the middleware only considers information about which nodes are actively reachable at any point in time. The hosts that receive advertisement messages add entries to their JNDI registry. Each entry is characterized by a lease (a mechanism similar to that present in Jini [15]). A lease represents the time of validity of a particular entry. If a lease is not refreshed (i.e, its life is not extended), it can expire and, consequently, the entry is deleted from the registry. In other words, the host assumes that the queue will be unreachable from that point in time. This may be caused, for example, if a host storing the queue becomes unreachable. A host that initiates a discovery process will find the topics and the queues present in its connected portion of the network in a straightforward manner. In order to deliver a message to a host that is not currently in reach1 , we use an asynchronous epidemic routing protocol that will be discussed in detail in Section 4.2. If two hosts are in the same cloud (i.e., a connected path exists between them), but no synchronous protocol is available, the messages are sent using the epidemic protocol. In this case, the delivery latency will be low as a result of the rapidity of propagation of the infection in the connected cloud (see also the simulation results in Section 5). Given the existence of an epidemic protocol, the discovery mechanism consists of advertising the queues to the hosts that are currently unreachable using analogous mechanisms. Publish-Subscribe Model In the publish-subscribe model, some of the hosts are similarly designated to hold topics and store subscriptions, as before. Topics are advertised through the registry in the same way as are queues, and a client wishing to subscribe to a topic must register with the client holding the topic. When a client wishes to send a message to the topic list, it sends it to the topic holder (in the same way as it would send a message to a queue). The topic holder then forwards the message to all subscribers, using the synchronous protocol if possible, the epidemic protocol otherwise. It is worth noting that we use a single message with multiple recipients, instead of multiple messages with multiple recipients. When a message is delivered to one of the subscribers, this recipient is deleted from the list. In order to delete the other possible replicas, we employ acknowledgment messages (discussed in Section 4.4), returned in the same way as a normal message. We have also adapted the concepts of durable and non durable subscriptions for ad-hoc settings. In fixed platforms, durable subscriptions are maintained during the disconnections of the clients, whether these are intentional or are the result of failures. In traditional systems, while a durable subscriber is disconnected from the server, it is responsible for storing messages. When the durable subscriber reconnects, the server sends it all unexpired messages. The problem is that, in our scenario, disconnections are the norm 1 In theory, it is not possible to send a message to a peer that has never been reachable in the past, since there can be no entry present in the registry. However, to overcome this possible limitation, we provide a primitive through which information can be added to the registry without using the normal channels. 123 Middleware 2004 Companion rather than the exception. In other words, we cannot consider disconnections as failures. For these reasons, we adopt a slightly different semantics. With respect to durable subscriptions, if a subscriber becomes disconnected, notifications are not stored but are sent using the epidemic protocol rather than the synchronous protocol. In other words, durable notifications remain valid during the possible disconnections of the subscriber. On the other hand, if a non-durable subscriber becomes disconnected, its subscription is deleted; in other words, during disconnections, notifications are not sent using the epidemic protocol but exploit only the synchronous protocol. If the topic becomes accessible to this host again, it must make another subscription in order to receive the notifications. Unsubscription messages are delivered in the same way as are subscription messages. It is important to note that durable subscribers have explicitly to unsubscribe from a topic in order to stop the notification process; however, all durable subscriptions have a predefined expiration time in order to cope with the cases of subscribers that do not meet again because of their movements or failures. This feature is clearly provided to limit the number of the unnecessary messages sent around the network. 4.2 Message Delivery using Epidemic Routing In this section, we examine one possible mechanism that will allow the delivery of messages in a partially connected network. The mechanism we discuss is intended for the purposes of demonstrating feasibility; more efficient communication mechanisms for this environment are themselves complex, and are the subject of another paper [13]. The asynchronous message delivery described above is based on a typical pure epidemic-style routing protocol [16]. A message that needs to be sent is replicated on each host in reach. In this way, copies of the messages are quickly spread through connected networks, like an infection. If a host becomes connected to another cloud of mobile nodes, during its movement, the message spreads through this collection of hosts. Epidemic-style replication of data and messages has been exploited in the past in many fields starting with the distributed database systems area [2]. Within epidemic routing, each host maintains a buffer containing the messages that it has created and the replicas of the messages generated by the other hosts. To improve the performance, a hash-table indexes the content of the buffer. When two hosts connect, the host with the smaller identifier initiates a so-called anti-entropy session, sending a list containing the unique identifiers of the messages that it currently stores. The other host evaluates this list and sends back a list containing the identifiers it is storing that are not present in the other host, together with the messages that the other does not have. The host that has started the session receives the list and, in the same way, sends the messages that are not present in the other host. Should buffer overflow occur, messages are dropped. The reliability offered by this protocol is typically best effort, since there is no guarantee that a message will eventually be delivered to its recipient. Clearly, the delivery ratio of the protocol increases proportionally to the maximum allowed delay time and the buffer size in each host (interesting simulation results may be found in [16]). 4.3 Adaptation of the JMS Message Model In this section, we will analyse the aspects of our adaptation of the specification related to the so-called JMS Message Model [5]. According to this, JMS messages are characterised by some properties defined using the header field, which contains values that are used by both clients and providers for their delivery. The aspects discussed in the remainder of this section are valid for both models (point to point and publish-subscribe). A JMS message can be persistent or non-persistent. According to the JMS specification, persistent messages must be delivered with a higher degree of reliability than the nonpersistent ones. However, it is worth noting that it is not possible to ensure once-and-only-once reliability for persistent messages as defined in the specification, since, as we discussed in the previous subsection, the underlying epidemic protocol can guarantee only best-effort delivery. However, clients maintain a list of the identifiers of the recently received messages to avoid the delivery of message duplicates. In other words, we provide the applications with at-mostonce reliability for both types of messages. In order to implement different levels of reliability, EMMA treats persistent and non-persistent messages differently, during the execution of the anti-entropy epidemic protocol. Since the message buffer space is limited, persistent messages are preferentially replicated using the available free space. If this is insufficient and non-persistent messages are present in the buffer, these are replaced. Only the successful deliveries of the persistent messages are notified to the senders. According to the JMS specification, it is possible to assign a priority to each message. The messages with higher priorities are delivered in a preferential way. As discussed above, persistent messages are prioritised above the non-persistent ones. Further selection is based on their priorities. Messages with higher priorities are treated in a preferential way. In fact, if there is not enough space to replicate all the persistent messages, a mechanism based on priorities is used to delete and replicate non-persistent messages (and, if necessary, persistent messages). Messages are deleted from the buffers using the expiration time value that can be set by senders. This is a way to free space in the buffers (one preferentially deletes older messages in cases of conflict); to eliminate stale replicas in the system; and to limit the time for which destinations must hold message identifiers to dispose of duplicates. 4.4 Reliability and Acknowledgment Mechanisms As already discussed, at-most-once message delivery is the best that can be achieved in terms of delivery semantics in partially connected ad-hoc settings. However, it is possible to improve the reliability of the system with efficient acknowledgment mechanisms. EMMA provides a mechanism for failure notification to applications if the acknowledgment is not received within a given timeout (that can be configured by application developers). This mechanism is the one that distinguishes the delivery of persistent and non-persistent messages in our JMS implementation: the deliveries of the former are notified to the senders, whereas the latter are not. We use acknowledgment messages not only to inform senders about the successful delivery of messages but also to delete the replicas of the delivered messages that are still present in the network. Each host maintains a list of the messages Middleware for Pervasive and ad-hoc Computing 124 successfully delivered that is updated as part of the normal process of information exchange between the hosts. The lists are exchanged during the first steps of the anti-entropic epidemic protocol with a certain predefined frequency. In the case of messages with multiple recipients, a list of the actual recipients is also stored. When a host receives the list, it checks its message buffer and updates it according to the following rules: (1) if a message has a single recipient and it has been delivered, it is deleted from the buffer; (2) if a message has multiple recipients, the identifiers of the delivered hosts are deleted from the associated list of recipients. If the resulting length of the list of recipients is zero, the message is deleted from the buffer. These lists have, clearly, finite dimensions and are implemented as circular queues. This simple mechanism, together with the use of expiration timestamps, guarantees that the old acknowledgment notifications are deleted from the system after a limited period of time. In order to improve the reliability of EMMA, a design mechanism for intelligent replication of queues and topics based on the context information could be developed. However this is not yet part of the current architecture of EMMA. 5. IMPLEMENTATION AND PRELIMINARY EVALUATION We implemented a prototype of our platform using the J2ME Personal Profile. The size of the executable is about 250KB including the JMS 1.1 jar file; this is a perfectly acceptable figure given the available memory of the current mobile devices on the market. We tested our prototype on HP IPaq PDAs running Linux, interconnected with WaveLan, and on a number of laptops with the same network interface. We also evaluated the middleware platform using the OMNET++ discrete event simulator [17] in order to explore a range of mobile scenarios that incorporated a more realistic number of hosts than was achievable experimentally. More specifically, we assessed the performance of the system in terms of delivery ratio and average delay, varying the density of population and the buffer size, and using persistent and non-persistent messages with different priorities. The simulation results show that the EMMA"s performance, in terms of delivery ratio and delay of persistent messages with higher priorities, is good. In general, it is evident that the delivery ratio is strongly related to the correct dimensioning of the buffers to the maximum acceptable delay. Moreover, the epidemic algorithms are able to guarantee a high delivery ratio if one evaluates performance over a time interval sufficient for the dissemination of the replicas of messages (i.e., the infection spreading) in a large portion of the ad-hoc network. One consequence of the dimensioning problem is that scalability may be seriously impacted in peer-to-peer middleware for mobile computing due to the resource poverty of the devices (limited memory to store temporarily messages) and the number of possible interconnections in ad-hoc settings. What is worse is that common forms of commercial and social organisation (six degrees of separation) mean that even modest TTL values on messages will lead to widespread flooding of epidemic messages. This problem arises because of the lack of intelligence in the epidemic protocol, and can be addressed by selecting carrier nodes for messages with greater care. The details of this process are, however, outside the scope of this paper (but may be found in [13]) and do not affect the foundation on which the EMMA middleware is based: the ability to deliver messages asynchronously. 6. CRITICAL VIEW OF THE STATE OF THE ART The design of middleware platforms for mobile computing requires researchers to answer new and fundamentally different questions; simply assuming the presence of wired portions of the network on which centralised functionality can reside is not generalisable. Thus, it is necessary to investigate novel design principles and to devise architectural patterns that differ from those traditionally exploited in the design of middleware for fixed systems. As an example, consider the recent cross-layering trend in ad-hoc networking [1]. This is a way of re-thinking software systems design, explicitly abandoning the classical forms of layering, since, although this separation of concerns afford portability, it does so at the expense of potential efficiency gains. We believe that it is possible to view our approach as an instance of cross-layering. In fact, we have added the epidemic network protocol at middleware level and, at the same time, we have used the existing synchronous network protocol if present both in delivering messages (traditional layering) and in informing the middleware about when messages may be delivered by revealing details of the forwarding tables (layer violation). For this reason, we prefer to consider them jointly as the communication layer of our platform together providing more efficient message delivery. Another interesting aspect is the exploitation of context and system information to improve the performance of mobile middleware platforms. Again, as a result of adopting a cross-layering methodology, we are able to build systems that gather information from the underlying operating system and communication components in order to allow for adaptation of behaviour. We can summarise this conceptual design approach by saying that middleware platforms must be not only context-aware (i.e., they should be able to extract and analyse information from the surrounding context) but also system-aware (i.e., they should be able to gather information from the software and hardware components of the mobile system). A number of middleware systems have been developed to support ad-hoc networking with the use of asynchronous communication (such as LIME, XMIDDLE, STEAM [11]). In particular, the STEAM platform is an interesting example of event-based middleware for ad-hoc networks, providing location-aware message delivery and an effective solution for event filtering. A discussion of JMS, and its mobile realisation, has already been conducted in Sections 4 and 2. The Swiss company Softwired has developed the first JMS middleware for mobile computing, called iBus Mobile [10]. The main components of this typically infrastructure-based architecture are the JMS provider, the so-called mobile JMS gateway, which is deployed on a fixed host and a lightweight JMS client library. The gateway is used for the communication between the application server and mobile hosts. The gateway is seen by the JMS provider as a normal JMS client. The JMS provider can be any JMS-enabled application server, such as BEA Weblogic. Pronto [19] is an example of mid125 Middleware 2004 Companion dleware system based on messaging that is specifically designed for mobile environments. The platform is composed of three classes of components: mobile clients implementing the JMS specification, gateways that control traffic, guaranteeing efficiency and possible user customizations using different plug-ins and JMS servers. Different configurations of these components are possible; with respect to mobile ad hoc networks applications, the most interesting is Serverless JMS. The aim of this configuration is to adapt JMS to a decentralized model. The publish-subscribe model exploits the efficiency and the scalability of the underlying IP multicast protocol. Unreliable and reliable message delivery services are provided: reliability is provided through a negative acknowledgment-based protocol. Pronto represents a good solution for infrastructure-based mobile networks but it does not adequately target ad-hoc settings, since mobile nodes rely on fixed servers for the exchange of messages. Other MOM implemented for mobile environments exist; however, they are usually straightforward extensions of existing middleware [8]. The only implementation of MOM specifically designed for mobile ad-hoc networks was developed at the University of Newcastle [18]. This work is again a JMS adaptation; the focus of that implementation is on group communication and the use of application level routing algorithms for topic delivery of messages. However, there are a number of differences in the focus of our work. The importance that we attribute to disconnections makes persistence a vital requirement for any middleware that needs to be used in mobile ad-hoc networks. The authors of [18] signal persistence as possible future work, not considering the fact that routing a message to a non-connected host will result in delivery failure. This is a remarkable limitation in mobile settings where unpredictable disconnections are the norm rather than the exception. 7. ROADMAP AND CONCLUSIONS Asynchronous communication is a useful communication paradigm for mobile ad-hoc networks, as hosts are allowed to come, go and pick up messages when convenient, also taking account of their resource availability (e.g., power, connectivity levels). In this paper we have described the state of the art in terms of MOM for mobile systems. We have also shown a proof of concept adaptation of JMS to the extreme scenario of partially connected mobile ad-hoc networks. We have described and discussed the characteristics and differences of our solution with respect to traditional JMS implementations and the existing adaptations for mobile settings. However, trade-offs between application-level routing and resource usage should also be investigated, as mobile devices are commonly power/resource scarce. A key limitation of this work is the poorly performing epidemic algorithm and an important advance in the practicability of this work requires an algorithm that better balances the needs of efficiency and message delivery probability. We are currently working on algorithms and protocols that, exploiting probabilistic and statistical techniques on the basis of small amounts of exchanged information, are able to improve considerably the efficiency in terms of resources (memory, bandwidth, etc) and the reliability of our middleware platform [13]. One futuristic research development, which may take these ideas of adaptation of messaging middleware for mobile environments further is the introduction of more mobility oriented communication extensions, for instance the support of geocast (i.e., the ability to send messages to specific geographical areas). 8. REFERENCES [1] M. Conti, G. Maselli, G. Turi, and S. Giordano. Cross-layering in Mobile ad-hoc Network Design. IEEE Computer, 37(2):48-51, February 2004. [2] A. Demers, D. Greene, C. Hauser, W. Irish, J. Larson, S. Shenker, H. Sturgis, D. Swinehart, and D. Terry. Epidemic Algorithms for Replicated Database Maintenance. In Sixth Symposium on Principles of Distributed Computing, pages 1-12, August 1987. [3] A. Doria, M. Uden, and D. P. Pandey. Providing connectivity to the Saami nomadic community. In Proceedings of the Second International Conference on Open Collaborative Design for Sustainable Innovation, December 2002. [4] M. Haahr, R. Cunningham, and V. Cahill. Supporting CORBA applications in a Mobile Environment. In 5th International Conference on Mobile Computing and Networking (MOBICOM99), pages 36-47. ACM, August 1999. [5] M. Hapner, R. Burridge, R. Sharma, J. Fialli, and K. Stout. Java Message Service Specification Version 1.1. Sun Microsystems, Inc., April 2002. http://java.sun.com/products/jms/. [6] J. Hart. WebSphere MQ: Connecting your applications without complex programming. IBM WebSphere Software White Papers, 2003. [7] S. Hayward and M. Pezzini. Marrying Middleware and Mobile Computing. Gartner Group Research Report, September 2001. [8] IBM. WebSphere MQ EveryPlace Version 2.0, November 2002. http://www-3.ibm.com/software/integration/wmqe/. [9] ITU. Connecting remote communities. Documents of the World Summit on Information Society, 2003. http://www.itu.int/osg/spu/wsis-themes. [10] S. Maffeis. Introducing Wireless JMS. Softwired AG, www.sofwired-inc.com, 2002. [11] C. Mascolo, L. Capra, and W. Emmerich. Middleware for Mobile Computing. In E. Gregori, G. Anastasi, and S. Basagni, editors, Advanced Lectures on Networking, volume 2497 of Lecture Notes in Computer Science, pages 20-58. Springer Verlag, 2002. [12] Microsoft. Microsoft Message Queuing (MSMQ) Version 2.0 Documentation. [13] M. Musolesi, S. Hailes, and C. Mascolo. Adaptive routing for intermittently connected mobile ad-hoc networks. Technical report, UCL-CS Research Note, July 2004. Submitted for Publication. [14] Sun Microsystems. Java Naming and Directory Interface (JNDI) Documentation Version 1.2. 2003. http://java.sun.com/products/jndi/. [15] Sun Microsystems. Jini Specification Version 2.0, 2003. http://java.sun.com/products/jini/. [16] A. Vahdat and D. Becker. Epidemic routing for Partially Connected ad-hoc Networks. Technical Report CS-2000-06, Department of Computer Science, Duke University, 2000. [17] A. Vargas. The OMNeT++ discrete event simulation system. In Proceedings of the European Simulation Multiconference (ESM"2001), Prague, June 2001. [18] E. Vollset, D. Ingham, and P. Ezhilchelvan. JMS on Mobile ad-hoc Networks. In Personal Wireless Communications (PWC), pages 40-52, Venice, September 2003. [19] E. Yoneki and J. Bacon. Pronto: Mobilegateway with publish-subscribe paradigm over wireless network. Technical Report 559, University of Cambridge, Computer Laboratory, February 2003. Middleware for Pervasive and ad-hoc Computing 126

mobile ad-hoc network;asynchronous messaging middleware;context awareness;epidemic protocol;message-oriented middleware;message orient middleware;asynchronous communication;middleware for mobile computing;mobile ad-hoc environment;cross-layering;group communication;application level routing;java messaging service;epidemic messaging middleware

train_C-75

Composition of a DIDS by Integrating Heterogeneous IDSs on Grids

This paper considers the composition of a DIDS (Distributed Intrusion Detection System) by integrating heterogeneous IDSs (Intrusion Detection Systems). A Grid middleware is used for this integration. In addition, an architecture for this integration is proposed and validated through simulation.

1. INTRODUCTION Solutions for integrating heterogeneous IDSs (Intrusion Detection Systems) have been proposed by several groups [6],[7],[11],[2]. Some reasons for integrating IDSs are described by the IDWG (Intrusion Detection Working Group) from the IETF (Internet Engineering Task Force) [12] as follows: • Many IDSs available in the market have strong and weak points, which generally make necessary the deployment of more than one IDS to provided an adequate solution. • Attacks and intrusions generally originate from multiple networks spanning several administrative domains; these domains usually utilize different IDSs. The integration of IDSs is then needed to correlate information from multiple networks to allow the identification of distributed attacks and or intrusions. • The interoperability/integration of different IDS components would benefit the research on intrusion detection and speed up the deployment of IDSs as commercial products. DIDSs (Distributed Intrusion Detection Systems) therefore started to emerge in early 90s [9] to allow the correlation of intrusion information from multiple hosts, networks or domains to detect distributed attacks. Research on DIDSs has then received much interest, mainly because centralised IDSs are not able to provide the information needed to prevent such attacks [13]. However, the realization of a DIDS requires a high degree of coordination. Computational Grids are appealing as they enable the development of distributed application and coordination in a distributed environment. Grid computing aims to enable coordinate resource sharing in dynamic groups of individuals and/or organizations. Moreover, Grid middleware provides means for secure access, management and allocation of remote resources; resource information services; and protocols and mechanisms for transfer of data [4]. According to Foster et al. [4], Grids can be viewed as a set of aggregate services defined by the resources that they share. OGSA (Open Grid Service Architecture) provides the foundation for this service orientation in computational Grids. The services in OGSA are specified through well-defined, open, extensible and platformindependent interfaces, which enable the development of interoperable applications. This article proposes a model for integration of IDSs by using computational Grids. The proposed model enables heterogeneous IDSs to work in a cooperative way; this integration is termed DIDSoG (Distributed Intrusion Detection System on Grid). Each of the integrated IDSs is viewed by others as a resource accessed through the services that it exposes. A Grid middleware provides several features for the realization of a DIDSoG, including [3]: decentralized coordination of resources; use of standard protocols and interfaces; and the delivery of optimized QoS (Quality of Service). The service oriented architecture followed by Grids (OGSA) allows the definition of interfaces that are adaptable to different platforms. Different implementations can be encapsulated by a service interface; this virtualisation allows the consistent access to resources in heterogeneous environments [3]. The virtualisation of the environment through service interfaces allows the use of services without the knowledge of how they are actually implemented. This characteristic is important for the integration of IDSs as the same service interfaces can be exposed by different IDSs. Grid middleware can thus be used to implement a great variety of services. Some functions provided by Grid middleware are [3]: (i) data management services, including access services, replication, and localisation; (ii) workflow services that implement coordinate execution of multiple applications on multiple resources; (iii) auditing services that perform the detection of frauds or intrusions; (iv) monitoring services which implement the discovery of sensors in a distributed environment and generate alerts under determined conditions; (v) services for identification of problems in a distributed environment, which implement the correlation of information from disparate and distributed logs. These services are important for the implementation of a DIDSoG. A DIDS needs services for the location of and access to distributed data from different IDSs. Auditing and monitoring services take care of the proper needs of the DIDSs such as: secure storage, data analysis to detect intrusions, discovery of distributed sensors, and sending of alerts. The correlation of distributed logs is also relevant because the detection of distributed attacks depends on the correlation of the alert information generated by the different IDSs that compose the DIDSoG. The next sections of this article are organized as follows. Section 2 presents related work. The proposed model is presented in Section 3. Section 4 describes the development and a case study. Results and discussion are presented in Section 5. Conclusions and future work are discussed in Section 6. 2. RELATED WORK DIDMA [5] is a flexible, scalable, reliable, and platformindependent DIDS. DIDMA architecture allows distributed analysis of events and can be easily extended by developing new agents. However, the integration with existing IDSs and the development of security components are presented as future work [5]. The extensibility of DIDS DIDMA and the integration with other IDSs are goals pursued by DIDSoG. The flexibility, scalability, platform independence, reliability and security components discussed in [5] are achieved in DIDSoG by using a Grid platform. More efficient techniques for analysis of great amounts of data in wide scale networks based on clustering and applicable to DIDSs are presented in [13]. The integration of heterogeneous IDSs to increase the variety of intrusion detection techniques in the environment is mentioned as future work [13] DIDSoG thus aims at integrating heterogeneous IDSs [13]. Ref. [10] presents a hierarchical architecture for a DIDS; information is collected, aggregated, correlated and analysed as it is sent up in the hierarchy. The architecture comprises of several components for: monitoring, correlation, intrusion detection by statistics, detection by signatures and answers. Components in the same level of the hierarchy cooperate with one another. The integration proposed by DIDSoG also follows a hierarchical architecture. Each IDS integrated to the DIDSoG offers functionalities at a given level of the hierarchy and requests functionalities from IDSs from another level. The hierarchy presented in [10] integrates homogeneous IDSs whereas the hierarchical architecture of DIDSoG integrates heterogeneous IDSs. There are proposals on integrating computational Grids and IDSs [6],[7],[11],[2]. Ref. [6] and [7] propose the use of Globus Toolkit for intrusion detection, especially for DoS (Denial of Service) and DDoS (Distributed Denial of Service) attacks; Globus is used due to the need to process great amounts of data to detect these kinds of attack. A two-phase processing architecture is presented. The first phase aims at the detection of momentary attacks, while the second phase is concerned with chronic or perennial attacks. Traditional IDSs or DIDSs are generally coordinated by a central point; a characteristic that leaves them prone to attacks. Leu et al. [6] point out that IDSs developed upon Grids platforms are less vulnerable to attacks because of the distribution provided for such platforms. Leu et al. [6],[7] have used tools to generate several types of attacks - including TCP, ICMP and UDP flooding - and have demonstrated through experimental results the advantages of applying computational Grids to IDSs. This work proposes the development of a DIDS upon a Grid platform. However, the resulting DIDS integrates heterogeneous IDSs whereas the DIDSs upon Grids presented by Leu et al. [6][7] do not consider the integration of heterogeneous IDSs. The processing in phases [6][7] is also contemplated by DIDSoG, which is enabled by the specification of several levels of processing allowed by the integration of heterogeneous IDSs. The DIDS GIDA (Grid Intrusion Detection Architecture) targets at the detection of intrusions in a Grid environment [11]. GridSim Grid simulator was used for the validation of DIDS GIDA. Homogeneous resources were used to simplify the development [11]. However, the possibility of applying heterogeneous detection systems is left for future work Another DIDS for Grids is presented by Choon and Samsudim [2]. Scenarios demonstrating how a DIDS can execute on a Grid environment are presented. DIDSoG does not aim at detecting intrusions in a Grid environment. In contrast, DIDSoG uses the Grid to compose a DIDS by integrating specific IDSs; the resulting DIDS could however be used to identify attacks in a Grid environment. Local and distributed attacks can be detected through the integration of traditional IDSs while attacks particular to Grids can be detected through the integration of Grid IDSs. 3. THE PROPOSED MODEL DIDSoG presents a hierarchy of intrusion detection services; this hierarchy is organized through a two-dimensional vector defined by Scope:Complexity. The IDSs composing DIDSoG can be organized in different levels of scope or complexity, depending on its functionalities, the topology of the target environment and expected results. Figure 1 presents a DIDSoG composed by different intrusion detection services (i.e. data gathering, data aggregation, data correlation, analysis, intrusion response and management) provided by different IDSs. The information flow and the relationship between the levels of scope and complexity are presented in this figure. Information about the environment (host, network or application) is collected by Sensors located both in user 1"s and user 2"s computers in domain 1. The information is sent to both simple Analysers that act on the information from a single host (level 1:1), and to aggregation and correlation services that act on information from multiple hosts from the same domain (level 2:1). Simple Analysers in the first scope level send the information to more complex Analysers in the next levels of complexity (level 1: N). When an Analyser detects an intrusion, it communicates with Countermeasure and Monitoring services registered to its scope. An Analyser can invoke a Countermeasure service that replies to a detected attack, or informs a Monitoring service about the ongoing attack, so the administrator can act accordingly. Aggregation and correlation resources in the second scope receive information from Sensors from different users" computers (user 1"s and user 2"s) in the domain 1. These resources process the received information and send it to the analysis resources registered to the first level of complexity in the second scope (level 2:1). The information is also sent to the aggregation and correlation resources registered in the first level of complexity in the next scope (level 3:1). User 1 Domain 1 Analysers Level 1:1 Local Sensors Analysers Level 1:N Aggreg. Correlation Level 2:1 User 2 Domain 1 Local Sensors Analysers Level 2:1 Analysers Level 2:N Aggreg. Correlation Level 3:1 Domain 2 Monitor Level 1 Monitor Level 2 Analysers Level 3:1 Analysers Level 3:N Monitor Level 3 Response Level 1 Response Level 2 Response Level 3 Fig. 1. How DIDSoG works. The analysis resources in the second scope act like the analysis resources in the first scope, directing the information to a more complex analysis resource and putting the Countermeasure and Monitoring resources in action in case of detected attacks. Aggregation and correlation resources in the third scope receive information from domains 1 and 2. These resources then carry out the aggregation and correlation of the information from different domains and send it to the analysis resources in the first level of complexity in the third scope (level 3:1). The information could also be sent to the aggregate service in the next scope in case of any resources registered to such level. The analysis resources in the third scope act similar to the analysis resources in the first and second scopes, except that the analysis resources in the third scope act on information from multiple domains. The functionalities of the registered resources in each of the scopes and complexity level can vary from one environment to another. The model allows the development of N levels of scope and complexity. Figure 2 presents the architecture of a resource participating in the DIDSoG. Initially, the resource registers itself to GIS (Grid Information Service) so other participating resources can query the services provided. After registering itself, the resource requests information about other intrusion detection resources registered to the GIS. A given resource of DIDSoG interacts with other resources, by receiving data from the Source Resources, processing it, and sending the results to the Destination Resources, therefore forming a grid of intrusion detection resources. Grid Resource BaseNative IDS Grid Origin Resources Grid Destination Resources Grid Information Service Descri ptor Connec tor Fig. 2. Architecture of a resource participating of the DIDSoG. A resource is made up of four components: Base, Connector, Descriptor and Native IDS. Native IDS corresponds to the IDS being integrated to the DIDSoG. This component process the data received from the Origin Resources and generates new data to be sent to the Destination Resources. A Native IDS component can be any tool processes information related to intrusion detection, including analysis, data gathering, data aggregation, data correlation, intrusion response or management. The Descriptor is responsible for the information that identifies a resource and its respective Destination Resources in the DIDSoG. Figure 3 presents the class diagram of the stored information by the Descriptor. The ResourceDescriptor class has Feature, Level, DataType and Target Resources type members. Feature class represents the functionalities that a resource has. Type, name and version attributes refer to the functions offered by the Native IDS component, its name and version, respectively. Level class identifies the level of target and complexity in which the resource acts. DataType class represents the data format that the resource accepts to receive. DataType class is specialized by classes Text, XML and Binary. Class XML contains the DTDFile attribute to specify the DTD file that validates the received XML. -ident -version -description ResourceDescriptor -featureType -name -version Feature 1 1 -type -version DataType -escope -complex Level 1 1 Text Binary -DTDFile XML 1 1 TargetResources 1 1 10..* -featureType Resource11 1 1 Fig. 3. Class Diagram of the Descriptor component. TargetResources class represents the features of the Destination Resources of a determined resource. This class aggregates Resource. The Resource class identifies the characteristics of a Destination Resource. This identification is made through the featureType attribute and the Level and DataType classes. A given resource analyses the information from Descriptors from other resources, and compares this information with the information specified in TargetResources to know to which resources to send the results of its processing. The Base component is responsible for the communication of a resource with other resources of the DIDSoG and with the Grid Information Service. It is this component that registers the resource and the queries other resources in the GIS. The Connector component is the link between Base and Native IDS. The information that Base receives from Origin Resources is passed to Connector component. The Connector component performs the necessary changes in the data so that it is understood by Native IDS, and sends this data to Native IDS for processing. The Connector component also has the responsibility of collecting the information processed by Native IDS, and making the necessary changes so the information can pass through the DIDSoG again. After these changes, Connector sends the information to the Base, which in turn sends it to the Destination Resources in accordance with the specifications of the Descriptor component. 4. IMPLEMENTATION We have used GridSim Toolkit 3 [1] for development and evaluation of the proposed model. We have used and extended GridSim features to model and simulate the resources and components of DIDSoG. Figure 4 presents the Class diagram of the simulated DIDSoG. The Simulation_DIDSoG class starts the simulation components. The Simulation_User class represents a user of DIDSoG. This class" function is to initiate the processing of a resource Sensor, from where the gathered information will be sent to other resources. DIDSoG_GIS keeps a registry of the DIDSoG resources.The DIDSoG_BaseResource class implements the Base component (see Figure 2). DIDSoG_BaseResource interacts with DIDSoG_Descriptor class, which represents the Descriptor component. The DIDSoG_Descriptor class is created from an XML file that specifies a resource descriptor (see Figure 3). DIDSoG_BaseResource DIDSoG_Descriptor 11 DIDSoG_GIS Simulation_User Simulation_DIDSoG 1 *1* 1 1 GridInformationService GridSim GridResource Fig. 4. Class Diagram of the simulatated DIDSoG. A Connector component must be developed for each Native IDS integrated to DIDSoG. The Connector component is implemented by creating a class derived from DIDSoG_BaseResource. The new class will implement new functionalities in accordance with the needs of the corresponding Native IDS. In the simulation environment, data collection resources, analysis, aggregation/correlation and generation of answers were integrated. Classes were developed to simulate the processing of each Native IDS components associated to the resources. For each simulated Native IDS a class derived from DIDSoG_BaseResource was developed. This class corresponds to the Connector component of the Native IDS and aims at the integrating the IDS to DIDSoG. A XML file describing each of the integrated resources is chosen by using the Connector component. The resulting relationship between the resources integrated to the DIDSoG, in accordance with the specification of its respective descriptors, is presented in Figure 5. The Sensor_1 and Sensor_2 resources generate simulated data in the TCPDump [8] format. The generated data is directed to Analyser_1 and Aggreg_Corr_1 resources, in the case of Sensor_1, and to Aggreg_Corr_1 in the case of Sensor_2, according to the specification of their descriptors. User_1 Analyser_ 1 Level 1:1 Sensor_1 Aggreg_ Corr_1 Level 2:1 User_2 Sensor_2 Analyser_2 Level 2:1 Analyser_3 Level 2:2 TCPDump TCPDump TCPDumpAg TCPDumpAg IDMEF IDMEF IDMEF TCPDump Countermeasure_1 Level 1 Countermeasure_2 Level 2 Fig. 5. Flow of the execution of the simulation. The Native IDS of Analyser_1 generates alerts for any attempt of connection to port 23. The data received from Analyser_1 had presented such features, generating an IDMEF (Intrusion Detection Message Exchange Format) alert [14]. The generated alert was sent to Countermeasure_1 resource, where a warning was dispatched to the administrator informing him of the alert received. The Aggreg_Corr_1 resource received the information generated by sensors 1 and 2. Its processing activities consist in correlating the source IP addresses with the received data. The resultant information of the processing of Aggreg_Corr_1 was directed to the Analyser_2 resource. The Native IDS component of the Analyser_2 generates alerts when a source tries to connect to the same port number of multiple destinations. This situation is identified by the Analyser_2 in the data received from Aggreg_Corr_1 and an alert in IDMEF format is then sent to the Countermeasures_2 resource. In addition to generating alerts in IDMEF format, Analyser_2 also directs the received data to the Analyser_3, in the level of complexity 2. The Native IDS component of Analyser_3 generates alerts when the transmission of ICMP messages from a given source to multiple destinations is detected. This situation is detected in the data received from Analyser_2, and an IDMEF alert is then sent to the Countermeasure_2 resource. The Countermeasure_2 resource receives the alerts generated by analysers 3 and 2, in accordance with the implementation of its Native IDS component. Warnings on alerts received are dispatched to the administrator. The simulation carried out demonstrates how DIDSoG works. Simulated data was generated to be the input for a grid of intrusion detection systems composed by several distinct resources. The resources carry out tasks such as data collection, aggregation and analysis, and generation of alerts and warnings in an integrated manner. 5. EXPERIMENT RESULTS The hierarchic organization of scope and complexity provides a high degree of flexibility to the model. The DIDSoG can be modelled in accordance with the needs of each environment. The descriptors define data flow desired for the resulting DIDS. Each Native IDS is integrated to the DIDSoG through a Connector component. The Connector component is also flexible in the DIDSoG. Adaptations, conversions of data types and auxiliary processes that Native IDSs need are provided by the Connector. Filters and generation of Specific logs for each Native IDS or environment can also be incorporated to the Connector. If the integration of a new IDS to an environment already configured is desired, it is enough to develop the Connector for the desired IDS and to specify the resource Descriptor. After the specification of the Connector and the Descriptor the new IDS is integrated to the DIDSoG. Through the definition of scopes, resources can act on data of different source groups. For example, scope 1 can be related to a given set of hosts, scope 2 to another set of hosts, while scope 3 can be related to hosts from scopes 1 and 2. Scopes can be defined according to the needs of each environment. The complexity levels allow the distribution of the processing between several resources inside the same scope. In an analysis task, for example, the search for simple attacks can be made by resources of complexity 1, whereas the search for more complex attacks, that demands more time, can be performed by resources of complexity 2. With this, the analysis of the data is made by two resources. The distinction between complexity levels can also be organized in order to integrate different techniques of intrusion detection. The complexity level 1 could be defined for analyses based on signatures, which are simpler techniques; the complexity level 2 for techniques based on behaviour, that require greater computational power; and the complexity level 3 for intrusion detection in applications, where the techniques are more specific and depend on more data. The division of scopes and the complexity levels make the processing of the data to be carried out in phases. No resource has full knowledge about the complete data processing flow. Each resource only knows the results of its processing and the destination to which it sends the results. Resources of higher complexity must be linked to resources of lower complexity. Therefore, the hierarchic structure of the DIDSoG is maintained, facilitating its extension and integration with other domains of intrusion detection. By carrying out a hierarchic relationship between the several chosen analysers for an environment, the sensor resource is not overloaded with the task to send the data to all the analysers. An initial analyser will exist (complexity level 1) to which the sensor will send its data, and this analyser will then direct the data to the next step of the processing flow. Another feature of the hierarchical organization is the easy extension and integration with other domains. If it is necessary to add a new host (sensor) to the DIDSoG, it is enough to plug it to the first hierarchy of resources. If it is necessary to add a new analyser, it will be in the scope of several domains, it is enough to relate it to another resource of same scope. The DIDSoG allows different levels to be managed by different entities. For example, the first scope can be managed by the local user of a host. The second scope, comprising several hosts of a domain can be managed by the administrator of the domain. A third entity can be responsible for managing the security of several domains in a joint way. This entity can act in the scope 3 independently from others. With the proposed model for integration of IDSs in Grids, the different IDSs of an environment (or multiple IDSs integrated) act in a cooperative manner improving the intrusion detection services, mainly in two aspects. First, the information from multiple sources are analysed in an integrated way to search for distributed attacks. This integration can be made under several scopes. Second, there is a great diversity of data aggregation techniques, data correlation and analysis, and intrusion response that can be applied to the same environment; these techniques can be organized under several levels of complexity. 6. CONCLUSION The integration of heterogeneous IDSs is important. However, the incompatibility and diversity of IDS solutions make such integration extremely difficult. This work thus proposed a model for composition of DIDS by integrating existing IDSs on a computational Grid platform (DIDSoG). IDSs in DIDSoG are encapsulated as Grid services for intrusion detection. A computational Grid platform is used for the integration by providing the basic requirements for communication, localization, resource sharing and security mechanisms. The components of the architecture of the DIDSoG were developed and evaluated using the GridSim Grid simulator. Services for communication and localization were used to carry out the integration between components of different resources. Based on the components of the architecture, several resources were modelled forming a grid of intrusion detection. The simulation demonstrated the usefulness of the proposed model. Data from the sensor resources was read and this data was used to feed other resources of DIDSoG. The integration of distinct IDSs could be observed through the simulated environment. Resources providing different intrusion detection services were integrated (e.g. analysis, correlation, aggregation and alert). The communication and localization services provided by GridSim were used to integrate components of different resources. Various resources were modelled following the architecture components forming a grid of intrusion detection. The components of DIDSoG architecture have served as base for the integration of the resources presented in the simulation. During the simulation, the different IDSs cooperated with one another in a distributed manner; however, in a coordinated way with an integrated view of the events, having, thus, the capability to detect distributed attacks. This capability demonstrates that the IDSs integrated have resulted in a DIDS. Related work presents cooperation between components of a specific DIDS. Some work focus on either the development of DIDSs on computational Grids or the application of IDSs to computational Grids. However, none deals with the integration of heterogeneous IDSs. In contrast, the proposed model developed and simulated in this work, can shed some light into the question of integration of heterogeneous IDSs. DIDSoG presents new research opportunities that we would like to pursue, including: deployment of the model in a more realistic environment such as a Grid; incorporation of new security services; parallel analysis of data by Native IDSs in multiple hosts. In addition to the integration of IDSs enabled by a grid middleware, the cooperation of heterogeneous IDSs can be viewed as an economic problem. IDSs from different organizations or administrative domains need incentives for joining a grid of intrusion detection services and for collaborating with other IDSs. The development of distributed strategy proof mechanisms for integration of IDSs is a challenge that we would like to tackle. 7. REFERENCES [1] Sulistio, A.; Poduvaly, G.; Buyya, R; and Tham, CK, Constructing A Grid Simulation with Differentiated Network Service Using GridSim, Proc. of the 6th International Conference on Internet Computing (ICOMP'05), June 27-30, 2005, Las Vegas, USA. [2] Choon, O. T.; Samsudim, A. Grid-based Intrusion Detection System. The 9th IEEE Asia-Pacific Conference Communications, September 2003. [3] Foster, I.; Kesselman, C.; Tuecke, S. The Physiology of the Grid: An Open Grid Service Architecture for Distributed System Integration. Draft June 2002. Available at http://www.globus.org/research/papers/ogsa.pdf. Access Feb. 2006. [4] Foster, Ian; Kesselman, Carl; Tuecke, Steven. The anatomy of the Grid: enabling scalable virtual organizations. International Journal of Supercomputer Applications, 2001. [5] Kannadiga, P.; Zulkernine, M. DIDMA: A Distributed Intrusion Detection System Using Mobile Agents. Proceedings of the IEEE Sixth International Conference on Software Engineering, Artificial Intelligence, Networking and Parallel/Distributed Computing, May 2005. [6] Leu, Fang-Yie, et al. Integrating Grid with Intrusion Detection. Proceedings of 19th IEEE AINA"05, March 2005. [7] Leu, Fang-Yie, et al. A Performance-Based Grid Intrusion Detection System. Proceedings of the 29th IEEE COMPSAC"05, July 2005. [8] McCanne, S.; Leres, C.; Jacobson, V.; TCPdump/Libpcap, http://www.tcpdump.org/, 1994. [9] Snapp, S. R. et al. DIDS (Distributed Intrusion Detection System) - Motivation, Architecture and An Early Prototype. Proceeding of the Fifteenth IEEE National Computer Security Conference. Baltimore, MD, October 1992. [10] Sterne, D.; et al. A General Cooperative Intrusion Detection Architecture for MANETs. Proceedings of the Third IEEE IWIA"05, March 2005. [11] Tolba, M. F. ; et al. GIDA: Toward Enabling Grid Intrusion Detection Systems. 5th IEEE International Symposium on Cluster Computing and the Grid, May 2005. [12] Wood, M. Intrusion Detection message exchange requirements. Draft-ietf-idwg-requirements-10, October 2002. Available at http://www.ietf.org/internet-drafts/draftietf-idwg-requirements-10.txt. Access March 2006. [13] Zhang, Yu-Fang; Xiong, Z.; Wang, X. Distributed Intrusion Detection Based on Clustering. Proceedings of IEEE International Conference Machine Learning and Cybernetics, August 2005. [14] Curry, D.; Debar, H. Intrusion Detection Message exchange format data model and Extensible Markup Language (XML) Document Type Definition. Draft-ietf-idwg-idmef-xml-10, March 2006. Available at http://www.ietf.org/internetdrafts/draft-ietf-idwg-idmef-xml-16.txt.

system integration;gridsim grid simulator;heterogeneous intrusion detection system;distributed intrusion detection system;grid middleware;intrusion detection system;grid service for intrusion detection;ids integration;grid;open grid service architecture;integration of ids;computational grid;grid intrusion detection architecture;intrusion detection service

train_C-76

Assured Service Quality by Improved Fault Management Service-Oriented Event Correlation

The paradigm shift from device-oriented to service-oriented management has also implications to the area of event correlation. Today"s event correlation mainly addresses the correlation of events as reported from management tools. However, a correlation of user trouble reports concerning services should also be performed. This is necessary to improve the resolution time and to reduce the effort for keeping the service agreements. We refer to such a type of correlation as service-oriented event correlation. The necessity to use this kind of event correlation is motivated in the paper. To introduce service-oriented event correlation for an IT service provider, an appropriate modeling of the correlation workflow and of the information is necessary. Therefore, we examine the process management frameworks IT Infrastructure Library (ITIL) and enhanced Telecom Operations Map (eTOM) for their contribution to the workflow modeling in this area. The different kinds of dependencies that we find in our general scenario are then used to develop a workflow for the service-oriented event correlation. The MNM Service Model, which is a generic model for IT service management proposed by the Munich Network Management (MNM) Team, is used to derive an appropriate information modeling. An example scenario, the Web Hosting Service of the Leibniz Supercomputing Center (LRZ), is used to demonstrate the application of service-oriented event correlation.

1. INTRODUCTION In huge networks a single fault can cause a burst of failure events. To handle the flood of events and to find the root cause of a fault, event correlation approaches like rule-based reasoning, case-based reasoning or the codebook approach have been developed. The main idea of correlation is to condense and structure events to retrieve meaningful information. Until now, these approaches address primarily the correlation of events as reported from management tools or devices. Therefore, we call them device-oriented. In this paper we define a service as a set of functions which are offered by a provider to a customer at a customer provider interface. The definition of a service is therefore more general than the definition of a Web Service, but a Web Service is included in this service definition. As a consequence, the results are applicable for Web Services as well as for other kinds of services. A service level agreement (SLA) is defined as a contract between customer and provider about guaranteed service performance. As in today"s IT environments the offering of such services with an agreed service quality becomes more and more important, this change also affects the event correlation. It has become a necessity for providers to offer such guarantees for a differentiation from other providers. To avoid SLA violations it is especially important for service providers to identify the root cause of a fault in a very short time or even act proactively. The latter refers to the case of recognizing the influence of a device breakdown on the offered services. As in this scenario the knowledge about services and their SLAs is used we call it service-oriented. It can be addressed from two directions. Top-down perspective: Several customers report a problem in a certain time interval. Are these trouble reports correlated? How to identify a resource as being the problem"s root cause? 183 Bottom-up perspective: A device (e.g., router, server) breaks down. Which services, and especially which customers, are affected by this fault? The rest of the paper is organized as follows. In Section 2 we describe how event correlation is performed today and present a selection of the state-of-the-art event correlation techniques. Section 3 describes the motivation for serviceoriented event correlation and its benefits. After having motivated the need for such type of correlation we use two well-known IT service management models to gain requirements for an appropriate workflow modeling and present our proposal for it (see Section 4). In Section 5 we present our information modeling which is derived from the MNM Service Model. An application of the approach for a web hosting scenario is performed in Section 6. The last section concludes the paper and presents future work. 2. TODAY"S EVENT CORRELATION TECHNIQUES In [11] the task of event correlation is defined as a conceptual interpretation procedure in the sense that a new meaning is assigned to a set of events that happen in a certain time interval. We can distinguish between three aspects for event correlation. Functional aspect: The correlation focuses on functions which are provided by each network element. It is also regarded which other functions are used to provide a specific function. Topology aspect: The correlation takes into account how the network elements are connected to each other and how they interact. Time aspect: When explicitly regarding time constraints, a start and end time has to be defined for each event. The correlation can use time relationships between the events to perform the correlation. This aspect is only mentioned in some papers [11], but it has to be treated in an event correlation system. In the event correlation it is also important to distinguish between the knowledge acquisition/representation and the correlation algorithm. Examples of approaches to knowledge acquisition/representation are Gruschke"s dependency graphs [6] and Ensel"s dependency detection by neural networks [3]. It is also possible to find the dependencies by analyzing interactions [7]. In addition, there is an approach to manage service dependencies with XML and to define a resource description framework [4]. To get an overview about device-oriented event correlation a selection of several event correlation techniques being used for this kind of correlation is presented. Model-based reasoning: Model-based reasoning (MBR) [15, 10, 20] represents a system by modeling each of its components. A model can either represent a physical entity or a logical entity (e.g., LAN, WAN, domain, service, business process). The models for physical entities are called functional model, while the models for all logical entities are called logical model. A description of each model contains three categories of information: attributes, relations to other models, and behavior. The event correlation is a result of the collaboration among models. As all components of a network are represented with their behavior in the model, it is possible to perform simulations to predict how the whole network will behave. A comparison in [20] showed that a large MBR system is not in all cases easy to maintain. It can be difficult to appropriately model the behavior for all components and their interactions correctly and completely. An example system for MBR is NetExpert[16] from OSI which is a hybrid MBR/RBR system (in 2000 OSI was acquired by Agilent Technologies). Rule-based reasoning: Rule-based reasoning (RBR) [15, 10] uses a set of rules for event correlation. The rules have the form conclusion if condition. The condition uses received events and information about the system, while the conclusion contains actions which can either lead to system changes or use system parameters to choose the next rule. An advantage of the approach is that the rules are more or less human-readable and therefore their effect is intuitive. The correlation has proved to be very fast in practice by using the RETE algorithm. In the literature [20, 1] it is claimed that RBR systems are classified as relatively inflexible. Frequent changes in the modeled IT environment would lead to many rule updates. These changes would have to be performed by experts as no automation has currently been established. In some systems information about the network topology which is needed for the event correlation is not used explicitly, but is encoded into the rules. This intransparent usage would make rule updates for topology changes quite difficult. The system brittleness would also be a problem for RBR systems. It means that the system fails if an unknown case occurs, because the case cannot be mapped onto similar cases. The output of RBR systems would also be difficult to predict, because of unforeseen rule interactions in a large rule set. According to [15] an RBR system is only appropriate if the domain for which it is used is small, nonchanging, and well understood. The GTE IMPACT system [11] is an example of a rulebased system. It also uses MBR (GTE has merged with Bell Atlantic in 1998 and is now called Verizon [19]). Codebook approach: The codebook approach [12, 21] has similarities to RBR, but takes a further step and encodes the rules into a correlation matrix. The approach starts using a dependency graph with two kinds of nodes for the modeling. The first kind of nodes are the faults (denoted as problems in the cited papers) which have to be detected, while the second kind of nodes are observable events (symptoms in the papers) which are caused by the faults or other events. The dependencies between the nodes are denoted as directed edges. It is possible to choose weights for the edges, e.g., a weight for the probability that 184 fault/event A causes event B. Another possible weighting could indicate time dependencies. There are several possibilities to reduce the initial graph. If, e.g., a cyclic dependency of events exists and there are no probabilities for the cycles" edges, all events can be treated as one event. After a final input graph is chosen, the graph is transformed into a correlation matrix where the columns contain the faults and the rows contain the events. If there is a dependency in the graph, the weight of the corresponding edge is put into the according matrix cell. In case no weights are used, the matrix cells get the values 1 for dependency and 0 otherwise. Afterwards, a simplification can be done, where events which do not help to discriminate faults are deleted. There is a trade-off between the minimization of the matrix and the robustness of the results. If the matrix is minimized as much as possible, some faults can only be distinguished by a single event. If this event cannot be reliably detected, the event correlation system cannot discriminate between the two faults. A measure how many event observation errors can be compensated by the system is the Hamming distance. The number of rows (events) that can be deleted from the matrix can differ very much depending on the relationships [15]. The codebook approach has the advantage that it uses long-term experience with graphs and coding. This experience is used to minimize the dependency graph and to select an optimal group of events with respect to processing time and robustness against noise. A disadvantage of the approach could be that similar to RBR frequent changes in the environment make it necessary to frequently edit the input graph. SMARTS InCharge [12, 17] is an example of such a correlation system. Case-based reasoning: In contrast to other approaches case-based reasoning (CBR) [14, 15] systems do not use any knowledge about the system structure. The knowledge base saves cases with their values for system parameters and successful recovery actions for these cases. The recovery actions are not performed by the CBR system in the first place, but in most cases by a human operator. If a new case appears, the CBR system compares the current system parameters with the system parameters in prior cases and tries to find a similar one. To identify such a match it has to be defined for which parameters the cases can differ or have to be the same. If a match is found, a learned action can be performed automatically or the operator can be informed with a recovery proposal. An advantage of this approach is that the ability to learn is an integral part of it which is important for rapid changing environments. There are also difficulties when applying the approach [15]. The fields which are used to find a similar case and their importance have to be defined appropriately. If there is a match with a similar case, an adaptation of the previous solution to the current one has to be found. An example system for CBR is SpectroRx from Cabletron Systems. The part of Cabletron that developed SpectroRx became an independent software company in 2002 and is now called Aprisma Management Technologies [2]. In this section four event correlation approaches were presented which have evolved into commercial event correlation systems. The correlation approaches have different focuses. MBR mainly deals with the knowledge acquisition and representation, while RBR and the codebook approach propose a correlation algorithm. The focus of CBR is its ability to learn from prior cases. 3. MOTIVATION OF SERVICE-ORIENTED EVENT CORRELATION Fig. 1 shows a general service scenario upon which we will discuss the importance of a service-oriented correlation. Several services like SSH, a web hosting service, or a video conference service are offered by a provider to its customers at the customer provider interface. A customer can allow several users to use a subscribed service. The quality and cost issues of the subscribed services between a customer and a provider are agreed in SLAs. On the provider side the services use subservices for their provisioning. In case of the services mentioned above such subservices are DNS, proxy service, and IP service. Both services and subservices depend on resources upon which they are provisioned. As displayed in the figure a service can depend on more than one resource and a resource can be used by one or more services. SSH DNS proxy IP service dependency resource dependency user a user b user c customer SLA web a video conf. SSH sun1 provider video conf. web services subservices resources Figure 1: Scenario To get a common understanding, we distinguish between different types of events: Resource event: We use the term resource event for network events and system events. A network event refers to events like node up/down or link up/down whereas system events refer to events like server down or authentication failure. Service event: A service event indicates that a service does not work properly. A trouble ticket which is generated from a customer report is a kind of such an 185 event. Other service events can be generated by the provider of a service, if the provider himself detects a service malfunction. In such a scenario the provider may receive service events from customers which indicate that SSH, web hosting service, and video conference service are not available. When referring to the service hierarchy, the provider can conclude in such a case that all services depend on DNS. Therefore, it seems more likely that a common resource which is necessary for this service does not work properly or is not available than to assume three independent service failures. In contrast to a resource-oriented perspective where all of the service events would have to be processed separately, the service events can be linked together. Their information can be aggregated and processed only once. If, e.g., the problem is solved, one common message to the customers that their services are available again is generated and distributed by using the list of linked service events. This is certainly a simplified example. However, it shows the general principle of identifying the common subservices and common resources of different services. It is important to note that the service-oriented perspective is needed to integrate service aspects, especially QoS aspects. An example of such an aspect is that a fault does not lead to a total failure of a service, but its QoS parameters, respectively agreed service levels, at the customer-provider interface might not be met. A degradation in service quality which is caused by high traffic load on the backbone is another example. In the resource-oriented perspective it would be possible to define events which indicate that there is a link usage higher than a threshold, but no mechanism has currently been established to find out which services are affected and whether a QoS violation occurs. To summarize, the reasons for the necessity of a serviceoriented event correlation are the following: Keeping of SLAs (top-down perspective): The time interval between the first symptom (recognized either by provider, network management tools, or customers) that a service does not perform properly and the verified fault repair needs to be minimized. This is especially needed with respect to SLAs as such agreements often contain guarantees like a mean time to repair. Effort reduction (top-down perspective): If several user trouble reports are symptoms of the same fault, fault processing should be performed only once and not several times. If the fault has been repaired, the affected customers should be informed about this automatically. Impact analysis (bottom-up perspective): In case of a fault in a resource, its influence on the associated services and affected customers can be determined. This analysis can be performed for short term (when there is currently a resource failure) or long term (e.g., network optimization) considerations. 4. WORKFLOW MODELING In the following we examine the established IT process management frameworks IT Infrastructure Library (ITIL) and enhanced Telecom Operations Map (eTOM). The aim is find out where event correlation can be found in the process structure and how detailed the frameworks currently are. After that we present our solution for a workflow modeling for the service-oriented event correlation. 4.1 IT Infrastructure Library (ITIL) The British Office of Government Commerce (OGC) and the IT Service Management Forum (itSMF) [9] provide a collection of best practices for IT processes in the area of IT service management which is called ITIL. The service management is described by 11 modules which are grouped into Service Support Set (provider internal processes) and Service Delivery Set (processes at the customer-provider interface). Each module describes processes, functions, roles, and responsibilities as well as necessary databases and interfaces. In general, ITIL describes contents, processes, and aims at a high abstraction level and contains no information about management architectures and tools. The fault management is divided into Incident Management process and Problem Management process. Incident Management: The Incident Management contains the service desk as interface to customers (e.g., receives reports about service problems). In case of severe errors structured queries are transferred to the Problem Management. Problem Management: The Problem Management"s tasks are to solve problems, take care of keeping priorities, minimize the reoccurrence of problems, and to provide management information. After receiving requests from the Incident Management, the problem has to be identified and information about necessary countermeasures is transferred to the Change Management. The ITIL processes describe only what has to be done, but contain no information how this can be actually performed. As a consequence, event correlation is not part of the modeling. The ITIL incidents could be regarded as input for the service-oriented event correlation, while the output could be used as a query to the ITIL Problem Management. 4.2 Enhanced Telecom Operations Map (eTOM) The TeleManagement Forum (TMF) [18] is an international non-profit organization from service providers and suppliers in the area of telecommunications services. Similar to ITIL a process-oriented framework has been developed at first, but the framework was designed for a narrower focus, i.e., the market of information and communications service providers. A horizontal grouping into processes for customer care, service development & operations, network & systems management, and partner/supplier is performed. The vertical grouping (fulfillment, assurance, billing) reflects the service life cycle. In the area of fault management three processes have been defined along the horizontal process grouping. Problem Handling: The purpose of this process is to receive trouble reports from customers and to solve them by using the Service Problem Management. The aim is also to keep the customer informed about the current status of the trouble report processing as well as about the general network status (e.g., planned maintenance). It is also a task of this process to inform the 186 QoS/SLA management about the impact of current errors on the SLAs. Service Problem Management: In this process reports about customer-affecting service failures are received and transformed. Their root causes are identified and a problem solution or a temporary workaround is established. The task of the Diagnose Problem subprocess is to find the root cause of the problem by performing appropriate tests. Nothing is said how this can be done (e.g., no event correlation is mentioned). Resource Trouble Management: A subprocess of the Resource Trouble Management is responsible for resource failure event analysis, alarm correlation & filtering, and failure event detection & reporting. Another subprocess is used to execute different tests to find a resource failure. There is also another subprocess which keeps track about the status of the trouble report processing. This subprocess is similar to the functionality of a trouble ticket system. The process description in eTOM is not very detailed. It is useful to have a check list which aspects for these processes have to be taken into account, but there is no detailed modeling of the relationships and no methodology for applying the framework. Event correlation is only mentioned in the resource management, but it is not used in the service level. 4.3 Workflow Modeling for the Service-Oriented Event Correlation Fig. 2 shows a general service scenario which we will use as basis for the workflow modeling for the service-oriented event correlation. We assume that the dependencies are already known (e.g., by using the approaches mentioned in Section 2). The provider offers different services which depend on other services called subservices (service dependency). Another kind of dependency exists between services/subservices and resources. These dependencies are called resource dependencies. These two kinds of dependencies are in most cases not used for the event correlation performed today. This resource-oriented event correlation deals only with relationships on the resource level (e.g., network topology). service dependency resources subservices provider services resource dependency Figure 2: Different kinds of dependencies for the service-oriented event correlation The dependencies depicted in Figure 2 reflect a situation with no redundancy in the service provisioning. The relationships can be seen as AND relationships. In case of redundancy, if e.g., a provider has 3 independent web servers, another modeling (see Figure 3) should be used (OR relationship). In such a case different relationships are possible. The service could be seen as working properly if one of the servers is working or a certain percentage of them is working. services ) AND relationship b) OR relationship resources Figure 3: Modeling of no redundancy (a) and of redundancy (b) As both ITIL and eTOM contain no description how event correlation and especially service-oriented event correlation should actually be performed, we propose the following design for such a workflow (see Fig. 4). The additional components which are not part of a device-oriented event correlation are depicted with a gray background. The workflow is divided into the phases fault detection, fault diagnosis, and fault recovery. In the fault detection phase resource events and service events can be generated from different sources. The resource events are issued during the use of a resource, e.g., via SNMP traps. The service events are originated from customer trouble reports, which are reported via the Customer Service Management (see below) access point. In addition to these two passive ways to get the events, a provider can also perform active tests. These tests can either deal with the resources (resource active probing) or can assume the role of a virtual customer and test a service or one of its subservices by performing interactions at the service access points (service active probing). The fault diagnosis phase is composed of three event correlation steps. The first one is performed by the resource event correlator which can be regarded as the event correlator in today"s commercial systems. Therefore, it deals only with resource events. The service event correlator does a correlation of the service events, while the aggregate event correlator finally performs a correlation of both resource and service events. If the correlation result in one of the correlation steps shall be improved, it is possible to go back to the fault detection phase and start the active probing to get additional events. These events can be helpful to confirm a correlation result or to reduce the list of possible root causes. After the event correlation an ordered list of possible root causes is checked by the resource management. When the root cause is found, the failure repair starts. This last step is performed in the fault recovery phase. The next subsections present different elements of the event correlation process. 4.4 Customer Service Management and Intelligent Assistant The Customer Service Management (CSM) access point was proposed by [13] as a single interface between customer 187 fault detection fault diagnosis resource active probing resource event resource event correlator resource management candidate list fault recovery resource usage service active probing intelligent assistant service event correlator aggregate event correlator service eventCSM AP Figure 4: Event correlation workflow and provider. Its functionality is to provide information to the customer about his subscribed services, e.g., reports about the fulfillment of agreed SLAs. It can also be used to subscribe services or to allow the customer to manage his services in a restricted way. Reports about problems with a service can be sent to the customer via CSM. The CSM is also contained in the MNM Service Model (see Section 5). To reduce the effort for the provider"s first level support, an Intelligent Assistant can be added to the CSM. The Intelligent Assistant structures the customer"s information about a service problem. The information which is needed for a preclassification of the problem is gathered from a list of questions to the customer. The list is not static as the current question depends on the answers to prior questions or from the result of specific tests. A decision tree is used to structure the questions and tests. The tests allow the customer to gain a controlled access to the provider"s management. At the LRZ a customer of the E-Mail Service can, e.g., use the Intelligent Assistant to perform ping requests to the mail server. But also more complex requests could be possible, e.g., requests of a combination of SNMP variables. 4.5 Active Probing Active probing is useful for the provider to check his offered services. The aim is to identify and react to problems before a customer notices them. The probing can be done from a customer point of view or by testing the resources which are part of the services. It can also be useful to perform tests of subservices (own subservices or subservices offered by suppliers). Different schedules are possible to perform the active probing. The provider could select to test important services and resources in regular time intervals. Other tests could be initiated by a user who traverses the decision tree of the Intelligent Assistant including active tests. Another possibility for the use of active probing is a request from the event correlator, if the current correlation result needs to be improved. The results of active probing are reported via service or resource events to the event correlator (or if the test was demanded by the Intelligent Assistant the result is reported to it, too). While the events that are received from management tools and customers denote negative events (something does not work), the events from active probing should also contain positive events for a better discrimination. 4.6 Event Correlator The event correlation should not be performed by a single event correlator, but by using different steps. The reason for this are the different characteristics of the dependencies (see Fig. 1). On the resource level there are only relationships between resources (network topology, systems configuration). An example for this could be a switch linking separate LANs. If the switch is down, events are reported that other network components which are located behind the switch are also not reachable. When correlating these events it can be figured out that the switch is the likely error cause. At this stage, the integration of service events does not seem to be helpful. The result of this step is a list of resources which could be the problem"s root cause. The resource event correlator is used to perform this step. In the service-oriented scenario there are also service and resource dependencies. As next step in the event correlation process the service events should be correlated with each other using the service dependencies, because the service dependencies have no direct relationship to the resource level. The result of this step, which is performed by the service event correlator, is a list of services/subservices which could contain a failure in a resource. If, e.g., there are service events from customers that two services do not work and both services depend on a common subservice, it seems more likely that the resource failure can be found inside the subservice. The output of this correlation is a list of services/subservices which could be affected by a failure in an associated resource. In the last step the aggregate event correlator matches the lists from resource event correlator and service event correlator to find the problem"s possible root cause. This is done by using the resource dependencies. The event correlation techniques presented in Section 2 could be used to perform the correlation inside the three event correlators. If the dependencies can be found precisely, an RBR or codebook approach seems to be appropriate. A case database (CBR) could be used if there are cases which could not be covered by RBR or the codebook approach. These cases could then be used to improve the modeling in a way that RBR or the codebook approach can deal with them in future correlations. 5. INFORMATION MODELING In this section we use a generic model for IT service management to derive the information necessary for the event correlation process. 5.1 MNM Service Model The MNM Service Model [5] is a generic model for IT service management. A distinction is made between customer side and provider side. The customer side contains the basic roles customer and user, while the provider side contains the role provider. The provider makes the service available to the customer side. The service as a whole is divided into usage which is accessed by the role user and management which is used by the role customer. The model consists of two main views. The Service View (see Fig. 5) shows a common perspective of the service for customer and provider. Everything that is only important 188 for the service realization is not contained in this view. For these details another perspective, the Realization View, is defined (see Fig. 6). customer domain supplies supplies provider domain «role» provider accesses uses concludes accesses implements observesrealizes provides directs substantiates usesuses manages implementsrealizes manages service concludes QoS parameters usage functionality service access point management functionality service implementation service management implementation service agreement customersideprovidersidesideindependent «role» user «role» customer CSM access point service client CSM client Figure 5: Service View The Service View contains the service for which the functionality is defined for usage as well as for management. There are two access points (service access point and CSM access point) where user and customer can access the usage and management functionality, respectively. Associated to each service is a list of QoS parameters which have to be met by the service at the service access point. The QoS surveillance is performed by the management. provider domain implements observesrealizes provides directs implementsrealizes accesses uses concludes accesses usesuses manages side independent side independent manages manages concludes acts as service implementation service management implementation manages uses acts as service logic sub-service client service client CSM client uses resources usesuses service management logic sub-service management client basic management functionality «role» customer «role» provider «role» user Figure 6: Realization View In the Realization View the service implementation and the service management implementation are described in detail. For both there are provider-internal resources and subservices. For the service implementation a service logic uses internal resources (devices, knowledge, staff) and external subservices to provide the service. Analogous, the service management implementation includes a service management logic using basic management functionalities [8] and external management subservices. The MNM Service Model can be used for a similar modeling of the used subservices, i.e., the model can be applied recursively. As the service-oriented event correlation has to use dependencies of a service from subservices and resources, the model is used in the following to derive the needed information for service events. 5.2 Information Modeling for Service Events Today"s event correlation deals mainly with events which are originated from resources. Beside a resource identifier these events contain information about the resource status, e.g., SNMP variables. To perform a service-oriented event correlation it is necessary to define events which are related to services. These events can be generated from the provider"s own service surveillance or from customer reports at the CSM interface. They contain information about the problems with the agreed QoS. In our information modeling we define an event superclass which contains common attributes (e.g., time stamp). Resource event and service event inherit from this superclass. Derived from the MNM Service Model we define the information necessary for a service event. Service: As a service event shall represent the problems of a single service, a unique identification of the affected service is contained here. Event description: This field has to contain a description of the problem. Depending on the interactions at the service access point (Service View) a classification of the problem into different categories should be defined. It should also be possible to add an informal description of the problem. QoS parameters: For each service QoS parameters (Service View) are defined between the provider and the customer. This field represents a list of these QoS parameters and agreed service levels. The list can help the provider to set the priority of a problem with respect to the service levels agreed. Resource list: This list contains the resources (Realization View) which are needed to provide the service. This list is used by the provider to check if one of these resources causes the problem. Subservice service event identification: In the service hierarchy (Realization View) the service, for which this service event has been issued, may depend on subservices. If there is a suspicion that one of these subservices causes the problem, child service events are issued from this service event for the subservices. In such a case this field contains links to the corresponding events. Other event identifications: In the event correlation process the service event can be correlated with other service events or with resource events. This field then contains links to other events which have been correlated to this service event. This is useful to, e.g., send a common message to all affected customers when their subscribed services are available again. Issuer"s identification: This field can either contain an identification of the customer who reported the problem, an identification of a service provider"s employee 189 (in case the failure has been detected by the provider"s own service active probing) or a link to a parent service event. The identification is needed, if there are ambiguities in the service event or the issuer should be informed (e.g., that the service is available again). The possible issuers refer to the basic roles (customer, provider) in the Service Model. Assignee: To keep track of the processing the name and address of the provider"s employee who is solving or solved the problem is also noted. This is a specialization of the provider role in the Service Model. Dates: This field contains key dates in the processing of the service event such as initial date, problem identification date, resolution date. These dates are important to keep track how quick the problems have been solved. Status: This field represents the service event"s actual status (e.g., active, suspended, solved). Priority: The priority shows which importance the service event has from the provider"s perspective. The importance is derived from the service agreement, especially the agreed QoS parameters (Service View). The fields date, status, and other service events are not derived directly from the Service Model, but are necessary for the event correlation process. 6. APPLICATION OF SERVICE-ORIENTED EVENT CORRELATION FOR A WEB HOSTING SCENARIO The Leibniz Supercomputing Center is the joint computing center for the Munich universities and research institutions. It also runs the Munich Scientific Network and offers related services. One of these services is the Virtual WWW Server, a web hosting offer for smaller research institutions. It currently has approximately 200 customers. A subservice of the Virtual WWW Server is the Storage Service which stores the static and dynamic web pages and uses caching techniques for a fast access. Other subservices are DNS and IP service. When a user accesses a hosted web site via one of the LRZ"s Virtual Private Networks the VPN service is also used. The resources of the Virtual WWW Server include a load balancer and 5 redundant servers. The network connections are also part of the resources as well as the Apache web server application running on the servers. Figure 7 shows the dependencies of the Virtual WWW Server. 6.1 Customer Service Management and Intelligent Assistant The Intelligent Assistant that is available at the Leibniz Supercomputing Center can currently be used for connectivity or performance problems or problems with the LRZ E-Mail Service. A selection of possible customer problem reports for the Virtual WWW Server is given in the following: • The hosted web site is not reachable. • The web site access is (too) slow. • The web site contains outdated content. server serverserver server server server server server server outgoing connection hosting of LRZ"s own pages content caching server emergency server webmail server dynamic web pages static web pages DNS ProxyIP Storage Resources: Services: Virtual WWW Server five redundant servers AFS NFS DBload balancer Figure 7: Dependencies of the Virtual WWW Server • The transfer of new content to the LRZ does not change the provided content. • The web site looks strange (e.g., caused by problems with HTML version) This customer reports have to be mapped onto failures in resources. For, e.g., an unreachable web site different root causes are possible like a DNS problem, connectivity problem, wrong configuration of the load balancer. 6.2 Active Probing In general, active probing can be used for services or resources. For the service active probing of the Virtual WWW Server a virtual customer could be installed. This customer does typical HTTP requests of web sites and compares the answer with the known content. To check the up-to-dateness of a test web site, the content could contain a time stamp. The service active probing could also include the testing of subservices, e.g., sending requests to the DNS. The resource active probing performs tests of the resources. Examples are connectivity tests, requests to application processes, and tests of available disk space. 6.3 Event Correlation for the Virtual WWW Server Figure 8 shows the example processing. At first, a customer who takes a look at his hosted web site reports that the content that he had changed is not displayed correctly. This report is transferred to the service management via the CSM interface. An Intelligent Assistant could be used to structure the customer report. The service management translates the customer report into a service event. Independent from the customer report the service provider"s own service active probing tries to change the content of a test web site. Because this is not possible, a service event is issued. Meanwhile, a resource event has been reported to the event correlator, because an access of the content caching server to one of the WWW servers failed. As there are no other events at the moment the resource event correlation 190 customer CSM service mgmt event correlator resource mgmt customer reports: "web site content not up−to−date" service active probing reports: "web site content change not possible" event: "retrieval of server content failed"event forward resource event correlation service event correlation aggregate event correlation link failure report event forward check WWW server check link result display link repair result display result forward customer report Figure 8: Example processing of a customer report cannot correlate this event to other events. At this stage it would be possible that the event correlator asks the resource management to perform an active probing of related resources. Both service events are now transferred to the service event correlator and are correlated. From the correlation of these events it seems likely that either the WWW server itself or the link to the WWW server is the problem"s root cause. A wrong web site update procedure inside the content caching server seems to be less likely as this would only explain the customer report and not the service active probing result. At this stage a service active probing could be started, but this does not seem to be useful as this correlation only deals with the Web Hosting Service and its resources and not with other services. After the separate correlation of both resource and service events, which can be performed in parallel, the aggregate event correlator is used to correlate both types of events. The additional resource event makes it seem much more likely that the problems are caused by a broken link to the WWW server or by the WWW server itself and not by the content caching server. In this case the event correlator asks the resource management to check the link and the WWW server. The decision between these two likely error causes can not be further automated here. Later, the resource management finds out that a broken link is the failure"s root cause. It informs the event correlator about this and it can be determined that this explains all previous events. Therefore, the event correlation can be stopped at this point. Depending on the provider"s customer relationship management the finding of the root cause and an expected repair time could be reported to the customers. After the link has been repaired, it is possible to report this event via the CSM interface. Even though many details of this event correlation process could also be performed differently, the example showed an important advantage of the service-oriented event correlation. The relationship between the service provisioning and the provider"s resources is explicitly modeled. This allows a mapping of the customer report onto the provider-internal resources. 6.4 Event Correlation for Different Services If a provider like the LRZ offers several services the serviceoriented event correlation can be used to reveal relationships that are not obvious in the first place. If the LRZ E-Mail Service and its events are viewed in relationship with the events for the Virtual WWW Server, it is possible to identify failures in common subservices and resources. Both services depend on the DNS which means that customer reports like I cannot retrieve new e-mail and The web site of my research institute is not available can have a common cause, e.g., the DNS does not work properly. 7. CONCLUSION AND FUTURE WORK In our paper we showed the need for a service-oriented event correlation. For an IT service provider this new kind of event correlation makes it possible to automatically map problems with the current service quality onto resource failures. This helps to find the failure"s root cause earlier and to reduce costs for SLA violations. In addition, customer reports can be linked together and therefore the processing effort can be reduced. To receive these benefits we presented our approach for performing the service-oriented event correlation as well as a modeling of the necessary correlation information. In the future we are going to apply our workflow and information modeling for services offered by the Leibniz Supercomputing Center going further into details. Several issues have not been treated in detail so far, e.g., the consequences for the service-oriented event correlation if a subservice is offered by another provider. If a service does not perform properly, it has to be determined whether this is caused by the provider himself or by the subservice. In the latter case appropriate information has to be exchanged between the providers via the CSM interface. Another issue is the use of active probing in the event correlation process which can improve the result, but can also lead to a correlation delay. Another important point is the precise definition of dependency which has also been left out by many other publications. To avoid having to much dependencies in a certain situation one could try to check whether the dependencies currently exist. In case of a download from a web site there is only a dependency from the DNS subservice at the beginning, but after the address is resolved a download failure is unlikely to have been caused by the DNS. Another possibility to reduce the dependencies is to divide a service into its possible user interactions (e.g., an e-mail service into transactions like get mail, sent mail, etc) and to define the dependencies for each user interaction. Acknowledgments The authors wish to thank the members of the Munich Network Management (MNM) Team for helpful discussions and valuable comments on previous versions of the paper. The MNM Team, directed by Prof. Dr. Heinz-Gerd Hegering, is a 191 group of researchers of the Munich Universities and the Leibniz Supercomputing Center of the Bavarian Academy of Sciences. Its web server is located at wwwmnmteam.informatik. uni-muenchen.de. 8. REFERENCES [1] K. Appleby, G. Goldszmidt, and M. Steinder. Yemanja - A Layered Event Correlation Engine for Multi-domain Server Farms. In Proceedings of the Seventh IFIP/IEEE International Symposium on Integrated Network Management, pages 329-344. IFIP/IEEE, May 2001. [2] Spectrum, Aprisma Corporation. http://www.aprisma.com. [3] C. Ensel. New Approach for Automated Generation of Service Dependency Models. In Network Management as a Strategy for Evolution and Development; Second Latin American Network Operation and Management Symposium (LANOMS 2001). IEEE, August 2001. [4] C. Ensel and A. Keller. An Approach for Managing Service Dependencies with XML and the Resource Description Framework. Journal of Network and Systems Management, 10(2), June 2002. [5] M. Garschhammer, R. Hauck, H.-G. Hegering, B. Kempter, M. Langer, M. Nerb, I. Radisic, H. Roelle, and H. Schmidt. Towards generic Service Management Concepts - A Service Model Based Approach. In Proceedings of the Seventh IFIP/IEEE International Symposium on Integrated Network Management, pages 719-732. IFIP/IEEE, May 2001. [6] B. Gruschke. Integrated Event Management: Event Correlation using Dependency Graphs. In Proceedings of the 9th IFIP/IEEE International Workshop on Distributed Systems: Operations & Management (DSOM 98). IEEE/IFIP, October 1998. [7] M. Gupta, A. Neogi, M. Agarwal, and G. Kar. Discovering Dynamic Dependencies in Enterprise Environments for Problem Determination. In Proceedings of the 14th IFIP/IEEE Workshop on Distributed Sytems: Operations and Management. IFIP/IEEE, October 2003. [8] H.-G. Hegering, S. Abeck, and B. Neumair. Integrated Management of Networked Systems - Concepts, Architectures and their Operational Application. Morgan Kaufmann Publishers, 1999. [9] IT Infrastructure Library, Office of Government Commerce and IT Service Management Forum. http://www.itil.co.uk. [10] G. Jakobson and M. Weissman. Alarm Correlation. IEEE Network, 7(6), November 1993. [11] G. Jakobson and M. Weissman. Real-time Telecommunication Network Management: Extending Event Correlation with Temporal Constraints. In Proceedings of the Fourth IEEE/IFIP International Symposium on Integrated Network Management, pages 290-301. IEEE/IFIP, May 1995. [12] S. Kliger, S. Yemini, Y. Yemini, D. Ohsie, and S. Stolfo. A Coding Approach to Event Correlation. In Proceedings of the Fourth IFIP/IEEE International Symposium on Integrated Network Management, pages 266-277. IFIP/IEEE, May 1995. [13] M. Langer, S. Loidl, and M. Nerb. Customer Service Management: A More Transparent View To Your Subscribed Services. In Proceedings of the 9th IFIP/IEEE International Workshop on Distributed Systems: Operations & Management (DSOM 98), Newark, DE, USA, October 1998. [14] L. Lewis. A Case-based Reasoning Approach for the Resolution of Faults in Communication Networks. In Proceedings of the Third IFIP/IEEE International Symposium on Integrated Network Management. IFIP/IEEE, 1993. [15] L. Lewis. Service Level Management for Enterprise Networks. Artech House, Inc., 1999. [16] NETeXPERT, Agilent Technologies. http://www.agilent.com/comms/OSS. [17] InCharge, Smarts Corporation. http://www.smarts.com. [18] Enhanced Telecom Operations Map, TeleManagement Forum. http://www.tmforum.org. [19] Verizon Communications. http://www.verizon.com. [20] H. Wietgrefe, K.-D. Tuchs, K. Jobmann, G. Carls, P. Froelich, W. Nejdl, and S. Steinfeld. Using Neural Networks for Alarm Correlation in Cellular Phone Networks. In International Workshop on Applications of Neural Networks to Telecommunications (IWANNT), May 1997. [21] S. Yemini, S. Kliger, E. Mozes, Y. Yemini, and D. Ohsie. High Speed and Robust Event Correlation. IEEE Communiations Magazine, 34(5), May 1996. 192

service level agreement;qos;fault management;process management framework;customer service management;service management;service-oriented management;event correlation;rule-based reasoning;case-based reasoning;service-oriented event correlation

train_C-77

Tracking Immediate Predecessors in Distributed Computations

A distributed computation is usually modeled as a partially ordered set of relevant events (the relevant events are a subset of the primitive events produced by the computation). An important causality-related distributed computing problem, that we call the Immediate Predecessors Tracking (IPT) problem, consists in associating with each relevant event, on the fly and without using additional control messages, the set of relevant events that are its immediate predecessors in the partial order. So, IPT is the on-the-fly computation of the transitive reduction (i.e., Hasse diagram) of the causality relation defined by a distributed computation. This paper addresses the IPT problem: it presents a family of protocols that provides each relevant event with a timestamp that exactly identifies its immediate predecessors. The family is defined by a general condition that allows application messages to piggyback control information whose size can be smaller than n (the number of processes). In that sense, this family defines message size-efficient IPT protocols. According to the way the general condition is implemented, different IPT protocols can be obtained. Two of them are exhibited.

1. INTRODUCTION A distributed computation consists of a set of processes that cooperate to achieve a common goal. A main characteristic of these computations lies in the fact that the processes do not share a common global memory, and communicate only by exchanging messages over a communication network. Moreover, message transfer delays are finite but unpredictable. This computation model defines what is known as the asynchronous distributed system model. It is particularly important as it includes systems that span large geographic areas, and systems that are subject to unpredictable loads. Consequently, the concepts, tools and mechanisms developed for asynchronous distributed systems reveal to be both important and general. Causality is a key concept to understand and master the behavior of asynchronous distributed systems [18]. More precisely, given two events e and f of a distributed computation, a crucial problem that has to be solved in a lot of distributed applications is to know whether they are causally related, i.e., if the occurrence of one of them is a consequence of the occurrence of the other. The causal past of an event e is the set of events from which e is causally dependent. Events that are not causally dependent are said to be concurrent. Vector clocks [5, 16] have been introduced to allow processes to track causality (and concurrency) between the events they produce. The timestamp of an event produced by a process is the current value of the vector clock of the corresponding process. In that way, by associating vector timestamps with events it becomes possible to safely decide whether two events are causally related or not. Usually, according to the problem he focuses on, a designer is interested only in a subset of the events produced by a distributed execution (e.g., only the checkpoint events are meaningful when one is interested in determining consistent global checkpoints [12]). It follows that detecting causal dependencies (or concurrency) on all the events of the distributed computation is not desirable in all applications [7, 15]. In other words, among all the events that may occur in a distributed computation, only a subset of them are relevant. In this paper, we are interested in the restriction of the causality relation to the subset of events defined as being the relevant events of the computation. Being a strict partial order, the causality relation is transitive. As a consequence, among all the relevant events that causally precede a given relevant event e, only a subset are its immediate predecessors: those are the events f such that there is no relevant event on any causal path from f to e. Unfortunately, given only the vector timestamp associated with an event it is not possible to determine which events of its causal past are its immediate predecessors. This comes from the fact that the vector timestamp associated with e determines, for each process, the last relevant event belong210 ing to the causal past of e, but such an event is not necessarily an immediate predecessor of e. However, some applications [4, 6] require to associate with each relevant event only the set of its immediate predecessors. Those applications are mainly related to the analysis of distributed computations. Some of those analyses require the construction of the lattice of consistent cuts produced by the computation [15, 16]. It is shown in [4] that the tracking of immediate predecessors allows an efficient on the fly construction of this lattice. More generally, these applications are interested in the very structure of the causal past. In this context, the determination of the immediate predecessors becomes a major issue [6]. Additionally, in some circumstances, this determination has to satisfy behavior constraints. If the communication pattern of the distributed computation cannot be modified, the determination has to be done without adding control messages. When the immediate predecessors are used to monitor the computation, it has to be done on the fly. We call Immediate Predecessor Tracking (IPT) the problem that consists in determining on the fly and without additional messages the immediate predecessors of relevant events. This problem consists actually in determining the transitive reduction (Hasse diagram) of the causality graph generated by the relevant events of the computation. Solving this problem requires tracking causality, hence using vector clocks. Previous works have addressed the efficient implementation of vector clocks to track causal dependence on relevant events. Their aim was to reduce the size of timestamps attached to messages. An efficient vector clock implementation suited to systems with fifo channels is proposed in [19]. Another efficient implementation that does not depend on channel ordering property is described in [11]. The notion of causal barrier is introduced in [2, 17] to reduce the size of control information required to implement causal multicast. However, none of these papers considers the IPT problem. This problem has been addressed for the first time (to our knowledge) in [4, 6] where an IPT protocol is described, but without correctness proof. Moreover, in this protocol, timestamps attached to messages are of size n. This raises the following question which, to our knowledge, has never been answered: Are there efficient vector clock implementation techniques that are suitable for the IPT problem?. This paper has three main contributions: (1) a positive answer to the previous open question, (2) the design of a family of efficient IPT protocols, and (3) a formal correctness proof of the associated protocols. From a methodological point of view the paper uses a top-down approach. It states abstract properties from which more concrete properties and protocols are derived. The family of IPT protocols is defined by a general condition that allows application messages to piggyback control information whose size can be smaller than the system size (i.e., smaller than the number of processes composing the system). In that sense, this family defines low cost IPT protocols when we consider the message size. In addition to efficiency, the proposed approach has an interesting design property. Namely, the family is incrementally built in three steps. The basic vector clock protocol is first enriched by adding to each process a boolean vector whose management allows the processes to track the immediate predecessor events. Then, a general condition is stated to reduce the size of the control information carried by messages. Finally, according to the way this condition is implemented, three IPT protocols are obtained. The paper is composed of seven sections. Sections 2 introduces the computation model, vector clocks and the notion of relevant events. Section 3 presents the first step of the construction that results in an IPT protocol in which each message carries a vector clock and a boolean array, both of size n (the number of processes). Section 4 improves this protocol by providing the general condition that allows a message to carry control information whose size can be smaller than n. Section 5 provides instantiations of this condition. Section 6 provides a simulation study comparing the behaviors of the proposed protocols. Finally, Section 7 concludes the paper. (Due to space limitations, proofs of lemmas and theorems are omitted. They can be found in [1].) 2. MODEL AND VECTOR CLOCK 2.1 Distributed Computation A distributed program is made up of sequential local programs which communicate and synchronize only by exchanging messages. A distributed computation describes the execution of a distributed program. The execution of a local program gives rise to a sequential process. Let {P1, P2, . . . , Pn} be the finite set of sequential processes of the distributed computation. Each ordered pair of communicating processes (Pi, Pj ) is connected by a reliable channel cij through which Pi can send messages to Pj. We assume that each message is unique and a process does not send messages to itself1 . Message transmission delays are finite but unpredictable. Moreover, channels are not necessarily fifo. Process speeds are positive but arbitrary. In other words, the underlying computation model is asynchronous. The local program associated with Pi can include send, receive and internal statements. The execution of such a statement produces a corresponding send/receive/internal event. These events are called primitive events. Let ex i be the x-th event produced by process Pi. The sequence hi = e1 i e2 i . . . ex i . . . constitutes the history of Pi, denoted Hi. Let H = ∪n i=1Hi be the set of events produced by a distributed computation. This set is structured as a partial order by Lamport"s happened before relation [14] (denoted hb →) and defined as follows: ex i hb → ey j if and only if (i = j ∧ x + 1 = y) (local precedence) ∨ (∃m : ex i = send(m) ∧ ey j = receive(m)) (msg prec.) ∨ (∃ ez k : ex i hb → ez k ∧ e z k hb → ey j ) (transitive closure). max(ex i , ey j ) is a partial function defined only when ex i and ey j are ordered. It is defined as follows: max(ex i , ey j ) = ex i if ey j hb → ex i , max(ex i , ey j ) = ey i if ex i hb → ey j . Clearly the restriction of hb → to Hi, for a given i, is a total order. Thus we will use the notation ex i < ey i iff x < y. Throughout the paper, we will use the following notation: if e ∈ Hi is not the first event produced by Pi, then pred(e) denotes the event immediately preceding e in the sequence Hi. If e is the first event produced by Pi, then pred(e) is denoted by ⊥ (meaning that there is no such event), and ∀e ∈ Hi : ⊥ < e. The partial order bH = (H, hb →) constitutes a formal model of the distributed computation it is associated with. 1 This assumption is only in order to get simple protocols. 211 P1 P2 P3 [1, 1, 2] [1, 0, 0] [3, 2, 1] [1, 1, 0] (2, 1) [0, 0, 1] (3, 1) [2, 0, 1] (1, 1) (1, 3)(1, 2) (2, 2) (2, 3) (3, 2) [2, 2, 1] [2, 3, 1] (1, 1) (1, 2) (1, 3) (2, 1) (2, 2) (2, 3) (3, 1) (3, 2) Figure 1: Timestamped Relevant Events and Immediate Predecessors Graph (Hasse Diagram) 2.2 Relevant Events For a given observer of a distributed computation, only some events are relevant2 [7, 9, 15]. An interesting example of what an observation is, is the detection of predicates on consistent global states of a distributed computation [3, 6, 8, 9, 13, 15]. In that case, a relevant event corresponds to the modification of a local variable involved in the global predicate. Another example is the checkpointing problem where a relevant event is the definition of a local checkpoint [10, 12, 20]. The left part of Figure 1 depicts a distributed computation using the classical space-time diagram. In this figure, only relevant events are represented. The sequence of relevant events produced by process Pi is denoted by Ri, and R = ∪n i=1Ri ⊆ H denotes the set of all relevant events. Let → be the relation on R defined in the following way: ∀ (e, f) ∈ R × R : (e → f) ⇔ (e hb → f). The poset (R, →) constitutes an abstraction of the distributed computation [7]. In the following we consider a distributed computation at such an abstraction level. Moreover, without loss of generality we consider that the set of relevant events is a subset of the internal events (if a communication event has to be observed, a relevant internal event can be generated just before a send and just after a receive communication event occurred). Each relevant event is identified by a pair (process id, sequence number) (see Figure 1). Definition 1. The relevant causal past of an event e ∈ H is the (partially ordered) subset of relevant events f such that f hb → e. It is denoted ↑ (e). We have ↑ (e) = {f ∈ R | f hb → e}. Note that, if e ∈ R then ↑ (e) = {f ∈ R | f → e}. In the computation described in Figure 1, we have, for the event e identified (2, 2): ↑ (e) = {(1, 1), (1, 2), (2, 1), (3, 1)}. The following properties are immediate consequences of the previous definitions. Let e ∈ H. CP1 If e is not a receive event then ↑ (e) = 8 < : ∅ if pred(e) = ⊥, ↑ (pred(e)) ∪ {pred(e)} if pred(e) ∈ R, ↑ (pred(e)) if pred(e) ∈ R. CP2 If e is a receive event (of a message m) then ↑ (e) = 8 >>< >>: ↑ (send(m)) if pred(e) = ⊥, ↑ (pred(e))∪ ↑ (send(m)) ∪ {pred(e)} if pred(e) ∈ R, ↑ (pred(e))∪ ↑ (send(m)) if pred(e) ∈ R. 2 Those events are sometimes called observable events. Definition 2. Let e ∈ Hi. For every j such that ↑ (e) ∩ Rj = ∅, the last relevant event of Pj with respect to e is: lastr(e, j) = max{f | f ∈↑ (e) ∩ Rj}. When ↑ (e) ∩ Rj = ∅, lastr(e, j) is denoted by ⊥ (meaning that there is no such event). Let us consider the event e identified (2,2) in Figure 1. We have lastr(e, 1) = (1, 2), lastr(e, 2) = (2, 1), lastr(e, 3) = (3, 1). The following properties relate the events lastr(e, j) and lastr(f, j) for all the predecessors f of e in the relation hb →. These properties follow directly from the definitions. Let e ∈ Hi. LR0 ∀e ∈ Hi: lastr(e, i) = 8 < : ⊥ if pred(e) = ⊥, pred(e) if pred(e) ∈ R, lastr(pred(e),i) if pred(e) ∈ R. LR1 If e is not a receipt event: ∀j = i : lastr(e, j) = lastr(pred(e),j). LR2 If e is a receive event of m: ∀j = i : lastr(e, j) = max(lastr(pred(e),j), lastr(send(m),j)). 2.3 Vector Clock System Definition As a fundamental concept associated with the causality theory, vector clocks have been introduced in 1988, simultaneously and independently by Fidge [5] and Mattern [16]. A vector clock system is a mechanism that associates timestamps with events in such a way that the comparison of their timestamps indicates whether the corresponding events are or are not causally related (and, if they are, which one is the first). More precisely, each process Pi has a vector of integers V Ci[1..n] such that V Ci[j] is the number of relevant events produced by Pj, that belong to the current relevant causal past of Pi. Note that V Ci[i] counts the number of relevant events produced so far by Pi. When a process Pi produces a (relevant) event e, it associates with e a vector timestamp whose value (denoted e.V C) is equal to the current value of V Ci. Vector Clock Implementation The following implementation of vector clocks [5, 16] is based on the observation that ∀i, ∀e ∈ Hi, ∀j : e.V Ci[j] = y ⇔ lastr(e, j) = ey j where e.V Ci is the value of V Ci just after the occurrence of e (this relation results directly from the properties LR0, LR1, and LR2). Each process Pi manages its vector clock V Ci[1..n] according to the following rules: VC0 V Ci[1..n] is initialized to [0, . . . , 0]. VC1 Each time it produces a relevant event e, Pi increments its vector clock entry V Ci[i] (V Ci[i] := V Ci[i] + 1) to 212 indicate it has produced one more relevant event, then Pi associates with e the timestamp e.V C = V Ci. VC2 When a process Pi sends a message m, it attaches to m the current value of V Ci. Let m.V C denote this value. VC3 When Pi receives a message m, it updates its vector clock as follows: ∀k : V Ci[k] := max(V Ci[k], m.V C[k]). 3. IMMEDIATE PREDECESSORS In this section, the Immediate Predecessor Tracking (IPT) problem is stated (Section 3.1). Then, some technical properties of immediate predecessors are stated and proved (Section 3.2). These properties are used to design the basic IPT protocol and prove its correctness (Section 3.3). This IPT protocol, previously presented in [4] without proof, is built from a vector clock protocol by adding the management of a local boolean array at each process. 3.1 The IPT Problem As indicated in the introduction, some applications (e.g., analysis of distributed executions [6], detection of distributed properties [7]) require to determine (on-the-fly and without additional messages) the transitive reduction of the relation → (i.e., we must not consider transitive causal dependency). Given two relevant events f and e, we say that f is an immediate predecessor of e if f → e and there is no relevant event g such that f → g → e. Definition 3. The Immediate Predecessor Tracking (IPT) problem consists in associating with each relevant event e the set of relevant events that are its immediate predecessors. Moreover, this has to be done on the fly and without additional control message (i.e., without modifying the communication pattern of the computation). As noted in the Introduction, the IPT problem is the computation of the Hasse diagram associated with the partially ordered set of the relevant events produced by a distributed computation. 3.2 Formal Properties of IPT In order to design a protocol solving the IPT problem, it is useful to consider the notion of immediate relevant predecessor of any event, whether relevant or not. First, we observe that, by definition, the immediate predecessor on Pj of an event e is necessarily the lastr(e, j) event. Second, for lastr(e, j) to be immediate predecessor of e, there must not be another lastr(e, k) event on a path between lastr(e, j) and e. These observations are formalized in the following definition: Definition 4. Let e ∈ Hi. The set of immediate relevant predecessors of e (denoted IP(e)), is the set of the relevant events lastr(e, j) (j = 1, . . . , n) such that ∀k : lastr(e, j) ∈↑ (lastr(e, k)). It follows from this definition that IP(e) ⊆ {lastr(e, j)|j = 1, . . . , n} ⊂↑ (e). When we consider Figure 1, The graph depicted in its right part describes the immediate predecessors of the relevant events of the computation defined in its left part, more precisely, a directed edge (e, f) means that the relevant event e is an immediate predecessor of the relevant event f (3 ). The following lemmas show how the set of immediate predecessors of an event is related to those of its predecessors in the relation hb →. They will be used to design and prove the protocols solving the IPT problem. To ease the reading of the paper, their proofs are presented in Appendix A. The intuitive meaning of the first lemma is the following: if e is not a receive event, all the causal paths arriving at e have pred(e) as next-to-last event (see CP1). So, if pred(e) is a relevant event, all the relevant events belonging to its relevant causal past are separated from e by pred(e), and pred(e) becomes the only immediate predecessor of e. In other words, the event pred(e) constitutes a reset w.r.t. the set of immediate predecessors of e. On the other hand, if pred(e) is not relevant, it does not separate its relevant causal past from e. Lemma 1. If e is not a receive event, IP(e) is equal to: ∅ if pred(e) = ⊥, {pred(e)} if pred(e) ∈ R, IP(pred(e)) if pred(e) ∈ R. The intuitive meaning of the next lemma is as follows: if e is a receive event receive(m), the causal paths arriving at e have either pred(e) or send(m) as next-to-last events. If pred(e) is relevant, as explained in the previous lemma, this event hides from e all its relevant causal past and becomes an immediate predecessor of e. Concerning the last relevant predecessors of send(m), only those that are not predecessors of pred(e) remain immediate predecessors of e. Lemma 2. Let e ∈ Hi be the receive event of a message m. If pred(e) ∈ Ri, then, ∀j, IP(e) ∩ Rj is equal to: {pred(e)} if j = i, ∅ if lastr(pred(e),j) ≥ lastr(send(m),j), IP(send(m)) ∩ Rj if lastr(pred(e),j) < lastr(send(m),j). The intuitive meaning of the next lemma is the following: if e is a receive event receive(m), and pred(e) is not relevant, the last relevant events in the relevant causal past of e are obtained by merging those of pred(e) and those of send(m) and by taking the latest on each process. So, the immediate predecessors of e are either those of pred(e) or those of send(m). On a process where the last relevant events of pred(e) and of send(m) are the same event f, none of the paths from f to e must contain another relevant event, and thus, f must be immediate predecessor of both events pred(e) and send(m). Lemma 3. Let e ∈ Hi be the receive event of a message m. If pred(e) ∈ Ri, then, ∀j, IP(e) ∩ Rj is equal to: IP(pred(e)) ∩ Rj if lastr(pred(e),j) > lastr(send(m),j), IP(send(m)) ∩ Rj if lastr(pred(e),j) < lastr(send(m),j) IP(pred(e))∩IP(send(m))∩Rj if lastr(pred(e),j) = lastr (send(m), j). 3.3 A Basic IPT Protocol The basic protocol proposed here associates with each relevant event e, an attribute encoding the set IP(e) of its immediate predecessors. From the previous lemmas, the set 3 Actually, this graph is the Hasse diagram of the partial order associated with the distributed computation. 213 IP(e) of any event e depends on the sets IP of the events pred(e) and/or send(m) (when e = receive(m)). Hence the idea to introduce a data structure allowing to manage the sets IPs inductively on the poset (H, hb →). To take into account the information from pred(e), each process manages a boolean array IPi such that, ∀e ∈ Hi the value of IPi when e occurs (denoted e.IPi) is the boolean array representation of the set IP(e). More precisely, ∀j : IPi[j] = 1 ⇔ lastr(e, j) ∈ IP(e). As recalled in Section 2.3, the knowledge of lastr(e,j) (for every e and every j) is based on the management of vectors V Ci. Thus, the set IP(e) is determined in the following way: IP(e) = {ey j | e.V Ci[j] = y ∧ e.IPi[j] = 1, j = 1, . . . , n} Each process Pi updates IPi according to the Lemmas 1, 2, and 3: 1. It results from Lemma 1 that, if e is not a receive event, the current value of IPi is sufficient to determine e.IPi. It results from Lemmas 2 and 3 that, if e is a receive event (e = receive(m)), then determining e.IPi involves information related to the event send(m). More precisely, this information involves IP(send(m)) and the timestamp of send(m) (needed to compare the events lastr(send(m),j) and lastr(pred(e),j), for every j). So, both vectors send(m).V Cj and send(m).IPj (assuming send(m) produced by Pj ) are attached to message m. 2. Moreover, IPi must be updated upon the occurrence of each event. In fact, the value of IPi just after an event e is used to determine the value succ(e).IPi. In particular, as stated in the Lemmas, the determination of succ(e).IPi depends on whether e is relevant or not. Thus, the value of IPi just after the occurrence of event e must keep track of this event. The following protocol, previously presented in [4] without proof, ensures the correct management of arrays V Ci (as in Section 2.3) and IPi (according to the Lemmas of Section 3.2). The timestamp associated with a relevant event e is denoted e.TS. R0 Initialization: Both V Ci[1..n] and IPi[1..n] are initialized to [0, . . . , 0]. R1 Each time it produces a relevant event e: - Pi associates with e the timestamp e.TS defined as follows e.TS = {(k, V Ci[k]) | IPi[k] = 1}, - Pi increments its vector clock entry V Ci[i] (namely it executes V Ci[i] := V Ci[i] + 1), - Pi resets IPi: ∀ = i : IPi[ ] := 0; IPi[i] := 1. R2 When Pi sends a message m to Pj, it attaches to m the current values of V Ci (denoted m.V C) and the boolean array IPi (denoted m.IP). R3 When it receives a message m from Pj , Pi executes the following updates: ∀k ∈ [1..n] : case V Ci[k] < m.V C[k] thenV Ci[k] := m.V C[k]; IPi[k] := m.IP[k] V Ci[k] = m.V C[k] then IPi[k] := min(IPi[k], m.IP[k]) V Ci[k] > m.V C[k] then skip endcase The proof of the following theorem directly follows from Lemmas 1, 2 and 3. Theorem 1. The protocol described in Section 3.3 solves the IPT problem: for any relevant event e, the timestamp e.TS contains the identifiers of all its immediate predecessors and no other event identifier. 4. A GENERAL CONDITION This section addresses a previously open problem, namely, How to solve the IPT problem without requiring each application message to piggyback a whole vector clock and a whole boolean array?. First, a general condition that characterizes which entries of vectors V Ci and IPi can be omitted from the control information attached to a message sent in the computation, is defined (Section 4.1). It is then shown (Section 4.2) that this condition is both sufficient and necessary. However, this general condition cannot be locally evaluated by a process that is about to send a message. Thus, locally evaluable approximations of this general condition must be defined. To each approximation corresponds a protocol, implemented with additional local data structures. In that sense, the general condition defines a family of IPT protocols, that solve the previously open problem. This issue is addressed in Section 5. 4.1 To Transmit or Not to Transmit Control Information Let us consider the previous IPT protocol (Section 3.3). Rule R3 shows that a process Pj does not systematically update each entry V Cj[k] each time it receives a message m from a process Pi: there is no update of V Cj[k] when V Cj[k] ≥ m.V C[k]. In such a case, the value m.V C[k] is useless, and could be omitted from the control information transmitted with m by Pi to Pj. Similarly, some entries IPj[k] are not updated when a message m from Pi is received by Pj. This occurs when 0 < V Cj[k] = m.V C[k] ∧ m.IP[k] = 1, or when V Cj [k] > m.V C[k], or when m.V C[k] = 0 (in the latest case, as m.IP[k] = IPi[k] = 0 then no update of IPj[k] is necessary). Differently, some other entries are systematically reset to 0 (this occurs when 0 < V Cj [k] = m.V C[k] ∧ m.IP[k] = 0). These observations lead to the definition of the condition K(m, k) that characterizes which entries of vectors V Ci and IPi can be omitted from the control information attached to a message m sent by a process Pi to a process Pj: Definition 5. K(m, k) ≡ (send(m).V Ci[k] = 0) ∨ (send(m).V Ci[k] < pred(receive(m)).V Cj[k]) ∨ ; (send(m).V Ci[k] = pred(receive(m)).V Cj[k]) ∧(send(m).IPi[k] = 1) . 4.2 A Necessary and Sufficient Condition We show here that the condition K(m, k) is both necessary and sufficient to decide which triples of the form (k, send(m).V Ci[k], send(m).IPi[k]) can be omitted in an outgoing message m sent by Pi to Pj. A triple attached to m will also be denoted (k, m.V C[k], m.IP[k]). Due to space limitations, the proofs of Lemma 4 and Lemma 5 are given in [1]. (The proof of Theorem 2 follows directly from these lemmas.) 214 Lemma 4. (Sufficiency) If K(m, k) is true, then the triple (k, m.V C[k], m.IP[k]) is useless with respect to the correct management of IPj[k] and V Cj [k]. Lemma 5. (Necessity) If K(m, k) is false, then the triple (k, m.V C[k], m.IP[k]) is necessary to ensure the correct management of IPj[k] and V Cj [k]. Theorem 2. When a process Pi sends m to a process Pj, the condition K(m, k) is both necessary and sufficient not to transmit the triple (k, send(m).V Ci[k], send(m).IPi[k]). 5. A FAMILY OF IPT PROTOCOLS BASED ON EVALUABLE CONDITIONS It results from the previous theorem that, if Pi could evaluate K(m, k) when it sends m to Pj, this would allow us improve the previous IPT protocol in the following way: in rule R2, the triple (k, V Ci[k], IPi[k]) is transmitted with m only if ¬K(m, k). Moreover, rule R3 is appropriately modified to consider only triples carried by m. However, as previously mentioned, Pi cannot locally evaluate K(m, k) when it is about to send m. More precisely, when Pi sends m to Pj , Pi knows the exact values of send(m).V Ci[k] and send(m).IPi[k] (they are the current values of V Ci[k] and IPi[k]). But, as far as the value of pred(receive(m)).V Cj[k] is concerned, two cases are possible. Case (i): If pred(receive(m)) hb → send(m), then Pi can know the value of pred(receive(m)).V Cj[k] and consequently can evaluate K(m, k). Case (ii): If pred(receive(m)) and send(m) are concurrent, Pi cannot know the value of pred(receive(m)).V Cj[k] and consequently cannot evaluate K(m, k). Moreover, when it sends m to Pj , whatever the case (i or ii) that actually occurs, Pi has no way to know which case does occur. Hence the idea to define evaluable approximations of the general condition. Let K (m, k) be an approximation of K(m, k), that can be evaluated by a process Pi when it sends a message m. To be correct, the condition K must ensure that, every time Pi should transmit a triple (k, V Ci[k], IPi[k]) according to Theorem 2 (i.e., each time ¬K(m, k)), then Pi transmits this triple when it uses condition K . Hence, the definition of a correct evaluable approximation: Definition 6. A condition K , locally evaluable by a process when it sends a message m to another process, is correct if ∀(m, k) : ¬K(m, k) ⇒ ¬K (m, k) or, equivalently, ∀(m, k) : K (m, k) ⇒ K(m, k). This definition means that a protocol evaluating K to decide which triples must be attached to messages, does not miss triples whose transmission is required by Theorem 2. Let us consider the constant condition (denoted K1), that is always false, i.e., ∀(m, k) : K1(m, k) = false. This trivially correct approximation of K actually corresponds to the particular IPT protocol described in Section 3 (in which each message carries a whole vector clock and a whole boolean vector). The next section presents a better approximation of K (denoted K2). 5.1 A Boolean Matrix-Based Evaluable Condition Condition K2 is based on the observation that condition K is composed of sub-conditions. Some of them can be Pj send(m) Pi V Ci[k] = x IPi[k] = 1 V Cj[k] ≥ x receive(m) Figure 2: The Evaluable Condition K2 locally evaluated while the others cannot. More precisely, K ≡ a ∨ α ∨ (β ∧ b), where a ≡ send(m).V Ci[k] = 0 and b ≡ send(m).IPi[k] = 1 are locally evaluable, whereas α ≡ send(m).V Ci[k] < pred(receive(m)).V Cj[k] and β ≡ send(m).V Ci[k] = pred(receive(m)).V Cj[k] are not. But, from easy boolean calculus, a∨((α∨β)∧b) =⇒ a∨α∨ (β ∧ b) ≡ K. This leads to condition K ≡ a ∨ (γ ∧ b), where γ = α ∨ β ≡ send(m).V Ci[k] ≤ pred(receive(m)).V Cj[k] , i.e., K ≡ (send(m).V Ci[k] ≤ pred(receive(m)).V Cj[k] ∧ send(m).IPi[k] = 1) ∨ send(m).V Ci[k] = 0. So, Pi needs to approximate the predicate send(m).V Ci[k] ≤ pred(receive(m)).V Cj[k]. To be correct, this approximation has to be a locally evaluable predicate ci(j, k) such that, when Pi is about to send a message m to Pj, ci(j, k) ⇒ (send(m).V Ci[k] ≤ pred(receive(m)).V Cj[k]). Informally, that means that, when ci(j, k) holds, the local context of Pi allows to deduce that the receipt of m by Pj will not lead to V Cj[k] update (Pj knows as much as Pi about Pk). Hence, the concrete condition K2 is the following: K2 ≡ send(m).V Ci[k] = 0 ∨ (ci(j, k) ∧ send(m).IPi[k] = 1). Let us now examine the design of such a predicate (denoted ci). First, the case j = i can be ignored, since it is assumed (Section 2.1) that a process never sends a message to itself. Second, in the case j = k, the relation send(m).V Ci[j] ≤ pred(receive(m)).V Cj [j] is always true, because the receipt of m by Pj cannot update V Cj[j]. Thus, ∀j = i : ci(j, j) must be true. Now, let us consider the case where j = i and j = k (Figure 2). Suppose that there exists an event e = receive(m ) with e < send(m), m sent by Pj and piggybacking the triple (k, m .V C[k], m .IP[k]), and m .V C[k] ≥ V Ci[k] (hence m .V C[k] = receive(m ).V Ci[k]). As V Cj[k] cannot decrease this means that, as long as V Ci[k] does not increase, for every message m sent by Pi to Pj we have the following: send(m).V Ci[k] = receive(m ).V Ci[k] = send(m ).V Cj[k] ≤ receive(m).V Cj [k], i.e., ci(j, k) must remain true. In other words, once ci(j, k) is true, the only event of Pi that could reset it to false is either the receipt of a message that increases V Ci[k] or, if k = i, the occurrence of a relevant event (that increases V Ci[i]). Similarly, once ci(j, k) is false, the only event that can set it to true is the receipt of a message m from Pj, piggybacking the triple (k, m .V C[k], m .IP[k]) with m .V C[k] ≥ V Ci[k]. In order to implement the local predicates ci(j, k), each process Pi is equipped with a boolean matrix Mi (as in [11]) such that M[j, k] = 1 ⇔ ci(j, k). It follows from the previous discussion that this matrix is managed according to the following rules (note that its i-th line is not significant (case j = i), and that its diagonal is always equal to 1): M0 Initialization: ∀ (j, k) : Mi[j, k] is initialized to 1. 215 M1 Each time it produces a relevant event e: Pi resets4 the ith column of its matrix: ∀j = i : Mi[j, i] := 0. M2 When Pi sends a message: no update of Mi occurs. M3 When it receives a message m from Pj , Pi executes the following updates: ∀ k ∈ [1..n] : case V Ci[k] < m.V C[k] then ∀ = i, j, k : Mi[ , k] := 0; Mi[j, k] := 1 V Ci[k] = m.V C[k] then Mi[j, k] := 1 V Ci[k] > m.V C[k] then skip endcase The following lemma results from rules M0-M3. The theorem that follows shows that condition K2(m, k) is correct. (Both are proved in [1].) Lemma 6. ∀i, ∀m sent by Pi to Pj, ∀k, we have: send(m).Mi[j, k] = 1 ⇒ send(m).V Ci[k] ≤ pred(receive(m)).V Cj [k]. Theorem 3. Let m be a message sent by Pi to Pj . Let K2(m, k) ≡ ((send(m).Mi[j, k] = 1) ∧ (send(m).IPi[k] = 1)∨(send(m).V Ci[k] = 0)). We have: K2(m, k) ⇒ K(m, k). 5.2 Resulting IPT Protocol The complete text of the IPT protocol based on the previous discussion follows. RM0 Initialization: - Both V Ci[1..n] and IPi[1..n] are set to [0, . . . , 0], and ∀ (j, k) : Mi[j, k] is set to 1. RM1 Each time it produces a relevant event e: - Pi associates with e the timestamp e.TS defined as follows: e.TS = {(k, V Ci[k]) | IPi[k] = 1}, - Pi increments its vector clock entry V Ci[i] (namely, it executes V Ci[i] := V Ci[i] + 1), - Pi resets IPi: ∀ = i : IPi[ ] := 0; IPi[i] := 1. - Pi resets the ith column of its boolean matrix: ∀j = i : Mi[j, i] := 0. RM2 When Pi sends a message m to Pj, it attaches to m the set of triples (each made up of a process id, an integer and a boolean): {(k, V Ci[k], IPi[k]) | (Mi[j, k] = 0 ∨ IPi[k] = 0) ∧ (V Ci[k] > 0)}. RM3 When Pi receives a message m from Pj , it executes the following updates: ∀(k,m.V C[k], m.IP[k]) carried by m: case V Ci[k] < m.V C[k] then V Ci[k] := m.V C[k]; IPi[k] := m.IP[k]; ∀ = i, j, k : Mi[ , k] := 0; 4 Actually, the value of this column remains constant after its first update. In fact, ∀j, Mi[j, i] can be set to 1 only upon the receipt of a message from Pj, carrying the value V Cj[i] (see R3). But, as Mj [i, i] = 1, Pj does not send V Cj[i] to Pi. So, it is possible to improve the protocol by executing this reset of the column Mi[∗, i] only when Pi produces its first relevant event. Mi[j, k] := 1 V Ci[k] = m.V C[k] then IPi[k] := min(IPi[k], m.IP[k]); Mi[j, k] := 1 V Ci[k] > m.V C[k] then skip endcase 5.3 A Tradeoff The condition K2(m, k) shows that a triple has not to be transmitted when (Mi[j, k] = 1 ∧ IPi[k] = 1) ∨ (V Ci[k] > 0). Let us first observe that the management of IPi[k] is governed by the application program. More precisely, the IPT protocol does not define which are the relevant events, it has only to guarantee a correct management of IPi[k]. Differently, the matrix Mi does not belong to the problem specification, it is an auxiliary variable of the IPT protocol, which manages it so as to satisfy the following implication when Pi sends m to Pj : (Mi[j, k] = 1) ⇒ (pred(receive(m)).V Cj [k] ≥ send(m).V Ci[k]). The fact that the management of Mi is governed by the protocol and not by the application program leaves open the possibility to design a protocol where more entries of Mi are equal to 1. This can make the condition K2(m, k) more often satisfied5 and can consequently allow the protocol to transmit less triples. We show here that it is possible to transmit less triples at the price of transmitting a few additional boolean vectors. The previous IPT matrix-based protocol (Section 5.2) is modified in the following way. The rules RM2 and RM3 are replaced with the modified rules RM2" and RM3" (Mi[∗, k] denotes the kth column of Mi). RM2" When Pi sends a message m to Pj, it attaches to m the following set of 4-uples (each made up of a process id, an integer, a boolean and a boolean vector): {(k, V Ci[k], IPi[k], Mi[∗, k]) | (Mi[j, k] = 0 ∨ IPi[k] = 0) ∧ V Ci[k] > 0}. RM3" When Pi receives a message m from Pj , it executes the following updates: ∀(k,m.V C[k], m.IP[k], m.M[1..n, k]) carried by m: case V Ci[k] < m.V C[k] then V Ci[k] := m.V C[k]; IPi[k] := m.IP[k]; ∀ = i : Mi[ , k] := m.M[ , k] V Ci[k] = m.V C[k] then IPi[k] := min(IPi[k], m.IP[k]); ∀ =i : Mi[ , k] := max(Mi[ , k], m.M[ , k]) V Ci[k] > m.V C[k] then skip endcase Similarly to the proofs described in [1], it is possible to prove that the previous protocol still satisfies the property proved in Lemma 6, namely, ∀i, ∀m sent by Pi to Pj, ∀k we have (send(m).Mi[j, k] = 1) ⇒ (send(m).V Ci[k] ≤ pred(receive(m)).V Cj[k]). 5 Let us consider the previously described protocol (Section 5.2) where the value of each matrix entry Mi[j, k] is always equal to 0. The reader can easily verify that this setting correctly implements the matrix. Moreover, K2(m, k) is then always false: it actually coincides with K1(k, m) (which corresponds to the case where whole vectors have to be transmitted with each message). 216 Intuitively, the fact that some columns of matrices M are attached to application messages allows a transitive transmission of information. More precisely, the relevant history of Pk known by Pj is transmitted to a process Pi via a causal sequence of messages from Pj to Pi. In contrast, the protocol described in Section 5.2 used only a direct transmission of this information. In fact, as explained Section 5.1, the predicate c (locally implemented by the matrix M) was based on the existence of a message m sent by Pj to Pi, piggybacking the triple (k, m .V C[k], m .IP[k]), and m .V C[k] ≥ V Ci[k], i.e., on the existence of a direct transmission of information (by the message m ). The resulting IPT protocol (defined by the rules RM0, RM1, RM2" and RM3") uses the same condition K2(m, k) as the previous one. It shows an interesting tradeoff between the number of triples (k, V Ci[k], IPi[k]) whose transmission is saved and the number of boolean vectors that have to be additionally piggybacked. It is interesting to notice that the size of this additional information is bounded while each triple includes a non-bounded integer (namely a vector clock value). 6. EXPERIMENTAL STUDY This section compares the behaviors of the previous protocols. This comparison is done with a simulation study. IPT1 denotes the protocol presented in Section 3.3 that uses the condition K1(m, k) (which is always equal to false). IPT2 denotes the protocol presented in Section 5.2 that uses the condition K2(m, k) where messages carry triples. Finally, IPT3 denotes the protocol presented in Section 5.3 that also uses the condition K2(m, k) but where messages carry additional boolean vectors. This section does not aim to provide an in-depth simulation study of the protocols, but rather presents a general view on the protocol behaviors. To this end, it compares IPT2 and IPT3 with regard to IPT1. More precisely, for IPT2 the aim was to evaluate the gain in terms of triples (k, V Ci[k], IPi[k]) not transmitted with respect to the systematic transmission of whole vectors as done in IPT1. For IPT3, the aim was to evaluate the tradeoff between the additional boolean vectors transmitted and the number of saved triples. The behavior of each protocol was analyzed on a set of programs. 6.1 Simulation Parameters The simulator provides different parameters enabling to tune both the communication and the processes features. These parameters allow to set the number of processes for the simulated computation, to vary the rate of communication (send/receive) events, and to alter the time duration between two consecutive relevant events. Moreover, to be independent of a particular topology of the underlying network, a fully connected network is assumed. Internal events have not been considered. Since the presence of the triples (k, V Ci[k], IPi[k]) piggybacked by a message strongly depends on the frequency at which relevant events are produced by a process, different time distributions between two consecutive relevant events have been implemented (e.g., normal, uniform, and Poisson distributions). The senders of messages are chosen according to a random law. To exhibit particular configurations of a distributed computation a given scenario can be provided to the simulator. Message transmission delays follow a standard normal distribution. Finally, the last parameter of the simulator is the number of send events that occurred during a simulation. 6.2 Parameter Settings To compare the behavior of the three IPT protocols, we performed a large number of simulations using different parameters setting. We set to 10 the number of processes participating to a distributed computation. The number of communication events during the simulation has been set to 10 000. The parameter λ of the Poisson time distribution (λ is the average number of relevant events in a given time interval) has been set so that the relevant events are generated at the beginning of the simulation. With the uniform time distribution, a relevant event is generated (in the average) every 10 communication events. The location parameter of the standard normal time distribution has been set so that the occurrence of relevant events is shifted around the third part of the simulation experiment. As noted previously, the simulator can be fed with a given scenario. This allows to analyze the worst case scenarios for IPT2 and IPT3. These scenarios correspond to the case where the relevant events are generated at the maximal frequency (i.e., each time a process sends or receives a message, it produces a relevant event). Finally, the three IPT protocols are analyzed with the same simulation parameters. 6.3 Simulation Results The results are displayed on the Figures 3.a-3.d. These figures plot the gain of the protocols in terms of the number of triples that are not transmitted (y axis) with respect to the number of communication events (x axis). From these figures, we observe that, whatever the time distribution followed by the relevant events, both IPT2 and IPT3 exhibit a behavior better than IPT1 (i.e., the total number of piggybacked triples is lower in IPT2 and IPT3 than in IPT1), even in the worst case (see Figure 3.d). Let us consider the worst scenario. In that case, the gain is obtained at the very beginning of the simulation and lasts as long as it exists a process Pj for which ∀k : V Cj[k] = 0. In that case, the condition ∀k : K(m, k) is satisfied. As soon as ∃k : V Cj[k] = 0, both IPT2 and IPT3 behave as IPT1 (the shape of the curve becomes flat) since the condition K(m, k) is no longer satisfied. Figure 3.a shows that during the first events of the simulation, the slope of curves IPT2 and IPT3 are steep. The same occurs in Figure 3.d (that depicts the worst case scenario). Then the slope of these curves decreases and remains constant until the end of the simulation. In fact, as soon as V Cj[k] becomes greater than 0, the condition ¬K(m, k) reduces to (Mi[j, k] = 0 ∨ IPi[k] = 0). Figure 3.b displays an interesting feature. It considers λ = 100. As the relevant events are taken only during the very beginning of the simulation, this figure exhibits a very steep slope as the other figures. The figure shows that, as soon as no more relevant events are taken, on average, 45% of the triples are not piggybacked by the messages. This shows the importance of matrix Mi. Furthermore, IPT3 benefits from transmitting additional boolean vectors to save triple transmissions. The Figures 3.a-3.c show that the average gain of IPT3 with respect to IPT2 is close to 10%. Finally, Figure 3.c underlines even more the importance 217 of matrix Mi. When very few relevant events are taken, IPT2 and IPT3 turn out to be very efficient. Indeed, this figure shows that, very quickly, the gain in number of triples that are saved is very high (actually, 92% of the triples are saved). 6.4 Lessons Learned from the Simulation Of course, all simulation results are consistent with the theoretical results. IPT3 is always better than or equal to IPT2, and IPT2 is always better than IPT1. The simulation results teach us more: • The first lesson we have learnt concerns the matrix Mi. Its use is quite significant but mainly depends on the time distribution followed by the relevant events. On the one hand, when observing Figure 3.b where a large number of relevant events are taken in a very short time, IPT2 can save up to 45% of the triples. However, we could have expected a more sensitive gain of IPT2 since the boolean vector IP tends to stabilize to [1, ..., 1] when no relevant events are taken. In fact, as discussed in Section 5.3, the management of matrix Mi within IPT2 does not allow a transitive transmission of information but only a direct transmission of this information. This explains why some columns of Mi may remain equal to 0 while they could potentially be equal to 1. Differently, as IPT3 benefits from transmitting additional boolean vectors (providing a transitive transmission information) it reaches a gain of 50%. On the other hand, when very few relevant events are taken in a large period of time (see Figure 3.c), the behavior of IPT2 and IPT3 turns out to be very efficient since the transmission of up to 92% of the triples is saved. This comes from the fact that very quickly the boolean vector IPi tends to stabilize to [1, ..., 1] and that matrix Mi contains very few 0 since very few relevant events have been taken. Thus, a direct transmission of the information is sufficient to quickly get matrices Mi equal to [1, ..., 1], . . . , [1, ..., 1]. • The second lesson concerns IPT3, more precisely, the tradeoff between the additional piggybacking of boolean vectors and the number of triples whose transmission is saved. With n = 10, adding 10 booleans to a triple does not substantially increases its size. The Figures 3.a-3.c exhibit the number of triples whose transmission is saved: the average gain (in number of triples) of IPT3 with respect to IPT2 is about 10%. 7. CONCLUSION This paper has addressed an important causality-related distributed computing problem, namely, the Immediate Predecessors Tracking problem. It has presented a family of protocols that provide each relevant event with a timestamp that exactly identify its immediate predecessors. The family is defined by a general condition that allows application messages to piggyback control information whose size can be smaller than n (the number of processes). In that sense, this family defines message size-efficient IPT protocols. According to the way the general condition is implemented, different IPT protocols can be obtained. Three of them have been described and analyzed with simulation experiments. Interestingly, it has also been shown that the efficiency of the protocols (measured in terms of the size of the control information that is not piggybacked by an application message) depends on the pattern defined by the communication events and the relevant events. Last but not least, it is interesting to note that if one is not interested in tracking the immediate predecessor events, the protocols presented in the paper can be simplified by suppressing the IPi booleans vectors (but keeping the boolean matrices Mi). The resulting protocols, that implement a vector clock system, are particularly efficient as far as the size of the timestamp carried by each message is concerned. Interestingly, this efficiency is not obtained at the price of additional assumptions (such as fifo channels). 8. REFERENCES [1] Anceaume E., H´elary J.-M. and Raynal M., Tracking Immediate Predecessors in Distributed Computations. Res. Report #1344, IRISA, Univ. Rennes (France), 2001. [2] Baldoni R., Prakash R., Raynal M. and Singhal M., Efficient ∆-Causal Broadcasting. Journal of Computer Systems Science and Engineering, 13(5):263-270, 1998. [3] Chandy K.M. and Lamport L., Distributed Snapshots: Determining Global States of Distributed Systems, ACM Transactions on Computer Systems, 3(1):63-75, 1985. [4] Diehl C., Jard C. and Rampon J.-X., Reachability Analysis of Distributed Executions, Proc. TAPSOFT"93, Springer-Verlag LNCS 668, pp. 629-643, 1993. [5] Fidge C.J., Timestamps in Message-Passing Systems that Preserve Partial Ordering, Proc. 11th Australian Computing Conference, pp. 56-66, 1988. [6] Fromentin E., Jard C., Jourdan G.-V. and Raynal M., On-the-fly Analysis of Distributed Computations, IPL, 54:267-274, 1995. [7] Fromentin E. and Raynal M., Shared Global States in Distributed Computations, JCSS, 55(3):522-528, 1997. [8] Fromentin E., Raynal M., Garg V.K. and Tomlinson A., On-the-Fly Testing of Regular Patterns in Distributed Computations. Proc. ICPP"94, Vol. 2:73-76, 1994. [9] Garg V.K., Principles of Distributed Systems, Kluwer Academic Press, 274 pages, 1996. [10] H´elary J.-M., Most´efaoui A., Netzer R.H.B. and Raynal M., Communication-Based Prevention of Useless Ckeckpoints in Distributed Computations. Distributed Computing, 13(1):29-43, 2000. [11] H´elary J.-M., Melideo G. and Raynal M., Tracking Causality in Distributed Systems: a Suite of Efficient Protocols. Proc. SIROCCO"00, Carleton University Press, pp. 181-195, L"Aquila (Italy), June 2000. [12] H´elary J.-M., Netzer R. and Raynal M., Consistency Issues in Distributed Checkpoints. IEEE TSE, 25(4):274-281, 1999. [13] Hurfin M., Mizuno M., Raynal M. and Singhal M., Efficient Distributed Detection of Conjunction of Local Predicates in Asynch Computations. IEEE TSE, 24(8):664-677, 1998. [14] Lamport L., Time, Clocks and the Ordering of Events in a Distributed System. Comm. ACM, 21(7):558-565, 1978. [15] Marzullo K. and Sabel L., Efficient Detection of a Class of Stable Properties. Distributed Computing, 8(2):81-91, 1994. [16] Mattern F., Virtual Time and Global States of Distributed Systems. Proc. Int. Conf. Parallel and Distributed Algorithms, (Cosnard, Quinton, Raynal, Robert Eds), North-Holland, pp. 215-226, 1988. [17] Prakash R., Raynal M. and Singhal M., An Adaptive Causal Ordering Algorithm Suited to Mobile Computing Environment. JPDC, 41:190-204, 1997. [18] Raynal M. and Singhal S., Logical Time: Capturing Causality in Distributed Systems. IEEE Computer, 29(2):49-57, 1996. [19] Singhal M. and Kshemkalyani A., An Efficient Implementation of Vector Clocks. IPL, 43:47-52, 1992. [20] Wang Y.M., Consistent Global Checkpoints That Contain a Given Set of Local Checkpoints. IEEE TOC, 46(4):456-468, 1997. 218 0 1000 2000 3000 4000 5000 6000 0 2000 4000 6000 8000 10000 gaininnumberoftriples communication events number IPT1 IPT2 IPT3 relevant events (a) The relevant events follow a uniform distribution (ratio=1/10) -5000 0 5000 10000 15000 20000 25000 30000 35000 40000 45000 50000 0 2000 4000 6000 8000 10000 gaininnumberoftriples communication events number IPT1 IPT2 IPT3 relevant events (b) The relevant events follow a Poisson distribution (λ = 100) 0 10000 20000 30000 40000 50000 60000 70000 80000 90000 100000 0 2000 4000 6000 8000 10000 gaininnumberoftriples communication events number IPT1 IPT2 IPT3 relevant events (c) The relevant events follow a normal distribution 0 50 100 150 200 250 300 350 400 450 1 10 100 1000 10000 gaininnumberoftriples communication events number IPT1 IPT2 IPT3 relevant events (d) For each pi, pi takes a relevant event and broadcast to all processes Figure 3: Experimental Results 219

immediate predecessor tracking;relevant event;causality track;transitive reduction;ipt protocol;timestamp;message-pass;hasse diagram;piggybacking;immediate predecessor;tracking causality;common global memory;message transfer delay;checkpointing problem;control information;vector timestamp;distributed computation;vector clock;channel ordering property

train_C-78

An Architectural Framework and a Middleware for Cooperating Smart Components

In a future networked physical world, a myriad of smart sensors and actuators assess and control aspects of their environments and autonomously act in response to it. Examples range in telematics, traffic management, team robotics or home automation to name a few. To a large extent, such systems operate proactively and independently of direct human control driven by the perception of the environment and the ability to organize respective computations dynamically. The challenging characteristics of these applications include sentience and autonomy of components, issues of responsiveness and safety criticality, geographical dispersion, mobility and evolution. A crucial design decision is the choice of the appropriate abstractions and interaction mechanisms. Looking to the basic building blocks of such systems we may find components which comprise mechanical components, hardware and software and a network interface, thus these components have different characteristics compared to pure software components. They are able to spontaneously disseminate information in response to events observed in the physical environment or to events received from other component via the network interface. Larger autonomous components may be composed recursively from these building blocks. The paper describes an architectural framework and a middleware supporting a component-based system and an integrated view on events-based communication comprising the real world events and the events generated in the system. It starts by an outline of the component-based system construction. The generic event architecture GEAR is introduced which describes the event-based interaction between the components via a generic event layer. The generic event layer hides the different communication channels including the interactions through the environment. An appropriate middleware is presented which reflects these needs and allows to specify events which have quality attributes to express temporal constraints. This is complemented by the notion of event channels which are abstractions of the underlying network and allow to enforce quality attributes. They are established prior to interaction to reserve the needed computational and network resources for highly predictable event dissemination.

1. INTRODUCTION In recent years we have seen the continuous improvement of technologies that are relevant for the construction of distributed embedded systems, including trustworthy visual, auditory, and location sensing [11], communication and processing. We believe that in a future networked physical world a new class of applications will emerge, composed of a myriad of smart sensors and actuators to assess and control aspects of their environments and autonomously act in response to it. The anticipated challenging characteristics of these applications include autonomy, responsiveness and safety criticality, large scale, geographical dispersion, mobility and evolution. In order to deal with these challenges, it is of fundamental importance to use adequate high-level models, abstractions and interaction paradigms. Unfortunately, when facing the specific characteristics of the target systems, the shortcomings of current architectures and middleware interaction paradigms become apparent. Looking to the basic building blocks of such systems we may find components which comprise mechanical parts, hardware, software and a network interface. However, classical event/object models are usually software oriented and, as such, when trans28 ported to a real-time, embedded systems setting, their harmony is cluttered by the conflict between, on the one side, send/receive of software events (message-based), and on the other side, input/output of hardware or real-world events, register-based. In terms of interaction paradigms, and although the use of event-based models appears to be a convenient solution [10, 22], these often lack the appropriate support for non-functional requirements like reliability, timeliness or security. This paper describes an architectural framework and a middleware, supporting a component-based system and an integrated view on event-based communication comprising the real world events and the events generated in the system. When choosing the appropriate interaction paradigm it is of fundamental importance to address the challenging issues of the envisaged sentient applications. Unlike classical approaches that confine the possible interactions to the application boundaries, i.e. to its components, we consider that the environment surrounding the application also plays a relevant role in this respect. Therefore, the paper starts by clarifying several issues concerning our view of the system, about the interactions that may take place and about the information flows. This view is complemented by providing an outline of the component-based system construction and, in particular, by showing that it is possible to compose larger applications from basic components, following an hierarchical composition approach. This provides the necessary background to introduce the Generic-Events Architecture (GEAR), which describes the event-based interaction between the components via a generic event layer while allowing the seamless integration of physical and computer information flows. In fact, the generic event layer hides the different communication channels, including the interactions through the environment. Additionally, the event layer abstraction is also adequate for the proper handling of the non-functional requirements, namely reliability and timeliness, which are particularly stringent in real-time settings. The paper devotes particular attention to this issue by discussing the temporal aspects of interactions and the needs for predictability. An appropriate middleware is presented which reflects these needs and allows to specify events which have quality attributes to express temporal constraints. This is complemented by the notion of Event Channels (EC), which are abstractions of the underlying network while being abstracted by the event layer. In fact, event channels play a fundamental role in securing the functional and non-functional (e.g. reliability and timeliness) properties of the envisaged applications, that is, in allowing the enforcement of quality attributes. They are established prior to interaction to reserve the needed computational and network resources for highly predictable event dissemination. The paper is organized as follows. In Section 3 we introduce the fundamental notions and abstractions that we adopt in this work to describe the interactions taking place in the system. Then, in Section 4, we describe the componentbased approach that allows composition of objects. GEAR is then described in Section 5 and Section 6 focuses on temporal aspects of the interactions. Section 7 describes the COSMIC middleware, which may be used to specify the interaction between sentient objects. A simple example to highlight the ideas presented in the paper appears in Section 8 and Section 9 concludes the paper. 2. RELATED WORK Our work considers a wired physical world in which a very large number of autonomous components cooperate. It is inspired by many research efforts in very different areas. Event-based systems in general have been introduced to meet the requirements of applications in which entities spontaneously generate information and disseminate it [1, 25, 22]. Intended for large systems and requiring quite complex infrastructures, these event systems do not consider stringent quality aspects like timeliness and dependability issues. Secondly, they are not created to support inter-operability between tiny smart devices with substantial resource constraints. In [10] a real-time event system for CORBA has been introduced. The events are routed via a central event server which provides scheduling functions to support the real-time requirements. Such a central component is not available in an infrastructure envisaged in our system architecture and the developed middleware TAO (The Ace Orb) is quite complex and unsuitable to be directly integrated in smart devices. There are efforts to implement CORBA for control networks, tailored to connect sensor and actuator components [15, 19]. They are targeted for the CAN-Bus [9], a popular network developed for the automotive industry. However, in these approaches the support for timeliness or dependability issues does not exist or is only very limited. A new scheme to integrate smart devices in a CORBA environment is proposed in [17] and has lead to the proposal of a standard by the Object Management Group (OMG) [26]. Smart transducers are organized in clusters that are connected to a CORBA system by a gateway. The clusters form isolated subnetworks. A special master node enforces the temporal properties in the cluster subnet. A CORBA gateway allows to access sensor data and write actuator data by means of an interface file system (IFS). The basic structure is similar to the WAN-of-CANs structure which has been introduced in the CORTEX project [4]. Islands of tight control may be realized by a control network and cooperate via wired or wireless networks covering a large number of these subnetworks. However, in contrast to the event channel model introduced in this paper, all communication inside a cluster relies on a single technical solution of a synchronous communication channel. Secondly, although the temporal behaviour of a single cluster is rigorously defined, no model to specify temporal properties for clusterto-CORBA or cluster-to-cluster interactions is provided. 3. INFORMATION FLOW AND INTERACTION MODEL In this paper we consider a component-based system model that incorporates previous work developed in the context of the IST CORTEX project [5]. As mentioned above, a fundamental idea underlying the approach is that applications can be composed of a large number of smart components that are able to sense their surrounding environment and interact with it. These components are referred to as sentient objects, a metaphor elaborated in CORTEX and inspired on the generic concept of sentient computing introduced in [12]. Sentient objects accept input events from a variety of different sources (including sensors, but not constrained to that), process them, and produce output events, whereby 29 they actuate on the environment and/or interact with other objects. Therefore, the following kinds of interactions can take place in the system: Environment-to-object interactions: correspond to a flow of information from the environment to application objects, reporting about the state of the former, and/or notifying about events taking place therein. Object-to-object interactions: correspond to a flow of information among sentient objects, serving two purposes. The first is related with complementing the assessment of each individual object about the state of the surrounding space. The second is related to collaboration, in which the object tries to influence other objects into contributing to a common goal, or into reacting to an unexpected situation. Object-to-environment interactions: correspond to a flow of information from an object to the environment, with the purpose of forcing a change in the state of the latter. Before continuing, we need to clarify a few issues with respect to these possible forms of interaction. We consider that the environment can be a producer or consumer of information while interacting with sentient objects. The environment is the real (physical) world surrounding an object, not necessarily close to the object or limited to certain boundaries. Quite clearly, the information produced by the environment corresponds to the physical representation of real-time entities, of which typical examples include temperature, distance or the state of a door. On the other hand, actuation on the environment implies the manipulation of these real-time entities, like increasing the temperature (applying more heat), changing the distance (applying some movement) or changing the state of the door (closing or opening it). The required transformations between system representations of these real-time entities and their physical representations is accomplished, generically, by sensors and actuators. We further consider that there may exist dumb sensors and actuators, which interact with the objects by disseminating or capturing raw transducer information, and smart sensors and actuators, with enhanced processing capabilities, capable of speaking some more elaborate event dialect (see Sections 5 and 6.1). Interaction with the environment is therefore done through sensors and actuators, which may, or may not be part of sentient objects, as discussed in Section 4.2. State or state changes in the environment are considered as events, captured by sensors (in the environment or within sentient objects) and further disseminated to other potentially interested sentient objects in the system. In consequence, it is quite natural to base the communication and interaction among sentient objects and with the environment on an event-based communication model. Moreover, typical properties of event-based models, such as anonymous and non-blocking communication, are highly desirable in systems where sentient objects can be mobile and where interactions are naturally very dynamic. A distinguishing aspect of our work from many of the existing approaches, is that we consider that sentient objects may indirectly communicate with each other through the environment, when they act on it. Thus the environment constitutes an interaction and communication channel and is in the control and awareness loop of the objects. In other words, when a sentient object actuates on the environment it will be able to observe the state changes in the environment by means of events captured by the sensors. Clearly, other objects might as well capture the same events, thus establishing the above-mentioned indirect communication path. In systems that involve interactions with the environment it is very important to consider the possibility of communication through the environment. It has been shown that the hidden channels developing through the latter (e.g., feedback loops) may hinder software-based algorithms ignoring them [30]. Therefore, any solution to the problem requires the definition of convenient abstractions and appropriate architectural constructs. On the other hand, in order to deal with the information flow through the whole computer system and environment in a seamless way, handling software and hardware events uniformly, it is also necessary to find adequate abstractions. As discussed in Section 5, the Generic-Events Architecture introduces the concept of Generic Event and an Event Layer abstraction which aim at dealing, among others, with these issues. 4. SENTIENT OBJECT COMPOSITION In this section we analyze the most relevant issues related with the sentient object paradigm and the construction of systems composed of sentient objects. 4.1 Component-based System Construction Sentient objects can take several different forms: they can simply be software-based components, but they can also comprise mechanical and/or hardware parts, amongst which the very sensorial apparatus that substantiates sentience, mixed with software components to accomplish their task. We refine this notion by considering a sentient object as an encapsulating entity, a component with internal logic and active processing elements, able to receive, transform and produce new events. This interface hides the internal hardware/software structure of the object, which may be complex, and shields the system from the low-level functional and temporal details of controlling a specific sensor or actuator. Furthermore, given the inherent complexity of the envisaged applications, the number of simultaneous input events and the internal size of sentient objects may become too large and difficult to handle. Therefore, it should be possible to consider the hierarchical composition of sentient objects so that the application logic can be separated across as few or as many of these objects as necessary. On the other hand, composition of sentient objects should normally be constrained by the actual hardware component"s structure, preventing the possibility of arbitrarily composing sentient objects. This is illustrated in Figure 1, where a sentient object is internally composed of a few other sentient objects, each of them consuming and producing events, some of which only internally propagated. Observing the figure, and recalling our previous discussion about the possible interactions, we identify all of them here: an object-to-environment interaction occurs between the object controlling a WLAN transmitter and some WLAN receiver in the environment; an environment-to-object interaction takes place when the object responsible for the GPS 30 G P S r e c e p t i o n W i r e l e s s t r a n s m i s s i o n D o p p l e r r a d a r P h y s i c a l f e e d b a c k O b j e c t ' s b o d y I n t e r n a l N e t w o r k Figure 1: Component-aware sentient object composition. signal reception uses the information transmitted by the satellites; finally, explicit object-to-object interactions occur internally to the container object, through an internal communication network. Additionally, it is interesting to observe that implicit communication can also occur, whether the physical feedback develops through the environment internal to the container object (as depicted) or through the environment external to this object. However, there is a subtle difference between both cases. While in the former the feedback can only be perceived by objects internal to the container, bounding the extent to which consistency must be ensured, such bounds do not exist in the latter. In fact, the notion of sentient object as an encapsulating entity may serve other purposes (e.g., the confinement of feedback and of the propagation of events), beyond the mere hierarchical composition of objects. To give a more concrete example of such component-aware object composition we consider a scenario of cooperating robots. Each robot is made of several components, corresponding, for instance, to axis and manipulator controllers. Together with the control software, each of these controllers may be a sentient object. On the other hand, a robot itself is a sentient object, composed of the objects materialized by the controllers, and the environment internal to its own structure, or body. This means that it should be possible to define cooperation activities using the events produced by robot sentient objects, without the need to know the internal structure of robots, or the events produced by body objects or by smart sensors within the body. From an engineering point of view, however, this also means that robot sentient object may have to generate new events that reflect its internal state, which requires the definition of a gateway to make the bridge between the internal and external environments. 4.2 Encapsulation and Scoping Now an important question is about how to represent and disseminate events in a large scale networked world. As we have seen above, any event generated by a sentient object could, in principle, be visible anywhere in the system and thus received by any other sentient object. However, there are substantial obstacles to such universal interactions, originating from the components heterogeneity in such a largescale setting. Firstly, the components may have severe performance constraints, particularly because we want to integrate smart sensors and actuators in such an architecture. Secondly, the bandwidth of the participating networks may vary largely. Such networks may be low power, low bandwidth fieldbuses, or more powerful wireless networks as well as high speed backbones. Thirdly, the networks may have widely different reliability and timeliness characteristics. Consider a platoon of cooperating vehicles. Inside a vehicle there may be a field-bus like CAN [8, 9], TTP/A [17] or LIN [20], with a comparatively low bandwidth. On the other hand, the vehicles are communicating with others in the platoon via a direct wireless link. Finally, there may be multiple platoons of vehicles which are coordinated by an additional wireless network layer. At the abstraction level of sentient objects, such heterogeneity is reflected by the notion of body-vs-environment. At the network level, we assume the WAN-of-CANs structure [27] to model the different networks. The notion of body and environment is derived from the recursively defined component-based object model. A body is similar to a cell membrane and represents a quality of service container for the sentient objects inside. On the network level, it may be associated with the components coupled by a certain CAN. A CAN defines the dissemination quality which can be expected by the cooperating objects. In the above example, a vehicle may be a sentient object, whose body is composed of the respective lower level objects (sensors and actuators) which are connected by the internal network (see Figure 1). Correspondingly, the platoon can be seen itself as an object composed of a collection of cooperating vehicles, its body being the environment encapsulated by the platoon zone. At the network level, the wireless network represents the respective CAN. However, several platoons united by their CANs may interact with each other and objects further away, through some wider-range, possible fixed networking substrate, hence the concept of WAN-of-CANs. The notions of body-environment and WAN-of-CANs are very useful when defining interaction properties across such boundaries. Their introduction obeyed to our belief that a single mechanism to provide quality measures for interactions is not appropriate. Instead, a high level construct for interaction across boundaries is needed which allows to specify the quality of dissemination and exploits the knowledge about body and environment to assess the feasibility of quality constraints. As we will see in the following section, the notion of an event channel represents this construct in our architecture. It disseminates events and allows the network independent specification of quality attributes. These attributes must be mapped to the respective properties of the underlying network structure. 5. A GENERIC EVENTS ARCHITECTURE In order to successfully apply event-based object-oriented models, addressing the challenges enumerated in the introduction of this paper, it is necessary to use adequate architectural constructs, which allow the enforcement of fundamental properties such as timeliness or reliability. We propose the Generic-Events Architecture (GEAR), depicted in Figure 2, which we briefly describe in what follows (for a more detailed description please refer to [29]). The L-shaped structure is crucial to ensure some of the properties described. Environment: The physical surroundings, remote and close, solid and etherial, of sentient objects. 31 C o m m ' sC o m m ' sC o m m ' s T r a n s l a t i o n L a y e r T r a n s l a t i o n L a y e r B o d y E n v i r o n m e n t B o d y E n v i r o n m e n t B o d y E n v i r o n m e n t ( i n c l u d i n g o p e r a t i o n a l n e t w o r k ) ( o f o b j e c t o r o b j e c t c o m p o u n d ) T r a n s l a t i o n L a y e r T r a n s l a t i o n S e n t i e n t O b j e c t S e n t i e n t O b j e c t S e n t i e n t O b j e c t R e g u l a r N e t w o r k c o n s u m ep r o d u c e E v e n t L a y e r E v e n t L a y e r E v e n t L a y e r S e n t i e n t O b j e c t Figure 2: Generic-Events architecture. Body: The physical embodiment of a sentient object (e.g., the hardware where a mechatronic controller resides, the physical structure of a car). Note that due to the compositional approach taken in our model, part of what is environment to a smaller object seen individually, becomes body for a larger, containing object. In fact, the body is the internal environment of the object. This architecture layering allows composition to take place seamlessly, in what concerns information flow. Inside a body there may also be implicit knowledge, which can be exploited to make interaction more efficient, like the knowledge about the number of cooperating entities, the existence of a specific communication network or the simple fact that all components are co-located and thus the respective events do not need to specify location in their context attributes. Such intrinsic information is not available outside a body and, therefore, more explicit information has to be carried by an event. Translation Layer: The layer responsible for physical event transformation from/to their native form to event channel dialect, between environment/body and an event channel. Essentially one doing observation and actuation operations on the lower side, and doing transactions of event descriptions on the other. On the lower side this layer may also interact with dumb sensors or actuators, therefore talking the language of the specific device. These interactions are done through operational networks (hence the antenna symbol in the figure). Event Layer: The layer responsible for event propagation in the whole system, through several Event Channels (EC):. In concrete terms, this layer is a kind of middleware that provides important event-processing services which are crucial for any realistic event-based system. For example, some of the services that imply the processing of events may include publishing, subscribing, discrimination (zoning, filtering, fusion, tracing), and queuing. Communication Layer: The layer responsible for wrapping events (as a matter of fact, event descriptions in EC dialect) into carrier event-messages, to be transported to remote places. For example, a sensing event generated by a smart sensor is wrapped in an event-message and disseminated, to be caught by whoever is concerned. The same holds for an actuation event produced by a sentient object, to be delivered to a remote smart actuator. Likewise, this may apply to an event-message from one sentient object to another. Dumb sensors and actuators do not send event-messages, since they are unable to understand the EC dialect (they do not have an event layer neither a communication layer- they communicate, if needed, through operational networks). Regular Network: This is represented in the horizontal axis of the block diagram by the communication layer, which encompasses the usual LAN, TCP/IP, and realtime protocols, desirably augmented with reliable and/or ordered broadcast and other protocols. The GEAR introduces some innovative ideas in distributed systems architecture. While serving an object model based on production and consumption of generic events, it treats events produced by several sources (environment, body, objects) in a homogeneous way. This is possible due to the use of a common basic dialect for talking about events and due to the existence of the translation layer, which performs the necessary translation between the physical representation of a real-time entity and the EC compliant format. Crucial to the architecture is the event layer, which uses event channels to propagate events through regular network infrastructures. The event layer is realized by the COSMIC middleware, as described in Section 7. 5.1 Information Flow in GEAR The flow of information (external environment and computational part) is seamlessly supported by the L-shaped architecture. It occurs in a number of different ways, which demonstrates the expressiveness of the model with regard to the necessary forms of information encountered in real-time cooperative and embedded systems. Smart sensors produce events which report on the environment. Body sensors produce events which report on the body. They are disseminated by the local event layer module, on an event channel (EC) propagated through the regular network, to any relevant remote event layer modules where entities showed an interest on them, normally, sentient objects attached to the respective local event layer modules. Sentient objects consume events they are interested in, process them, and produce other events. Some of these events are destined to other sentient objects. They are published on an EC using the same EC dialect that serves, e.g., sensor originated events. However, these events are semantically of a kind such that they are to be subscribed by the relevant sentient objects, for example, the sentient objects composing a robot controller system, or, at a higher level, the sentient objects composing the actual robots in 32 a cooperative application. Smart actuators, on the other hand, merely consume events produced by sentient objects, whereby they accept and execute actuation commands. Alternatively to talking to other sentient objects, sentient objects can produce events of a lower level, for example, actuation commands on the body or environment. They publish these exactly the same way: on an event channel through the local event layer representative. Now, if these commands are of concern to local actuator units (e.g., body, including internal operational networks), they are passed on to the local translation layer. If they are of concern to a remote smart actuator, they are disseminated through the distributed event layer, to reach the former. In any case, if they are also of interest to other entities, such as other sentient objects that wish to be informed of the actuation command, then they are also disseminated through the EC to these sentient objects. A key advantage of this architecture is that event-messages and physical events can be globally ordered, if necessary, since they all pass through the event layer. The model also offers opportunities to solve a long lasting problem in realtime, computer control, and embedded systems: the inconsistency between message passing and the feedback loop information flow subsystems. 6. TEMPORAL ASPECTS OF THE INTERACTIONS Any interaction needs some form of predictability. If safety critical scenarios are considered as it is done in CORTEX, temporal aspects become crucial and have to be made explicit. The problem is how to define temporal constraints and how to enforce them by appropriate resource usage in a dynamic ad-hoc environment. In an system where interactions are spontaneous, it may be also necessary to determine temporal properties dynamically. To do this, the respective temporal information must be stated explicitly and available during run-time. Secondly, it is not always ensured that temporal properties can be fulfilled. In these cases, adaptations and timing failure notification must be provided [2, 28]. In most real-time systems, the notion of a deadline is the prevailing scheme to express and enforce timeliness. However, a deadline only weakly reflect the temporal characteristics of the information which is handled. Secondly, a deadline often includes implicit knowledge about the system and the relations between activities. In a rather well defined, closed environment, it is possible to make such implicit assumptions and map these to execution times and deadlines. E.g. the engineer knows how long a vehicle position can be used before the vehicle movement outdates this information. Thus he maps this dependency between speed and position on a deadline which then assures that the position error can be assumed to be bounded. In a open environment, this implicit mapping is not possible any more because, as an obvious reason, the relation between speed and position, and thus the error bound, cannot easily be reverse engineered from a deadline. Therefore, our event model includes explicit quality attributes which allow to specify the temporal attributes for every individual event. This is of course an overhead compared to the use of implicit knowledge, but in a dynamic environment such information is needed. To illustrate the problem, consider the example of the position of a vehicle. A position is a typical example for time, value entity [30]. Thus, the position is useful if we can determine an error bound which is related to time, e.g. if we want a position error below 10 meters to establish a safety property between cooperating cars moving with 5 m/sec, the position has a validity time of 2 seconds. In a time, value entity entity we can trade time against the precision of the value. This is known as value over time and time over value [18]. Once having established the time-value relation and captured in event attributes, subscribers of this event can locally decide about the usefulness of an information. In the GEAR architecture temporal validity is used to reason about safety properties in a event-based system [29]. We will briefly review the respective notions and see how they are exploited in our COSMIC event middleware. Consider the timeline of generating an event representing some real-time entity [18] from its occurrence to the notification of a certain sentient object (Figure 3). The real-time entity is captured at the sensor interface of the system and has to be transformed in a form which can be treated by a computer. During the time interval t0 the sensor reads the real-time entity and a time stamp is associated with the respective value. The derived time, value entity represents an observation. It may be necessary to perform substantial local computations to derive application relevant information from the raw sensor data. However, it should be noted that the time stamp of the observation is associated with the capture time and thus independent from further signal processing and event generation. This close relationship between capture time and the associated value is supported by smart sensors described above. The processed sensor information is assembled in an event data structure after ts to be published to an event channel. As is described later, the event includes the time stamp of generation and the temporal validity as attributes. The temporal validity is an application defined measure for the expiration of a time, value . As we explained in the example of a position above, it may vary dependent on application parameters. Temporal validity is a more general concept than that of a deadline. It is independent of a certain technical implementation of a system. While deadlines may be used to schedule the respective steps in an event generation and dissemination, a temporal validity is an intrinsic property of a time, value entity carried in an event. A temporal validity allows to reason about the usefulness of information and is beneficial even in systems in which timely dissemination of events cannot be enforced because it enables timing failure detection at the event consumer. It is obvious that deadlines or periods can be derived from the temporal validity of an event. To set a deadline, knowledge of an implementation, worst case execution times or message dissemination latencies is necessary. Thus, in the timeline of Figure 3 every interval may have a deadline. Event dissemination through soft real-time channels in COSMIC exploits the temporal validity to define dissemination deadlines. Quality attributes can be defined, for instance, in terms of validity interval, omission degree pairs. These allow to characterize the usefulness of the event for a certain application, in a certain context. Because of that, quality attributes of an event clearly depend on higher level issues, such as the nature of the sentient object or of the smart sensor that produced the event. For instance, an event containing an indication of some vehicle speed must have different quality attributes depending on the kind of vehicle 33 real-world event observation: <time stamp, value> event generated ready to be transmitted event received notification , to t event producer communication network event consumer event channel push <event> , ts , tm , tt , tn , t o : t i m e t o o b t a i n a n o b s e r v a t i o n , t s : t i m e t o p r o c e s s s e n s o r r e a d i n g , t m : t i m e t o a s s e m b l e a n e v e n t m e s s a g e , t t : t i m e t o t r a n s f e r t h e e v e n t o n t h e r e g u l a r n e t w o r k , t n : t i m e f o r n o t i f i c a t i o n o n t h e c o n s u m e r s i t e Figure 3: Event processing and dissemination. from which it originated, or depending on its current speed. The same happens with the position event of the car example above, whose validity depends on the current speed and on a predefined required precision. However, since quality attributes are strictly related with the semantics of the application or, at least, with some high level knowledge of the purpose of the system (from which the validity of the information can be derived), the definition of these quality attributes may be done by exploiting the information provided at the programming interface. Therefore, it is important to understand how the system programmer can specify non-functional requirements at the API, and how these requirements translate into quality attributes assigned to events. While temporal validity is identified as an intrinsic event property, which is exploited to decide on the usefulness of data at a certain point in time, it is still necessary to provide a communication facility which can disseminate the event before the validity is expired. In a WAN-of-CANs network structure we have to cope with very different network characteristics and quality of service properties. Therefore, when crossing the network boundaries the quality of service guarantees available in a certain network will be lost and it will be very hard, costly and perhaps impossible to achieve these properties in the next larger area of the WAN-of CANs structure. CORTEX has a couple of abstractions to cope with this situation (network zones, body/environment) which have been discussed above. From the temporal point of view we need a high level abstraction like the temporal validity for the individual event now to express our quality requirements of the dissemination over the network. The bound, coverage pair, introduced in relation with the TCB [28] seems to be an appropriate approach. It considers the inherent uncertainty of networks and allows to trade the quality of dissemination against the resources which are needed. In relation with the event channel model discussed later, the bound, coverage pair allows to specify the quality properties of an event channel independently of specific technical issues. Given the typical environments in which sentient applications will operate, where it is difficult or even impossible to provide timeliness or reliability guarantees, we proposed an alternative way to handle non-functional application requirements, in relation with the TCB approach [28]. The proposed approach exploits intrinsic characteristics of applications, such as fail-safety, or time-elasticity, in order to secure QoS specifications of the form bound, coverage . Instead of constructing systems that rely on guaranteed bounds, the idea is to use (possibly changing) bounds that are secured with a constant probability all over the execution. This obviously requires an application to be able to adapt to changing conditions (and/or changing bounds) or, if this is not possible, to be able to perform some safety procedures when the operational conditions degrade to an unbearable level. The bounds we mentioned above refer essentially to timeliness bounds associated to the execution of local or distributed activities, or combinations thereof. From these bounds it is then possible to derive the quality attributes, in particular validity intervals, that characterize the events published in the event channel. 6.1 The Role of Smart Sensors and Actuators Smart devices encapsulate hardware, software and mechanical components and provide information and a set of well specified functions and which are closely related to the interaction with the environment. The built-in computational components and the network interface enable the implementation of a well-defined high level interface that does not just provide raw transducer data, but a processed, application-related set of events. Moreover, they exhibit an autonomous spontaneous behaviour. They differ from general purpose nodes because they are dedicated to a certain functionality which complies to their sensing and actuating capabilities while general purpose node may execute any program. Concerning the sentient object model, smart sensors and actuators may be basic sentient objects themselves, consuming events from the real-world environment and producing the respective generic events for the system"s event layer or, 34 vice versa consuming a generic event and converting it to a real-world event by an actuation. Smart components therefore constitute the periphery, i.e. the real-world interface of a more complex sentient object. The model of sentient objects also constitutes the framework to built more complex virtual sensors by relating multiple (primary, i.e. sensors which directly sense a physical entity) sensors. Smart components translate events of the environment to an appropriate form available at the event layer or, vice versa, transform a system event into an actuation. For smart components we can assume that: • Smart components have dedicated resources to perform a specific function. • These resources are not used for other purposes during normal real-time operation. • No local temporal conflicts occur that will change the observable temporal behaviour. • The functions of a component can usually only be changed during a configuration procedure which is not performed when the component is involved in critical operations. • An observation of the environment as a time,value pair can be obtained with a bounded jitter in time. Many predictability and scheduling problems arise from the fact, that very low level timing behaviours have to be handled on a single processor. Here, temporal encapsulation of activities is difficult because of the possible side effects when sharing a single processor resource. Consider the control of a simple IR-range detector which is used for obstacle avoidance. Dependent on its range and the speed of a vehicle, it has to be polled to prevent the vehicle from crashing into an obstacle. On a single central processor, this critical activity has to be coordinated with many similar, possibly less critical functions. It means that a very fine grained schedule has to be derived based purely on the artifacts of the low level device control. In a smart sensor component, all this low level timing behaviour can be optimized and encapsulated. Thus we can assume temporal encapsulation similar to information hiding in the functional domain. Of course, there is still the problem to guarantee that an event will be disseminated and recognized in due time by the respective system components, but this relates to application related events rather than the low artifacts of a device timing. The main responsibility to provide timeliness guarantees is shifted to the event layer where these events are disseminated. Smart sensors thus lead to network centric system model. The network constitute the shared resource which has to be scheduled in a predictable way. The COSMIC middleware introduced in the next section is an approach to provide predictable event dissemination for a network of smart sensors and actuators. 7. AN EVENT MODEL ANDMIDDLEWARE FOR COOPERATING SMART DEVICES An event model and a middleware suitable for smart components must support timely and reliable communication and also must be resource efficient. COSMIC (COoperating Smart devices) is aimed at supporting the interaction between those components according to the concepts introduced so far. Based on the model of a WAN-of-CANs, we assume that the components are connected to some form of CAN as a fieldbus or a special wireless sensor network which provides specific network properties. E.g. a fieldbus developed for control applications usually includes mechanisms for predictable communication while other networks only support a best effort dissemination. A gateway connects these CANs to the next level in the network hierarchy. The event system should allow the dynamic interaction over a hierarchy of such networks and comply with the overall CORTEX generic event model. Events are typed information carriers and are disseminated in a publisher/ subscriber style [24, 7], which is particularly suitable because it supports generative, anonymous communication [3] and does not create any artificial control dependencies between producers of information and the consumers. This decoupling in space (no references or names of senders or receivers are needed for communication) and the flow decoupling (no control transfer occurs with a data transfer) are well known [24, 7, 14] and crucial properties to maintain autonomy of components and dynamic interactions. It is obvious that not all networks can provide the same QoS guarantees and secondly, applications may have widely differing requirements for event dissemination. Additionally, when striving for predictability, resources have to be reserved and data structures must be set up before communication takes place. Thus, these things can not predictably be made on the fly while disseminating an event. Therefore, we introduced the notion of an event channel to cope with differing properties and requirements and have an object to which we can assign resources and reservations. The concept of an event channel is not new [10, 25], however, it has not yet been used to reflect the properties of the underlying heterogeneous communication networks and mechanisms as described by the GEAR architecture. Rather, existing event middleware allows to specify the priorities or deadlines of events handled in an event server. Event channels allow to specify the communication properties on the level of the event system in a fine grained way. An event channel is defined by: event channel := subject, quality attributeList, handlers The subject determines the types of events event which may be issued to the channel. The quality attributes model the properties of the underlying communication network and dissemination scheme. These attributes include latency specifications, dissemination constraints and reliability parameters. The notion of zones which represent a guaranteed quality of service in a subnetwork support this approach. Our goal is to handle the temporal specifications as bound, coverage pairs [28] orthogonal to the more technical questions of how to achieve a certain synchrony property of the dissemination infrastructure. Currently, we support quality attributes of event channels in a CAN-Bus environment represented by explicit synchrony classes. The COSMIC middleware maps the channel properties to lower level protocols of the regular network. Based on our previous work on predictable protocols for the CAN-Bus, COSMIC defines an abstract network which provides hard, soft and non real-time message classes [21]. Correspondingly, we distinguish three event channel classes according to their synchrony properties: hard real-time channels, soft real-time channels and non-real-time channels. Hard real-time channels (HRTC) guarantee event propagation within the defined time constraints in the presence 35 of a specified number of omission faults. HRTECs are supported by a reservation scheme which is similar to the scheme used in time-triggered protocols like TTP [16][31], TTP/A [17], and TTCAN [8]. However, a substantial advantage over a TDMA scheme is that due to CAN-Bus properties, bandwidth which was reserved but is not needed by a HRTEC can be used by less critical traffic [21]. Soft real-time channels (SRTC) exploit the temporal validity interval of events to derive deadlines for scheduling. The validity interval defines the point in time after which an event becomes temporally inconsistent. Therefore, in a real-time system an event is useless after this point and may me discarded. The transmission deadline (DL) is defined as the latest point in time when a message has to be transmitted and is specified in a time interval which is derived from the expiration time: tevent ready < DL < texpiration − ∆notification texpiration defines the point in time when the temporal validity expires. ∆notification is the expected end-to-end latency which includes the transfer time over the network and the time the event may be delayed by the local event handling in the nodes. As said before, event deadlines are used to schedule the dissemination by SRTECs. However, deadlines may be missed in transient overload situations or due to arbitrary arrival times of events. On the publisher side the application"s exception handler is called whenever the event deadline expires before event transmission. At this point in time the event is also not expected to arrive at the subscriber side before the validity expires. Therefore, the event is removed from the sending queue. On the subscriber side the expiration time is used to schedule the delivery of the event. If the event cannot be delivered until its expiration time it is removed from the respective queues allocated by the COSMIC middleware. This prevents the communication system to be loaded by outdated messages. Non-real-time channels do not assume any temporal specification and disseminate events in a best effort manner. An instance of an event channel is created locally, whenever a publisher makes an announcement for publication or a subscriber subscribes for an event notification. When a publisher announces publication, the respective data structures of an event channel are created by the middleware. When a subscriber subscribes to an event channel, it may specify context attributes of an event which are used to filter events locally. E.g. a subscriber may only be interested in events generated at a certain location. Additionally the subscriber specifies quality properties of the event channel. A more detailed description of the event channels can be found in [13]. Currently, COSMIC handles all event channels which disseminate events beyond the CAN network boundary as non real-time event channels. This is mainly because we use the TCP/IP protocol to disseminate events over wireless links or to the standard Ethernet. However, there are a number of possible improvements which can easily be integrated in the event channel model. The Timely Computing Base (TCB) [28] can be exploited for timing failure detection and thus would provide awareness for event dissemination in environments where timely delivery of events cannot be enforced. Additionally, there are wireless protocols which can provide timely and reliable message delivery [6, 23] which may be exploited for the respective event channel classes. Events are the information carriers which are exchanged between sentient objects through event channels. To cope with the requirements of an ad-hoc environment, an event includes the description of the context in which it has been generated and quality attributes defining requirements for dissemination. This is particularly important in an open, dynamic environment where an event may travel over multiple networks. An event instance is specified as: event := subject, context attributeList, quality attributeList, contents A subject defines the type of the event and is related to the event contents. It supports anonymous communication and is used to route an event. The subject has to match to the subject of the event channel through which the event is disseminated. Attributes are complementary to the event contents. They describe individual functional and non-functional properties of the event. The context attributes describe the environment in which the event has been generated, e.g. a location, an operational mode or a time of occurrence. The quality attributes specify timeliness and dependability aspects in terms of validity interval, omission degree pairs. The validity interval defines the point in time after which an event becomes temporally inconsistent [18]. As described above, the temporal validity can be mapped to a deadline. However, usually a deadline is an engineering artefact which is used for scheduling while the temporal validity is a general property of a time, value entity. In a environment where a deadline cannot be enforced, a consumer of an event eventually must decide whether the event still is temporally consistent, i.e. represents a valid time, value entity. 7.1 The Architecture of the COSMIC Middleware On the architectural level, COSMIC distinguish three layers roughly depicted in Figure 4. Two of them, the event layer and the abstract network layer are implemented by the COSMIC middleware. The event layer provides the API for the application and realizes the abstraction of event and event channels. The abstract network implements real-time message classes and adapts the quality requirements to the underlying real network. An event channel handler resides in every node. It supports the programming interface and provides the necessary data structures for event-based communication. Whenever an object subscribes to a channel or a publisher announces a channel, the event channel handler is involved. It initiates the binding of the channel"s subject, which is represented by a network independent unique identifier to an address of the underlying abstract network to enable communication [14]. The event channel handler then tightly cooperates with the respective handlers of the abstract network layer to disseminate events or receive event notifications. It should be noted that the QoS properties of the event layer in general depend on what the abstract network layer can provide. Thus, it may not always be possible to e.g. support hard real-time event channels because the abstract network layer cannot provide the respective guarantees. In [13], we describe the protocols and services of the abstract network layer particularly for the CAN-Bus. As can be seen in Figure 4, the hard real-time (HRT) message class is supported by a dedicated handler which is able to provide the time triggered message dissemination. 36  event notifications HRT-msg list SRT-msg queue NRT-msg queue HRT-msg calendar HRTC Handler S/NRTC Handler Abstract Network Layer CAN Layer RX Buffer TX Buffer RX, TX, error interrupts Event Channel Specs. Event Layer send messages exception notification exceptions, notifications ECH: Event Channel Handler p u b l i s h a n n o u n c e s u b s c r i b e b i n d i n g p r o t o c o l c o n f i g . p r o t o c o l Global Time Service event notifications HRT-msg list SRT-msg queue NRT-msg queue HRT-msg calendar HRTC Handler S/NRTC Handler Abstract Network Layer CAN Layer RX Buffer TX Buffer RX, TX, error interrupts Event Channel Specs. Event Layer send messages exception notification exceptions, notifications ECH: Event Channel Handler p u b l i s h a n n o u n c e s u b s c r i b e b i n d i n g p r o t o c o l c o n f i g . p r o t o c o l Global Time Service Figure 4: Architecture layers of COSMIC. The HRT handler maintains the HRT message list, which contains an entry for each local HRT message to be sent. The entry holds the parameters for the message, the activation status and the binding information. Messages are scheduled on the bus according to the HRT message calendar which comprises the precise start time for each time slot allocated for a message. Soft real-time message queues order outgoing messages according to their transmission deadlines derived from the temporal validity interval. If the transmission deadline is exceeded, the event message is purged out of the queue. The respective application is notified via the exception notification interface and can take actions like trying to publish the event again or publish it to a channel of another class. Incoming event messages are ordered according to their temporal validity. If an event message arrive, the respective applications are notified. At the moment, an outdated message is deleted from the queue and if the queue runs out of space, the oldest message is discarded. However, there are other policies possible depending on event attributes and available memory space. Non real-time messages are FIFO ordered in a fixed size circular buffer. 7.2 Status of COSMIC The goal for developing COSMIC was to provide a platform to seamlessly integrate smart tiny components in a large system. Therefore, COSMIC should run also on the small, resource constraint devices which are built around 16Bit or even 8-Bit micro-controllers. The distributed COSMIC middleware has been implemented and tested on various platforms. Under RT-Linux, we support the real-time channels over the CAN Bus as described above. The RTLinux version runs on Pentium processors and is currently evaluated before we intent to port it to a smart sensor or actuator. For the interoperability in a WAN-of-CANs environment, we only provide non real-time channels at the moment. This version includes a gateway between the CANbus and a TCP/IP network. It allows us to use a standard wireless 802.11 network. The non real-time version of COSMIC is available on Linux, RT-Linux and on the microcontroller families C167 (Infineon) and 68HC908 (Motorola). Both micro-controllers have an on-board CAN controller and thus do not require additional hardware components for the network. The memory footprint of COSMIC is about 13 Kbyte on a C167 and slightly more on the 68HC908 where it fits into the on-board flash memory without problems. Because only a few channels are required on such a smart sensor or actuator component, the requirement of RAM (which is a scarce resource on many single chip systems) to hold the dynamic data structures of a channel is low. The COSMIC middleware makes it very easy to include new smart sensors in an existing system. Particularly, the application running on a smart sensor to condition and process the raw physical data must not be aware of any low level network specific details. It seamlessly interacts with other components of the system exclusively via event channels. The demo example, briefly described in the next chapter, is using a distributed infrastructure of tiny smart sensors and actuators directly cooperating via event channels over heterogeneous networks. 8. AN ILLUSTRATIVE EXAMPLE A simple example for many important properties of the proposed system showing the coordination through the environment and events disseminated over the network is the demo of two cooperating robots depicted in Figure 5. Each robot is equipped with smart distance sensors, speed sensors, acceleration sensors and one of the robots (the guide (KURT2) in front (Figure 5)) has a tracking camera allowing to follow a white line. The robots form a WAN-of-CANs system in which their local CANs are interconnected via a wireless 802.11 network. COSMIC provides the event layer for seamless interaction. The blind robot (N.N.) is searching the guide randomly. Whenever the blind robot detects (by its front distance sensors) an obstacle, it checks whether this may be the guide. For this purpose, it dynamically subscribes to the event channel disseminating distance events from rear distance sensors of the guide(s) and compares these with the distance events from its local front sensors. If the distance is approximately the same it infers that it is really behind a guide. Now N.N. also subscribes to the event channels of the tracking camera and the speed sensors 37 Figure 5: Cooperating robots. to follow the guide. The demo application highlights the following properties of the system: 1. Dynamic interaction of robots which is not known in advance. In principle, any two a priori unknown robots can cooperate. All what publishers and subscribers have to know to dynamically interact in this environment is the subject of the respective event class. A problem will be to receive only the events of the robot which is closest. A robot identity does not help much to solve this problem. Rather, the position of the event generation entity which is captured in the respective attributes can be evaluated to filter the relevant event out of the event stream. A suitable wireless protocol which uses proximity to filter events has been proposed by Meier and Cahill [22] in the CORTEX project. 2. Interaction through the environment. The cooperation between the robots is controlled by sensing the distance between the robots. If the guide detects that the distance grows, it slows down. Respectively, if the blind robot comes too close it reduces its speed. The local distance sensors produce events which are disseminated through a low latency, highly predictable event channel. The respective reaction time can be calculated as function of the speed and the distance of the robots and define a dynamic dissemination deadline for events. Thus, the interaction through the environment will secure the safety properties of the application, i.e. the follower may not crash into the guide and the guide may not loose the follower. Additionally, the robots have remote subscriptions to the respective distance events which are used to check it with the local sensor readings to validate that they really follow the guide which they detect with their local sensors. Because there may be longer latencies and omissions, this check occasionally will not be possible. The unavailability of the remote events will decrease the quality of interaction and probably and slow down the robots, but will not affect safety properties. 3. Cooperative sensing. The blind robot subscribes to the events of the line tracking camera. Thus it can see through the eye of the guide. Because it knows the distance to the guide and the speed as well, it can foresee the necessary movements. The proposed system provides the architectural framework for such a cooperation. The respective sentient object controlling the actuation of the robot receives as input the position and orientation of the white line to be tracked. In the case of the guide robot, this information is directly delivered as a body event with a low latency and a high reliability over the internal network. For the follower robot, the information comes also via an event channel but with different quality attributes. These quality attributes are reflected in the event channel description. The sentient object controlling the actuation of the follower is aware of the increased latency and higher probability of omission. 9. CONCLUSION AND FUTURE WORK The paper addresses problems of building large distributed systems interacting with the physical environment and being composed from a huge number of smart components. We cannot assume that the network architecture in such a system is homogeneous. Rather multiple edge- networks are fused to a hierarchical, heterogeneous wide area network. They connect the tiny sensors and actuators perceiving the environment and providing sentience to the application. Additionally, mobility and dynamic deployment of components require the dynamic interaction without fixed, a priori known addressing and routing schemes. The work presented in the paper is a contribution towards the seamless interaction in such an environment which should not be restricted by technical obstacles. Rather it should be possible to control the flow of information by explicitly specifying functional and temporal dissemination constraints. The paper presented the general model of a sentient object to describe composition, encapsulation and interaction in such an environment and developed the Generic Event Architecture GEAR which integrates the interaction through the environment and the network. While appropriate abstractions and interaction models can hide the functional heterogeneity of the networks, it is impossible to hide the quality differences. Therefore, one of the main concerns is to define temporal properties in such an open infrastructure. The notion of an event channel has been introduced which allows to specify quality aspects explicitly. They can be verified at subscription and define a boundary for event dissemination. The COSMIC middleware is a first attempt to put these concepts into operation. COSMIC allows the interoperability of tiny components over multiple network boundaries and supports the definition of different real-time event channel classes. There are many open questions that emerged from our work. One direction of future research will be the inclusion of real-world communication channels established between sensors and actuators in the temporal analysis and the ordering of such events in a cause-effect chain. Additionally, the provision of timing failure detection for the adaptation of interactions will be in the focus of our research. To reduce network traffic and only disseminate those events to the subscribers which they are really interested in and which have a chance to arrive timely, the encapsulation and scoping schemes have to be transformed into respective multi-level filtering rules. The event attributes which describe aspects of the context and temporal constraints for the dissemination will be exploited for this purpose. Finally, it is intended to integrate the results in the COSMIC middleware to enable experimental assessment. 38 10. REFERENCES [1] J. Bacon, K. Moody, J. Bates, R. Hayton, C. Ma, A. McNeil, O. Seidel, and M. Spiteri. Generic support for distributed applications. IEEE Computer, 33(3):68-76, 2000. [2] L. B. Becker, M. Gergeleit, S. Schemmer, and E. Nett. Using a flexible real-time scheduling strategy in a distributed embedded application. In Proc. of the 9th IEEE International Conference on Emerging Technologies and Factory Automation (ETFA), Lisbon, Portugal, Sept. 2003. [3] N. Carriero and D. Gelernter. Linda in context. Communications of the ACM, 32(4):444-458, apr 1989. [4] A. Casimiro (Ed.). Preliminary definition of cortex system architecture. CORTEX project, IST-2000-26031, Deliverable D4, Apr. 2002. [5] CORTEX project Annex 1, Description of Work. Technical report, CORTEX project, IST-2000-26031, Oct. 2000. http://cortex.di.fc.ul.pt. [6] R. Cunningham and V. Cahill. Time bounded medium access control for ad-hoc networks. In Proceedings of the Second ACM International Workshop on Principles of Mobile Computing (POMC"02), pages 1-8, Toulouse, France, Oct. 2002. ACM Press. [7] P. T. Eugster, P. Felber, R. Guerraoui, and A.-M. Kermarrec. The many faces of publish/subscribe. Technical Report DSC ID:200104, EPFL, Lausanne, Switzerland, 2001. [8] T. F¨uhrer, B. M¨uller, W. Dieterle, F. Hartwich, R. Hugel, and M.Walther. Time triggered communication on CAN, 2000. http://www.can-cia.org/can/ttcan/fuehrer.pdf. [9] R. B. GmbH. CAN Specification Version 2.0. Technical report, Sept. 1991. [10] T. Harrison, D. Levine, and D. Schmidt. The design and performance of a real-time corba event service. In Proceedings of the 1997 Conference on Object Oriented Programming Systems, Languages and Applications (OOPSLA), pages 184-200, Atlanta, Georgia, USA, 1997. ACM Press. [11] J. Hightower and G. Borriello. Location systems for ubiquitous computing. IEEE Computer, 34(8):57-66, aug 2001. [12] A. Hopper. The Clifford Paterson Lecture, 1999 Sentient Computing. Philosophical Transactions of the Royal Society London, 358(1773):2349-2358, Aug. 2000. [13] J. Kaiser, C. Mitidieri, C. Brudna, and C. Pereira. COSMIC: A Middleware for Event-Based Interaction on CAN. In Proc. 2003 IEEE Conference on Emerging Technologies and Factory Automation, Lisbon, Portugal, Sept. 2003. [14] J. Kaiser and M. Mock. Implementing the real-time publisher/subscriber model on the controller area network (CAN). In Proceedings of the 2nd International Symposium on Object-oriented Real-time distributed Computing (ISORC99), Saint-Malo, France, May 1999. [15] K. Kim, G. Jeon, S. Hong, T. Kim, and S. Kim. Integrating subscription-based and connection-oriented communications into the embedded CORBA for the CAN Bus. In Proceedings of the IEEE Real-time Technology and Application Symposium, May 2000. [16] H. Kopetz and G. Gr¨unsteidl. TTP - A Time-Triggered Protocol for Fault-Tolerant Real-Time Systems. Technical Report rr-12-92, Institut f¨ur Technische Informatik, Technische Universit¨at Wien, Treilstr. 3/182/1, A-1040 Vienna, Austria, 1992. [17] H. Kopetz, M. Holzmann, and W. Elmenreich. A Universal Smart Transducer Interface: TTP/A. International Journal of Computer System, Science Engineering, 16(2), Mar. 2001. [18] H. Kopetz and P. Ver´ıssimo. Real-time and Dependability Concepts. In S. J. Mullender, editor, Distributed Systems, 2nd Edition, ACM-Press, chapter 16, pages 411-446. Addison-Wesley, 1993. [19] S. Lankes, A. Jabs, and T. Bemmerl. Integration of a CAN-based connection-oriented communication model into Real-Time CORBA. In Workshop on Parallel and Distributed Real-Time Systems, Nice, France, Apr. 2003. [20] Local Interconnect Network: LIN Specification Package Revision 1.2. Technical report, Nov. 2000. [21] M. Livani, J. Kaiser, and W. Jia. Scheduling hard and soft real-time communication in the controller area network. Control Engineering, 7(12):1515-1523, 1999. [22] R. Meier and V. Cahill. Steam: Event-based middleware for wireless ad-hoc networks. In Proceedings of the International Workshop on Distributed Event-Based Systems (ICDCS/DEBS"02), pages 639-644, Vienna, Austria, 2002. [23] E. Nett and S. Schemmer. Reliable real-time communication in cooperative mobile applications. IEEE Transactions on Computers, 52(2):166-180, Feb. 2003. [24] B. Oki, M. Pfluegl, A. Seigel, and D. Skeen. The information bus - an architecture for extensible distributed systems. Operating Systems Review, 27(5):58-68, 1993. [25] O. M. G. (OMG). CORBAservices: Common Object Services Specification - Notification Service Specification, Version 1.0, 2000. [26] O. M. G. (OMG). Smart transducer interface, initial submission, June 2001. [27] P. Ver´ıssimo, V. Cahill, A. Casimiro, K. Cheverst, A. Friday, and J. Kaiser. Cortex: Towards supporting autonomous and cooperating sentient entities. In Proceedings of European Wireless 2002, Florence, Italy, Feb. 2002. [28] P. Ver´ıssimo and A. Casimiro. The Timely Computing Base model and architecture. Transactions on Computers - Special Issue on Asynchronous Real-Time Systems, 51(8):916-930, Aug. 2002. [29] P. Ver´ıssimo and A. Casimiro. Event-driven support of real-time sentient objects. In Proceedings of the 8th IEEE International Workshop on Object-oriented Real-time Dependable Systems, Guadalajara, Mexico, Jan. 2003. [30] P. Ver´ıssimo and L. Rodrigues. Distributed Systems for System Architects. Kluwer Academic Publishers, 2001. 39

smart sensor;sensor and actuator;corba;dissemination quality;sentient object;event channel;real-time entity;generic event architecture;gear architecture;soft real-time channel;temporal constraint;event-based system;cosmic middleware;middleware architecture;cortex;temporal validity;event-base communication;componentbase system;sentient computing

train_C-79

A Cross-Layer Approach to Resource Discovery and Distribution in Mobile ad-hoc Networks

This paper describes a cross-layer approach to designing robust P2P system over mobile ad-hoc networks. The design is based on simple functional primitives that allow routing at both P2P and network layers to be integrated to reduce overhead. With these primitives, the paper addresses various load balancing techniques. Preliminary simulation results are also presented.

1. INTRODUCTION Mobile ad-hoc networks (MANETs) consist of mobile nodes that autonomously establish connectivity via multi-hop wireless communications. Without relying on any existing, pre-configured network infrastructure or centralized control, MANETs are useful in situations where impromptu communication facilities are required, such as battlefield communications and disaster relief missions. As MANET applications demand collaborative processing and information sharing among mobile nodes, resource (service) discovery and distribution have become indispensable capabilities. One approach to designing resource discovery and distribution schemes over MANETs is to construct a peer-to-peer (P2P) system (or an overlay) which organizes peers of the system into a logical structure, on top of the actual network topology. However, deploying such P2P systems over MANETs may result in either a large number of flooding operations triggered by the reactive routing process, or inefficiency in terms of bandwidth utilization in proactive routing schemes. Either way, constructing an overlay will potentially create a scalability problem for large-scale MANETs. Due to the dynamic nature of MANETs, P2P systems should be robust by being scalable and adaptive to topology changes. These systems should also provide efficient and effective ways for peers to interact, as well as other desirable application specific features. This paper describes a design paradigm that uses the following two functional primitives to design robust resource discovery and distribution schemes over MANETs. 1. Positive/negative feedback. Query packets are used to explore a route to other peers holding resources of interest. Optionally, advertisement packets are sent out to advertise routes from other peers about available resources. When traversing a route, these control packets measure goodness of the route and leave feedback information on each node along the way to guide subsequent control packets to appropriate directions. 2. Sporadic random walk. As the network topology and/or the availability of resources change, existing routes may become stale while better routes become available. Sporadic random walk allows a control packet to explore different paths and opportunistically discover new and/or better routes. Adopting this paradigm, the whole MANET P2P system operates as a collection of autonomous entities which consist of different types of control packets such as query and advertisement packets. These packets work collaboratively, but indirectly, to achieve common tasks, such as resource discovery, routing, and load balancing. With collaboration among these entities, a MANET P2P system is able to ‘learn" the network dynamics by itself and adjust its behavior accordingly, without the overhead of organizing peers into an overlay. The remainder of this paper is organized as follows. Related work is described in the next section. Section 3 describes the resource discovery scheme. Section 4 describes the resource distribution scheme. The replica invalidation scheme is described in Section 5, followed by it performance evaluation in Section 6. Section 7 concludes the paper. 2. RELATED WORK For MANETs, P2P systems can be classified based on the design principle, into layered and cross-layer approaches. A layered approach adopts a P2P-like [1] solution, where resource discovery is facilitated as an application layer protocol and query/reply messages are delivered by the underlying MANET routing protocols. For instance, Konark [2] makes use of a underlying multicast protocol such that service providers and queriers advertise and search services via a predefined multicast group, respectively. Proem [3] is a high-level mobile computing platform for P2P systems over MANETs. It defines a transport protocol that sits on top of the existing TCP/IP stack, hence relying on an existing routing protocol to operate. With limited control over how control and data packets are routed in the network, it is difficult to avoid the inefficiency of the general-purpose routing protocols which are often reactive and flooding-based. In contrast, cross-layer approaches either relies on its own routing mechanism or augments existing MANET routing algorithms to support resource discovery. 7DS [4], which is the pioneering work deploying P2P system on mobile devices, exploits data locality and node mobility to dissemination data in a single-hop fashion. Hence, long search latency may be resulted as a 7DS node can get data of interest only if the node that holds the data is in its radio coverage. Mohan et al. [5] propose an adaptive service discovery algorithm that combines both push and pull models. Specifically, a service provider/querier broadcasts advertisement/query only when the number of nodes advertising or querying, which is estimated by received control packets, is below a threshold during a period of time. In this way, the number of control packets on the network is constrained, thus providing good scalability. Despite the mechanism to reduce control packets, high overhead may still be unavoidable, especially when there are many clients trying to locate different services, due to the fact that the algorithm relies on flooding, For resource replication, Yin and Cao [6] design and evaluate cooperative caching techniques for MANETs. Caching, however, is performed reactively by intermediate nodes when a querier requests data from a server. Data items or resources are never pushed into other nodes proactively. Thanedar et al. [7] propose a lightweight content replication scheme using an expanding ring technique. If a server detects the number of requests exceed a threshold within a time period, it begins to replicate its data onto nodes capable of storing replicas, whose hop counts from the server are of certain values. Since data replication is triggered by the request frequency alone, it is possible that there are replicas unnecessarily created in a large scope even though only nodes within a small range request this data. Our proposed resource replication mechanism, in contrast, attempts to replicate a data item in appropriate areas, instead of a large area around the server, where the item is requested frequently. 3. RESOURCE DISCOVERY We propose a cross-layer, hybrid resource discovery scheme that relies on the multiple interactions of query, reply and advertisement packets. We assume that each resource is associated with a unique ID1 . Initially, when a node wants to discover a resource, it deploys query packets, which carry the corresponding resource ID, and randomly explore the network to search for the requested resource. Upon receiving such a query packet, a reply packet is generated by the node providing the requested resource. Advertisement packets can also be used to proactively inform other nodes about what resources are available at each node. In addition to discovering the ‘identity" of the node providing the requested resource, it may be also necessary to discover a ‘route" leading to this node for further interaction. To allow intermediate nodes to make a decision on where to forward query packets, each node maintains two tables: neighbor 1 The assumption of unique ID is made for brevity in exposition, and resources could be specified via attribute-value assertions. table and pheromone table. The neighbor table maintains a list of all current neighbors obtained via a neighbor discovery protocol. The pheromone table maintains the mapping of a resource ID and a neighbor ID to a pheromone value. This table is initially empty, and is updated by a reply packet generated by a successful query. Figure 1 illustrates an example of a neighbor table and a pheromone table maintained by node A having four neighbors. When node A receives a query packet searching for a resource, it makes a decision to which neighbor it should forward the query packet by computing the desirability of each of the neighbors that have not been visited before by the same query packet. For a resource ID r, the desirability of choosing a neighbor n, δ(r,n), is obtained from the pheromone value of the entry whose neighbor and resource ID fields are n and r, respectively. If no such entry exists in the pheromone table, δ(r,n) is set to zero. Once the desirabilities of all valid next hops have been calculated, they are normalized to obtain the probability of choosing each neighbor. In addition, a small probability is also assigned to those neighbors with zero desirability to exercise the sporadic random walk primitive. Based on these probabilities, a next hop is selected to forward the query packet to. When a query packet encounters a node with a satisfying resource, a reply packet is returned to the querying node. The returning reply packet also updates the pheromone table at each node on its return trip by increasing the pheromone value in the entry whose resource ID and neighbor ID fields match the ID of the discovered resource and the previous hop, respectively. If such an entry does not exist, a new entry is added into the table. Therefore, subsequent query packets looking for the same resource, when encountering this pheromone information, are then guided toward the same destination with a small probability of taking an alternate path. Since the hybrid discovery scheme neither relies on a MANET routing protocol nor arranges nodes into a logical overlay, query packets are to traverse the actual network topology. In dense networks, relatively large nodal degrees can have potential impacts on this random exploring mechanism. To address this issue, the hybrid scheme also incorporates proactive advertisement in addition to the reactive query. To perform proactive advertisement, each node periodically deploys an advertising packet containing a list of its available resources" IDs. These packets will traverse away from the advertising node in a random walk manner up to a limited number of hops and advertise resource information to surrounding nodes in the same way as reply packets. In the hybrid scheme, an increase of pheromone serves as a positive feedback which indirectly guides query packets looking for similar resources. Intuitively, the amount of pheromone increased is inversely proportional to the distance the reply packet has traveled back, and other metrics, such as quality of the resource, could contribute to this amount as well. Each node also performs an implicit negative feedback for resources that have not been given a positive feedback for some time by regularly decreasing the pheromone in all of its pheromone table entries over time. In addition, pheromone can be reduced by an explicit negative response, for instance, a reply packet returning from a node that is not willing to provide a resource due to excessive workload. As a result, load balancing can be achieved via positive and negative feedback. A node serving too many nodes can either return fewer responses to query packets or generate negative responses. 2 The 3rd International Conference on Mobile Technology, Applications and Systems - Mobility 2006 Figure 1: Example illustrating neighbor and pheromone tables maintained by node A: (a) wireless connectivity around A showing that it currently has four neighbors, (b) A"s neighbor table, and (c) a possible pheromone table of A Figure 2: Sample scenarios illustrating the three mechanisms supporting load-balancing: (a) resource replication, (b) resource relocation, and (c) resource division 4. RESOURCE DISTRIBUTION In addition to resource discovery, a querying node usually attempts to access and retrieve the contents of a resource after a successful discovery. In certain situations, it is also beneficial to make a resource readily available at multiple nodes when the resource can be relocated and/or replicated, such as data files. Furthermore, in MANETs, we should consider not only the amount of load handled by a resource provider, but also the load on those intermediate nodes that are located on the communication paths between the provider and other nodes as well. Hence, we describe a cross-layer, hybrid resource distribution scheme to achieve load balancing by incorporating the functionalities of resource relocation, resource replication, and resource division. 4.1 Resource Replication Multiple replicas of a resource in the network help prevent a single node, as well as nodes surrounding it, from being overloaded by a large number of requests and data transfers. An example is when a node has obtained a data file from another node, the requesting node and the intermediate nodes can cache the file and start sharing that file with other surrounding nodes right away. In addition, replicable resources can also be proactively replicated at other nodes which are located in certain strategic areas. For instance, to help nodes find a resource quickly, we could replicate the resource so that it becomes reachable by random walk for a specific number of hops from any node with some probability, as depicted in Figure 2(a). To realize this feature, the hybrid resource distribution scheme employs a different type of control packet, called resource replication packet, which is responsible for finding an appropriate place to create a replica of a resource. A resource replication packet of type R is deployed by a node that is providing the resource R itself. Unlike a query packet which follows higher pheromone upstream toward a resource it is looking for, a resource replication packet tends to be propelled away from similar resources by moving itself downstream toward weaker pheromone. When a resource replication packet finds itself in an area with sufficiently low pheromone, it makes a decision whether it should continue exploring or turn back. The decision depends on conditions such as current workload and/or remaining energy of the node being visited, as well as popularity of the resource itself. 4.2 Resource Relocation In certain situations, a resource may be required to transfer from one node to another. For example, a node may no longer want to possess a file due to the shortage of storage space, but it cannot simply delete the file since other nodes may still need it in the future. In this case, the node can choose to create replicas of the file by the aforementioned resource replication mechanism and then delete its own copy. Let us consider a situation where a majority of nodes requesting for a resource are located far away from a resource provider, as shown on the top of Figure 2(b). If the resource R is relocatable, it is preferred to be relocated to another area that is closer to those nodes, similar to the bottom of the same figure. Hence network bandwidth is more efficiently utilized. The 3rd Conference on Mobile Technology, Applications and Systems - Mobility 2006 3 The hybrid resource distribution scheme incorporates resource relocation algorithms that are adaptive to user requests and aim to reduce communication overhead. Specifically, by following the same pheromone maintenance concept, the hybrid resource distribution scheme introduces another type of pheromone which corresponds to user requests, instead of resources. This type of pheromone, called request pheromone, is setup by query packets that are in their exploring phases (not returning ones) to guide a resource to a new location. 4.3 Resource Division Certain types of resources can be divided into smaller subresources (e.g., a large file being broken into smaller files) and distributed to multiple locations to avoid overloading a single node, as depicted in Figure 2(c). The hybrid resource distribution scheme incorporates a resource division mechanism that operates at a thin layer right above all the other mechanisms described earlier. The resource division mechanism is responsible for decomposing divisible resources into sub-resources, and then adds an extra keyword to distinguish each sub-resource from one another. Therefore, each of these sub-resources will be seen by the other mechanisms as one single resource which can be independently discovered, replicated, and relocated. The resource division mechanism is also responsible for combining data from these subresources together (e.g., merging pieces of a file) and delivering the final result to the application. 5. REPLICA INVALIDATION Although replicas improve accessibility and balance load, replica invalidation becomes a critical issue when nodes caching updatable resources may concurrently update their own replicas, which renders replicas held by other nodes obsolete. Most existing solutions to the replica invalidation problem either impose constrains that only the data source could perform update and invalidate other replicas, or resort to network-wide flooding which results in heavy network traffic and leads to scalability problem, or both. The lack of infrastructure supports and frequent topology changes in MANETs further challenge the issue. We apply the same cross-layer paradigm to invalidating replicas in MANETs which allows concurrent updates performed by multiple replicas. To coordinate concurrent updates and disseminate replica invalidations, a special infrastructure, called validation mesh or mesh for short, is adaptively maintained among nodes possessing ‘valid" replicas of a resource. Once a node has updated its replica, an invalidation packet will only be disseminated over the validation mesh to inform other replica-possessing nodes that their replicas become invalid and should be deleted. The structure (topology) of the validation mesh keeps evolving (1) when nodes request and cache a resource, (2) when nodes update their respective replicas and invalidate other replicas, and (3) when nodes move. To accommodate the dynamics, our scheme integrates the components of swarm intelligence to adaptively maintain the validation mesh without relying on any underlying MANET routing protocol. In particular, the scheme takes into account concurrent updates initiated by multiple nodes to ensure the consistency among replicas. In addition, version number is used to distinguish new from old replicas when invalidating any stale replica. Simulation results show that the proposed scheme effectively facilitates concurrent replica updates and efficiently perform replica invalidation without incurring network-wide flooding. Figure 3 depicts the idea of ‘validation mesh" which maintains connectivity among nodes holding valid replicas of a resource to avoid network-wide flooding when invalidating replicas. Figure 3: Examples showing maintenance of validation mesh There are eight nodes in the sample network, and we start with only node A holding the valid file, as shown in Figure 3(a). Later on, node G issues a query packet for the file and eventually obtains the file from A via nodes B and D. Since intermediate nodes are allowed to cache forwarded data, nodes B, D, and G will now hold valid replicas of the file. As a result, a validation mesh is established among nodes A, B, D, and G, as depicted in Figure 3(b). In Figure 3(c), another node, H, has issued a query packet for the same file and obtained it from node B"s cache via node E. At this point, six nodes hold valid replicas and are connected through the validation mesh. Now we assume node G updates its replica of the file and informs the other nodes by sending an invalidation packet over the validation mesh. Consequently, all other nodes except G remove their replicas of the file from their storage and the validation mesh is torn down. However, query forwarding pheromone, as denoted by the dotted arrows in Figure 3(d), is setup at these nodes via the ‘reverse paths" in which the invalidation packets have traversed, so that future requests for this file will be forwarded to node G. In Figure 3(e), node H makes a new request for the file again. This time, its query packet follows the pheromone toward node G, where the updated file can be obtained. Eventually, a new validation mesh is established over nodes G, B, D, E, and H. To maintain a validation mesh among the nodes holding valid replicas, one of them is designated to be the focal node. Initially, the node that originally holds the data is the focal node. As nodes update replicas, the node that last (or most recently) updates a 4 The 3rd International Conference on Mobile Technology, Applications and Systems - Mobility 2006 corresponding replica assumes the role of focal node. We also name nodes, such as G and H, who originate requests to replicate data as clients, and nodes B, D, and E who locally cache passing data as data nodes. For instance, in Figures 3(a), 3(b), and 3(c), node A is the focal node; in Figures 3(d), 3(e), and 3(f), node G becomes the focal node. In addition, to accommodate newly participating nodes and mobility of nodes, the focal node periodically floods the validation mesh with a keep-alive packet, so that nodes who can hear this packet are considered themselves to be part of the validation mesh. If a node holding a valid/updated replica doesn"t hear a keep-alive packet for a certain time interval, it will deploy a search packet using the resource discovery mechanism described in Section 3 to find another node (termed attachment point) currently on the validation mesh so that it can attach itself to. Once an attachment point is found, a search_reply packet is returned to the disconnected node who originated the search. Intermediate nodes who forward the search_reply packet will become part of the validation mesh as well. To illustrate the effect of node mobility, in Figure 3(f), node H has moved to a location where it is not directly connected to the mesh. Via the resource discovery mechanism, node H relies on an intermediate node F to connect itself to the mesh. Here, node F, although part of the validation mesh, doesn"t hold data replica, and hence is termed nondata node. Client and data node who keep hearing the keep-alive packets from the focal node act as if they are holding a valid replica, so that they can reply to query packets, like node B in Figure 3(c) replying a request from node H. While a disconnected node attempting to discover an attachment point to reattach itself to the mesh, the disconnected node can"t reply to a query packet. For instance, in Figure 3(f), node H does not reply to any query packet before it reattaches itself to the mesh. Although validation mesh provides a conceptual topology that (1) connects all replicas together, (2) coordinates concurrent updates, and (3) disseminates invalidation packets, the technical issue is how such a mesh topology could be effectively and efficiently maintained and evolved when (a) nodes request and cache a resource, (b) when nodes update their respective replicas and invalidate other replicas, and (c) when nodes move. Without relying on any MANET routing protocols, the two primitives work together to facilitate efficient search and adaptive maintenance. 6. PERFORMANCE EVALUATION We have conducted simulation experiments using the QualNet simulator to evaluate the performance of the described resource discovery, resource distribution, and replica invalidation schemes. However, due to space limitation only the performance of the replica invalidation is reported. In our experiments, eighty nodes are uniformly distributed over a terrain of size 1000×1000 m2 . Each node has a communication range of approximately 250 m over a 2 Mbps wireless channel, using IEEE802.11 as the MAC layer. We use the random-waypoint mobility model with a pause time of 1 second. Nodes may move at the minimum and maximum speeds of 1 m/s and 5 m/s, respectively. Table 1 lists other parameter settings used in the simulation. Initially, there is one resource server node in network. Two nodes are randomly picked up every 10 seconds as clients. Every β seconds, we check the number of nodes, N, which have gotten data. Then we randomly pickup Min(γ,N) nodes from them to initiate data update. Each experiment is run for 10 minutes. Table 1: Simulation Settings HOP_LIMIT 10 ADVERTISE_HOP_LIMIT 1 KEEPALIVE_INTERVAL 3 second NUM_SEARCH 1 ADVERTISE_INTERVAL 5 second EXPIRATION_INTERVAL 10 second Average Query Generation Rate 2 query/ 10 sec Max # of Concurrent Update (γ) 2 Frequency of Update (β) 3s We evaluate the performance under different mobility speed, the density, the maximum number of concurrent update nodes, and update frequency using two metrics: • Average overhead per update measures the average number of packets transmitted per update in the network. • Average delay per update measures how long our approach takes to finish an update on average. All figures shown present the results with a 70% confidence interval. Figure 4: Overhead vs. speed for 80 nodes Figure 5: Overhead vs. density Figure 6: Overhead vs. max #concurrent updates Figure 7: Overhead vs. freq. Figure 8: Delay vs. speed Figure 9: Delay vs. density The 3rd Conference on Mobile Technology, Applications and Systems - Mobility 2006 5 Figure 10: Delay vs. max #concurrent updates Figure 11: Delay vs. freq. Figures 4, 5, 6, and 7 show the overhead versus various parameter values. In Figure 4, the overhead increases as the speed increase, which is due to the fact that as the speed increase, nodes move out of mesh more frequently, and will send out more search packets. However, the overhead is not high, and even in speed 10m/sec, the overhead is below 100 packets. In contrast, the packets will be expected to be more than 200 packets at various speeds when flooding is used. Figure 5 shows that the overhead almost remains the same under various densities. That is attributed to only flooding over the mesh instead of the whole network. The size of mesh doesn"t vary much on various densities, so that the overhead doesn"t vary much. Figure 6 shows that overhead also almost remains the same under various maximum number of concurrent updates. That"s because one more node just means one more flood over the mesh during update process so that the impact is limited. Figure 7 shows that if updates happen more frequently, the overhead is higher. This is because the more quickly updates happen, (1) there will be more keep_alive message over the mesh between two updates, and (2) nodes move out of mesh more frequently and send out more search packets. Figures 8, 9, 10, and 11 show the delay versus various parameter values. From Figure 8, we know the delay increases as the speed increases, which is due to the fact that with increasing speed, clients will move out of mesh with higher probability. When these clients want to update data, they will spend time to first search the mesh. The faster the speed, the more time clients need to spend to search the mesh. Figure 9 shows that delay is negligibly affected by the density. Delay decreases slightly as the number of nodes increases, due to the fact that the more nodes in the network, the more nodes receives the advertisement packets which helps the search packet find the target so that the delay of update decreases. Figure 10 shows that delay decreases slightly as the maximum number of concurrent updates increases. The larger the maximum number of concurrent updates is, the more nodes are picked up to do update. Then with higher probability, one of these nodes is still in mesh and finishes the update immediately (don"t need to search mesh first), which decreases the delay. Figure 11 shows how the delay varies with the update frequency. When updates happen more frequently, the delay will higher. Because the less frequently, the more time nodes in mesh have to move out of mesh then they need to take time to search the mesh when they do update, which increases the delay. The simulation results show that the replica invalidation scheme can significantly reduce the overhead with an acceptable delay. 7. CONCLUSION To facilitate resource discovery and distribution over MANETs, one approach is to designing peer-to-peer (P2P) systems over MANETs which constructs an overlay by organizing peers of the system into a logical structure on the top of MANETs" physical topology. However, deploying overlay over MANETs may result in either a large number of flooding operations triggered by the routing process, or inefficiency in terms of bandwidth usage. Specifically, overlay routing relies on the network-layer routing protocols. In the case of a reactive routing protocol, routing on the overlay may cause a large number of flooded route discovery message since the routing path in each routing step must be discovered on demand. On the other hand, if a proactive routing protocol is adopted, each peer has to periodically broadcast control messages, which leads to poor efficiency in terms of bandwidth usage. Either way, constructing an overlay will potentially suffer from the scalability problem. The paper describes a design paradigm that uses the functional primitives of positive/negative feedback and sporadic random walk to design robust resource discovery and distribution schemes over MANETs. In particular, the scheme offers the features of (1) cross-layer design of P2P systems, which allows the routing process at both the P2P and the network layers to be integrated to reduce overhead, (2) scalability and mobility support, which minimizes the use of global flooding operations and adaptively combines proactive resource advertisement and reactive resource discovery, and (3) load balancing, which facilitates resource replication, relocation, and division to achieve load balancing. 8. REFERENCES [1] A. Oram, Peer-to-Peer: Harnessing the Power of Disruptive Technologies. O"Reilly, March 2000. [2] S. Helal, N. Desai, V. Verma, and C. Lee, Konark - A Service Discovery and Delivery Protocol for ad-hoc Networks, in the Third IEEE Conference on Wireless Communication Networks (WCNC), New Orleans, Louisiana, 2003. [3] G. Krotuem, Proem: A Peer-to-Peer Computing Platform for Mobile ad-hoc Networks, in Advanced Topic Workshop Middleware for Mobile Computing, Germany, 2001. [4] M. Papadopouli and H. Schulzrinne, A Performance Analysis of 7DS a Peer-to-Peer Data Dissemination and Prefetching Tool for Mobile Users, in Advances in wired and wireless communications, IEEE Sarnoff Symposium Digest, Ewing, NJ, 2001, (Best student paper & poster award). [5] U. Mohan, K. Almeroth, and E. Belding-Royer, Scalable Service Discovery in Mobile ad-hoc Networks, in IFIP Networking Conference, Athens, Greece, May 2004. [6] L. Yin and G. Cao, Supporting Cooperative Caching in Ad Hoc Networks, in IEEE INFOCOM, 2004. [7] V. Thanedar, K. Almeroth, and E. Belding-Royer, A Lightweight Content Replication Scheme for Mobile ad-hoc Environments, in IFIP Networking Conference, Athens, Greece, May 2004. 6 The 3rd International Conference on Mobile Technology, Applications and Systems - Mobility 2006

manet routing protocol;resource discovery;query packet;mobile ad-hoc network;manet p2p system;manet;neighbor discovery protocol;concurrent update;replica invalidation;invalidation packet;route discovery message;validation mesh;hybrid discovery scheme;negative feedback

train_C-80

Consistency-preserving Caching of Dynamic Database Content

With the growing use of dynamic web content generated from relational databases, traditional caching solutions for throughput and latency improvements are ineffective. We describe a middleware layer called Ganesh that reduces the volume of data transmitted without semantic interpretation of queries or results. It achieves this reduction through the use of cryptographic hashing to detect similarities with previous results. These benefits do not require any compromise of the strict consistency semantics provided by the back-end database. Further, Ganesh does not require modifications to applications, web servers, or database servers, and works with closed-source applications and databases. Using two benchmarks representative of dynamic web sites, measurements of our prototype show that it can increase end-to-end throughput by as much as twofold for non-data intensive applications and by as much as tenfold for data intensive ones.

1. INTRODUCTION An increasing fraction of web content is dynamically generated from back-end relational databases. Even when database content remains unchanged, temporal locality of access cannot be exploited because dynamic content is not cacheable by web browsers or by intermediate caching servers such as Akamai mirrors. In a multitiered architecture, each web request can stress the WAN link between the web server and the database. This causes user experience to be highly variable because there is no caching to insulate the client from bursty loads. Previous attempts in caching dynamic database content have generally weakened transactional semantics [3, 4] or required application modifications [15, 34]. We report on a new solution that takes the form of a databaseagnostic middleware layer called Ganesh. Ganesh makes no effort to semantically interpret the contents of queries or their results. Instead, it relies exclusively on cryptographic hashing to detect similarities with previous results. Hash-based similarity detection has seen increasing use in distributed file systems [26, 36, 37] for improving performance on low-bandwidth networks. However, these techniques have not been used for relational databases. Unlike previous approaches that use generic methods to detect similarity, Ganesh exploits the structure of relational database results to yield superior performance improvement. One faces at least three challenges in applying hash-based similarity detection to back-end databases. First, previous work in this space has traditionally viewed storage content as uninterpreted bags of bits with no internal structure. This allows hash-based techniques to operate on long, contiguous runs of data for maximum effectiveness. In contrast, relational databases have rich internal structure that may not be as amenable to hash-based similarity detection. Second, relational databases have very tight integrity and consistency constraints that must not be compromised by the use of hash-based techniques. Third, the source code of commercial databases is typically not available. This is in contrast to previous work which presumed availability of source code. Our experiments show that Ganesh, while conceptually simple, can improve performance significantly at bandwidths representative of today"s commercial Internet. On benchmarks modeling multitiered web applications, the throughput improvement was as high as tenfold for data-intensive workloads. For workloads that were not data-intensive, throughput improvements of up to twofold were observed. Even when bandwidth was not a constraint, Ganesh had low overhead and did not hurt performance. Our experiments also confirm that exploiting the structure present in database results is crucial to this performance improvement. 2. BACKGROUND 2.1 Dynamic Content Generation As the World Wide Web has grown, many web sites have decentralized their data and functionality by pushing them to the edges of the Internet. Today, eBusiness systems often use a three-tiered architecture consisting of a front-end web server, an application server, and a back-end database server. Figure 1 illustrates this architecture. The first two tiers can be replicated close to a concentration of clients at the edge of the Internet. This improves user experience by lowering end-to-end latency and reducing exposure WWW 2007 / Track: Performance and Scalability Session: Scalable Systems for Dynamic Content 311 Back-End Database Server Front-End Web and Application Servers Figure 1: Multi-Tier Architecture to backbone traffic congestion. It can also increase the availability and scalability of web services. Content that is generated dynamically from the back-end database cannot be cached in the first two tiers. While databases can be easily replicated in a LAN, this is infeasible in a WAN because of the difficult task of simultaneously providing strong consistency, availability, and tolerance to network partitions [7]. As a result, databases tend to be centralized to meet the strong consistency requirements of many eBusiness applications such as banking, finance, and online retailing [38]. Thus, the back-end database is usually located far from many sets of first and second-tier nodes [2]. In the absence of both caching and replication, WAN bandwidth can easily become a limiting factor in the performance and scalability of data-intensive applications. 2.2 Hash-Based Systems Ganesh"s focus is on efficient transmission of results by discovering similarities with the results of previous queries. As SQL queries can generate large results, hash-based techniques lend themselves well to the problem of efficiently transferring these large results across bandwidth constrained links. The use of hash-based techniques to reduce the volume of data transmitted has emerged as a common theme of many recent storage systems, as discussed in Section 8.2. These techniques rely on some basic assumptions. Cryptographic hash functions are assumed to be collision-resistant. In other words, it is computationally intractable to find two inputs that hash to the same output. The functions are also assumed to be one-way; that is, finding an input that results in a specific output is computationally infeasible. Menezes et al. [23] provide more details about these assumptions. The above assumptions allow hash-based systems to assume that collisions do not occur. Hence, they are able to treat the hash of a data item as its unique identifier. A collection of data items effectively becomes content-addressable, allowing a small hash to serve as a codeword for a much larger data item in permanent storage or network transmission. The assumption that collisions are so rare as to be effectively non-existent has recently come under fire [17]. However, as explained by Black [5], we believe that these issues do not form a concern for Ganesh. All communication is between trusted parts of the system and an adversary has no way to force Ganesh to accept invalid data. Further, Ganesh does not depend critically on any specific hash function. While we currently use SHA-1, replacing it with a different hash function would be simple. There would be no impact on performance as stronger hash functions (e.g. SHA256) only add a few extra bytes and the generated hashes are still orders of magnitude smaller than the data items they represent. No re-hashing of permanent storage is required since Ganesh only uses hashing on volatile data. 3. DESIGN AND IMPLEMENTATION Ganesh exploits redundancy in the result stream to avoid transmitting result fragments that are already present at the query site. Redundancy can arise naturally in many different ways. For example, a query repeated after a certain interval may return a different result because of updates to the database; however, there may be significant commonality between the two results. As another example, a user who is refining a search may generate a sequence of queries with overlapping results. When Ganesh detects redundancy, it suppresses transmission of the corresponding result fragments. Instead, it transmits a much smaller digest of those fragments and lets the query site reconstruct the result through hash lookup in a cache of previous results. In effect, Ganesh uses computation at the edges to reduce Internet communication. Our description of Ganesh focuses on four aspects. We first explain our approach to detecting similarity in query results. Next, we discuss how the Ganesh architecture is completely invisible to all components of a multi-tier system. We then describe Ganesh"s proxy-based approach and the dataflow for detecting similarity. 3.1 Detecting Similarity One of the key design decisions in Ganesh is how similarity is detected. There are many potential ways to decompose a result into fragments. The optimal way is, of course, the one that results in the smallest possible object for transmission for a given query"s results. Finding this optimal decomposition is a difficult problem because of the large space of possibilities and because the optimal choice depends on many factors such as the contents of the query"s result, the history of recent results, and the cache management algorithm. When an object is opaque, the use of Rabin fingerprints [8, 30] to detect common data between two objects has been successfully shown in the past by systems such as LBFS [26] and CASPER [37]. Rabin fingerprinting uses a sliding window over the data to compute a rolling hash. Assuming that the hash function is uniformly distributed, a chunk boundary is defined whenever the lower order bits of the hash value equal some predetermined value. The number of lower order bits used defines the average chunk size. These sub-divided chunks of the object become the unit of comparison for detecting similarity between different objects. As the locations of boundaries found by using Rabin fingerprints is stochastically determined, they usually fail to align with any structural properties of the underlying data. The algorithm therefore deals well with in-place updates, insertions and deletions. However, it performs poorly in the presence of any reordering of data. Figure 2 shows an example where two results, A and B, consisting of three rows, have the same data but have different sort attributes. In the extreme case, Rabin fingerprinting might be unable to find any similar data due to the way it detects chunk boundaries. Fortunately, Ganesh can use domain specific knowledge for more precise boundary detection. The information we exploit is that a query"s result reflects the structure of a relational database where all data is organized as tables and rows. It is therefore simple to check for similarity with previous results at two granularities: first the entire result, and then individual rows. The end of a row in a result serves as a natural chunk boundary. It is important to note that using the tabular structure in results only involves shallow interpretation of the data. Ganesh does not perform any deeper semantic interpretation such as understanding data types, result schema, or integrity constraints. Tuning Rabin fingerprinting for a workload can also be difficult. If the average chunk size is too large, chunks can span multiple result rows. However, selecting a smaller average chunk size increases the amount of metadata required to the describe the results. WWW 2007 / Track: Performance and Scalability Session: Scalable Systems for Dynamic Content 312 Figure 2: Rabin Fingerprinting vs. Ganesh"s Chunking This, in turn, would decrease the savings obtained via its use. Rabin fingerprinting also needs two computationally-expensive passes over the data: once to determine chunk boundaries and one again to generate cryptographic hashes for the chunks. Ganesh only needs a single pass for hash generation as the chunk boundaries are provided by the data"s natural structure. The performance comparison in Section 6 shows that Ganesh"s row-based algorithm outperforms Rabin fingerprinting. Given that previous work has already shown that Rabin fingerprinting performs better than gzip [26], we do not compare Ganesh to compression algorithms in this paper. 3.2 Transparency The key factor influencing our design was the need for Ganesh to be completely transparent to all components of a typical eBusiness system: web servers, application servers, and database servers. Without this, Ganesh stands little chance of having a significant real-world impact. Requiring modifications to any of the above components would raise the barrier for entry of Ganesh into an existing system, and thus reduce its chances of adoption. Preserving transparency is simplified by the fact that Ganesh is purely a performance enhancement, not a functionality or usability enhancement. We chose agent interposition as the architectural approach to realizing our goal. This approach relies on the existence of a compact programming interface that is already widely used by target software. It also relies on a mechanism to easily add new code without disrupting existing module structure. These conditions are easily met in our context because of the popularity of Java as the programming language for eBusiness systems. The Java Database Connectivity (JDBC) API [32] allows Java applications to access a wide variety of databases and even other tabular data repositories such as flat files. Access to these data sources is provided by JDBC drivers that translate between the JDBC API and the database communication mechanism. Figure 3(a) shows how JDBC is typically used in an application. As the JDBC interface is standardized, one can substitute one JDBC driver for another without application modifications. The JDBC driver thus becomes the natural module to exploit for code interposition. As shown in Figure 3(b), the native JDBC driver is replaced with a Ganesh JDBC driver that presents the same standardized interface. The Ganesh driver maintains an in-memory cache of result fragments from previous queries and performs reassembly of results. At the database, we add a new process called the Ganesh proxy. This proxy, which can be shared by multiple front-end nodes, consists of two parts: code to detect similarity in result fragments and the original native JDBC driver that communicates with the database. The use of a proxy at the database makes Ganesh database-agnostic and simplifies prototyping and experimentation. Ganesh is thus able to work with a wide range of databases and applications, requiring no modifications to either. 3.3 Proxy-Based Caching The native JDBC driver shown in Figure 3(a) is a lightweight code component supplied by the database vendor. Its main funcClient Database Web and Application Server Native JDBC Driver WAN (a) Native Architecture Client Database Ganesh Proxy Native JDBC Driver WAN Web and Application Server Ganesh JDBC Driver (b) Ganesh"s Interposition-based Architecture Figure 3: Native vs. Ganesh Architecture tion is to mediate communication between the application and the remote database. It forwards queries, buffers entire results, and responds to application requests to view parts of results. The Ganesh JDBC driver shown in Figure 3(b) presents the application with an interface identical to that provided by the native driver. It provides the ability to reconstruct results from compact hash-based descriptions sent by the proxy. To perform this reconstruction, the driver maintains an in-memory cache of recentlyreceived results. This cache is only used as a source of result fragments in reconstructing results. No attempt is made by the Ganesh driver or proxy to track database updates. The lack of cache consistency does not hurt correctness as a description of the results is always fetched from the proxy - at worst, there will be no performance benefit from using Ganesh. Stale data will simply be paged out of the cache over time. The Ganesh proxy accesses the database via the native JDBC driver, which remains unchanged between Figures 3(a) and (b). The database is thus completely unaware of the existence of the proxy. The proxy does not examine any queries received from the Ganesh driver but passes them to the native driver. Instead, the proxy is responsible for inspecting database output received from the native driver, detecting similar results, and generating hash-based encodings of these results whenever enough similarity is found. While this architecture does not decrease the load on a database, as mentioned earlier in Section 2.1, it is much easier to replicate databases for scalability in a LAN than in a WAN. To generate a hash-based encoding, the proxy must be aware of what result fragments are available in the Ganesh driver"s cache. One approach is to be optimistic, and to assume that all result fragments are available. This will result in the smallest possible initial transmission of a result. However, in cases where there is little overlap with previous results, the Ganesh driver will have to make many calls to the proxy during reconstruction to fetch missing result fragments. To avoid this situation, the proxy loosely tracks the state of the Ganesh driver"s cache. Since both components are under our control, it is relatively simple to do this without resorting to gray-box techniques or explicit communication for maintaining cache coherence. Instead, the proxy simulates the Ganesh driver"s cache management algorithm and uses this to maintain a list of hashes for which the Ganesh driver is likely to possess the result fragments. In case of mistracking, there will be no loss of correctness but there will be extra round-trip delays to fetch the missing fragments. If the client detects loss of synchronization with the proxy, it can ask the proxy to reset the state shared between them. Also note that the proxy does not need to keep the result fragments themselves, only their hashes. This allows the proxy to remain scalable even when it is shared by many front-end nodes. WWW 2007 / Track: Performance and Scalability Session: Scalable Systems for Dynamic Content 313 Object Output Stream Convert ResultSet Object Input Stream Convert ResultSet All Data Recipe ResultSet All Data ResultSet Network Ganesh Proxy Ganesh JDBC Driver Result Set Recipe Result Set Yes Yes No No GaneshInputStream GaneshOutputStream Figure 4: Dataflow for Result Handling 3.4 Encoding and Decoding Results The Ganesh proxy receives database output as Java objects from the native JDBC driver. It examines this output to see if a Java object of type ResultSet is present. The JDBC interface uses this data type to store results of database queries. If a ResultSet object is found, it is shrunk as discussed below. All other Java objects are passed through unmodified. As discussed in Section 3.1, the proxy uses the row boundaries defined in the ResultSet to partition it into fragments consisting of single result rows. All ResultSet objects are converted into objects of a new type called RecipeResultSet. We use the term recipe for this compact description of a database result because of its similarity to a file recipe in the CASPER file system [37]. The conversion replaces each result fragment that is likely to be present in the Ganesh driver"s cache by a SHA-1 hash of that fragment. Previously unseen result fragments are retained verbatim. The proxy also retains hashes for the new result fragments as they will be present in the driver"s cache in the future. Note that the proxy only caches hashes for result fragments and does not cache recipes. The proxy constructs a RecipeResultSet by checking for similarity at the entire result and then the row level. If the entire result is predicted to be present in the Ganesh driver"s cache, the RecipeResultSet is simply a single hash of the entire result. Otherwise, it contains hashes for those rows predicted to be present in that cache; all other rows are retained verbatim. If the proxy estimates an overall space savings, it will transmit the RecipeResultSet. Otherwise the original ResultSet is transmitted. The RecipeResultSet objects are transformed back into ResultSet objects by the Ganesh driver. Figure 4 illustrates ResultSet handling at both ends. Each SHA-1 hash found in a RecipeResultSet is looked up in the local cache of result fragments. On a hit, the hash is replaced by the corresponding fragment. On a miss, the driver contacts the Ganesh proxy to fetch the fragment. All previously unseen result fragments that were retained verbatim by the proxy are hashed and added to the result cache. There should be very few misses if the proxy has accurately tracked the Ganesh driver"s cache state. A future optimization would be to batch the fetch of missing fragments. This would be valuable when there are many small missing fragments in a high-latency WAN. Once the transformation is complete, the fully reconstructed ResultSet object is passed up to the application. 4. EXPERIMENTAL VALIDATION Three questions follow from the goals and design of Ganesh: • First, can performance can be improved significantly by exploiting similarity across database results? Benchmark Dataset Details 500,000 Users, 12,000 Stories BBOARD 2.0 GB 3,298,000 Comments AUCTION 1.3 GB 1,000,000 Users, 34,000 Items Table 1: Benchmark Dataset Details • Second, how important is Ganesh"s structural similarity detection relative to Rabin fingerprinting"s similarity detection? • Third, is the overhead of the proxy-based design acceptable? Our evaluation answers these question through controlled experiments with the Ganesh prototype. This section describes the benchmarks used, our evaluation procedure, and the experimental setup. Results of the experiments are presented in Sections 5, 6, and 7. 4.1 Benchmarks Our evaluation is based on two benchmarks [18] that have been widely used by other researchers to evaluate various aspects of multi-tier [27] and eBusiness architectures [9]. The first benchmark, BBOARD, is modeled after Slashdot, a technology-oriented news site. The second benchmark, AUCTION, is modeled after eBay, an online auction site. In both benchmarks, most content is dynamically generated from information stored in a database. Details of the datasets used can be found in Table 1. 4.1.1 The BBOARD Benchmark The BBOARD benchmark, also known as RUBBoS [18], models Slashdot, a popular technology-oriented web site. Slashdot aggregates links to news stories and other topics of interest found elsewhere on the web. The site also serves as a bulletin board by allowing users to comment on the posted stories in a threaded conversation form. It is not uncommon for a story to gather hundreds of comments in a matter of hours. The BBOARD benchmark is similar to the site and models the activities of a user, including readonly operations such as browsing the stories of the day, browsing story categories, and viewing comments as well as write operations such as new user registration, adding and moderating comments, and story submission. The benchmark consists of three different phases: a short warmup phase, a runtime phase representing the main body of the workload, and a short cool-down phase. In this paper we only report results from the runtime phase. The warm-up phase is important in establishing dynamic system state, but measurements from that phase are not significant for our evaluation. The cool-down phase is solely for allowing the benchmark to shut down. The warm-up, runtime, and cool-down phases are 2, 15, and 2 minutes respectively. The number of simulated clients were 400, 800, 1200, and 1600. The benchmark is available in a Java Servlets and PHP version and has different datasets; we evaluated Ganesh using the Java Servlets version and the Expanded dataset. The BBOARD benchmark defines two different workloads. The first, the Authoring mix, consists of 70% read-only operations and 30% read-write operations. The second, the Browsing mix, contains only read-only operations and does not update the database. 4.1.2 The AUCTION Benchmark The AUCTION benchmark, also known as RUBiS [18], models eBay, the online auction site. The eBay web site is used to buy and sell items via an auction format. The main activities of a user include browsing, selling, or bidding for items. Modeling the activities on this site, this benchmark includes read-only activities such as browsing items by category and by region, as well as read-write WWW 2007 / Track: Performance and Scalability Session: Scalable Systems for Dynamic Content 314 NetEm Router Ganesh Proxy Clients Web and Application Server Database Server Figure 5: Experimental Setup activities such as bidding for items, buying and selling items, and leaving feedback. As with BBOARD, the benchmark consists of three different phases. The warm-up, runtime, and cool-down phases for this experiment are 1.5, 15, and 1 minutes respectively. We tested Ganesh with four client configurations where the number of test clients was set to 400, 800, 1200, and 1600. The benchmark is available in a Enterprise Java Bean (EJB), Java Servlets, and PHP version and has different datasets; we evaluated Ganesh with the Java Servlets version and the Expanded dataset. The AUCTION benchmark defines two different workloads. The first, the Bidding mix, consists of 70% read-only operations and 30% read-write operations. The second, the Browsing mix, contains only read-only operations and does not update the database. 4.2 Experimental Procedure Both benchmarks involve a synthetic workload of clients accessing a web server. The number of clients emulated is an experimental parameter. Each emulated client runs an instance of the benchmark in its own thread, using a matrix to transition between different benchmark states. The matrix defines a stochastic model with probabilities of transitioning between the different states that represent typical user actions. An example transition is a user logging into the AUCTION system and then deciding on whether to post an item for sale or bid on active auctions. Each client also models user think time between requests. The think time is modeled as an exponential distribution with a mean of 7 seconds. We evaluate Ganesh along two axes: number of clients and WAN bandwidth. Higher loads are especially useful in understanding Ganesh"s performance when the CPU or disk of the database server or proxy is the limiting factor. A previous study has shown that approximately 50% of the wide-area Internet bottlenecks observed had an available bandwidth under 10 Mb/s [1]. Based on this work, we focus our evaluation on the WAN bandwidth of 5 Mb/s with 66 ms of round-trip latency, representative of severely constrained network paths, and 20 Mb/s with 33 ms of round-trip latency, representative of a moderately constrained network path. We also report Ganesh"s performance at 100 Mb/s with no added round-trip latency. This bandwidth, representative of an unconstrained network, is especially useful in revealing any potential overhead of Ganesh in situations where WAN bandwidth is not the limiting factor. For each combination of number of clients and WAN bandwidth, we measured results from the two configurations listed below: • Native: This configuration corresponds to Figure 3(a). Native avoids Ganesh"s overhead in using a proxy and performing Java object serialization. • Ganesh: This configuration corresponds to Figure 3(b). For a given number of clients and WAN bandwidth, comparing these results to the corresponding Native results gives the performance benefit due to the Ganesh middleware system. The metric used to quantify the improvement in throughput is the number of client requests that can be serviced per second. The metric used to quantify Ganesh"s overhead is the average response time for a client request. For all of the experiments, the Ganesh driver used by the application server used a cache size of 100,000 items1 . The proxy was effective in tracking the Ganesh driver"s cache state; for all of our experiments the miss rate on the driver never exceeded 0.7%. 4.3 Experimental Setup The experimental setup used for the benchmarks can be seen in Figure 5. All machines were 3.2 GHz Pentium 4s (with HyperThreading enabled.) With the exception of the database server, all machines had 2 GB of SDRAM and ran the Fedora Core Linux distribution. The database server had 4 GB of SDRAM. We used Apache"s Tomcat as both the application server that hosted the Java Servlets and the web server. Both benchmarks used Java Servlets to generate the dynamic content. The database server used the open source MySQL database. For the native JDBC drivers, we used the Connector/J drivers provided by MySQL. The application server used Sun"s Java Virtual Machine as the runtime environment for the Java Servlets. The sysstat tool was used to monitor the CPU, network, disk, and memory utilization on all machines. The machines were connected by a switched gigabit Ethernet network. As shown in Figure 5, the front-end web and application server was separated from the proxy and database server by a NetEm router [16]. This router allowed us to control the bandwidth and latency settings on the network. The NetEm router is a standard PC with two network cards running the Linux Traffic Control and Network Emulation software. The bandwidth and latency constraints were only applied to the link between the application server and the database for the native case and between the application server and the proxy for the Ganesh case. There is no communication between the application server and the database with Ganesh as all data flows through the proxy. As our focus was on the WAN link between the application server and the database, there were no constraints on the link between the simulated test clients and the web server. 5. THROUGHPUT AND RESPONSE TIME In this section, we address the first question raised in Section 4: Can performance can be improved significantly by exploiting similarity across database results? To answer this question, we use results from the BBOARD and AUCTION benchmarks. We use two metrics to quantify the performance improvement obtainable through the use of Ganesh: throughput, from the perspective of the web server, and average response time, from the perspective of the client. Throughput is measured in terms of the number of client requests that can be serviced per second. 5.1 BBOARD Results and Analysis 5.1.1 Authoring Mix Figures 6 (a) and (b) present the average number of requests serviced per second and the average response time for these requests as perceived by the clients for BBOARD"s Authoring Mix. As Figure 6 (a) shows, Native easily saturates the 5 Mb/s link. At 400 clients, the Native solution delivers 29 requests/sec with an average response time of 8.3 seconds. Native"s throughput drops with an increase in test clients as clients timeout due to congestion at the application server. Usability studies have shown that response times above 10 seconds cause the user to move on to 1 As Java lacks a sizeof() operator, Java caches therefore limit their size based on the number of objects. The size of cache dumps taken at the end of the experiments never exceeded 212 MB. WWW 2007 / Track: Performance and Scalability Session: Scalable Systems for Dynamic Content 315 0 50 100 150 200 250 400 800 1200 1600 400 800 1200 1600 400 800 1200 1600 5 Mb/s 20 Mb/s 100 Mb/s Test Clients Requests/sec Native Ganesh 0.001 0.01 0.1 1 10 100 400 800 1200 1600 400 800 1200 1600 400 800 1200 1600 5 Mb/s 20 Mb/s 100 Mb/s Test Clients Avg.Resp.Time(sec) Native Ganesh Note Logscale (a) Throughput: Authoring Mix (b) Response Time: Authoring Mix 0 50 100 150 200 250 400 800 1200 1600 400 800 1200 1600 400 800 1200 1600 5 Mb/s 20 Mb/s 100 Mb/s Test Clients Requests/sec Native Ganesh 0.001 0.01 0.1 1 10 100 400 800 1200 1600 400 800 1200 1600 400 800 1200 1600 5 Mb/s 20 Mb/s 100 Mb/s Test Clients Avg.Resp.Time(sec) Native Ganesh Note Logscale (c) Throughput: Browsing Mix (d) Response Time: Browsing Mix Mean of three trials. The maximum standard deviation for throughput and response time was 9.8% and 11.9% of the corresponding mean. Figure 6: BBOARD Benchmark - Throughput and Average Response Time other tasks [24]. Based on these numbers, increasing the number of test clients makes the Native system unusable. Ganesh at 5 Mb/s, however, delivers a twofold improvement with 400 test clients and a fivefold improvement at 1200 clients. Ganesh"s performance drops slightly at 1200 and 1600 clients as the network is saturated. Compared to Native, Figure 6 (b) shows that Ganesh"s response times are substantially lower with sub-second response times at 400 clients. Figure 6 (a) also shows that for 400 and 800 test clients Ganesh at 5 Mb/s has the same throughput and average response time as Native at 20 Mb/s. Only at 1200 and 1600 clients does Native at 20 Mb/s deliver higher throughput than Ganesh at 5 Mb/s. Comparing both Ganesh and Native at 20 Mb/s, we see that Ganesh is no longer bandwidth constrained and delivers up to a twofold improvement over Native at 1600 test clients. As Ganesh does not saturate the network with higher test client configurations, at 1600 test clients, its average response time is 0.1 seconds rather than Native"s 7.7 seconds. As expected, there are no visible gains from Ganesh at the higher bandwidth of 100 Mb/s where the network is no longer the bottleneck. Ganesh, however, still tracks Native in terms of throughput. 5.1.2 Browsing Mix Figures 6 (c) and (d) present the average number of requests serviced per second and the average response time for these requests as perceived by the clients for BBOARD"s Browsing Mix. Regardless of the test client configuration, Figure 6 (c) shows that Native"s throughput at 5 Mb/s is limited to 10 reqs/sec. Ganesh at 5 Mb/s with 400 test clients, delivers more than a sixfold increase in throughput. The improvement increases to over a elevenfold increase at 800 test clients before Ganesh saturates the network. Further, Figure 6 (d) shows that Native"s average response time of 35 seconds at 400 test clients make the system unusable. These high response times further increase with the addition of test clients. Even with the 1600 test client configuration Ganesh delivers an acceptable average response time of 8.2 seconds. Due to the data-intensive nature of the Browsing mix, Ganesh at 5 Mb/s surprisingly performs much better than Native at 20 Mb/s. Further, as shown in Figure 6 (d), while the average response time for Native at 20 Mb/s is acceptable at 400 test clients, it is unusable with 800 test clients with an average response time of 15.8 seconds. Like the 5 Mb/s case, this response time increases with the addition of extra test clients. Ganesh at 20 Mb/s and both Native and Ganesh at 100 Mb/s are not bandwidth limited. However, performance plateaus out after 1200 test clients due to the database CPU being saturated. 5.1.3 Filter Variant We were surprised by the Native performance from the BBOARD benchmark. At the bandwidth of 5 Mb/s, Native performance was lower than what we had expected. It turned out the benchmark code that displays stories read all the comments associated with the particular story from the database and only then did some postprocessing to select the comments to be displayed. While this is exactly the behavior of SlashCode, the code base behind the Slashdot web site, we decided to modify the benchmark to perform some pre-filtering at the database. This modified benchmark, named the Filter Variant, models a developer who applies optimizations at the SQL level to transfer less data. In the interests of brevity, we only briefly summarize the results from the Authoring mix. For the Authoring mix, at 800 test clients at 5 Mb/s, Figure 7 (a) shows that Native"s throughput increase by 85% when compared to the original benchmark while Ganesh"s improvement is smaller at 15%. Native"s performance drops above 800 clients as the test clients time out due to high response times. The most significant gain for Native is seen at 20 Mb/s. At 1600 test clients, when compared to the original benchmark, Native sees a 73% improvement in throughput and a 77% reduction in average response time. While WWW 2007 / Track: Performance and Scalability Session: Scalable Systems for Dynamic Content 316 0 50 100 150 200 250 400 800 1200 1600 400 800 1200 1600 400 800 1200 1600 5 Mb/s 20 Mb/s 100 Mb/s Test Clients Requests/sec Native Ganesh 0.001 0.01 0.1 1 10 100 400 800 1200 1600 400 800 1200 1600 400 800 1200 1600 5 Mb/s 20 Mb/s 100 Mb/s Test Clients Avg.Resp.Time(sec) Native Ganesh Note Logscale (a) Throughput: Authoring Mix (b) Response Time: Authoring Mix Mean of three trials. The maximum standard deviation for throughput and response time was 7.2% and 11.5% of the corresponding mean. Figure 7: BBOARD Benchmark - Filter Variant - Throughput and Average Response Time Ganesh sees no improvement when compared to the original, it still processes 19% more requests/sec than Native. Thus, while the optimizations were more helpful to Native, Ganesh still delivers an improvement in performance. 5.2 AUCTION Results and Analysis 5.2.1 Bidding Mix Figures 8 (a) and (b) present the average number of requests serviced per second and the average response time for these requests as perceived by the clients for AUCTION"s Bidding Mix. As mentioned earlier, the Bidding mix consists of a mixture of read and write operations. The AUCTION benchmark is not as data intensive as BBOARD. Therefore, most of the gains are observed at the lower bandwidth of 5 Mb/s. Figure 8 (a) shows that the increase in throughput due to Ganesh ranges from 8% at 400 test clients to 18% with 1600 test clients. As seen in Figure 8 (b), the average response times for Ganesh are significantly lower than Native ranging from a decrease of 84% at 800 test clients to 88% at 1600 test clients. Figure 8 (a) also shows that with a fourfold increase of bandwidth from 5 Mb/s to 20 Mb/s, Native is no longer bandwidth constrained and there is no performance difference between Ganesh and Native. With the higher test client configurations, we did observe that the bandwidth used by Ganesh was lower than Native. Ganesh might still be useful in these non-constrained scenarios if bandwidth is purchased on a metered basis. Similar results are seen for the 100 Mb/s scenario. 5.2.2 Browsing Mix For AUCTION"s Browsing Mix, Figures 8 (c) and (d) present the average number of requests serviced per second and the average response time for these requests as perceived by the clients. Again, most of the gains are observed at lower bandwidths. At 5 Mb/s, Native and Ganesh deliver similar throughput and response times with 400 test clients. While the throughput for both remains the same at 800 test clients, Figure 8 (d) shows that Ganesh"s average response time is 62% lower than Native. Native saturates the link at 800 clients and adding extra test clients only increases the average response time. Ganesh, regardless of the test client configuration, is not bandwidth constrained and maintains the same response time. At 1600 test clients, Figure 8 (c) shows that Ganesh"s throughput is almost twice that of Native. At the higher bandwidths of 20 and 100 Mb/s, neither Ganesh nor Native is bandwidth limited and deliver equivalent throughput and response times. Benchmark Orig. Size Ganesh Size Rabin Size SelectSort1 223.6 MB 5.4 MB 219.3 MB SelectSort2 223.6 MB 5.4 MB 223.6 MB Table 2: Similarity Microbenchmarks 6. STRUCTURAL VS. RABIN SIMILARITY In this section, we address the second question raised in Section 4: How important is Ganesh"s structural similarity detection relative to Rabin fingerprinting-based similarity detecting? To answer this question, we used microbenchmarks and the BBOARD and AUCTION benchmarks. As Ganesh always performed better than Rabin fingerprinting, we only present a subset of the results here in the interests of brevity. 6.1 Microbenchmarks Two microbenchmarks show an example of the effects of data reordering on Rabin fingerprinting algorithm. In the first microbenchmark, SelectSort1, a query with a specified sort order selects 223.6 MB of data spread over approximately 280 K rows. The query is then repeated with a different sort attribute. While the same number of rows and the same data is returned, the order of rows is different. In such a scenario, one would expect a large amount of similarity to be detected between both results. As Table 2 shows, Ganesh"s row-based algorithm achieves a 97.6% reduction while the Rabin fingerprinting algorithm, with the average chunk size parameter set to 4 KB, only achieves a 1% reduction. The reason, as shown earlier in Figure 2, is that with Rabin fingerprinting, the spans of data between two consecutive boundaries usually cross row boundaries. With the order of the rows changing in the second result and the Rabin fingerprints now spanning different rows, the algorithm is unable to detect significant similarity. The small gain seen is mostly for those single rows that are large enough to be broken into multiple chunks. SelectSort2, another micro-benchmark executed the same queries but increased the minimum chunk size of the Rabin fingerprinting algorithm. As can be seen in Table 2, even the small gain from the previous microbenchmark disappears as the minimum chunk size was greater than the average row size. While one can partially address these problems by dynamically varying the parameters of the Rabin fingerprinting algorithm, this can be computationally expensive, especially in the presence of changing workloads. 6.2 Application Benchmarks We ran the BBOARD benchmark described in Section 4.1.1 on two versions of Ganesh: the first with Rabin fingerprinting used as WWW 2007 / Track: Performance and Scalability Session: Scalable Systems for Dynamic Content 317 0 50 100 150 200 250 300 350 400 800 1200 1600 400 800 1200 1600 400 800 1200 1600 5 Mb/s 20 Mb/s 100 Mb/s Test Clients Requests/sec Native Ganesh 0.001 0.01 0.1 1 10 400 800 1200 1600 400 800 1200 1600 400 800 1200 1600 5 Mb/s 20 Mb/s 100 Mb/s Test Clients Avg.Resp.Time(sec) Native Ganesh Note Logscale (a) Throughput: Bidding Mix (b) Response Time: Bidding Mix 0 50 100 150 200 250 300 350 400 800 1200 1600 400 800 1200 1600 400 800 1200 1600 5 Mb/s 20 Mb/s 100 Mb/s Test Clients Requests/sec Native Ganesh 0.001 0.01 0.1 1 10 400 800 1200 1600 400 800 1200 1600 400 800 1200 1600 5 Mb/s 20 Mb/s 100 Mb/s Test Clients Avg.Resp.Time(sec) Native Ganesh Note Logscale (c) Throughput: Browsing Mix (d) Response Time: Browsing Mix Mean of three trials. The maximum standard deviation for throughput and response time was 2.2% and 11.8% of the corresponding mean. Figure 8: AUCTION Benchmark - Throughput and Average Response Time the chunking algorithm and the second with Ganesh"s row-based algorithm. Rabin"s results for the Browsing Mix are normalized to Ganesh"s results and presented in Figure 9. As Figure 9 (a) shows, at 5 Mb/s, independent of the test client configuration, Rabin significantly underperforms Ganesh. This happens because of a combination of two reasons. First, as outlined in Section 3.1, Rabin finds less similarity as it does not exploit the result"s structural information. Second, this benchmark contained some queries that generated large results. In this case, Rabin, with a small average chunk size, generated a large number of objects that evicted other useful data from the cache. In contrast, Ganesh was able to detect these large rows and correspondingly increase the size of the chunks. This was confirmed as cache statistics showed that Ganesh"s hit ratio was roughly three time that of Rabin. Throughput measurements at 20 Mb/s were similar with the exception of Rabin"s performance with 400 test clients. In this case, Ganesh was not network limited and, in fact, the throughput was the same as 400 clients at 5 Mb/s. Rabin, however, took advantage of the bandwidth increase from 5 to 20 Mb/s to deliver a slightly better performance. At 100 Mb/s, Rabin"s throughput was almost similar to Ganesh as bandwidth was no longer a bottleneck. The normalized response time, presented in Figure 9 (b), shows similar trends. At 5 and 20 Mb/s, the addition of test clients decreases the normalized response time as Ganesh"s average response time increases faster than Rabin"s. However, at no point does Rabin outperform Ganesh. Note that at 400 and 800 clients at 100 Mb/s, Rabin does have a higher overhead even when it is not bandwidth constrained. As mentioned in Section 3.1, this is due to the fact that Rabin has to hash each ResultSet twice. The overhead disappears with 1200 and 1600 clients as the database CPU is saturated and limits the performance of both Ganesh and Rabin. 7. PROXY OVERHEAD In this section, we address the third question raised in Section 4: Is the overhead of Ganesh"s proxy-based design acceptable? To answer this question, we concentrate on its performance at the higher bandwidths. Our evaluation in Section 5 showed that Ganesh, when compared to Native, can deliver a substantial throughput improvement at lower bandwidths. It is only at higher bandwidths that latency, measured by the average response time for a client request, and throughput, measured by the number of client requests that can be serviced per second, overheads would be visible. Looking at the Authoring mix of the original BBOARD benchmark, there are no visible gains from Ganesh at 100 Mb/s. Ganesh, however, still tracks Native in terms of throughput. While the average response time is higher for Ganesh, the absolute difference is in between 0.01 and 0.04 seconds and would be imperceptible to the end-user. The Browsing mix shows an even smaller difference in average response times. The results from the filter variant of the BBOARD benchmarks are similar. Even for the AUCTION benchmark, the difference between Native and Ganesh"s response time at 100 Mb/s was never greater than 0.02 seconds. The only exception to the above results was seen in the filter variant of the BBOARD benchmark where Ganesh at 1600 test clients added 0.85 seconds to the average response time. Thus, even for much faster networks where the WAN link is not the bottleneck, Ganesh always delivers throughput equivalent to Native. While some extra latency is added by the proxy-based design, it is usually imperceptible. 8. RELATED WORK To the best of our knowledge, Ganesh is the first system that combines the use of hash-based techniques with caching of database results to improve throughput and response times for applications with dynamic content. We also believe that it is also the first system to demonstrate the benefits of using structural information for WWW 2007 / Track: Performance and Scalability Session: Scalable Systems for Dynamic Content 318 0.0 0.2 0.4 0.6 0.8 1.0 400 800 1200 1600 400 800 1200 1600 400 800 1200 1600 5 Mb/s 20 Mb/s 100 Mb/s Test Clients Norm.Throughput 31.8 3.8 2.8 2.3 23.8 32.8 5.8 3.6 1.8 2.1 1.1 1.0 0 5 10 15 20 25 30 35 400 800 1200 1600 400 800 1200 1600 400 800 1200 1600 5 Mb/s 20 Mb/s 100 Mb/s Test Clients Norm.ResponseTime (a) Normalized Throughput: Higher is better (b) Normalized Response Time: Higher is worse For throughput, a normalized result greater than 1 implies that Rabin is better, For response time, a normalized result greater than 1 implies that Ganesh is better. Mean of three trials. The maximum standard deviation for throughput and response time was 9.1% and 13.9% of the corresponding mean. Figure 9: Normalized Comparison of Ganesh vs. Rabin - BBOARD Browsing Mix detecting similarity. In this section, we first discuss alternative approaches to caching dynamic content and then examine other uses of hash-based primitives in distributed systems. 8.1 Caching Dynamic Content At the database layer, a number of systems have advocated middletier caching where parts of the database are replicated at the edge or server [3, 4, 20]. These systems either cache entire tables in what is essentially a replicated database or use materialized views from previous query replies [19]. They require tight integration with the back-end database to ensure a time bound on the propagation of updates. These systems are also usually targeted towards workloads that do not require strict consistency and can tolerate stale data. Further, unlike Ganesh, some of these mid-tier caching solutions [2, 3], suffer from the complexity of having to participate in query planing and distributed query processing. Gao et al. [15] propose using a distributed object replication architecture where the data store"s consistency requirements are adapted on a per-application basis. These solutions require substantial developer resources and detailed understanding of the application being modified. While systems that attempt to automate the partitioning and replication of an application"s database exist [34], they do not provide full transaction semantics. In comparison, Ganesh does not weaken any of the semantics provided by the underlying database. Recent work in the evaluation of edge caching options for dynamic web sites [38] has suggested that, without careful planning, employing complex offloading strategies can hurt performance. Instead, the work advocates for an architecture in which all tiers except the database should be offloaded to the edge. Our evaluation of Ganesh has shown that it would benefit these scenarios. To improve database scalability, C-JDBC [10], SSS [22], and Ganymed [28] also advocate the use of an interposition-based architecture to transparently cluster and replicate databases at the middleware level. The approaches of these architectures and Ganesh are complementary and they would benefit each other. Moving up to the presentation layer, there has been widespread adoption of fragment-based caching [14], which improves cache utilization by separately caching different parts of generated web pages. While fragment-based caching works at the edge, a recent proposal has proposed moving web page assembly to the clients to optimize content delivery [31]. While Ganesh is not used at the presentation layer, the same principles have been applied in Duplicate Transfer Detection [25] to increase web cache efficiency as well as for web access across bandwidth limited links [33]. 8.2 Hash-based Systems The past few years have seen the emergence of many systems that exploit hash-based techniques. At the heart of all these systems is the idea of detecting similarity in data without requiring interpretation of that data. This simple yet elegant idea relies on cryptographic hashing, as discussed earlier in Section 2. Successful applications of this idea span a wide range of storage systems. Examples include peer-to-peer backup of personal computing files [11], storage-efficient archiving of data [29], and finding similar files [21]. Spring and Wetherall [35] apply similar principles at the network level. Using synchronized caches at both ends of a network link, duplicated data is replaced by smaller tokens for transmission and then restored at the remote end. This and other hash-based systems such as the CASPER [37] and LBFS [26] filesystems, and Layer-2 bandwidth optimizers such as Riverbed and Peribit use Rabin fingerprinting [30] to discover spans of commonality in data. This approach is especially useful when data items are modified in-place through insertions, deletions, and updates. However, as Section 6 shows, the performance of this technique can show a dramatic drop in the presence of data reordering. Ganesh instead uses row boundaries as dividers for detecting similarity. The most aggressive use of hash-based techniques is by systems that use hashes as the primary identifiers for objects in persistent storage. Storage systems such as CFS [12] and PAST [13] that have been built using distributed hash tables fall into this category. Single Instance Storage [6] and Venti [29] are other examples of such systems. As discussed in Section 2.2, the use of cryptographic hashes for addressing persistent data represents a deeper level of faith in their collision-resistance than that assumed by Ganesh. If time reveals shortcomings in the hash algorithm, the effort involved in correcting the flaw is much greater. In Ganesh, it is merely a matter of replacing the hash algorithm. 9. CONCLUSION The growing use of dynamic web content generated from relational databases places increased demands on WAN bandwidth. Traditional caching solutions for bandwidth and latency reduction are often ineffective for such content. This paper shows that the impact of WAN accesses to databases can be substantially reduced through the Ganesh architecture without any compromise of the database"s strict consistency semantics. The essence of the Ganesh architecture is the use of computation at the edges to reduce communication through the Internet. Ganesh is able to use cryptographic hashes to detect similarity with previous results and send WWW 2007 / Track: Performance and Scalability Session: Scalable Systems for Dynamic Content 319 compact recipes of results rather than full results. Our design uses interposition to achieve complete transparency: clients, application servers, and database servers are all unaware of Ganesh"s presence and require no modification. Our experimental evaluation confirms that Ganesh, while conceptually simple, can be highly effective in improving throughput and response time. Our results also confirm that exploiting the structure present in database results to detect similarity is crucial to this performance improvement. 10. REFERENCES [1] AKELLA, A., SESHAN, S., AND SHAIKH, A. An empirical evaluation of wide-area internet bottlenecks. In Proc. 3rd ACM SIGCOMM Conference on Internet Measurement (Miami Beach, FL, USA, Oct. 2003), pp. 101-114. [2] ALTINEL, M., BORNH ¨OVD, C., KRISHNAMURTHY, S., MOHAN, C., PIRAHESH, H., AND REINWALD, B. Cache tables: Paving the way for an adaptive database cache. In Proc. of 29th VLDB (Berlin, Germany, 2003), pp. 718-729. [3] ALTINEL, M., LUO, Q., KRISHNAMURTHY, S., MOHAN, C., PIRAHESH, H., LINDSAY, B. G., WOO, H., AND BROWN, L. Dbcache: Database caching for web application servers. In Proc. 2002 ACM SIGMOD (2002), pp. 612-612. [4] AMIRI, K., PARK, S., TEWARI, R., AND PADMANABHAN, S. Dbproxy: A dynamic data cache for web applications. In Proc. IEEE International Conference on Data Engineering (ICDE) (Mar. 2003). [5] BLACK, J. Compare-by-hash: A reasoned analysis. In Proc. 2006 USENIX Annual Technical Conference (Boston, MA, May 2006), pp. 85-90. [6] BOLOSKY, W. J., CORBIN, S., GOEBEL, D., , AND DOUCEUR, J. R. Single instance storage in windows 2000. In Proc. 4th USENIX Windows Systems Symposium (Seattle, WA, Aug. 2000), pp. 13-24. [7] BREWER, E. A. Lessons from giant-scale services. IEEE Internet Computing 5, 4 (2001), 46-55. [8] BRODER, A., GLASSMAN, S., MANASSE, M., AND ZWEIG, G. Syntactic clustering of the web. In Proc. 6th International WWW Conference (1997). [9] CECCHET, E., CHANDA, A., ELNIKETY, S., MARGUERITE, J., AND ZWAENEPOEL, W. Performance comparison of middleware architectures for generating dynamic web content. In Proc. Fourth ACM/IFIP/USENIX International Middleware Conference (Rio de Janeiro, Brazil, June 2003). [10] CECCHET, E., MARGUERITE, J., AND ZWAENEPOEL, W. C-JDBC: Flexible database clustering middleware. In Proc. 2004 USENIX Annual Technical Conference (Boston, MA, June 2004). [11] COX, L. P., MURRAY, C. D., AND NOBLE, B. D. Pastiche: Making backup cheap and easy. In OSDI: Symposium on Operating Systems Design and Implementation (2002). [12] DABEK, F., KAASHOEK, M. F., KARGER, D., MORRIS, R., AND STOICA, I. Wide-area cooperative storage with CFS. In 18th ACM Symposium on Operating Systems Principles (Banff, Canada, Oct. 2001). [13] DRUSCHEL, P., AND ROWSTRON, A. PAST: A large-scale, persistent peer-to-peer storage utility. In HotOS VIII (Schloss Elmau, Germany, May 2001), pp. 75-80. [14] Edge side includes. http://www.esi.org. [15] GAO, L., DAHLIN, M., NAYATE, A., ZHENG, J., AND IYENGAR, A. Application specific data replication for edge services. In WWW "03: Proc. Twelfth International Conference on World Wide Web (2003), pp. 449-460. [16] HEMMINGER, S. Netem - emulating real networks in the lab. In Proc. 2005 Linux Conference Australia (Canberra, Australia, Apr. 2005). [17] HENSON, V. An analysis of compare-by-hash. In Proc. 9th Workshop on Hot Topics in Operating Systems (HotOS IX) (May 2003), pp. 13-18. [18] Jmob benchmarks. http://jmob.objectweb.org/. [19] LABRINIDIS, A., AND ROUSSOPOULOS, N. Balancing performance and data freshness in web database servers. In Proc. 29th VLDB Conference (Sept. 2003). [20] LARSON, P.-A., GOLDSTEIN, J., AND ZHOU, J. Transparent mid-tier database caching in sql server. In Proc. 2003 ACM SIGMOD (2003), pp. 661-661. [21] MANBER, U. Finding similar files in a large file system. In Proc. USENIX Winter 1994 Technical Conference (San Fransisco, CA, 17-21 1994), pp. 1-10. [22] MANJHI, A., AILAMAKI, A., MAGGS, B. M., MOWRY, T. C., OLSTON, C., AND TOMASIC, A. Simultaneous scalability and security for data-intensive web applications. In Proc. 2006 ACM SIGMOD (June 2006), pp. 241-252. [23] MENEZES, A. J., VANSTONE, S. A., AND OORSCHOT, P. C. V. Handbook of Applied Cryptography. CRC Press, 1996. [24] MILLER, R. B. Response time in man-computer conversational transactions. In Proc. AFIPS Fall Joint Computer Conference (1968), pp. 267-277. [25] MOGUL, J. C., CHAN, Y. M., AND KELLY, T. Design, implementation, and evaluation of duplicate transfer detection in http. In Proc. First Symposium on Networked Systems Design and Implementation (San Francisco, CA, Mar. 2004). [26] MUTHITACHAROEN, A., CHEN, B., AND MAZIERES, D. A low-bandwidth network file system. In Proc. 18th ACM Symposium on Operating Systems Principles (Banff, Canada, Oct. 2001). [27] PFEIFER, D., AND JAKSCHITSCH, H. Method-based caching in multi-tiered server applications. In Proc. Fifth International Symposium on Distributed Objects and Applications (Catania, Sicily, Italy, Nov. 2003). [28] PLATTNER, C., AND ALONSO, G. Ganymed: Scalable replication for transactional web applications. In Proc. 5th ACM/IFIP/USENIX International Conference on Middleware (2004), pp. 155-174. [29] QUINLAN, S., AND DORWARD, S. Venti: A new approach to archival storage. In Proc. FAST 2002 Conference on File and Storage Technologies (2002). [30] RABIN, M. Fingerprinting by random polynomials. In Harvard University Center for Research in Computing Technology Technical Report TR-15-81 (1981). [31] RABINOVICH, M., XIAO, Z., DOUGLIS, F., AND KALMANEK, C. Moving edge side includes to the real edge - the clients. In Proc. 4th USENIX Symposium on Internet Technologies and Systems (Seattle, WA, Mar. 2003). [32] REESE, G. Database Programming with JDBC and Java, 1st ed. O"Reilly, June 1997. [33] RHEA, S., LIANG, K., AND BREWER, E. Value-based web caching. In Proc. Twelfth International World Wide Web Conference (May 2003). [34] SIVASUBRAMANIAN, S., ALONSO, G., PIERRE, G., AND VAN STEEN, M. Globedb: Autonomic data replication for web applications. In WWW "05: Proc. 14th International World-Wide Web conference (May 2005). [35] SPRING, N. T., AND WETHERALL, D. A protocol-independent technique for eliminating redundant network traffic. In Proc. of ACM SIGCOMM (Aug. 2000). [36] TOLIA, N., HARKES, J., KOZUCH, M., AND SATYANARAYANAN, M. Integrating portable and distributed storage. In Proc. 3rd USENIX Conference on File and Storage Technologies (San Francisco, CA, Mar. 2004). [37] TOLIA, N., KOZUCH, M., SATYANARAYANAN, M., KARP, B., PERRIG, A., AND BRESSOUD, T. Opportunistic use of content addressable storage for distributed file systems. In Proc. 2003 USENIX Annual Technical Conference (San Antonio, TX, June 2003), pp. 127-140. [38] YUAN, C., CHEN, Y., AND ZHANG, Z. Evaluation of edge caching/offloading for dynamic content delivery. In WWW "03: Proc. Twelfth International Conference on World Wide Web (2003), pp. 461-471. WWW 2007 / Track: Performance and Scalability Session: Scalable Systems for Dynamic Content 320

natural chunk boundary;read-write operation;bboard benchmark;content addressable storage;redundancy;relational database system;database cache;relational database;reciperesultset;temporal locality;caching dynamic database content;jdbc driver;bandwidth optimization;wide area network;database content;proxy;resultset object;hash-based technique

train_C-81

Adaptive Duty Cycling for Energy Harvesting Systems

Harvesting energy from the environment is feasible in many applications to ameliorate the energy limitations in sensor networks. In this paper, we present an adaptive duty cycling algorithm that allows energy harvesting sensor nodes to autonomously adjust their duty cycle according to the energy availability in the environment. The algorithm has three objectives, namely (a) achieving energy neutral operation, i.e., energy consumption should not be more than the energy provided by the environment, (b) maximizing the system performance based on an application utility model subject to the above energyneutrality constraint, and (c) adapting to the dynamics of the energy source at run-time. We present a model that enables harvesting sensor nodes to predict future energy opportunities based on historical data. We also derive an upper bound on the maximum achievable performance assuming perfect knowledge about the future behavior of the energy source. Our methods are evaluated using data gathered from a prototype solar energy harvesting platform and we show that our algorithm can utilize up to 58% more environmental energy compared to the case when harvesting-aware power management is not used.

1. INTRODUCTION Energy supply has always been a crucial issue in designing battery-powered wireless sensor networks because the lifetime and utility of the systems are limited by how long the batteries are able to sustain the operation. The fidelity of the data produced by a sensor network begins to degrade once sensor nodes start to run out of battery power. Therefore, harvesting energy from the environment has been proposed to supplement or completely replace battery supplies to enhance system lifetime and reduce the maintenance cost of replacing batteries periodically. However, metrics for evaluating energy harvesting systems are different from those used for battery powered systems. Environmental energy is distinct from battery energy in two ways. First it is an inexhaustible supply which, if appropriately used, can allow the system to last forever, unlike the battery which is a limited resource. Second, there is an uncertainty associated with its availability and measurement, compared to the energy stored in the battery which can be known deterministically. Thus, power management methods based on battery status are not always applicable to energy harvesting systems. In addition, most power management schemes designed for battery-powered systems only account for the dynamics of the energy consumers (e.g., CPU, radio) but not the dynamics of the energy supply. Consequently, battery powered systems usually operate at the lowest performance level that meets the minimum data fidelity requirement in order to maximize the system life. Energy harvesting systems, on the other hand, can provide enhanced performance depending on the available energy. In this paper, we will study how to adapt the performance of the available energy profile. There exist many techniques to accomplish performance scaling at the node level, such as radio transmit power adjustment [1], dynamic voltage scaling [2], and the use of low power modes [3]. However, these techniques require hardware support and may not always be available on resource constrained sensor nodes. Alternatively, a common performance scaling technique is duty cycling. Low power devices typically provide at least one low power mode in which the node is shut down and the power consumption is negligible. In addition, the rate of duty cycling is directly related to system performance metrics such as network latency and sampling frequency. We will use duty cycle adjustment as the primitive performance scaling technique in our algorithms. 2. RELATED WORK Energy harvesting has been explored for several different types of systems, such as wearable computers [4], [5], [6], sensor networks [7], etc. Several technologies to extract energy from the environment have been demonstrated including solar, motion-based, biochemical, vibration-based [8], [9], [10], [11], and others are being developed [12], [13]. While several energy harvesting sensor node platforms have been prototyped [14], [15], [16], there is a need for systematic power management techniques that provide performance guarantees during system operation. The first work to take environmental energy into account for data routing was [17], followed by [18]. While these works did demonstrate that environment aware decisions improve performance compared to battery aware decisions, their objective was not to achieve energy neutral operation. Our proposed techniques attempt to maximize system performance while maintaining energy-neutral operation. 3. SYSTEM MODEL The energy usage considerations in a harvesting system vary significantly from those in a battery powered system, as mentioned earlier. We propose the model shown in Figure 1 for designing energy management methods in a harvesting system. The functions of the various blocks shown in the figure are discussed below. The precise methods used in our system to achieve these functions will be discussed in subsequent sections. Harvested Energy Tracking: This block represents the mechanisms used to measure the energy received from the harvesting device, such as the solar panel. Such information is useful for determining the energy availability profile and adapting system performance based on it. Collecting this information requires that the node hardware be equipped with the facility to measure the power generated from the environment, and the Heliomote platform [14] we used for evaluating the algorithms has this capability. Energy Generation Model: For wireless sensor nodes with limited storage and processing capabilities to be able to use the harvested energy data, models that represent the essential components of this information without using extensive storage are required. The purpose of this block is to provide a model for the energy available to the system in a form that may be used for making power management decisions. The data measured by the energy tracking block is used here to predict future energy availability. A good prediction model should have a low prediction error and provide predicted energy values for durations long enough to make meaningful performance scaling decisions. Further, for energy sources that exhibit both long-term and short-term patterns (e.g., diurnal and climate variations vs. weather patterns for solar energy), the model must be able to capture both characteristics. Such a model can also use information from external sources such as local weather forecast service to improve its accuracy. Energy Consumption Model: It is also important to have detailed information about the energy usage characteristics of the system, at various performance levels. For general applicability of our design, we will assume that only one sleep mode is available. We assume that the power consumption in the sleep and active modes is known. It may be noted that for low power systems with more advanced capabilities such as dynamic voltage scaling (DVS), multiple low power modes, and the capability to shut down system components selectively, the power consumption in each of the states and the resultant effect on application performance should be known to make power management decisions. Energy Storage Model: This block represents the model for the energy storage technology. Since all the generated energy may not be used instantaneously, the harvesting system will usually have some energy storage technology. Storage technologies (e.g., batteries and ultra-capacitors) are non-ideal, in that there is some energy loss while storing and retrieving energy from them. These characteristics must be known to efficiently manage energy usage and storage. This block also includes the system capability to measure the residual stored energy. Most low power systems use batteries to store energy and provide residual battery status. This is commonly based on measuring the battery voltage which is then mapped to the residual battery energy using the known charge to voltage relationship for the battery technology in use. More sophisticated methods which track the flow of energy into and out of the battery are also available. Harvesting-aware Power Management: The inputs provided by the previously mentioned blocks are used here to determine the suitable power management strategy for the system. Power management could be carried to meet different objectives in different applications. For instance, in some systems, the harvested energy may marginally supplement the battery supply and the objective may be to maximize the system lifetime. A more interesting case is when the harvested energy is used as the primary source of energy for the system with the objective of achieving indefinitely long system lifetime. In such cases, the power management objective is to achieve energy neutral operation. In other words, the system should only use as much energy as harvested from the environment and attempt to maximize performance within this available energy budget. 4. THEORETICALLY OPTIMAL POWER MANAGEMENT We develop the following theory to understand the energy neutral mode of operation. Let us define Ps(t) as the energy harvested from the environment at time t, and the energy being consumed by the load at that time is Pc(t). Further, we model the non-ideal storage buffer by its round-trip efficiency η (strictly less than 1) and a constant leakage power Pleak. Using this notation, applying the rule of energy conservation leads to the following inequality: 0 0 00 [ ( ) ( )] [ ( ) ( )] 0 T T T s c c s leakP t P t dt P t P t dt P dtB η + + − − − ≥+ −∫ ∫ ∫ (1) where B0 is the initial battery level and the function [X]+ = X if X > 0 and zero otherwise. DEFINITION 1 (ρ,σ1,σ2) function: A non-negative, continuous and bounded function P (t) is said to be a (ρ,σ1,σ2) function if and only if for any value of finite real number T , the following are satisfied: 2 1( ) T T P t dt Tρ σ ρ σ− ≤ ≤ +∫ (2) This function can be used to model both energy sources and loads. If the harvested energy profile Ps(t) is a (ρ1,σ1,σ2) function, then the average rate of available energy over long durations becomes ρ1, and the burstiness is bounded by σ1 and σ2 . Similarly, Pc(t) can be modeled as a (ρ2,σ3) function, when ρ2 and σ3 are used to place an upper bound on power consumption (the inequality on the right side) while there are no minimum power consumption constraints. The condition for energy neutrality, equation (1), leads to the following theorem, based on the energy production, consumption, and energy buffer models discussed above. THEOREM 1 (ENERGY NEUTRAL OPERATION): Consider a harvesting system in which the energy production profile is characterized by a (ρ1, σ1, σ2) function, the load is characterized by a (ρ2, σ3) function and the energy buffer is characterized by parameters η for storage efficiency, and Pleak for leakage power. The following conditions are sufficient for the system to achieve energy neutrality: ρ2 ≤ ηρ1 − Pleak (3) B0 ≥ ησ2 + σ3 (4) B ≥ B0 (5) where B0 is the initial energy stored in the buffer and provides a lower bound on the capacity of the energy buffer B. The proof is presented in our prior work [19]. To adjust the duty cycle D using our performance scaling algorithm, we assume the following relation between duty cycle and the perceived utility of the system to the user: Suppose the utility of the application to the user is represented by U(D) when the system operates at a duty cycle D. Then, min 1 min max 2 max ( ) 0, ( ) , ( ) , U D if D D U D k D if D D D U D k if D D β = < = + ≤ ≤ = > This is a fairly general and simple model and the specific values of Dmin and Dmax may be determined as per application requirements. As an example, consider a sensor node designed to detect intrusion across a periphery. In this case, a linear increase in duty cycle translates into a linear increase in the detection probability. The fastest and the slowest speeds of the intruders may be known, leading to a minimum and Harvested Energy Tracking Energy Consumption Model Energy Storage Model Harvestingaware Power Mangement Energy Generation Model LOAD Figure 1. System model for an energy harvesting system. 181 maximum sensing delay tolerable, which results in the relevant Dmax and Dmin for the sensor node. While there may be cases where the relationship between utility and duty cycle may be non-linear, in this paper, we restrict our focus on applications that follow this linear model. In view of the above models for the system components and the required performance, the objective of our power management strategy is adjust the duty cycle D(i) dynamically so as to maximize the total utility U(D) over a period of time, while ensuring energy neutral operation for the sensor node. Before discussing the performance scaling methods for harvesting aware duty cycle adaptation, let us first consider the optimal power management strategy that is possible for a given energy generation profile. For the calculation of the optimal strategy, we assume complete knowledge of the energy availability profile at the node, including the availability in the future. The calculation of the optimal is a useful tool for evaluating the performance of our proposed algorithm. This is particularly useful for our algorithm since no prior algorithms are available to serve as a baseline for comparison. Suppose the time axis is partitioned into discrete slots of duration ΔT, and the duty cycle adaptation calculation is carried out over a window of Nw such time slots. We define the following energy profile variables, with the index i ranging over {1,…, Nw}: Ps(i) is the power output from the harvested source in time slot i, averaged over the slot duration, Pc is the power consumption of the load in active mode, and D(i) is the duty cycle used in slot i, whose value is to be determined. B(i) is the residual battery energy at the beginning of slot i. Following this convention, the battery energy left after the last slot in the window is represented by B(Nw+1). The values of these variables will depend on the choice of D(i). The energy used directly from the harvested source and the energy stored and used from the battery must be accounted for differently. Figure 2 shows two possible cases for Ps(i) in a time slot. Ps(i) may either be less than or higher than Pc , as shown on the left and right respectively. When Ps(i) is lower than Pc, some of the energy used by the load comes from the battery, while when Ps(i) is higher than Pc, all the energy used is supplied directly from the harvested source. The crosshatched area shows the energy that is available for storage into the battery while the hashed area shows the energy drawn from the battery. We can write the energy used from the battery in any slot i as: ( ) ( ) ( ) ( )[ ] ( ) ( ){ } ( ) ( )[ ]1 1c cs s sB i B i TD i P P i TP i D i TD i P i Pη η + + − + = Δ − − Δ − − − (6) In equation (6), the first term on the right hand side measures the energy drawn from the battery when Ps(i) < Pc, the next term measures the energy stored into the battery when the node is in sleep mode, and the last term measures the energy stored into the battery in active mode if Ps(i) > Pc. For energy neutral operation, we require the battery at the end of the window of Nw slots to be greater than or equal to the starting battery. Clearly, battery level will go down when the harvested energy is not available and the system is operated from stored energy. However, the window Nw is judiciously chosen such that over that duration, we expect the environmental energy availability to complete a periodic cycle. For instance, in the case of solar energy harvesting, Nw could be chosen to be a twenty-four hour duration, corresponding to the diurnal cycle in the harvested energy. This is an approximation since an ideal choice of the window size would be infinite, but a finite size must be used for analytical tractability. Further, the battery level cannot be negative at any time, and this is ensured by having a large enough initial battery level B0 such that node operation is sustained even in the case of total blackout during a window period. Stating the above constraints quantitatively, we can express the calculation of the optimal duty cycles as an optimization problem below: 1 max ( ) wN i D i = ∑ (7) ( ) ( ) ( ) ( ) ( ) ( ){ } ( ) ( )1 1c s s s cB i B i TD i P P i TP i D i TD i P i Pη η + + ⎡ ⎤ ⎡ ⎤− + = Δ − − Δ − − −⎣ ⎦ ⎣ ⎦ (8) 0(1)B B= (9) 0( 1)wB N B+ ≥ (10) min w( ) i {1,...,N }D i D≥ ∀ ∈ (11) max w( ) i {1,...,N }D i D≤ ∀ ∈ (12) The solution to the optimization problem yields the duty cycles that must be used in every slot and the evolution of residual battery over the course of Nw slots. Note that while the constraints above contain the non-linear function [x]+ , the quantities occurring within that function are all known constants. The variable quantities occur only in linear terms and hence the above optimization problem can be solved using standard linear programming techniques, available in popular optimization toolboxes. 5. HARVESTING-AWARE POWER MANAGEMENT We now present a practical algorithm for power management that may be used for adapting the performance based on harvested energy information. This algorithm attempts to achieve energy neutral operation without using knowledge of the future energy availability and maximizes the achievable performance within that constraint. The harvesting-aware power management strategy consists of three parts. The first part is an instantiation of the energy generation model which tracks past energy input profiles and uses them to predict future energy availability. The second part computes the optimal duty cycles based on the predicted energy, and this step uses our computationally tractable method to solve the optimization problem. The third part consists of a method to dynamically adapt the duty cycle in response to the observed energy generation profile in real time. This step is required since the observed energy generation may deviate significantly from the predicted energy availability and energy neutral operation must be ensured with the actual energy received rather than the predicted values. 5.1. Energy Prediction Model We use a prediction model based on Exponentially Weighted Moving-Average (EWMA). The method is designed to exploit the diurnal cycle in solar energy but at the same time adapt to the seasonal variations. A historical summary of the energy generation profile is maintained for this purpose. While the storage data size is limited to a vector length of Nw values in order to minimize the memory overheads of the power management algorithm, the window size is effectively infinite as each value in the history window depends on all the observed data up to that instant. The window size is chosen to be 24 hours and each time slot is taken to be 30 minutes as the variation in generated power by the solar panel using this setting is less than 10% between each adjacent slots. This yields Nw = 48. Smaller slot durations may be used at the expense of a higher Nw. The historical summary maintained is derived as follows. On a typical day, we expect the energy generation to be similar to the energy generation at the same time on the previous days. The value of energy generated in a particular slot is maintained as a weighted average of the energy received in the same time-slot during all observed days. The weights are exponential, resulting in decaying contribution from older Figure 2. Two possible cases for energy calculations Slot i Slot k Pc Pc P(i) P(i) Active Sleep 182 data. More specifically, the historical average maintained for each slot is given by: 1 (1 )k k kx x xα α−= + − where α is the value of the weighting factor, kx is the observed value of energy generated in the slot, and 1kx − is the previously stored historical average. In this model, the importance of each day relative to the previous one remains constant because the same weighting factor was used for all days. The average value derived for a slot is treated as an estimate of predicted energy value for the slot corresponding to the subsequent day. This method helps the historical average values adapt to the seasonal variations in energy received on different days. One of the parameters to be chosen in the above prediction method is the parameter α, which is a measure of rate of shift in energy pattern over time. Since this parameter is affected by the characteristics of the energy and sensor node location, the system should have a training period during which this parameter will be determined. To determine a good value of α, we collected energy data over 72 days and compared the average error of the prediction method for various values of α. The error based on the different values of α is shown in Figure 3. This curve suggests an optimum value of α = 0.15 for minimum prediction error and this value will be used in the remainder of this paper. 5.2. Low-complexity Solution The energy values predicted for the next window of Nw slots are used to calculated the desired duty cycles for the next window, assuming the predicted values match the observed values in the future. Since our objective is to develop a practical algorithm for embedded computing systems, we present a simplified method to solve the linear programming problem presented in Section 4. To this end, we define the sets S and D as follows: { } { } | ( ) 0 | ( ) 0 s c c s S i P i P D i P P i = − ≥ = − > The two sets differ by the condition that whether the node operation can be sustained entirely from environmental energy. In the case that energy produced from the environment is not sufficient, battery will be discharged to supplement the remaining energy. Next we sum up both sides of (6) over the entire Nw window and rewrite it with the new notation. 1 1 1 1 ( )[ ( )] ( ) ( ) ( ) ( )[ ( ) ] Nw Nw Nw i i c s s s s c i i D i i i S B B TD i P P i TP i TP i D i TD i P i Pη η η+ = ∈ = = ∈ − = Δ − − Δ + Δ − Δ −∑ ∑ ∑ ∑ ∑ The term on the left hand side is actually the battery energy used over the entire window of Nw slots, which can be set to 0 for energy neutral operation. After some algebraic manipulation, this yields: 1 1 ( ) ( ) ( ) 1 ( ) Nw c s s c i i D i S P P i D i P i P D i η η= ∈ ∈ ⎛ ⎞⎛ ⎞ = + − +⎜ ⎟⎜ ⎟ ⎝ ⎠⎝ ⎠ ∑ ∑ ∑ (13) The term on the left hand side is the total energy received in Nw slots. The first term on the right hand side can be interpreted as the total energy consumed during the D slots and the second term is the total energy consumed during the S slots. We can now replace three constraints (8), (9), and (10) in the original problem with (13), restating the optimization problem as follows: 1 max ( ) wN i D i = ∑ 1 1 ( ) ( ) ( ) 1 ( ) Nw c s s c i i D i S P P i D i P i P D i η η= ∈ ∈ ⎛ ⎞⎛ ⎞ = + − +⎜ ⎟⎜ ⎟ ⎝ ⎠⎝ ⎠ ∑ ∑ ∑ min max D(i) D {1,...,Nw) D(i) D {1,...,Nw) i i ≥ ∀ ∈ ≤ ∀ ∈ This form facilitates a low complexity solution that doesn"t require a general linear programming solver. Since our objective is to maximize the total system utility, it is preferable to set the duty cycle to Dmin for time slots where the utility per unit energy is the least. On the other hand, we would also like the time slots with the highest Ps to operate at Dmax because of better efficiency of using energy directly from the energy source. Combining these two characteristics, we define the utility co-efficient for each slot i as follows: 1 ( ) 1 1 ( ) 1 c c s P for i S W i P P i for i D η η ∈⎧ ⎪ = ⎛ ⎞⎛ ⎞⎨ + − ∈⎜ ⎟⎜ ⎟⎪ ⎝ ⎠⎝ ⎠⎩ where W(i) is a representation of how efficient the energy usage in a particular time slot i is. A larger W(i) indicates more system utility per unit energy in slot i and vice versa. The algorithm starts by assuming D(i) =Dmin for i = {1…NW} because of the minimum duty cycle requirement, and computes the remaining system energy R by: 1 1 ( ) ( ) ( ) 1 ( ) (14) Nw c s s c i i D i S P R P i D i P i P D i η η= ∈ ∈ ⎛ ⎞⎛ ⎞ = − + − −⎜ ⎟⎜ ⎟ ⎝ ⎠⎝ ⎠ ∑ ∑ ∑ A negative R concludes that the optimization problem is infeasible, meaning the system cannot achieve energy neutrality even at the minimum duty cycle. In this case, the system designer is responsible for increasing the environment energy availability (e.g., by using larger solar panels). If R is positive, it means the system has excess energy that is not being used, and this may be allocated to increase the duty cycle beyond Dmin for some slots. Since our objective is to maximize the total system utility, the most efficient way to allocate the excess energy is to assign duty cycle Dmax to the slots with the highest W(i). So, the coefficients W(i) are arranged in decreasing order and duty cycle Dmax is assigned to the slots beginning with the largest coefficients until the excess energy available, R (computed by (14) in every iteration), is insufficient to assign Dmax to another slot. The remaining energy, RLast, is used to increase the duty cycle to some value between Dmin and Dmax in the slot with the next lower coefficient. Denoting this slot with index j, the duty cycle is given by: D(j)= min / ( ( ) ) / ( ) Last c Last s c s R P if j D DR if j S P j P P jη ∈⎧ ⎫ ⎪ ⎪ +⎨ ⎬ ∈⎪ ⎪− −⎩ ⎭ The above solution to the optimization problem requires only simple arithmetic calculations and one sorting step which can be easily implemented on an embedded platform, as opposed to implementing a general linear program solver. 5.3. Slot-by-slot continual duty cycle adaptiation. The observed energy values may vary greatly from the predicted ones, such as due to the effect of clouds or other sudden changes. It is thus important to adapt the duty cycles calculated using the predicted values, to the actual energy measurements in real time to ensure energy neutrality. Denote the initial duty cycle assignments for each time slot i computed using the predicted energy values as D(i) = {1, ...,Nw}. First we compute the difference between predicted power level Ps(i) and actual power level observed, Ps"(i) in every slot i. Then, the excess energy in slot i, denoted by X, can be obtained as follows: ( ) '( ) '( ) 1 ( ) '( ) ( )[ ( ) '( )](1 ) '( ) s s s c s s s s s c P i P i if P i P X P i P i D i P i P i if P i P η − >⎧ ⎪ = ⎨ − − − − ≤⎪ ⎩ 0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5 2.6 2.7 2.8 2.9 3 alpha AvgError(mA) Figure 3. Choice of prediction parameter. 183 The upper term accounts for the energy difference when actual received energy is more than the power drawn by the load. On the other hand, if the energy received is less than Pc, we will need to account for the extra energy used from the battery by the load, which is a function of duty cycle used in time slot i and battery efficiency factor η. When more energy is received than predicted, X is positive and that excess energy is available for use in the subsequent solutes, while if X is negative, that energy must be compensated from subsequent slots. CASE I: X<0. In this case, we want to reduce the duty cycles used in the future slots in order to make up for this shortfall of energy. Since our objective function is to maximize the total system utility, we have to reduce the duty cycles for time slots with the smallest normalized utility coefficient, W(i). This is accomplished by first sorting the coefficient W(j) ,where j>i. in decreasing order, and then iteratively reducing Dj to Dmin until the total reduction in energy consumption is the same as X. CASE II: X>0. Here, we want to increase the duty cycles used in the future to utilize this excess energy we received in recent time slot. In contrast to Case I, the duty cycles of future time slots with highest utility coefficient W(i) should be increased first in order to maximize the total system utility. Suppose the duty cycle is changed by d in slot j. Define a quantity R(j,d) as follows: ⎪ ⎩ ⎪ ⎨ ⎧ <=⎟ ⎟ ⎠ ⎞ ⎜ ⎜ ⎝ ⎛ ⎟⎟ ⎠ ⎞ ⎜⎜ ⎝ ⎛ −+ >⋅ = lji l ljl PPifP P d PPifP djR 1 1 d ),( ηη The precise procedure to adapt the duty cycle to account for the above factors is presented in Algorithm 1. This calculation is performed at the end of every slot to set the duty cycle for the next slot. We claim that our duty cycling algorithm is energy neutral because an surplus of energy at the previous time slot will always translate to additional energy opportunity for future time slots, and vice versa. The claim may be violated in cases of severe energy shortages especially towards the end of window. For example, a large deficit in energy supply can"t be restored if there is no future energy input until the end of the window. In such case, this offset will be carried over to the next window so that long term energy neutrality is still maintained. 6. EVALUATION Our adaptive duty cycling algorithm was evaluated using an actual solar energy profile measured using a sensor node called Heliomote, capable of harvesting solar energy [14]. This platform not only tracks the generated energy but also the energy flow into and out of the battery to provide an accurate estimate of the stored energy. The energy harvesting platform was deployed in a residential area in Los Angeles from the beginning of June through the middle of August for a total of 72 days. The sensor node used is a Mica2 mote running at a fixed 40% duty cycle with an initially full battery. Battery voltage and net current from the solar panels are sampled at a period of 10 seconds. The energy generation profile for that duration, measured by tracking the output current from the solar cell is shown in Figure 4, both on continuous and diurnal scales. We can observe that although the energy profile varies from day to day, it still exhibits a general pattern over several days. 0 10 20 30 40 50 60 70 0 10 20 30 40 50 60 70 Day mA 0 5 10 15 20 0 10 20 30 40 50 60 70 Hour mA 6.1. Prediction Model We first evaluate the performance of the prediction model, which is judged by the amount of absolute error it made between the predicted and actual energy profile. Figure 5 shows the average error of each time slot in mA over the entire 72 days. Generally, the amount of error is larger during the day time because that"s when the factor of weather can cause deviations in received energy, while the prediction made for night time is mostly correct. 6.2. Adaptive Duty cycling algorithm Prior methods to optimize performance while achieving energy neutral operation using harvested energy are scarce. Instead, we compare the performance of our algorithm against two extremes: the theoretical optimal calculated assuming complete knowledge about future energy availability and a simple approach which attempts to achieve energy neutrality using a fixed duty cycle without accounting for battery inefficiency. The optimal duty cycles are calculated for each slot using the future knowledge of actual received energy for that slot. For the simple approach, the duty cycle is kept constant within each day and is Figure 4 Solar Energy Profile (Left: Continuous, Right: Diurnal) Input: D: Initial duty cycle, X: Excess energy due to error in the prediction, P: Predicted energy profile, i: index of current time slot Output: D: Updated duty cycles in one or more subsequent slots AdaptDutyCycle() Iteration: At each time slot do: if X > 0 Wsorted = W{1, ...,Nw} sorted in decending order. Q := indices of Wsorted for k = 1 to |Q| if Q(k) ≤ i or D(Q(k)) ≥ Dmax //slot is already passed continue if R(Q(k), Dmax − D(Q(k))) < X D(Q(k)) = Dmax X = X − R(j, Dmax − D(Q(k))) else //X insufficient to increase duty cycle to Dmax if P (Q(k)) > Pl D(Q(k)) = D(Q(k)) + X/Pl else D(Q(k)) = D(Q(k)) + ( ( ))(1 1 ))c s X P P Q kη η+ − if X < 0 Wsorted = W{1, ...,Nw} sorted in ascending order. Q := indices of Wsorted for k = 1 to |Q| if Q(k) ≤ I or D(Q(k)) ≤ Dmin continue if R(Q(k), Dmax − D(Q(k))) > X D(Q(k)) = Dmin X = X − R(j, Dmin − D(Q(k))) else if P (Q(k)) > Pc D(Q(k)) = D(Q(k)) + X/Pc else D(Q(k)) = D(Q(k)) + ( ( ))(1 1 ))c s X P P Q kη η+ − ALGORITHM 1 Pseudocode for the duty-cycle adaptation algorithm Figure 5. Average Predictor Error in mA 0 5 10 15 20 25 0 2 4 6 8 10 12 Time(H) abserror(mA) 184 computed by taking the ratio of the predicted energy availability and the maximum usage, and this guarantees that the senor node will never deplete its battery running at this duty cycle. {1.. ) ( )s w c i Nw D P i N Pη ∈ = ⋅ ⋅∑ We then compare the performance of our algorithm to the two extremes with varying battery efficiency. Figure 6 shows the results, using Dmax = 0.8 and Dmin = 0.3. The battery efficiency was varied from 0.5 to 1 on the x-axis and solar energy utilizations achieved by the three algorithms are shown on the y-axis. It shows the fraction of net received energy that is used to perform useful work rather than lost due to storage inefficiency. As can be seen from the figure, battery efficiency factor has great impact on the performance of the three different approaches. The three approaches all converges to 100% utilization if we have a perfect battery (η=1), that is, energy is not lost by storing it into the batteries. When battery inefficiency is taken into account, both the adaptive and optimal approach have much better solar energy utilization rate than the simple one. Additionally, the result also shows that our adaptive duty cycle algorithm performs extremely close to the optimal. 0.4 0.5 0.6 0.7 0.8 0.9 1 0.4 0.5 0.6 0.7 0.8 0.9 1 Eta-batery roundtrip efficiency SolarEnergyUtilization(%) Optimal Adaptive Simple We also compare the performance of our algorithm with different values of Dmin and Dmax for η=0.7, which is typical of NiMH batteries. These results are shown in Table 1 as the percentage of energy saved by the optimal and adaptive approaches, and this is the energy which would normally be wasted in the simple approach. The figures and table indicate that our real time algorithm is able to achieve a performance very close to the optimal feasible. In addition, these results show that environmental energy harvesting with appropriate power management can achieve much better utilization of the environmental energy. Dmax Dmin 0.8 0.05 0.8 0.1 0.8 0.3 0.5 0.2 0.9 0.2 1.0 0.2 Adaptive 51.0% 48.2% 42.3% 29.4% 54.7% 58.7% Optimal 52.3% 49.6% 43.7% 36.7% 56.6% 60.8% 7. CONCLUSIONS We discussed various issues in power management for systems powered using environmentally harvested energy. Specifically, we designed a method for optimizing performance subject to the constraint of energy neutral operation. We also derived a theoretically optimal bound on the performance and showed that our proposed algorithm operated very close to the optimal. The proposals were evaluated using real data collected using an energy harvesting sensor node deployed in an outdoor environment. Our method has significant advantages over currently used methods which are based on a conservative estimate of duty cycle and can only provide sub-optimal performance. However, this work is only the first step towards optimal solutions for energy neutral operation. It is designed for a specific power scaling method based on adapting the duty cycle. Several other power scaling methods, such as DVS, submodule power switching and the use of multiple low power modes are also available. It is thus of interest to extend our methods to exploit these advanced capabilities. 8. ACKNOWLEDGEMENTS This research was funded in part through support provided by DARPA under the PAC/C program, the National Science Foundation (NSF) under award #0306408, and the UCLA Center for Embedded Networked Sensing (CENS). Any opinions, findings, conclusions or recommendations expressed in this paper are those of the authors and do not necessarily reflect the views of DARPA, NSF, or CENS. REFERENCES [1] R Ramanathan, and R Hain, Toplogy Control of Multihop Wireless Networks Using Transmit Power Adjustment in Proc. Infocom. Vol 2. 26-30 pp. 404-413. March 2000 [2] T.A. Pering, T.D. Burd, and R. W. Brodersen, The simulation and evaluation of dynamic voltage scaling algorithms, in Proc. ACM ISLPED, pp. 76-81, 1998 [3] L. Benini and G. De Micheli, Dynamic Power Management: Design Techniques and CAD Tools. Kluwer Academic Publishers, Norwell, MA, 1997. [4] John Kymisis, Clyde Kendall, Joseph Paradiso, and Neil Gershenfeld. Parasitic power harvesting in shoes. In ISWC, pages 132-139. IEEE Computer Society press, October 1998. [5] Nathan S. Shenck and Joseph A. Paradiso. Energy scavenging with shoemounted piezoelectrics. IEEE Micro, 21(3):30ñ42, May-June 2001. [6] T Starner. Human-powered wearable computing. IBM Systems Journal, 35(3-4), 1996. [7] Mohammed Rahimi, Hardik Shah, Gaurav S. Sukhatme, John Heidemann, and D. Estrin. Studying the feasibility of energy harvesting in a mobile sensor network. In ICRA, 2003. [8] ChrisMelhuish. The ecobot project. www.ias.uwe.ac.uk/energy autonomy/EcoBot web page.html. [9] Jan M.Rabaey, M. Josie Ammer, Julio L. da Silva Jr., Danny Patel, and Shad Roundy. Picoradio supports ad-hoc ultra-low power wireless networking. IEEE Computer, pages 42-48, July 2000. [10] Joseph A. Paradiso and Mark Feldmeier. A compact, wireless, selfpowered pushbutton controller. In ACM Ubicomp, pages 299-304, Atlanta, GA, USA, September 2001. Springer-Verlag Berlin Heidelberg. [11] SE Wright, DS Scott, JB Haddow, andMA Rosen. The upper limit to solar energy conversion. volume 1, pages 384 - 392, July 2000. [12] Darpa energy harvesting projects. http://www.darpa.mil/dso/trans/energy/projects.html. [13] Werner Weber. Ambient intelligence: industrial research on a visionary concept. In Proceedings of the 2003 international symposium on Low power electronics and design, pages 247-251. ACM Press, 2003. [14] V Raghunathan, A Kansal, J Hsu, J Friedman, and MB Srivastava, "Design Considerations for Solar Energy Harvesting Wireless Embedded Systems," (IPSN/SPOTS), April 2005. [15] Xiaofan Jiang, Joseph Polastre, David Culler, Perpetual Environmentally Powered Sensor Networks, (IPSN/SPOTS), April 25-27, 2005. [16] Chulsung Park, Pai H. Chou, and Masanobu Shinozuka, "DuraNode: Wireless Networked Sensor for Structural Health Monitoring," to appear in Proceedings of the 4th IEEE International Conference on Sensors, Irvine, CA, Oct. 31 - Nov. 1, 2005. [17] Aman Kansal and Mani B. Srivastava. An environmental energy harvesting framework for sensor networks. In International symposium on Low power electronicsand design, pages 481-486. ACM Press, 2003. [18] Thiemo Voigt, Hartmut Ritter, and Jochen Schiller. Utilizing solar power in wireless sensor networks. In LCN, 2003. [19] A. Kansal, J. Hsu, S. Zahedi, and M. B. Srivastava. Power management in energy harvesting sensor networks. Technical Report TR-UCLA-NESL200603-02, Networked and Embedded Systems Laboratory, UCLA, March 2006. Figure 6. Duty Cycles achieved with respect to η TABLE 1. Energy Saved by adaptive and optimal approach. 185

sampling frequency;energy harvest;energy harvesting system;duty cycling rate;low power design;harvesting-aware power management;solar panel;duty cycle;performance scaling;power scaling;energy neutral operation;sensor network;energy tracking;storage buffer;environmental energy;power management;network latency

train_C-83

Concept and Architecture of a Pervasive Document Editing and Managing System

Collaborative document processing has been addressed by many approaches so far, most of which focus on document versioning and collaborative editing. We address this issue from a different angle and describe the concept and architecture of a pervasive document editing and managing system. It exploits database techniques and real-time updating for sophisticated collaboration scenarios on multiple devices. Each user is always served with upto-date documents and can organize his work based on document meta data. For this, we present our conceptual architecture for such a system and discuss it with an example.

1. INTRODUCTION Text documents are a valuable resource for virtually any enterprise and organization. Documents like papers, reports and general business documentations contain a large part of today"s (business) knowledge. Documents are mostly stored in a hierarchical folder structure on file servers and it is difficult to organize them in regard to classification, versioning etc., although it is of utmost importance that users can find, retrieve and edit up-to-date versions of documents whenever they want and, in a user-friendly way. 1.1 Problem Description With most of the commonly used word-processing applications documents can be manipulated by only one user at a time: tools for pervasive collaborative document editing and management, are rarely deployed in today"s world. Despite the fact, that people strive for location- and time- independence, the importance of pervasive collaborative work, i.e. collaborative document editing and management is totally neglected. Documents could therefore be seen as a vulnerable source in today"s world, which demands for an appropriate solution: The need to store, retrieve and edit these documents collaboratively anytime, everywhere and with almost every suitable device and with guaranteed mechanisms for security, consistency, availability and access control, is obvious. In addition, word processing systems ignore the fact that the history of a text document contains crucial information for its management. Such meta data includes creation date, creator, authors, version, location-based information such as time and place when/where a user reads/edits a document and so on. Such meta data can be gathered during the documents creation process and can be used versatilely. Especially in the field of pervasive document management, meta data is of crucial importance since it offers totally new ways of organizing and classifying documents: On the one hand, the user"s actual situation influences the user"s objectives. Meta data could be used to give the user the best possible view on the documents, dependent of his actual information. On the other hand, as soon as the user starts to work, i.e. reads or edits a document, new meta data can be gathered in order to make the system more adaptable and in a sense to the users situation and, to offer future users a better view on the documents. As far as we know, no system exists, that satisfies the aforementioned requirements. A very good overview about realtime communication and collaboration system is described in [7]. We therefore strive for a pervasive document editing and management system, which enables pervasive (and collaborative) document editing and management: users should be able to read and edit documents whenever, wherever, with whomever and with whatever device. In this paper, we present collaborative database-based real-time word processing, which provides pervasive document editing and management functionality. It enables the user to work on documents collaboratively and offers sophisticated document management facility: the user is always served with up-to-date documents and can organize and manage documents on the base of meta data. Additionally document data is treated as ‘first class citizen" of the database as demanded in [1]. 1.2 Underlying Concepts The concept of our pervasive document editing and management system requires an appropriate architectural foundation. Our concept and implementation are based on the TeNDaX [3] collaborative database-based document editing and management system, which enables pervasive document editing and managing. TeNDaX is a Text Native Database eXtension. It enables the storage of text in databases in a native form so that editing text is finally represented as real-time transactions. Under the term ‘text editing" we understand the following: writing and deleting text (characters), copying & pasting text, defining text layout & structure, inserting notes, setting access rights, defining business processes, inserting tables, pictures, and so on i.e. all the actions regularly carried out by word processing users. With ‘real-time transaction" we mean that editing text (e.g. writing a character/word) invokes one or several database transactions so that everything, which is typed appears within the editor as soon as these objects are stored persistently. Instead of creating files and storing them in a file system, the content and all of the meta data belonging to the documents is stored in a special way in the database, which enables very fast real-time transactions for all editing tasks [2]. The database schema and the above-mentioned transactions are created in such a way that everything can be done within a multiuser environment, as is usual done by database technology. As a consequence, many of the achievements (with respect to data organization and querying, recovery, integrity and security enforcement, multi-user operation, distribution management, uniform tool access, etc.) are now, by means of this approach, also available for word processing. 2. APPROACH Our pervasive editing and management system is based on the above-mentioned database-based TeNDaX approach, where document data is stored natively in the database and supports pervasive collaborative text editing and document management. We define the pervasive document editing and management system, as a system, where documents can easily be accessed and manipulated everywhere (within the network), anytime (independently of the number of users working on the same document) and with any device (desktop, notebook, PDA, mobile phone etc.). DB 3 RTSC 4 RTSC 1 RTSC 2 RTSC 3 AS 1 AS 3 DB 1 DB 2 AS 2 AS 4 DB 4 A B C D E F G Figure 1. TeNDaX Application Architecture In contrast to documents stored locally on the hard drive or on a file server, our system automatically serves the user with the up-to-date version of a document and changes done on the document are stored persistently in the database and immediately propagated to all clients who are working on the same document. Additionally, meta data gathered during the whole document creation process enables sophisticated document management. With the TeXt SQL API as abstract interface, this approach can be used by any tool and for any device. The system is built on the following components (see Figure 1): An editor in Java implements the presentation layer (A-G in Figure 1). The aim of this layer is the integration in a well-known wordprocessing application such as OpenOffice. The business logic layer represents the interface between the database and the word-processing application. It consists of the following three components: The application server (marked as AS 1-4 in Figure 1) enables text editing within the database environment and takes care of awareness, security, document management etc., all within a collaborative, real-time and multi-user environment. The real-time server component (marked as RTSC 14 in Figure 1) is responsible for the propagation of information, i.e. updates between all of the connected editors. The storage engine (data layer) primarily stores the content of documents as well as all related meta data within the database Databases can be distributed in a peer-to-peer network (DB 1-4 in Figure 1).. In the following, we will briefly present the database schema, the editor and the real-time server component as well as the concept of dynamic folders, which enables sophisticated document management on the basis of meta data. 2.1 Application Architecture A database-based real-time collaborative editor allows the same document to be opened and edited simultaneously on the same computer or over a network of several computers and mobile devices. All concurrency issues, as well as message propagation, are solved within this approach, while multiple instances of the same document are being opened [3]. Each insert or delete action is a database transaction and as such, is immediately stored persistently in the database and propagated to all clients working on the same document. 2.1.1 Database Schema As it was mentioned earlier that text is stored in a native way. Each character of a text document is stored as a single object in the database [3]. When storing text in such a native form, the performance of the employed database system is of crucial importance. The concept and performance issues of such a text database are described in [3], collaborative layouting in [2], dynamic collaborative business processes within documents in [5], the text editing creation time meta data model in [6] and the relation to XML databases in [7]. Figure 2 depicts the core database schema. By connecting a client to the database, a Session instance is created. One important attribute of the Session is the DocumentSession. This attribute refers to DocumentSession instances, which administrates all opened documents. For each opened document, a DocumentSession instance is created. The DocumentSession is important for the realtime server component, which, in case of a 42 is beforeis after Char (ID) has TextElement (ID) starts with is used by InternalFile (ID) is in includes created at has inserted by inserted is active ir ir CharacterValue (Unicode) has List (ID) starts starts with ends ends with FileSize has User (ID) last read by last written by created at created by Style DTD (ID) is used by uses uses is used by Authors arehas Description Password Picture UserColors UserListSecurity has has has has has has FileNode (ID) references/isreferencedby is dynamic DynStructure NodeDetails has has is NodeType is parent of has parent has Role (ID) created at created created by Name has Description is user Name has has main role FileNodeAccessMatrix (ID) has is AccessMatrix read option grand option write option contains has access Times opened … times with … by contains/ispartof ir ir is...andincludes Lineage (ID) references is after is before CopyPaste (ID) references is in is copy of is a copy from hasCopyPaste (ID) is activeLength has Str (Stream) has inserted by / inserted RegularChar StartChar EndChar File ExternalFile is from URL Type (extension) is of Title has DocumentSession (ID) is opened by has opened has opened Session (ID) isconnectedwith launched by VersionNumber uses has read option grand option write option ends with is used by is in has is unique DTD (Stream) has has Name Column (ID) has set on On/off isvisible…for false LanguageProfile (ID) has contains Name Profile Marking (ID) has parent internal is copy from hasRank is onPosition starts with ends with is logical style is itemized is italic is enumerated is underline is is part of Alignment Size has Font has hasColor is bold has uses ElementName StylesheetName isused by Process (ID) is running by OS is web session MainRoles Roles has has Timestamp (Date, Time) created at Timestamp (Date, Time) Timestamp (Date, Time) Timestamp (Date, Time) Timestamp (Date, Time)created at Type has Port IP has has MessagePropagator (ID) Picture (Stream) Name Picture (ID) has contains LayoutBlock WorkflowBlockLogicalBlock contains BlockDataType has property BlockData is of WorkflowInstance (ID) isin TaskInstance (ID) has parent Timestamp (Date, Time) Timestamp (Date, Time) Timestamp (Date, Time) Timestamp (Date, Time) last modified at completed at started at created at is on has Name created by has attached Comment Typeis of Timestamp (Date, Time) Timestamp (Date, Time) Timestamp (Date, Time) created at started at << last modified at is Category Editors has Status has Timestamp (Date, Time) << status last modified Timestamp (Date, Time) is due at DueType has Timezone has Notes has SecurityLevel hasset Timestamp (Date, Time) << is completed at isfollowedby Task (Code) Description has Indent references hasbeenopenedat...by Timestamp RedoHistory is before is after references hasCharCounter is inhas has Offset ActionID (Code) Timestamp (Date, Time) invoked at invoked by Version (ID) isbuild from has created byarchived has Comment Timestamp (Date, Time) <<createdat UndoHistory (ID) starts ends has Name created by Name has is before is after << references CharCounter has is in created at Timestamp is active created by is used by Offset has created at Timestamp Index (ID) lastmodifiedby Lexicon (ID) isof Frequency is occurring is stop word Term is is in ends with starts with << original starts with WordNumber SentenceNumber ParagraphNumber Citatons has is in is is in istemporary is in has Structure has ElementPath createdat Timestamp << describes SpiderBuild (ID) is updated is deleted Timestamp (Date, Time) <<lastupdatedat has validated structure <<neededtoindex Time (ms) IndexUpdate nextupdatein hasindexed isrunningbyOS lastupdate enabled Timestamp Time (s) Documents StopCharacter Description Character Value (ASCII) is sentence stop is paragraph stop Name has is is OptionsSettings show information show warningsshow exceptions do lineage recording do internal lineage recording ask for unknown source show intra document lineage information are set for X X X VirtualBorder (ID) isonhas {1, 2} {1, 2} ir ir UserMode (Code) UserMode (Code) Figure 2. TeNDaX Database Schema (Object Role Modeling Diagram) change on a document done by a client, is responsible for sending update information to all the clients working on the same document. The DocumentId in the class DocumentSession points to a FileNode instance, and corresponds to the ID of the opened document. Instances of the class FileNode either represent a folder node or a document node. The folder node corresponds to a folder of a file system and the document node to that of a file. Instances of the class Char represent the characters of a document. The value of a character is stored in the attribute CharacterValue. The sequence is defined by the attributes After and Before of the class Char. Particular instances of Char mark the beginning and the end of a document. The methods InsertChars and RemoveChars are used to add and delete characters. 2.1.2 Editor As seen above, each document is natively stored in the database. Our editor does not have a replica of one part of the native text database in the sense of database replicas. Instead, it has a so-called image as its replica. Even if several authors edit the same text at the same time, they work on one unique document at all times. The system guarantees this unique view. Editing a document involves a number of steps: first, getting the required information out of the image, secondly, invoking the corresponding methods within the database, thirdly, changing the image, and fourthly, informing all other clients about the changes. 2.1.3 Real-Time Server Component The real-time server component is responsible for the real-time propagation of any changes on a document done within an editor to all the editors who are working or have opened the same document. When an editor connects to the application server, which in turn connects to the database, the database also establishes a connection to the real-time server component (if there isn"t already a connection). The database system informs the real-time server component about each new editor session (session), which the realtime server component administrates in his SessionManager. Then, the editor as well connects to the real-time server component. The real-time server component adds the editor socket to the client"s data structure in the SessionManager and is then ready to communicate. Each time a change on a document from an editor is persistently stored in the database, the database sends a message to the real-time server component, which in turns, sends the changes to all the 43 editors working on the same document. Therefore, a special communication protocol is used: the update protocol. Update Protocol The real-time server component uses the update protocol to communicate with the database and the editors. Messages are sent from the database to the real-time server component, which sends the messages to the affected editors. The update protocol consists of different message types. Messages consist of two packages: package one contains information for the real-time server component whereas package two is passed to the editors and contains the update information, as depicted in Figure 3. || RTSC || Parameter | … | Parameter|| || Editor Data || Protocol between database system and real-time server component Protocol between real -time server component and editors Figure 3. Update Protocol In the following, two message types are presented: ||u|sessionId,...,sessionId||||editor data|| u: update message, sessionId: Id of the client session With this message type the real-time server component sends the editor data package to all editors specified in the sessionId list. ||ud|fileId||||editor data|| ud: update document message, fileId: Id of the file With this message type, the real-time server component sends the editor data to all editors who have opened the document with the indicated file-Id. Class Model Figure 4 depicts the class model as well as the environment of the real-time server component. The environment consists mainly of the editor and the database, but any other client application that could make use of the real-time server component can connect. ConnectionListener: This class is responsible for the connection to the clients, i.e. to the database and the editors. Depending on the connection type (database or editor) the connection is passed to an EditorWorker instance or DatabaseMessageWorker instance respectively. EditorWorker: This class manages the connections of type ‘editor". The connection (a socket and its input and output stream) is stored in the SessionManager. SessionManager: This class is similar to an ‘in-memory database": all editor session information, e.g. the editor sockets, which editor has opened which document etc. are stored within this data structure. DatabaseMessageWorker: This class is responsible for the connections of type ‘database". At run-time, only one connection exists for each database. Update messages from the database are sent to the DatabaseMessageWorker and, with the help of additional information from the SessionManager, sent to the corresponding clients. ServiceClass: This class offers a set of methods for reading, writing and logging messages. tdb.mp.editor tdb.mp.database tdb.mp.mgmt EditorWorker DatabaseMessageWorker SessionManager MessageHandler ConnectionListener ServiceClass MessageQueue tdb.mp.listener tdb.mp.service junit.tests 1 * 1 * 1 * 1 * 1* 1 * Editors Datenbanksystem 1 2 1 * 1 * 1 * TCP/IP Figure 4. Real-Time Server Component Class Diagram 2.1.4 Dynamic Folders As mentioned above, every editing action invoked by a user is immediately transferred to the database. At the same time, more information about the current transaction is gathered. As all information is stored in the database, one character can hold a multitude of information, which can later be used for the retrieval of documents. Meta data is collected at character level, from document structure (layout, workflow, template, semantics, security, workflow and notes), on the level of a document section and on the level of the whole document [6]. All of the above-mentioned meta data is crucial information for creating content and knowledge out of word processing documents. This meta data can be used to create an alternative storage system for documents. In any case, it is not an easy task to change users" familiarity to the well known hierarchical file system. This is also the main reason why we do not completely disregard the classical file system, but rather enhance it. Folders which correspond to the classical hierarchical file system will be called static folders. Folders where the documents are organized according to meta data, will be called dynamic folders. As all information is stored in the database, the file system, too, is based on the database. The dynamic folders build up sub-trees, which are guided by the meta data selected by the user. Thus, the first step in using a dynamic folder is the definition of how it should be built. For each level of a dynamic folder, exactly one meta data item is used to. The following example illustrates the steps which have to be taken in order to define a dynamic folder, and the meta data which should be used. As a first step, the meta data which will be used for the dynamic folder must be chosen (see Table 1): The sequence of the meta data influences the structure of the folder. Furthermore, for each meta data used, restrictions and granularity must be defined by the user; if no restrictions are defined, all accessible documents are listed. The granularity therefore influences the number of sub-folders which will be created for the partitioning of the documents. 44 As the user enters the tree structure of the dynamic folder, he can navigate through the branches to arrive at the document(s) he is looking for. The directory names indicate which meta data determines the content of the sub-folder in question. At each level, the documents, which have so far been found to match the meta data, can be inspected. Table 1. Defining dynamic folders (example) Level Meta data Restrictions Granularity 1 Creator Only show documents which have been created by the users Leone or Hodel or Gall One folder per creator 2 Current location Only show documents which where read at my current location One folder per task status 3 Authors Only show documents where at least 40% was written by user ‘Leone" Each 20% one folder ad-hoc changes of granularity and restrictions are possible in order to maximize search comfort for the user. It is possible to predefine dynamic folders for frequent use, e.g. a location-based folder, as well as to create and modify dynamic folders on an ad-hoc basis. Furthermore, the content of such dynamic folders can change from one second to another, depending on the changes made by other users at that moment. 3. VALIDATION The proposed architecture is validated on the example of a character insertion. Insert operations are the mostly used operations in a (collaborative) editing system. The character insertion is based on the TeNDaX Insert Algorithm which is formally described in the following. The algorithm is simplified for this purpose. 3.1 Insert Characters Algorithm The symbol c stands for the object character, p stands for the previous character, n stands for the next character of a character object c and the symbol l stands for a list of character objects. c = character p=previous character n = next character l = list of characters The symbol c1 stands for the first character in the list l, ci stands for a character in the list l at the position i, whereas i is a value between 1 and the length of the list l, and cn stands for the last character in the list l. c1 = first character in list l ci = character at position i in list l cn = last character in list l The symbol β stands for the special character that marks the beginning of a document and ε stands for the special character that marks the end of a document. β=beginning of document ε=end of document The function startTA starts a transaction. startTA = start transaction The function commitTA commits a transaction that was started. commitTA = commit transaction The function checkWriteAccess checks if the write access for a document session s is granted. checkWriteAccess(s) = check if write access for document session s is granted The function lock acquires an exclusive lock for a character c and returns 1 for a success and 0 for no success. lock(c) = acquire the lock for character c success : return 1, no success : return 0 The function releaseLocks releases all locks that a transaction has acquired so far. releaseLocks = release all locks The function getPrevious returns the previous character and getNext returns the next character of a character c. getPrevious(c) = return previous character of character c getNext(c) = return next character of character c The function linkBefore links a preceding character p with a succeeding character x and the function linkAfter links a succeeding character n with a preceding character y. linkBefore(p,x) = link character p to character x linkAfter(n,y) = link character n to character y The function updateString links a character p with the first character c1 of a character list l and a character n with the last character cn of a character list l updateString(l, p, n) = linkBefore(p cl)∧ linkAfter(n, cn ) The function insertChar inserts a character c in the table Char with the fields After set to a character p and Before set to a character n. insertChar(c, p, n) = linkAfter(c,p) ∧ linkBefore(c,n) ∧ linkBefore(p,c) ∧ linkAfter(n,c) The function checkPreceding determines the previous character's CharacterValue of a character c and if the previous character's status is active. checkPreceding(c) = return status and CharacterValue of the previous character The function checkSucceeding determines the next character's CharacterValue of a character c and if the next character's status is active. 45 checkSucceeding(c) = return status and CharacterValue of the next character The function checkCharValue determines the CharacterValue of a character c. checkCharValue(c) = return CharacterValue of character c The function sendUpdate sends an update message (UpdateMessage) from the database to the real-time server component. sendUpdate(UpdateMessage) The function Read is used in the real-time server component to read the UpdateMessage. Read(UpdateInformationMessage) The function AllocatEditors checks on the base of the UpdateMessage and the SessionManager, which editors have to be informed. AllocateEditors(UpdateInformationMessage, SessionManager) = returns the affected editors The function SendMessage(EditorData) sends the editor part of the UpdateMessage to the editors SendMessage(EditorData) In TeNDaX, the Insert Algorithm is implemented in the class method InsertChars of the class Char which is depicted in Figure 2. The relevant parameters for the definitions beneath, are introduced in the following list: - nextCharacterOID: OID of the character situated next to the string to be inserted - previousCharacterOID: OID of the character situated previously to the string to be inserted - characterOIDs (List): List of character which have to be inserted Thus, the insertion of characters can be defined stepwise as follows: Start a transaction. startTA Select the character that is situated before the character that follows the string to be inserted. getPrevious(nextCharacterOID) = PrevChar(prevCharOID) ⇐ Π After ϑOID = nextCharacterOID(Char)) Acquire the lock for the character that is situated in the document before the character that follows the string which shall be inserted. lock(prevCharId) At this time the list characterOIDs contains the characters c1 to cn that shall be inserted. characterOIDs={ c1, …, cn } Each character of the string is inserted at the appropriate position by linking the preceding and the succeeding character to it. For each character ci of characterOIDs: insertChar(ci, p, n) Whereas ci ∈ { c1,…, cn } Check if the preceding and succeeding characters are active or if it is the beginning or the end of the document. checkPreceding(prevCharOID) = IsOK(IsActive, CharacterValue) ⇐ Π IsActive, CharacterValue (ϑ OID = nextCharacterOID(Char)) checkSucceeding(nextCharacterOID) = IsOK(IsActive, CharacterValue)⇐ Π IsActive, CharacterValue (ϑ OID = nextCharacterOID(Char)) Update characters before and after the string to be inserted. updateString(characterOIDs, prevCharOID, nextCharacterOID) Release all locks and commit Transaction. releaseLocks commitTA Send update information to the real-time server component sendUpdate(UpdatenMessage) Read update message and inform affected editors of the change Read(UpdateMessage) Allocate Editors(UpdateMessage, SessionManager) SendMessage(EditorData) 3.2 Insert Characters Example Figure 1 gives a snapshot the system, i.e. of its architecture: four databases are distributed over a peer-to-peer network. Each database is connected to an application server (AS) and each application server is connected to a real-time server component (RTSC). Editors are connected to one or more real-time server components and to the corresponding databases. Considering that editor A (connected to database 1 and 4) and editor B (connected to database 1 and 2) are working on the same document stored in database 1. Editor B now inserts a character into this document. The insert operation is passed to application server 1, which in turns, passes it to the database 1, where an insert operation is invoked; the characters are inserted according to the algorithm discussed in the previous section. After the insertion, database 1 sends an update message (according to the update protocol discussed before) to real-time server component 1 (via AS 1). RTCS 1 combines the received update information with the information in his SessionManager and sends the editor data to the affected editors, in this case to editor A and B, where the changes are immediately shown. Occurring collaboration conflicts are solved and described in [3]. 4. SUMMARY With the approach presented in this paper and the implemented prototype, we offer real-time collaborative editing and management of documents stored in a special way in a database. With this approach we provide security, consistency and availability of documents and consequently offer pervasive document editing and management. Pervasive document editing and management is enabled due to the proposed architecture with the embedded real46 time server component, which propagates changes to a document immediately and consequently offers up-to-date documents. Document editing and managing is consequently enabled anywhere, anytime and with any device. The above-descried system is implemented in a running prototype. The system will be tested soon in line with a student workshop next autumn. REFERENCES [1] Abiteboul, S., Agrawal, R., et al.: The Lowell Database Research Self Assessment. Massachusetts, USA, 2003. [2] Hodel, T. B., Businger, D., and Dittrich, K. R.: Supporting Collaborative Layouting in Word Processing. IEEE International Conference on Cooperative Information Systems (CoopIS), Larnaca, Cyprus, IEEE, 2004. [3] Hodel, T. B. and Dittrich, K. R.: "Concept and prototype of a collaborative business process environment for document processing." Data & Knowledge Engineering 52, Special Issue: Collaborative Business Process Technologies(1): 61120, 2005. [4] Hodel, T. B., Dubacher, M., and Dittrich, K. R.: Using Database Management Systems for Collaborative Text Editing. ACM European Conference of Computersupported Cooperative Work (ECSCW CEW 2003), Helsinki, Finland, 2003. [5] Hodel, T. B., Gall, H., and Dittrich, K. R.: Dynamic Collaborative Business Processes within Documents. ACM Special Interest Group on Design of Communication (SIGDOC) , Memphis, USA, 2004. [6] Hodel, T. B., R. Hacmac, and Dittrich, K. R.: Using Text Editing Creation Time Meta Data for Document Management. Conference on Advanced Information Systems Engineering (CAiSE'05), Porto, Portugal, Springer Lecture Notes, 2005. [7] Hodel, T. B., Specker, F. and Dittrich, K. R.: Embedded SOAP Server on the Operating System Level for ad-hoc Automatic Real-Time Bidirectional Communication. Information Resources Management Association (IRMA), San Diego, USA, 2005. [8] O"Kelly, P.: Revolution in Real-Time Communication and Collaboration: For Real This Time. Application Strategies: In-Depth Research Report. Burton Group, 2005. 47

collaborative layouting;restriction;hierarchical file system;computer support collaborative work;cscw;computer supported collaborative work;real-time server component;text editing;collaborative document;pervasive document edit and management system;real-time transaction;character insertion;granularity;pervasive document editing and managing system;collaborative document processing;business logic layer

train_C-84

Selfish Caching in Distributed Systems: A Game-Theoretic Analysis

We analyze replication of resources by server nodes that act selfishly, using a game-theoretic approach. We refer to this as the selfish caching problem. In our model, nodes incur either cost for replicating resources or cost for access to a remote replica. We show the existence of pure strategy Nash equilibria and investigate the price of anarchy, which is the relative cost of the lack of coordination. The price of anarchy can be high due to undersupply problems, but with certain network topologies it has better bounds. With a payment scheme the game can always implement the social optimum in the best case by giving servers incentive to replicate.

1. INTRODUCTION Wide-area peer-to-peer file systems [2,5,22,32,33], peer-to-peer caches [15, 16], and web caches [6, 10] have become popular over the last few years. Caching1 of files in selected servers is widely used to enhance the performance, availability, and reliability of these systems. However, most such systems assume that servers cooperate with one another by following protocols optimized for overall system performance, regardless of the costs incurred by each server. In reality, servers may behave selfishly - seeking to maximize their own benefit. For example, parties in different administrative domains utilize their local resources (servers) to better support clients in their own domains. They have obvious incentives to cache objects2 that maximize the benefit in their domains, possibly at the expense of globally optimum behavior. It has been an open question whether these caching scenarios and protocols maintain their desirable global properties (low total social cost, for example) in the face of selfish behavior. In this paper, we take a game-theoretic approach to analyzing the problem of caching in networks of selfish servers through theoretical analysis and simulations. We model selfish caching as a non-cooperative game. In the basic model, the servers have two possible actions for each object. If a replica of a requested object is located at a nearby node, the server may be better off accessing the remote replica. On the other hand, if all replicas are located too far away, the server is better off caching the object itself. Decisions about caching the replicas locally are arrived at locally, taking into account only local costs. We also define a more elaborate payment model, in which each server bids for having an object replicated at another site. Each site now has the option of replicating an object and collecting the related bids. Once all servers have chosen a strategy, each game specifies a configuration, that is, the set of servers that replicate the object, and the corresponding costs for all servers. Game theory predicts that such a situation will end up in a Nash equilibrium, that is, a set of (possibly randomized) strategies with the property that no player can benefit by changing its strategy while the other players keep their strategies unchanged [28]. Foundational considerations notwithstanding, it is not easy to accept randomized strategies as the behavior of rational agents in a distributed system (see [28] for an extensive discussion) - but this is what classical game theory can guarantee. In certain very fortunate situations, however (see [9]), the existence of pure (that is, deterministic) Nash equilibria can be predicted. With or without randomization, however, the lack of coordination inherent in selfish decision-making may incur costs well beyond what would be globally optimum. This loss of efficiency is 1 We will use caching and replication interchangeably. 2 We use the term object as an abstract entity that represents files and other data objects. 21 quantified by the price of anarchy [21]. The price of anarchy is the ratio of the social (total) cost of the worst possible Nash equilibrium to the cost of the social optimum. The price of anarchy bounds the worst possible behavior of a selfish system, when left completely on its own. However, in reality there are ways whereby the system can be guided, through seeding or incentives, to a preselected Nash equilibrium. This optimistic version of the price of anarchy [3] is captured by the smallest ratio between a Nash equilibrium and the social optimum. In this paper we address the following questions : • Do pure strategy Nash equilibria exist in the caching game? • If pure strategy Nash equilibria do exist, how efficient are they (in terms of the price of anarchy, or its optimistic counterpart) under different placement costs, network topologies, and demand distributions? • What is the effect of adopting payments? Will the Nash equilibria be improved? We show that pure strategy Nash equilibria always exist in the caching game. The price of anarchy of the basic game model can be O(n), where n is the number of servers; the intuitive reason is undersupply. Under certain topologies, the price of anarchy does have tighter bounds. For complete graphs and stars, it is O(1). For D-dimensional grids, it is O(n D D+1 ). Even the optimistic price of anarchy can be O(n). In the payment model, however, the game can always implement a Nash equilibrium that is same as the social optimum, so the optimistic price of anarchy is one. Our simulation results show several interesting phases. As the placement cost increases from zero, the price of anarchy increases. When the placement cost first exceeds the maximum distance between servers, the price of anarchy is at its highest due to undersupply problems. As the placement cost further increases, the price of anarchy decreases, and the effect of replica misplacement dominates the price of anarchy. The rest of the paper is organized as follows. In Section 2 we discuss related work. Section 3 discusses details of the basic game and analyzes the bounds of the price of anarchy. In Section 4 we discuss the payment game and analyze its price of anarchy. In Section 5 we describe our simulation methodology and study the properties of Nash equilibria observed. We discuss extensions of the game and directions for future work in Section 6. 2. RELATED WORK There has been considerable research on wide-area peer-to-peer file systems such as OceanStore [22], CFS [5], PAST [32], FARSITE [2], and Pangaea [33], web caches such as NetCache [6] and SummaryCache [10], and peer-to-peer caches such as Squirrel [16]. Most of these systems use caching for performance, availability, and reliability. The caching protocols assume obedience to the protocol and ignore participants" incentives. Our work starts from the assumption that servers are selfish and quantifies the cost of the lack of coordination when servers behave selfishly. The placement of replicas in the caching problem is the most important issue. There is much work on the placement of web replicas, instrumentation servers, and replicated resources. All protocols assume obedience and ignore participants" incentives. In [14], Gribble et al. discuss the data placement problem in peer-to-peer systems. Ko and Rubenstein propose a self-stabilizing, distributed graph coloring algorithm for the replicated resource placement [20]. Chen, Katz, and Kubiatowicz propose a dynamic replica placement algorithm exploiting underlying distributed hash tables [4]. Douceur and Wattenhofer describe a hill-climbing algorithm to exchange replicas for reliability in FARSITE [8]. RaDar is a system that replicates and migrates objects for an Internet hosting service [31]. Tang and Chanson propose a coordinated en-route web caching that caches objects along the routing path [34]. Centralized algorithms for the placement of objects, web proxies, mirrors, and instrumentation servers in the Internet have been studied extensively [18,19,23,30]. The facility location problem has been widely studied as a centralized optimization problem in theoretical computer science and operations research [27]. Since the problem is NP-hard, approximation algorithms based on primal-dual techniques, greedy algorithms, and local search have been explored [17, 24, 26]. Our caching game is different from all of these in that the optimization process is performed among distributed selfish servers. There is little research in non-cooperative facility location games, as far as we know. Vetta [35] considers a class of problems where the social utility is submodular (submodularity means decreasing marginal utility). In the case of competitive facility location among corporations he proves that any Nash equilibrium gives an expected social utility within a factor of 2 of optimal plus an additive term that depends on the facility opening cost. Their results are not directly applicable to our problem, however, because we consider each server to be tied to a particular location, while in their model an agent is able to open facilities in multiple locations. Note that in that paper the increase of the price of anarchy comes from oversupply problems due to the fact that competing corporations can open facilities at the same location. On the other hand, the significant problems in our game are undersupply and misplacement. In a recent paper, Goemans et al. analyze content distribution on ad-hoc wireless networks using a game-theoretic approach [12]. As in our work, they provide monetary incentives to mobile users for caching data items, and provide tight bounds on the price of anarchy and speed of convergence to (approximate) Nash equilibria. However, their results are incomparable to ours because their payoff functions neglect network latencies between users, they consider multiple data items (markets), and each node has a limited budget to cache items. Cost sharing in the facility location problem has been studied using cooperative game theory [7, 13, 29]. Goemans and Skutella show strong connections between fair cost allocations and linear programming relaxations for facility location problems [13]. P´al and Tardos develop a method for cost-sharing that is approximately budget-balanced and group strategyproof and show that the method recovers 1/3 of the total cost for the facility location game [29]. Devanur, Mihail, and Vazirani give a strategyproof cost allocation for the facility location problem, but cannot achieve group strategyproofness [7]. 3. BASIC GAME The caching problem we study is to find a configuration that meets certain objectives (e.g., minimum total cost). Figure 1 shows examples of caching among four servers. In network (a), A stores an object. Suppose B wants to access the object. If it is cheaper to access the remote replica than to cache it, B accesses the remote replica as shown in network (b). In network (c), C wants to access the object. If C is far from A, C caches the object instead of accessing the object from A. It is possible that in an optimal configuration it would be better to place replicas in A and B. Understanding the placement of replicas by selfish servers is the focus of our study. The caching problem is abstracted as follows. There is a set N of n servers and a set M of m objects. The distance between servers can be represented as a distance matrix D (i.e., dij is the distance 22 Server Server Server Server A B C D (a) Server Server Server Server A B C D (b) Server Server Server Server A B C D (c) Figure 1: Caching. There are four servers labeled A, B, C, and D. The rectangles are object replicas. In (a), A stores an object. If B incurs less cost accessing A"s replica than it would caching the object itself, it accesses the object from A as in (b). If the distance cost is too high, the server caches the object itself, as C does in (c). This figure is an example of our caching game model. from server i to server j). D models an underlying network topology. For our analysis we assume that the distances are symmetric and the triangle inequality holds on the distances (for all servers i, j, k: dij + djk ≥ dik). Each server has demand from clients that is represented by a demand matrix W (i.e., wij is the demand of server i for object j). When a server caches objects, the server incurs some placement cost that is represented by a matrix α (i.e., αij is a placement cost of server i for object j). In this study, we assume that servers have no capacity limit. As we discuss in the next section, this fact means that the caching behavior with respect to each object can be examined separately. Consequently, we can talk about configurations of the system with respect to a given object: DEFINITION 1. A configuration X for some object O is the set of servers replicating this object. The goal of the basic game is to find configurations that are achieved when servers optimize their cost functions locally. 3.1 Game Model We take a game-theoretic approach to analyzing the uncapacitated caching problem among networked selfish servers. We model the selfish caching problem as a non-cooperative game with n players (servers/nodes) whose strategies are sets of objects to cache. In the game, each server chooses a pure strategy that minimizes its cost. Our focus is to investigate the resulting configuration, which is the Nash equilibrium of the game. It should be emphasized that we consider only pure strategy Nash equilibria in this paper. The cost model is an important part of the game. Let Ai be the set of feasible strategies for server i, and let Si ∈ Ai be the strategy chosen by server i. Given a strategy profile S = (S1, S2, ..., Sn), the cost incurred by server i is defined as: Ci(S) = j∈Si αij + j /∈Si wij di (i,j). (1) where αij is the placement cost of object j, wij is the demand that server i has for object j, (i, j) is the closest server to i that caches object j, and dik is the distance between i and k. When no server caches the object, we define distance cost di (i,j) to be dM -large enough that at least one server will choose to cache the object. The placement cost can be further divided into first-time installation cost and maintenance cost: αij = k1i + k2i UpdateSizej ObjectSizej 1 T Pj k wkj , (2) where k1i is the installation cost, k2i is the relative weight between the maintenance cost and the installation cost, Pj is the ratio of the number of writes over the number of reads and writes, UpdateSizej is the size of an update, ObjectSizej is the size of the object, and T is the update period. We see tradeoffs between different parameters in this equation. For example, placing replicas becomes more expensive as UpdateSizej increases, Pj increases, or T decreases. However, note that by varying αij itself we can capture the full range of behaviors in the game. For our analysis, we use only αij . Since there is no capacity limit on servers, we can look at each single object as a separate game and combine the pure strategy equilibria of these games to obtain a pure strategy equilibrium of the multi-object game. Fabrikant, Papadimitriou, and Talwar discuss this existence argument: if two games are known to have pure equilibria, and their cost functions are cross-monotonic, then their union is also guaranteed to have pure Nash equilibria, by a continuity argument [9]. A Nash equilibrium for the multi-object game is the cross product of Nash equilibria for single-object games. Therefore, we can focus on the single object game in the rest of this paper. For single object selfish caching, each server i has two strategies - to cache or not to cache. The object under consideration is j. We define Si to be 1 when server i caches j and 0 otherwise. The cost incurred by server i is Ci(S) = αij Si + wij di (i,j)(1 − Si). (3) We refer to this game as the basic game. The extent to which Ci(S) represents actual cost incurred by server i is beyond the scope of this paper; we will assume that an appropriate cost function of the form of Equation 3 can be defined. 3.2 Nash Equilibrium Solutions In principle, we can start with a random configuration and let this configuration evolve as each server alters its strategy and attempts to minimize its cost. Game theory is interested in stable solutions called Nash equilibria. A pure strategy Nash equilibrium is reached when no server can benefit by unilaterally changing its strategy. A Nash equilibrium3 (S∗ i , S∗ −i) for the basic game specifies a configuration X such that ∀i ∈ N, i ∈ X ⇔ S∗ i = 1. Thus, we can consider a set E of all pure strategy Nash equilibrium configurations: X ∈ E ⇔ ∀i ∈ N, ∀Si ∈ Ai, Ci(S∗ i , S∗ −i) ≤ Ci(Si, S∗ −i) (4) By this definition, no server has incentive to deviate in the configurations since it cannot reduce its cost. For the basic game, we can easily see that: X ∈ E ⇔ ∀i ∈ N, ∃j ∈ X s.t. dji ≤ α and ∀j ∈ X, ¬∃k ∈ X s.t. dkj < α (5) The first condition guarantees that there is a server that places the replica within distance α of each server i. If the replica is not placed 3 The notation for strategy profile (S∗ i , S∗ −i) separates node i s strategy (S∗ i ) from the strategies of other nodes (S∗ −i). 23 A B1−α 0 0 0 0 0 0 0 0 0 0 2 n nodes 2 n nodes (a) A B1−α 0 0 0 0 0 0 0 0 0 0 2 n nodes 2 n nodes (b) A B1−α 2 n nodes 2 n nodes n2 n2 n2 n2 n2 n2 n2 n2 n2 n2 (c) Figure 2: Potential inefficiency of Nash equilibria illustrated by two clusters of n 2 servers. The intra-cluster distances are all zero and the distance between clusters is α − 1, where α is the placement cost. The dark nodes replicate the object. Network (a) shows a Nash equilibrium in the basic game, where one server in a cluster caches the object. Network (b) shows the social optimum where two replicas, one for each cluster, are placed. The price of anarchy is O(n) and even the optimistic price of anarchy is O(n). This high price of anarchy comes from the undersupply of replicas due to the selfish nature of servers. Network (c) shows a Nash equilibrium in the payment game, where two replicas, one for each cluster, are placed. Each light node in each cluster pays 2/n to the dark node, and the dark node replicates the object. Here, the optimistic price of anarchy is one. at i, then it is placed at another server within distance α of i, so i has no incentive to cache. If the replica is placed at i, then the second condition ensures there is no incentive to drop the replica because no two servers separated by distance less than α both place replicas. 3.3 Social Optimum The social cost of a given strategy profile is defined as the total cost incurred by all servers, namely: C(S) = n−1 i=0 Ci(S) (6) where Ci(S) is the cost incurred by server i given by Equation 1. The social optimum cost, referred to as C(SO) for the remainder of the paper, is the minimum social cost. The social optimum cost will serve as an important base case against which to measure the cost of selfish caching. We define C(SO) as: C(SO) = min S C(S) (7) where S varies over all possible strategy profiles. Note that in the basic game, this means varying configuration X over all possible configurations. In some sense, C(SO) represents the best possible caching behavior - if only nodes could be convinced to cooperate with one another. The social optimum configuration is a solution of a mini-sum facility location problem, which is NP-hard [11]. To find such configurations, we formulate an integer programming problem: minimize Èi Èj ¢αij xij + Èk wij dikyijk £ subject to ∀i, j Èk yijk = I(wij) ∀i, j, k xij − ykji ≥ 0 ∀i, j xij ∈ {0, 1} ∀i, j, k yijk ∈ {0, 1} (8) Here, xij is 1 if server i replicates object j and 0 otherwise; yijk is 1 if server i accesses object j from server k and 0 otherwise; I(w) returns 1 if w is nonzero and 0 otherwise. The first constraint specifies that if server i has demand for object j, then it must access j from exactly one server. The second constraint ensures that server i replicates object j if any other server accesses j from i. 3.4 Analysis To analyze the basic game, we first give a proof of the existence of pure strategy Nash equilibria. We discuss the price of anarchy in general and then on specific underlying topologies. In this analysis we use simply α in place of αij , since we deal with a single object and we assume placement cost is the same for all servers. In addition, when we compute the price of anarchy, we assume that all nodes have the same demand (i.e., ∀i ∈ N wij = 1). THEOREM 1. Pure strategy Nash equilibria exist in the basic game. PROOF. We show a constructive proof. First, initialize the set V to N. Then, remove all nodes with zero demand from V . Each node x defines βx, where βx = α wxj . Furthermore, let Z(y) = {z : dzy ≤ βz, z ∈ V }; Z(y) represents all nodes z for which y lies within βz from z. Pick a node y ∈ V such that βy ≤ βx for all x ∈ V . Place a replica at y and then remove y and all z ∈ Z(y) from V . No such z can have incentive to replicate the object because it can access y"s replica at lower (or equal) cost. Iterate this process of placing replicas until V is empty. Because at each iteration y is the remaining node with minimum β, no replica will be placed within distance βy of any such y by this process. The resulting configuration is a pure-strategy Nash equilibrium of the basic game. The Price of Anarchy (POA): To quantify the cost of lack of coordination, we use the price of anarchy [21] and the optimistic price of anarchy [3]. The price of anarchy is the ratio of the social costs of the worst-case Nash equilibrium and the social optimum, and the optimistic price of anarchy is the ratio of the social costs of the best-case Nash equilibrium and the social optimum. We show general bounds on the price of anarchy. Throughout our discussion, we use C(SW ) to represent the cost of worst case Nash equilibrium, C(SO) to represent the cost of social optimum, and PoA to represent the price of anarchy, which is C(SW ) C(SO) . The worst case Nash equilibrium maximizes the total cost under the constraint that the configuration meets the Nash condition. Formally, we can define C(SW ) as follows. C(SW ) = max X∈E (α|X| + i min j∈X dij) (9) where minj∈X dij is the distance to the closest replica (including i itself) from node i and X varies through Nash equilibrium configurations. Bounds on the Price of Anarchy: We show bounds of the price of anarchy varying α. Let dmin = min(i,j)∈N×N,i=j dij and dmax = max(i,j)∈N×N dij . We see that if α ≤ dmin, PoA = 1 24 Topology PoA Complete graph 1 Star ≤ 2 Line O( √ n) D-dimensional grid O(n D D+1 ) Table 1: PoA in the basic game for specific topologies trivially, since every server caches the object for both Nash equilibrium and social optimum. When α > dmax, there is a transition in Nash equilibria: since the placement cost is greater than any distance cost, only one server caches the object and other servers access it remotely. However, the social optimum may still place multiple replicas. Since α ≤ C(SO) ≤ α+minj∈N Èi dij when α > dmax, we obtain α+maxj∈N Èi dij α+minj∈N Èi dij ≤ PoA ≤ α+maxj∈N Èi dij α . Note that depending on the underlying topology, even the lower bound of PoA can be O(n). Finally, there is a transition when α > maxj∈N Èi dij. In this case, PoA = α+maxj∈N Èi dij α+minj∈N Èi dij and it is upper bounded by 2. Figure 2 shows an example of the inefficiency of a Nash equilibrium. In the network there are two clusters of servers whose size is n 2 . The distance between two clusters is α − 1 where α is the placement cost. Figure 2(a) shows a Nash equilibrium where one server in a cluster caches the object. In this case, C(SW ) = α + (α − 1)n 2 , since all servers in the other cluster accesses the remote replica. However, the social optimum places two replicas, one for each cluster, as shown in Figure 2(b). Therefore, C(SO) = 2α. PoA = α+(α−1) n 2 2α , which is O(n). This bad price of anarchy comes from an undersupply of replicas due to the selfish nature of the servers. Note that all Nash equilibria have the same cost; thus even the optimistic price of anarchy is O(n). In Appendix A, we analyze the price of anarchy with specific underlying topologies and show that PoA can have tighter bounds than O(n) for the complete graph, star, line, and D-dimensional grid. In these topologies, we set the distance between directly connected nodes to one. We describe the case where α > 1, since PoA = 1 trivially when α ≤ 1. A summary of the results is shown in Table 1. 4. PAYMENT GAME In this section, we present an extension to the basic game with payments and analyze the price of anarchy and the optimistic price of anarchy of the game. 4.1 Game Model The new game, which we refer to as the payment game, allows each player to offer a payment to another player to give the latter incentive to replicate the object. The cost of replication is shared among the nodes paying the server that replicates the object. The strategy for each player i is specified by a triplet (vi, bi, ti) ∈ {N, Ê+, Ê+}. vi specifies the player to whom i makes a bid, bi ≥ 0 is the value of the bid, and ti ≥ 0 denotes a threshold for payments beyond which i will replicate the object. In addition, we use Ri to denote the total amount of bids received by a node i (Ri = Èj:vj =i bj). A node i replicates the object if and only if Ri ≥ ti, that is, the amount of bids it receives is greater than or equal to its threshold. Let Ii denote the corresponding indicator variable, that is, Ii equals 1 if i replicates the object, and 0 otherwise. We make the rule that if a node i makes a bid to another node j and j replicates the object, then i must pay j the amount bi. If j does not replicate the object, i does not pay j. Given a strategy profile, the outcome of the game is the set of tuples {(Ii, vi, bi, Ri)}. Ii tells us whether player i replicates the object or not, bi is the payment player i makes to player vi, and Ri is the total amount of bids received by player i. To compute the payoffs given the outcome, we must now take into account the payments a node makes, in addition to the placement costs and access costs of the basic game. By our rules, a server node i pays bi to node vi if vi replicates the object, and receives a payment of Ri if it replicates the object itself. Its net payment is biIvi − RiIi. The total cost incurred by each node is the sum of its placement cost, access cost, and net payment. It is defined as Ci(S) = αij Ii + wij di (i,j)(1 − Ii) + biIvi − RiIi. (10) The cost of social optimum for the payment game is same as that for the basic game, since the net payments made cancel out. 4.2 Analysis In analyzing the payment model, we first show that a Nash equilibrium in the basic game is also a Nash equilibrium in the payment game. We then present an important positive result - in the payment game the socially optimal configuration can always be implemented by a Nash equilibrium. We know from the counterexample in Figure 2 that this is not guaranteed in the the basic game. In this analysis we use α to represent αij . THEOREM 2. Any configuration that is a pure strategy Nash equilibrium in the basic game is also a pure strategy Nash equilibrium in the payment game. Therefore, the price of anarchy of the payment game is at least that of the basic game. PROOF. Consider any Nash equilibrium configuration in the basic game. For each node i replicating the object, set its threshold ti to 0; everyone else has threshold α. Also, for all i, bi = 0. A node that replicates the object does not have incentive to change its strategy: changing the threshold does not decrease its cost, and it would have to pay at least α to access a remote replica or incentivize a nearby node to cache. Therefore it is better off keeping its threshold and bid at 0 and replicating the object. A node that is not replicating the object can access the object remotely at a cost less than or equal to α. Lowering its threshold does not decrease its cost, since all bi are zero. The payment necessary for another server to place a replica is at least α. No player has incentive to deviate, so the current configuration is a Nash equilibrium. In fact, Appendix B shows that the PoA of the payment game can be more than that of the basic game in a given topology. Now let us look at what happens to the example shown in Figure 2 in the best case. Suppose node B"s neighbors each decide to pay node B an amount 2/n. B does not have an incentive to deviate, since accessing the remote replica does not decrease its cost. The same argument holds for A because of symmetry in the graph. Since no one has an incentive to deviate, the configuration is a Nash equilibrium. Its total cost is 2α, the same as in the socially optimal configuration shown in Figure 2(b). Next we prove that indeed the payment game always has a strategy profile that implements the socially optimal configuration as a Nash equilibrium. We first present the following observation, which is used in the proof, about thresholds in the payment game. OBSERVATION 1. If node i replicates the object, j is the nearest node to i among the other nodes that replicate the object, and dij < α in a Nash equilibrium, then i should have a threshold at 25 least (α − dij). Otherwise, it cannot collect enough payment to compensate for the cost of replicating the object and is better off accessing the replica at j. THEOREM 3. In the payment game, there is always a pure strategy Nash equilibrium that implements the social optimum configuration. The optimistic price of anarchy in the payment game is therefore always one. PROOF. Consider the socially optimal configuration φopt. Let No be the set of nodes that replicate the object and Nc = N − No be the rest of the nodes. Also, for each i in No, let Qi denote the set of nodes that access the object from i, not including i itself. In the socially optimal configuration, dij ≤ α for all j in Qi. We want to find a set of payments and thresholds that makes this configuration implementable. The idea is to look at each node i in No and distribute the minimum payment needed to make i replicate the object among the nodes that access the object from i. For each i in No, and for each j in Qi, we define δj = min{α, min k∈No−{i} djk} − dji (11) Note that δj is the difference between j"s cost for accessing the replica at i and j"s next best option among replicating the object and accessing some replica other than i. It is clear that δj ≥ 0. CLAIM 1. For each i ∈ No, let be the nearest node to i in No. Then, Èj∈Qi δj ≥ α − di . PROOF. (of claim) Assume the contrary, that is, Èj∈Qi δj < α − di . Consider the new configuration φnew wherein i does not replicate and each node in Qi chooses its next best strategy (either replicating or accessing the replica at some node in No − {i}). In addition, we still place replicas at each node in No − {i}. It is easy to see that cost of φopt minus cost of φnew is at least: (α + j∈Qi dij) − (di + j∈Qi min{α, min k∈No−{i} dik}) = α − di − j∈Qi δj > 0, which contradicts the optimality of φopt. We set bids as follows. For each i in No, bi = 0 and for each j in Qi, j bids to i (i.e., vj = i) the amount: bj = max{0, δj − i/(|Qi| + 1)}, j ∈ Qi (12) where i = Èj∈Qi δj − α + di ≥ 0 and |Qi| is the cardinality of Qi. For the thresholds, we have: ti = α if i ∈ Nc;Èj∈Qi bj if i ∈ No. (13) This fully specifies the strategy profile of the nodes, and it is easy to see that the outcome is indeed the socially optimal configuration. Next, we verify that the strategies stipulated constitute a Nash equilibrium. Having set ti to α for i in Nc means that any node in N is at least as well off lowering its threshold and replicating as bidding α to some node in Nc to make it replicate, so we may disregard the latter as a profitable strategy. By observation 1, to ensure that each i in No does not deviate, we require that if is the nearest node to i in No, then Èj∈Qi bj is at least (α − di ). Otherwise, i will raise ti above Èj∈Qi bj so that it does not replicate and instead accesses the replica at . We can easily check that j∈Qi bj ≥ j∈Qi δj − |Qi| i |Qi| + 1 = α − di + i |Qi| + 1 ≥ α − di . 1 1.5 2 2.5 3 3.5 4 4.5 5 5.5 0 20 40 60 80 100 120 140 160 180 200 1 10 100 C(NE)/C(SO) AverageNumberofReplicas alpha PoA Ratio OPoA Replica (SO) Replica (NE) Figure 3: We present P oA, Ratio, and OP oA results for the basic game, varying α on a 100-node line topology, and we show number of replicas placed by the Nash equilibria and by the optimal solution. We see large peaks in P oA and OP oA at α = 100, where a phase transition causes an abrupt transition in the lines. Therefore, each node i ∈ No does not have incentive to change ti since i loses its payments received or there is no change, and i does not have incentive to bi since it replicates the object. Each node j in Nc has no incentive to change tj since changing tj does not reduce its cost. It also does not have incentive to reduce bj since the node where j accesses does not replicate and j has to replicate the object or to access the next closest replica, which costs at least the same from the definition of bj . No player has incentive to deviate, so this strategy profile is a Nash equilibrium. 5. SIMULATION We run simulations to compare Nash equilibria for the singleobject caching game with the social optimum computed by solving the integer linear program described in Equation 8 using Mosek [1]. We examine price of anarchy (PoA), optimistic price of anarchy (OPoA), and the average ratio of the costs of Nash equilibria and social optima (Ratio), and when relevant we also show the average numbers of replicas placed by the Nash equilibrium (Replica(NE)) and the social optimum (Replica(SO)). The PoA and OPoA are taken from the worst and best Nash equilibria, respectively, that we observe over the runs. Each data point in our figures is based on 1000 runs, randomly varying the initial strategy profile and player order. The details of the simulations including protocols and a discussion of convergence are presented in Appendix C. In our evaluation, we study the effects of variation in four categories: placement cost, underlying topology, demand distribution, and payments. As we vary the placement cost α, we directly influence the tradeoff between caching and not caching. In order to get a clear picture of the dependency of PoA on α in a simple case, we first analyze the basic game with a 100-node line topology whose edge distance is one. We also explore transit-stub topologies generated using the GTITM library [36] and power-law topologies (Router-level BarabasiAlbert model) generated using the BRITE topology generator [25]. For these topologies, we generate an underlying physical graph of 3050 physical nodes. Both topologies have similar minimum, average, and maximum physical node distances. The average distance is 0.42. We create an overlay of 100 server nodes and use the same overlay for all experiments with the given topology. In the game, each server has a demand whose distribution is Bernoulli(p), where p is the probability of having demand for the object; the default unless otherwise specified is p = 1.0. 26 0.8 1 1.2 1.4 1.6 1.8 2 2.2 2.4 2.6 2.8 3 0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2 1 10 100 C(NE)/C(SO) AverageNumberofReplicas alpha PoA Ratio OPoA Replica (SO) Replica (NE) (a) 0.8 1 1.2 1.4 1.6 1.8 2 2.2 2.4 2.6 2.8 3 0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2 1 10 100 C(NE)/C(SO) AverageNumberofReplicas alpha PoA Ratio OPoA Replica (SO) Replica (NE) (b) Figure 4: Transit-stub topology: (a) basic game, (b) payment game. We show the P oA, Ratio, OP oA, and the number of replicas placed while varying α between 0 and 2 with 100 servers on a 3050-physical-node transit-stub topology. 0.8 1 1.2 1.4 1.6 1.8 2 2.2 2.4 0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2 1 10 100 C(NE)/C(SO) AverageNumberofReplicas alpha PoA Ratio OPoA Replica (SO) Replica (NE) (a) 0.8 1 1.2 1.4 1.6 1.8 2 2.2 2.4 0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2 1 10 100 C(NE)/C(SO) AverageNumberofReplicas alpha PoA Ratio OPoA Replica (SO) Replica (NE) (b) Figure 5: Power-law topology: (a) basic game, (b) payment game. We show the P oA, Ratio, OP oA, and the number of replicas placed while varying α between 0 and 2 with 100 servers on a 3050-physical-node power-law topology. 5.1 Varying Placement Cost Figure 3 shows PoA, OPoA, and Ratio, as well as number of replicas placed, for the line topology as α varies. We observe two phases. As α increases the PoA rises quickly to a peak at 100. After 100, there is a gradual decline. OPoA and Ratio show behavior similar to PoA. These behaviors can be explained by examining the number of replicas placed by Nash equilibria and by optimal solutions. We see that when α is above one, Nash equilibrium solutions place fewer replicas than optimal on average. For example, when α is 100, the social optimum places four replicas, but the Nash equilibrium places only one. The peak in PoA at α = 100 occurs at the point for a 100-node line where the worst-case cost of accessing a remote replica is slightly less than the cost of placing a new replica, so selfish servers will never place a second replica. The optimal solution, however, places multiple replicas to decrease the high global cost of access. As α continues to increase, the undersupply problem lessens as the optimal solution places fewer replicas. 5.2 Different Underlying Topologies In Figure 4(a) we examine an overlay graph on the more realistic transit-stub topology. The trends for the PoA, OPoA, and Ratio are similar to the results for the line topology, with a peak in PoA at α = 0.8 due to maximal undersupply. In Figure 5(a) we examine an overlay graph on the power-law topology. We observe several interesting differences between the power-law and transit-stub results. First, the PoA peaks at a lower level in the power-law graph, around 2.3 (at α = 0.9) while the peak PoA in the transit-stub topology is almost 3.0 (at α = 0.8). After the peak, PoA and Ratio decrease more slowly as α increases. OPoA is close to one for the whole range of α values. This can be explained by the observation in Figure 5(a) that there is no significant undersupply problem here like there was in the transit-stub graph. Indeed the high PoA is due mostly to misplacement problems when α is from 0.7 to 2.0, since there is little decrease in PoA when the number of replicas in social optimum changes from two to one. The OPoA is equal to one in the figure when the same number of replicas are placed. 5.3 Varying Demand Distribution Now we examine the effects of varying the demand distribution. The set of servers with demand is random for p < 1, so we calculate the expected PoA by averaging over 5 trials (each data point is based on 5000 runs). We run simulations for demand levels of p ∈ {0.2, 0.6, 1.0} as α is varied on the 100 servers on top of the transit-stub graph. We observe that as demand falls, so does expected PoA. As p decreases, the number of replicas placed in the social optimum decreases, but the number in Nash equilibria changes little. Furthermore, when α exceeds the overlay diameter, the number in Nash equilibria stays constant when p varies. Therefore, lower p leads to a lesser undersupply problem, agreeing with intuition. We do not present the graph due to space limitations and redundancy; the PoA for p = 1.0 is identical to PoA in Figure 4(a), and the lines for p = 0.6 and p = 0.2 are similar but lower and flatter. 27 5.4 Effects of Payment Finally, we discuss the effects of payments on the efficiency of Nash equilibria. The results are presented in Figure 4(b) and Figure 5(b). As shown in the analysis, the simulations achieve OPoA close to one (it is not exactly one because of randomness in the simulations). The Ratio for the payment game is much lower than the Ratio for the basic game, since the protocol for the payment game tends to explore good regions in the space of Nash equilibria. We observe in Figure 4 that for α ≥ 0.4, the average number of replicas of Nash equilibria gets closer with payments to that of the social optimum than it does without. We observe in Figure 5 that more replicas are placed with payments than without when α is between 0.7 and 1.3, the only range of significant undersupply in the power-law case. The results confirm that payments give servers incentive to replicate the object and this leads to better equilibria. 6. DISCUSSION AND FUTURE WORK We suggest several interesting extensions and directions. One extension is to consider multiple objects in the capacitated caching game, in which servers have capacity limits when placing objects. Since caching one object affects the ability to cache another, there is no separability of a multi-object game into multiple single object games. As studied in [12], one way to formulate this problem is to find the best response of a server by solving a knapsack problem and to compute Nash equilibria. In our analyses, we assume that all nodes have the same demand. However, nodes could have different demand depending on objects. We intend to examine the effects of heterogeneous demands (or heterogeneous placement costs) analytically. We also want to look at the following aggregation effect. Suppose there are n − 1 clustered nodes with distance of α−1 from a node hosting a replica. All nodes have demands of one. In that case, the price of anarchy is O(n). However, if we aggregate n − 1 nodes into one node with demand n − 1, the price of anarchy becomes O(1), since α should be greater than (n − 1)(α − 1) to replicate only one object. Such aggregation can reduce the inefficiency of Nash equilibria. We intend to compute the bounds of the price of anarchy under different underlying topologies such as random graphs or growthrestricted metrics. We want to investigate whether there are certain distance constraints that guarantee O(1) price of anarchy. In addition, we want to run large-scale simulations to observe the change in the price of anarchy as the network size increases. Another extension is to consider server congestion. Suppose the distance is the network distance plus γ × (number of accesses) where γ is an extra delay when an additional server accesses the replica. Then, when α > γ, it can be shown that PoA is bounded by α γ . As γ increases, the price of anarchy bound decreases, since the load of accesses is balanced across servers. While exploring the caching problem, we made several observations that seem counterintuitive. First, the PoA in the payment game can be worse than the PoA in the basic game. Another observation we made was that the number of replicas in a Nash equilibrium can be more than the number of replicas in the social optimum even without payments. For example, a graph with diameter slightly more than α may have a Nash equilibrium configuration with two replicas at the two ends. However, the social optimum may place one replica at the center. We leave the investigation of more examples as an open issue. 7. CONCLUSIONS In this work we introduce a novel non-cooperative game model to characterize the caching problem among selfish servers without any central coordination. We show that pure strategy Nash equilibria exist in the game and that the price of anarchy can be O(n) in general, where n is the number of servers, due to undersupply problems. With specific topologies, we show that the price of anarchy can have tighter bounds. More importantly, with payments, servers are incentivized to replicate and the optimistic price of anarchy is always one. Non-cooperative caching is a more realistic model than cooperative caching in the competitive Internet, hence this work is an important step toward viable federated caching systems. 8. ACKNOWLEDGMENTS We thank Kunal Talwar for enlightening discussions regarding this work. 9. REFERENCES [1] http://www.mosek.com. [2] A. Adya et al. FARSITE: Federated, Available, and Reliable Storage for an Incompletely Trusted Environment. In Proc. of USENIX OSDI, 2002. [3] E. Anshelevich, A. Dasgupta, E. Tardos, and T. Wexler. Near-optimal Network Design with Selfish Agents. In Proc. of ACM STOC, 2003. [4] Y. Chen, R. H. Katz, and J. D. Kubiatowicz. SCAN: A Dynamic, Scalable, and Efficient Content Distribution Network. In Proc. of Intl. Conf. on Pervasive Computing, 2002. [5] F. Dabek et al. Wide-area Cooperative Storage with CFS. In Proc. of ACM SOSP, Oct. 2001. [6] P. B. Danzig. NetCache Architecture and Deploment. In Computer Networks and ISDN Systems, 1998. [7] N. Devanur, M. Mihail, and V. Vazirani. Strategyproof cost-sharing Mechanisms for Set Cover and Facility Location Games. In Proc. of ACM EC, 2003. [8] J. R. Douceur and R. P. Wattenhofer. Large-Scale Simulation of Replica Placement Algorithms for a Serverless Distributed File System. In Proc. of MASCOTS, 2001. [9] A. Fabrikant, C. H. Papadimitriou, and K. Talwar. The Complexity of Pure Nash Equilibria. In Proc. of ACM STOC, 2004. [10] L. Fan, P. Cao, J. Almeida, and A. Z. Broder. Summary Cache: A Scalable Wide-area Web Cache Sharing Protocol. IEEE/ACM Trans. on Networking, 8(3):281-293, 2000. [11] M. R. Garey and D. S. Johnson. Computers and Intractability: A Guide to the Theory of NP-Completeness. W. H. Freeman and Co., 1979. [12] M. X. Goemans, L. Li, V. S. Mirrokni, and M. Thottan. Market Sharing Games Applied to Content Distribution in ad-hoc Networks. In Proc. of ACM MOBIHOC, 2004. [13] M. X. Goemans and M. Skutella. Cooperative Facility Location Games. In Proc. of ACM-SIAM SODA, 2000. [14] S. Gribble et al. What Can Databases Do for Peer-to-Peer? In WebDB Workshop on Databases and the Web, June 2001. [15] K. P. Gummadi et al. Measurement, Modeling, and Analysis of a Peer-to-Peer File-Sharing Workload. In Proc. of ACM SOSP, October 2003. [16] S. Iyer, A. Rowstron, and P. Druschel. Squirrel: A Decentralized Peer-to-Peer Web Cache. In Proc. of ACM PODC, 2002. [17] K. Jain and V. V. Vazirani. Primal-Dual Approximation Algorithms for Metric Facility Location and k-Median Problems. In Proc. of IEEE FOCS, 1999. 28 [18] S. Jamin et al. On the Placement of Internet Instrumentation. In Proc. of IEEE INFOCOM, pages 295-304, 2000. [19] S. Jamin et al. Constrained Mirror Placement on the Internet. In Proc. of IEEE INFOCOM, pages 31-40, 2001. [20] B.-J. Ko and D. Rubenstein. A Distributed, Self-stabilizing Protocol for Placement of Replicated Resources in Emerging Networks. In Proc. of IEEE ICNP, 2003. [21] E. Koutsoupias and C. Papadimitriou. Worst-Case Equilibria. In STACS, 1999. [22] J. Kubiatowicz et al. OceanStore: An Architecture for Global-scale Persistent Storage. In Proc. of ACM ASPLOS. ACM, November 2000. [23] B. Li, M. J. Golin, G. F. Italiano, and X. Deng. On the Optimal Placement of Web Proxies in the Internet. In Proc. of IEEE INFOCOM, 1999. [24] M. Mahdian, Y. Ye, and J. Zhang. Improved Approximation Algorithms for Metric Facility Location Problems. In Proc. of Intl. Workshop on Approximation Algorithms for Combinatorial Optimization Problems, 2002. [25] A. Medina, A. Lakhina, I. Matta, and J. Byers. BRITE: Universal Topology Generation from a User"s Perspective. Technical Report 2001-003, 1 2001. [26] R. R. Mettu and C. G. Plaxton. The Online Median Problem. In Proc. of IEEE FOCS, 2000. [27] P. B. Mirchandani and R. L. Francis. Discrete Location Theory. Wiley-Interscience Series in Discrete Mathematics and Optimization, 1990. [28] M. J. Osborne and A. Rubinstein. A Course in Game Theory. MIT Press, 1994. [29] M. Pal and E. Tardos. Group Strategyproof Mechanisms via Primal-Dual Algorithms. In Proc. of IEEE FOCS, 2003. [30] L. Qiu, V. N. Padmanabhan, and G. M. Voelker. On the Placement of Web Server Replicas. In Proc. of IEEE INFOCOM, 2001. [31] M. Rabinovich, I. Rabinovich, R. Rajaraman, and A. Aggarwal. A Dynamic Object Replication and Migration Protocol for an Internet Hosting Service. In Proc. of IEEE ICDCS, 1999. [32] A. Rowstron and P. Druschel. Storage Management and Caching in PAST, A Large-scale, Persistent Peer-to-peer Storage Utility. In Proc. of ACM SOSP, October 2001. [33] Y. Saito, C. Karamanolis, M. Karlsson, and M. Mahalingam. Taming Aggressive Replication in the Pangaea Wide-Area File System. In Proc. of USENIX OSDI, 2002. [34] X. Tang and S. T. Chanson. Coordinated En-route Web Caching. In IEEE Trans. Computers, 2002. [35] A. Vetta. Nash Equilibria in Competitive Societies, with Applications to Facility Location, Traffic Routing, and Auctions. In Proc. of IEEE FOCS, 2002. [36] E. W. Zegura, K. L. Calvert, and S. Bhattacharjee. How to Model an Internetwork. In Proc. of IEEE INFOCOM, 1996. APPENDIX A. ANALYZING SPECIFIC TOPOLOGIES We now analyze the price of anarchy (PoA) for the basic game with specific underlying topologies and show that PoA can have better bounds. We look at complete graph, star, line, and Ddimensional grid. In all these topologies, we set the distance between two directly connected nodes to one. We describe the case where α > 1, since PoA = 1 trivially when α ≤ 1. A BC D3 α 3 α 4 3α 4 α 4 α Figure 6: Example where the payment game has a Nash equilibrium which is worse than any Nash equilibrium in the basic game. The unlabeled distances between the nodes in the cluster are all 1. The thresholds of white nodes are all α and the thresholds of dark nodes are all α/4. The two dark nodes replicate the object in this payment game Nash equilibrium. For a complete graph, PoA = 1, and for a star, PoA ≤ 2. For a complete graph, when α > 1, both Nash equilibria and social optima place one replica at one server, so PoA = 1. For star, when 1 < α < 2, the worst case Nash equilibrium places replicas at all leaf nodes. However, the social optimum places one replica at the center node. Therefore, PoA = (n−1)α+1 α+(n−1) ≤ 2(n−1)+1 1+(n−1) ≤ 2. When α > 2, the worst case Nash equilibrium places one replica at a leaf node and the other nodes access the remote replica, and the social optimum places one replica at the center. PoA = α+1+2(n−2) α+(n−1) = 1 + n α+(n−1) ≤ 2. For a line, the price of anarchy is O( √ n). When 1 < α < n, the worst case Nash equilibrium places replicas every 2α so that there is no overlap between areas covered by two adjacent servers that cache the object. The social optimum places replicas at least every √ 2α. The placement of replicas for the social optimum is as follows. Suppose there are two replicas separated by distance d. By placing an additional replica in the middle, we want to have the reduction of distance to be at least α. The distance reduction is d/2 + 2{((d/2 − 1) − 1) + ((d/2 − 2) − 2) + ... + ((d/2 − d/4) − d/4)} ≥ d2 /8. d should be at most 2 √ 2α. Therefore, the distance between replicas in the social optimum is at most √ 2α. C(SW ) = α(n−1) 2α + α(α+1) 2 (n−1) 2α = Θ(αn). C(SO) ≥ α n−1√ 2α + 2 √ 2α/2( √ 2α/2+1) 2 n−1√ 2α . C(SO) = Ω( √ αn). Therefore, PoA = O( √ α). When α > n − 1, the worst case Nash equilibrium places one replica at a leaf node and C(SW ) = α + (n−1)n 2 . However, the social optimum still places replicas every √ 2α. If we view PoA as a continuous function of α and compute a derivative of PoA, the derivative becomes 0 when α is Θ(n2 ), which means the function decreases as α increases from n. Therefore, PoA is maximum when α is n, and PoA = Θ(n2 ) Ω( √ nn) = O( √ n). When α > (n−1)n 2 , the social optimum also places only one replica, and PoA is trivially bounded by 2. This result holds for the ring and it can be generalized to the D-dimensional grid. As the dimension in the grid increases, the distance reduction of additional replica placement becomes Ω(dD+1 ) where d is the distance between two adjacent replicas. Therefore, PoA = Θ(n2) Ω(n 1 D+1 n) = O(n D D+1 ). B. PAYMENT CAN DO WORSE Consider the network in Figure 6 where α > 1+α/3. Any Nash equilibrium in the basic game model would have exactly two replicas - one in the left cluster, and one in the right. It is easy to verify that the worst placement (in terms of social cost) of two replicas occurs when they are placed at nodes A and B. This placement can be achieved as a Nash equilibrium in the payment game, but not in the basic game since A and B are a distance 3α/4 apart. 29 Algorithm 1 Initialization for the Basic Game L1 = a random subset of servers for each node i in N do if i ∈ L1 then Si = 1 ; replicate the object else Si = 0 Algorithm 2 Move Selection of i for the Basic Game Cost1 = α Cost2 = minj∈X−{i} dij ; X is the current configuration Costmin = min{Cost1, Cost2} if Costnow > Costmin then if Costmin == Cost1 then Si = 1 else Si = 0 C. NASH DYNAMICS PROTOCOLS The simulator initializes the game according to the given parameters and a random initial strategy profile and then iterates through rounds. Initially the order of player actions is chosen randomly. In each round, each server performs the Nash dynamics protocol that adjusts its strategies greedily in the chosen order. When a round passes without any server changing its strategy, the simulation ends and a Nash equilibrium is reached. In the basic game, we pick a random initial subset of servers to replicate the object as shown in Algorithm 1. After the initialization, each player runs the move selection procedure described in Algorithm 2 (in algorithms 2 and 4, Costnow represents the current cost for node i). This procedure chooses greedily between replication and non-replication. It is not hard to see that this Nash dynamics protocol converges in two rounds. In the payment game, we pick a random initial subset of servers to replicate the object by setting their thresholds to 0. In addition, we initialize a second random subset of servers to replicate the object with payments from other servers. The details are shown in Algorithm 3. After the initialization, each player runs the move selection procedure described in Algorithm 4. This procedure chooses greedily between replication and accessing a remote replica, with the possibilities of receiving and making payments, respectively. In the protocol, each node increases its threshold value by incr if it does not replicate the object. By this ramp up procedure, the cost of replicating an object is shared fairly among the nodes that access a replica from a server that does cache. If incr is small, cost is shared more fairly, and the game tends to reach equilibria that encourages more servers to store replicas, though the convergence takes longer. If incr is large, the protocol converges quickly, but it may miss efficient equilibria. In the simulations we set incr to 0.1. Most of our A B C a b c α/3+1 2α/3−1 2α/3 Figure 7: An example where the Nash dynamics protocol does not converge in the payment game. Algorithm 3 Initialization for the Payment Game L1 = a random subset of servers for each node i in N do bi = 0 if i ∈ L1 then ti = 0 ; replicate the object else ti = α L2 = {} for each node i in N do if coin toss == head then Mi = {j : d(j, i) < mink∈L1∪L2 d(j, k)} if Mi != ∅ then for each node j ∈ Mi do bj = max{ α+ Èk∈Mi d(i,k) |Mi| − d(i, j), 0} L2 = L2 ∪ {i} Algorithm 4 Move Selection of i for the Payment Game Cost1 = α − Ri Cost2 = minj∈N−{i}{tj − Rj + dij } Costmin = min{Cost1, Cost2} if Costnow > Costmin then if Costmin == Cost1 then ti = Ri else ti = Ri + incr vi = argminj{tj − Rj + dij} bi = tvi − Rvi simulation runs converged, but there were a very few cases where the simulation did not converge due to the cycles of dynamics. The protocol does not guarantee convergence within a certain number of rounds like the protocol for the basic game. We provide an example graph and an initial condition such that the Nash dynamics protocol does not converge in the payment game if started from this initial condition. The graph is represented by a shortest path metric on the network shown in Figure 7. In the starting configuration, only A replicates the object, and a pays it an amount α/3 to do so. The thresholds for A, B and C are α/3 each, and the thresholds for a, b and c are 2α/3. It is not hard to verify that the Nash dynamics protocol will never converge if we start with this condition. The Nash dynamics protocol for the payment game needs further investigation. The dynamics protocol for the payment game should avoid cycles of actions to achieve stabilization of the protocol. Finding a self-stabilizing dynamics protocol is an interesting problem. In addition, a fixed value of incr cannot adapt to changing environments. A small value of incr can lead to efficient equilibria, but it can take long time to converge. An important area for future research is looking at adaptively changing incr. 30

caching problem;nash equilibrium;remote replica;social utility;submodularity;peer-to-peer file system;demand distribution;game-theoretic model;cache;anarchy price;aggregation effect;network topology;game-theoretic approach;distribute system;group strategyproofness;instrumentation server;price of anarchy;primal-dual technique;peer-to-peer system

train_H-35

AdaRank: A Boosting Algorithm for Information Retrieval

In this paper we address the issue of learning to rank for document retrieval. In the task, a model is automatically created with some training data and then is utilized for ranking of documents. The goodness of a model is usually evaluated with performance measures such as MAP (Mean Average Precision) and NDCG (Normalized Discounted Cumulative Gain). Ideally a learning algorithm would train a ranking model that could directly optimize the performance measures with respect to the training data. Existing methods, however, are only able to train ranking models by minimizing loss functions loosely related to the performance measures. For example, Ranking SVM and RankBoost train ranking models by minimizing classification errors on instance pairs. To deal with the problem, we propose a novel learning algorithm within the framework of boosting, which can minimize a loss function directly defined on the performance measures. Our algorithm, referred to as AdaRank, repeatedly constructs ‘weak rankers" on the basis of re-weighted training data and finally linearly combines the weak rankers for making ranking predictions. We prove that the training process of AdaRank is exactly that of enhancing the performance measure used. Experimental results on four benchmark datasets show that AdaRank significantly outperforms the baseline methods of BM25, Ranking SVM, and RankBoost.

1. INTRODUCTION Recently ‘learning to rank" has gained increasing attention in both the fields of information retrieval and machine learning. When applied to document retrieval, learning to rank becomes a task as follows. In training, a ranking model is constructed with data consisting of queries, their corresponding retrieved documents, and relevance levels given by humans. In ranking, given a new query, the corresponding retrieved documents are sorted by using the trained ranking model. In document retrieval, usually ranking results are evaluated in terms of performance measures such as MAP (Mean Average Precision) [1] and NDCG (Normalized Discounted Cumulative Gain) [15]. Ideally, the ranking function is created so that the accuracy of ranking in terms of one of the measures with respect to the training data is maximized. Several methods for learning to rank have been developed and applied to document retrieval. For example, Herbrich et al. [13] propose a learning algorithm for ranking on the basis of Support Vector Machines, called Ranking SVM. Freund et al. [8] take a similar approach and perform the learning by using boosting, referred to as RankBoost. All the existing methods used for document retrieval [2, 3, 8, 13, 16, 20] are designed to optimize loss functions loosely related to the IR performance measures, not loss functions directly based on the measures. For example, Ranking SVM and RankBoost train ranking models by minimizing classification errors on instance pairs. In this paper, we aim to develop a new learning algorithm that can directly optimize any performance measure used in document retrieval. Inspired by the work of AdaBoost for classification [9], we propose to develop a boosting algorithm for information retrieval, referred to as AdaRank. AdaRank utilizes a linear combination of ‘weak rankers" as its model. In learning, it repeats the process of re-weighting the training sample, creating a weak ranker, and calculating a weight for the ranker. We show that AdaRank algorithm can iteratively optimize an exponential loss function based on any of IR performance measures. A lower bound of the performance on training data is given, which indicates that the ranking accuracy in terms of the performance measure can be continuously improved during the training process. AdaRank offers several advantages: ease in implementation, theoretical soundness, efficiency in training, and high accuracy in ranking. Experimental results indicate that AdaRank can outperform the baseline methods of BM25, Ranking SVM, and RankBoost, on four benchmark datasets including OHSUMED, WSJ, AP, and .Gov. Tuning ranking models using certain training data and a performance measure is a common practice in IR [1]. As the number of features in the ranking model gets larger and the amount of training data gets larger, the tuning becomes harder. From the viewpoint of IR, AdaRank can be viewed as a machine learning method for ranking model tuning. Recently, direct optimization of performance measures in learning has become a hot research topic. Several methods for classification [17] and ranking [5, 19] have been proposed. AdaRank can be viewed as a machine learning method for direct optimization of performance measures, based on a different approach. The rest of the paper is organized as follows. After a summary of related work in Section 2, we describe the proposed AdaRank algorithm in details in Section 3. Experimental results and discussions are given in Section 4. Section 5 concludes this paper and gives future work. 2. RELATED WORK 2.1 Information Retrieval The key problem for document retrieval is ranking, specifically, how to create the ranking model (function) that can sort documents based on their relevance to the given query. It is a common practice in IR to tune the parameters of a ranking model using some labeled data and one performance measure [1]. For example, the state-ofthe-art methods of BM25 [24] and LMIR (Language Models for Information Retrieval) [18, 22] all have parameters to tune. As the ranking models become more sophisticated (more features are used) and more labeled data become available, how to tune or train ranking models turns out to be a challenging issue. Recently methods of ‘learning to rank" have been applied to ranking model construction and some promising results have been obtained. For example, Joachims [16] applies Ranking SVM to document retrieval. He utilizes click-through data to deduce training data for the model creation. Cao et al. [4] adapt Ranking SVM to document retrieval by modifying the Hinge Loss function to better meet the requirements of IR. Specifically, they introduce a Hinge Loss function that heavily penalizes errors on the tops of ranking lists and errors from queries with fewer retrieved documents. Burges et al. [3] employ Relative Entropy as a loss function and Gradient Descent as an algorithm to train a Neural Network model for ranking in document retrieval. The method is referred to as ‘RankNet". 2.2 Machine Learning There are three topics in machine learning which are related to our current work. They are ‘learning to rank", boosting, and direct optimization of performance measures. Learning to rank is to automatically create a ranking function that assigns scores to instances and then rank the instances by using the scores. Several approaches have been proposed to tackle the problem. One major approach to learning to rank is that of transforming it into binary classification on instance pairs. This ‘pair-wise" approach fits well with information retrieval and thus is widely used in IR. Typical methods of the approach include Ranking SVM [13], RankBoost [8], and RankNet [3]. For other approaches to learning to rank, refer to [2, 11, 31]. In the pair-wise approach to ranking, the learning task is formalized as a problem of classifying instance pairs into two categories (correctly ranked and incorrectly ranked). Actually, it is known that reducing classification errors on instance pairs is equivalent to maximizing a lower bound of MAP [16]. In that sense, the existing methods of Ranking SVM, RankBoost, and RankNet are only able to minimize loss functions that are loosely related to the IR performance measures. Boosting is a general technique for improving the accuracies of machine learning algorithms. The basic idea of boosting is to repeatedly construct ‘weak learners" by re-weighting training data and form an ensemble of weak learners such that the total performance of the ensemble is ‘boosted". Freund and Schapire have proposed the first well-known boosting algorithm called AdaBoost (Adaptive Boosting) [9], which is designed for binary classification (0-1 prediction). Later, Schapire & Singer have introduced a generalized version of AdaBoost in which weak learners can give confidence scores in their predictions rather than make 0-1 decisions [26]. Extensions have been made to deal with the problems of multi-class classification [10, 26], regression [7], and ranking [8]. In fact, AdaBoost is an algorithm that ingeniously constructs a linear model by minimizing the ‘exponential loss function" with respect to the training data [26]. Our work in this paper can be viewed as a boosting method developed for ranking, particularly for ranking in IR. Recently, a number of authors have proposed conducting direct optimization of multivariate performance measures in learning. For instance, Joachims [17] presents an SVM method to directly optimize nonlinear multivariate performance measures like the F1 measure for classification. Cossock & Zhang [5] find a way to approximately optimize the ranking performance measure DCG [15]. Metzler et al. [19] also propose a method of directly maximizing rank-based metrics for ranking on the basis of manifold learning. AdaRank is also one that tries to directly optimize multivariate performance measures, but is based on a different approach. AdaRank is unique in that it employs an exponential loss function based on IR performance measures and a boosting technique. 3. OUR METHOD: ADARANK 3.1 General Framework We first describe the general framework of learning to rank for document retrieval. In retrieval (testing), given a query the system returns a ranking list of documents in descending order of the relevance scores. The relevance scores are calculated with a ranking function (model). In learning (training), a number of queries and their corresponding retrieved documents are given. Furthermore, the relevance levels of the documents with respect to the queries are also provided. The relevance levels are represented as ranks (i.e., categories in a total order). The objective of learning is to construct a ranking function which achieves the best results in ranking of the training data in the sense of minimization of a loss function. Ideally the loss function is defined on the basis of the performance measure used in testing. Suppose that Y = {r1, r2, · · · , r } is a set of ranks, where denotes the number of ranks. There exists a total order between the ranks r r −1 · · · r1, where ‘ " denotes a preference relationship. In training, a set of queries Q = {q1, q2, · · · , qm} is given. Each query qi is associated with a list of retrieved documents di = {di1, di2, · · · , di,n(qi)} and a list of labels yi = {yi1, yi2, · · · , yi,n(qi)}, where n(qi) denotes the sizes of lists di and yi, dij denotes the jth document in di, and yij ∈ Y denotes the rank of document di j. A feature vector xij = Ψ(qi, di j) ∈ X is created from each query-document pair (qi, di j), i = 1, 2, · · · , m; j = 1, 2, · · · , n(qi). Thus, the training set can be represented as S = {(qi, di, yi)}m i=1. The objective of learning is to create a ranking function f : X → , such that for each query the elements in its corresponding document list can be assigned relevance scores using the function and then be ranked according to the scores. Specifically, we create a permutation of integers π(qi, di, f) for query qi, the corresponding list of documents di, and the ranking function f. Let di = {di1, di2, · · · , di,n(qi)} be identified by the list of integers {1, 2, · · · , n(qi)}, then permutation π(qi, di, f) is defined as a bijection from {1, 2, · · · , n(qi)} to itself. We use π( j) to denote the position of item j (i.e., di j). The learning process turns out to be that of minimizing the loss function which represents the disagreement between the permutation π(qi, di, f) and the list of ranks yi, for all of the queries. Table 1: Notations and explanations. Notations Explanations qi ∈ Q ith query di = {di1, di2, · · · , di,n(qi)} List of documents for qi yi j ∈ {r1, r2, · · · , r } Rank of di j w.r.t. qi yi = {yi1, yi2, · · · , yi,n(qi)} List of ranks for qi S = {(qi, di, yi)}m i=1 Training set xij = Ψ(qi, dij) ∈ X Feature vector for (qi, di j) f(xij) ∈ Ranking model π(qi, di, f) Permutation for qi, di, and f ht(xi j) ∈ tth weak ranker E(π(qi, di, f), yi) ∈ [−1, +1] Performance measure function In the paper, we define the rank model as a linear combination of weak rankers: f(x) = T t=1 αtht(x), where ht(x) is a weak ranker, αt is its weight, and T is the number of weak rankers. In information retrieval, query-based performance measures are used to evaluate the ‘goodness" of a ranking function. By query based measure, we mean a measure defined over a ranking list of documents with respect to a query. These measures include MAP, NDCG, MRR (Mean Reciprocal Rank), WTA (Winners Take ALL), and Precision@n [1, 15]. We utilize a general function E(π(qi, di, f), yi) ∈ [−1, +1] to represent the performance measures. The first argument of E is the permutation π created using the ranking function f on di. The second argument is the list of ranks yi given by humans. E measures the agreement between π and yi. Table 1 gives a summary of notations described above. Next, as examples of performance measures, we present the definitions of MAP and NDCG. Given a query qi, the corresponding list of ranks yi, and a permutation πi on di, average precision for qi is defined as: AvgPi = n(qi) j=1 Pi( j) · yij n(qi) j=1 yij , (1) where yij takes on 1 and 0 as values, representing being relevant or irrelevant and Pi( j) is defined as precision at the position of dij: Pi( j) = k:πi(k)≤πi(j) yik πi(j) , (2) where πi( j) denotes the position of di j. Given a query qi, the list of ranks yi, and a permutation πi on di, NDCG at position m for qi is defined as: Ni = ni · j:πi(j)≤m 2yi j − 1 log(1 + πi( j)) , (3) where yij takes on ranks as values and ni is a normalization constant. ni is chosen so that a perfect ranking π∗ i "s NDCG score at position m is 1. 3.2 Algorithm Inspired by the AdaBoost algorithm for classification, we have devised a novel algorithm which can optimize a loss function based on the IR performance measures. The algorithm is referred to as ‘AdaRank" and is shown in Figure 1. AdaRank takes a training set S = {(qi, di, yi)}m i=1 as input and takes the performance measure function E and the number of iterations T as parameters. AdaRank runs T rounds and at each round it creates a weak ranker ht(t = 1, · · · , T). Finally, it outputs a ranking model f by linearly combining the weak rankers. At each round, AdaRank maintains a distribution of weights over the queries in the training data. We denote the distribution of weights Input: S = {(qi, di, yi)}m i=1, and parameters E and T Initialize P1(i) = 1/m. For t = 1, · · · , T • Create weak ranker ht with weighted distribution Pt on training data S . • Choose αt αt = 1 2 · ln m i=1 Pt(i){1 + E(π(qi, di, ht), yi)} m i=1 Pt(i){1 − E(π(qi, di, ht), yi)} . • Create ft ft(x) = t k=1 αkhk(x). • Update Pt+1 Pt+1(i) = exp{−E(π(qi, di, ft), yi)} m j=1 exp{−E(π(qj, dj, ft), yj)} . End For Output ranking model: f(x) = fT (x). Figure 1: The AdaRank algorithm. at round t as Pt and the weight on the ith training query qi at round t as Pt(i). Initially, AdaRank sets equal weights to the queries. At each round, it increases the weights of those queries that are not ranked well by ft, the model created so far. As a result, the learning at the next round will be focused on the creation of a weak ranker that can work on the ranking of those ‘hard" queries. At each round, a weak ranker ht is constructed based on training data with weight distribution Pt. The goodness of a weak ranker is measured by the performance measure E weighted by Pt: m i=1 Pt(i)E(π(qi, di, ht), yi). Several methods for weak ranker construction can be considered. For example, a weak ranker can be created by using a subset of queries (together with their document list and label list) sampled according to the distribution Pt. In this paper, we use single features as weak rankers, as will be explained in Section 3.6. Once a weak ranker ht is built, AdaRank chooses a weight αt > 0 for the weak ranker. Intuitively, αt measures the importance of ht. A ranking model ft is created at each round by linearly combining the weak rankers constructed so far h1, · · · , ht with weights α1, · · · , αt. ft is then used for updating the distribution Pt+1. 3.3 Theoretical Analysis The existing learning algorithms for ranking attempt to minimize a loss function based on instance pairs (document pairs). In contrast, AdaRank tries to optimize a loss function based on queries. Furthermore, the loss function in AdaRank is defined on the basis of general IR performance measures. The measures can be MAP, NDCG, WTA, MRR, or any other measures whose range is within [−1, +1]. We next explain why this is the case. Ideally we want to maximize the ranking accuracy in terms of a performance measure on the training data: max f∈F m i=1 E(π(qi, di, f), yi), (4) where F is the set of possible ranking functions. This is equivalent to minimizing the loss on the training data min f∈F m i=1 (1 − E(π(qi, di, f), yi)). (5) It is difficult to directly optimize the loss, because E is a noncontinuous function and thus may be difficult to handle. We instead attempt to minimize an upper bound of the loss in (5) min f∈F m i=1 exp{−E(π(qi, di, f), yi)}, (6) because e−x ≥ 1 − x holds for any x ∈ . We consider the use of a linear combination of weak rankers as our ranking model: f(x) = T t=1 αtht(x). (7) The minimization in (6) then turns out to be min ht∈H,αt∈ + L(ht, αt) = m i=1 exp{−E(π(qi, di, ft−1 + αtht), yi)}, (8) where H is the set of possible weak rankers, αt is a positive weight, and ( ft−1 + αtht)(x) = ft−1(x) + αtht(x). Several ways of computing coefficients αt and weak rankers ht may be considered. Following the idea of AdaBoost, in AdaRank we take the approach of ‘forward stage-wise additive modeling" [12] and get the algorithm in Figure 1. It can be proved that there exists a lower bound on the ranking accuracy for AdaRank on training data, as presented in Theorem 1. T 1. The following bound holds on the ranking accuracy of the AdaRank algorithm on training data: 1 m m i=1 E(π(qi, di, fT ), yi) ≥ 1 − T t=1 e−δt min 1 − ϕ(t)2, where ϕ(t) = m i=1 Pt(i)E(π(qi, di, ht), yi), δt min = mini=1,··· ,m δt i, and δt i = E(π(qi, di, ft−1 + αtht), yi) − E(π(qi, di, ft−1), yi) −αtE(π(qi, di, ht), yi), for all i = 1, 2, · · · , m and t = 1, 2, · · · , T. A proof of the theorem can be found in appendix. The theorem implies that the ranking accuracy in terms of the performance measure can be continuously improved, as long as e−δt min 1 − ϕ(t)2 < 1 holds. 3.4 Advantages AdaRank is a simple yet powerful method. More importantly, it is a method that can be justified from the theoretical viewpoint, as discussed above. In addition AdaRank has several other advantages when compared with the existing learning to rank methods such as Ranking SVM, RankBoost, and RankNet. First, AdaRank can incorporate any performance measure, provided that the measure is query based and in the range of [−1, +1]. Notice that the major IR measures meet this requirement. In contrast the existing methods only minimize loss functions that are loosely related to the IR measures [16]. Second, the learning process of AdaRank is more efficient than those of the existing learning algorithms. The time complexity of AdaRank is of order O((k+T)·m·n log n), where k denotes the number of features, T the number of rounds, m the number of queries in training data, and n is the maximum number of documents for queries in training data. The time complexity of RankBoost, for example, is of order O(T · m · n2 ) [8]. Third, AdaRank employs a more reasonable framework for performing the ranking task than the existing methods. Specifically in AdaRank the instances correspond to queries, while in the existing methods the instances correspond to document pairs. As a result, AdaRank does not have the following shortcomings that plague the existing methods. (a) The existing methods have to make a strong assumption that the document pairs from the same query are independently distributed. In reality, this is clearly not the case and this problem does not exist for AdaRank. (b) Ranking the most relevant documents on the tops of document lists is crucial for document retrieval. The existing methods cannot focus on the training on the tops, as indicated in [4]. Several methods for rectifying the problem have been proposed (e.g., [4]), however, they do not seem to fundamentally solve the problem. In contrast, AdaRank can naturally focus on training on the tops of document lists, because the performance measures used favor rankings for which relevant documents are on the tops. (c) In the existing methods, the numbers of document pairs vary from query to query, resulting in creating models biased toward queries with more document pairs, as pointed out in [4]. AdaRank does not have this drawback, because it treats queries rather than document pairs as basic units in learning. 3.5 Differences from AdaBoost AdaRank is a boosting algorithm. In that sense, it is similar to AdaBoost, but it also has several striking differences from AdaBoost. First, the types of instances are different. AdaRank makes use of queries and their corresponding document lists as instances. The labels in training data are lists of ranks (relevance levels). AdaBoost makes use of feature vectors as instances. The labels in training data are simply +1 and −1. Second, the performance measures are different. In AdaRank, the performance measure is a generic measure, defined on the document list and the rank list of a query. In AdaBoost the corresponding performance measure is a specific measure for binary classification, also referred to as ‘margin" [25]. Third, the ways of updating weights are also different. In AdaBoost, the distribution of weights on training instances is calculated according to the current distribution and the performance of the current weak learner. In AdaRank, in contrast, it is calculated according to the performance of the ranking model created so far, as shown in Figure 1. Note that AdaBoost can also adopt the weight updating method used in AdaRank. For AdaBoost they are equivalent (cf., [12] page 305). However, this is not true for AdaRank. 3.6 Construction of Weak Ranker We consider an efficient implementation for weak ranker construction, which is also used in our experiments. In the implementation, as weak ranker we choose the feature that has the optimal weighted performance among all of the features: max k m i=1 Pt(i)E(π(qi, di, xk), yi). Creating weak rankers in this way, the learning process turns out to be that of repeatedly selecting features and linearly combining the selected features. Note that features which are not selected in the training phase will have a weight of zero. 4. EXPERIMENTAL RESULTS We conducted experiments to test the performances of AdaRank using four benchmark datasets: OHSUMED, WSJ, AP, and .Gov. Table 2: Features used in the experiments on OHSUMED, WSJ, and AP datasets. C(w, d) represents frequency of word w in document d; C represents the entire collection; n denotes number of terms in query; | · | denotes the size function; and id f(·) denotes inverse document frequency. 1 wi∈q d ln(c(wi, d) + 1) 2 wi∈q d ln( |C| c(wi,C) + 1) 3 wi∈q d ln(id f(wi)) 4 wi∈q d ln(c(wi,d) |d| + 1) 5 wi∈q d ln(c(wi,d) |d| · id f(wi) + 1) 6 wi∈q d ln(c(wi,d)·|C| |d|·c(wi,C) + 1) 7 ln(BM25 score) 0.2 0.3 0.4 0.5 0.6 MAP NDCG@1 NDCG@3 NDCG@5 NDCG@10 BM25 Ranking SVM RarnkBoost AdaRank.MAP AdaRank.NDCG Figure 2: Ranking accuracies on OHSUMED data. 4.1 Experiment Setting Ranking SVM [13, 16] and RankBoost [8] were selected as baselines in the experiments, because they are the state-of-the-art learning to rank methods. Furthermore, BM25 [24] was used as a baseline, representing the state-of-the-arts IR method (we actually used the tool Lemur1 ). For AdaRank, the parameter T was determined automatically during each experiment. Specifically, when there is no improvement in ranking accuracy in terms of the performance measure, the iteration stops (and T is determined). As the measure E, MAP and NDCG@5 were utilized. The results for AdaRank using MAP and NDCG@5 as measures in training are represented as AdaRank.MAP and AdaRank.NDCG, respectively. 4.2 Experiment with OHSUMED Data In this experiment, we made use of the OHSUMED dataset [14] to test the performances of AdaRank. The OHSUMED dataset consists of 348,566 documents and 106 queries. There are in total 16,140 query-document pairs upon which relevance judgments are made. The relevance judgments are either ‘d" (definitely relevant), ‘p" (possibly relevant), or ‘n"(not relevant). The data have been used in many experiments in IR, for example [4, 29]. As features, we adopted those used in document retrieval [4]. Table 2 shows the features. For example, tf (term frequency), idf (inverse document frequency), dl (document length), and combinations of them are defined as features. BM25 score itself is also a feature. Stop words were removed and stemming was conducted in the data. We randomly divided queries into four even subsets and conducted 4-fold cross-validation experiments. We tuned the parameters for BM25 during one of the trials and applied them to the other trials. The results reported in Figure 2 are those averaged over four trials. In MAP calculation, we define the rank ‘d" as relevant and 1 http://www.lemurproject.com Table 3: Statistics on WSJ and AP datasets. Dataset # queries # retrieved docs # docs per query AP 116 24,727 213.16 WSJ 126 40,230 319.29 0.40 0.45 0.50 0.55 0.60 MAP NDCG@1 NDCG@3 NDCG@5 NDCG@10 BM25 Ranking SVM RankBoost AdaRank.MAP AdaRank.NDCG Figure 3: Ranking accuracies on WSJ dataset. the other two ranks as irrelevant. From Figure 2, we see that both AdaRank.MAP and AdaRank.NDCG outperform BM25, Ranking SVM, and RankBoost in terms of all measures. We conducted significant tests (t-test) on the improvements of AdaRank.MAP over BM25, Ranking SVM, and RankBoost in terms of MAP. The results indicate that all the improvements are statistically significant (p-value < 0.05). We also conducted t-test on the improvements of AdaRank.NDCG over BM25, Ranking SVM, and RankBoost in terms of NDCG@5. The improvements are also statistically significant. 4.3 Experiment with WSJ and AP Data In this experiment, we made use of the WSJ and AP datasets from the TREC ad-hoc retrieval track, to test the performances of AdaRank. WSJ contains 74,520 articles of Wall Street Journals from 1990 to 1992, and AP contains 158,240 articles of Associated Press in 1988 and 1990. 200 queries are selected from the TREC topics (No.101 ∼ No.300). Each query has a number of documents associated and they are labeled as ‘relevant" or ‘irrelevant" (to the query). Following the practice in [28], the queries that have less than 10 relevant documents were discarded. Table 3 shows the statistics on the two datasets. In the same way as in section 4.2, we adopted the features listed in Table 2 for ranking. We also conducted 4-fold cross-validation experiments. The results reported in Figure 3 and 4 are those averaged over four trials on WSJ and AP datasets, respectively. From Figure 3 and 4, we can see that AdaRank.MAP and AdaRank.NDCG outperform BM25, Ranking SVM, and RankBoost in terms of all measures on both WSJ and AP. We conducted t-tests on the improvements of AdaRank.MAP and AdaRank.NDCG over BM25, Ranking SVM, and RankBoost on WSJ and AP. The results indicate that all the improvements in terms of MAP are statistically significant (p-value < 0.05). However only some of the improvements in terms of NDCG@5 are statistically significant, although overall the improvements on NDCG scores are quite high (1-2 points). 4.4 Experiment with .Gov Data In this experiment, we further made use of the TREC .Gov data to test the performance of AdaRank for the task of web retrieval. The corpus is a crawl from the .gov domain in early 2002, and has been used at TREC Web Track since 2002. There are a total 0.40 0.45 0.50 0.55 MAP NDCG@1 NDCG@3 NDCG@5 NDCG@10 BM25 Ranking SVM RankBoost AdaRank.MAP AdaRank.NDCG Figure 4: Ranking accuracies on AP dataset. 0.1 0.2 0.3 0.4 0.5 0.6 0.7 MAP NDCG@1 NDCG@3 NDCG@5 NDCG@10 BM25 Ranking SVM RankBoost AdaRank.MAP AdaRank.NDCG Figure 5: Ranking accuracies on .Gov dataset. Table 4: Features used in the experiments on .Gov dataset. 1 BM25 [24] 2 MSRA1000 [27] 3 PageRank [21] 4 HostRank [30] 5 Relevance Propagation [23] (10 features) of 1,053,110 web pages with 11,164,829 hyperlinks in the data. The 50 queries in the topic distillation task in the Web Track of TREC 2003 [6] were used. The ground truths for the queries are provided by the TREC committee with binary judgment: relevant or irrelevant. The number of relevant pages vary from query to query (from 1 to 86). We extracted 14 features from each query-document pair. Table 4 gives a list of the features. They are the outputs of some well-known algorithms (systems). These features are different from those in Table 2, because the task is different. Again, we conducted 4-fold cross-validation experiments. The results averaged over four trials are reported in Figure 5. From the results, we can see that AdaRank.MAP and AdaRank.NDCG outperform all the baselines in terms of all measures. We conducted ttests on the improvements of AdaRank.MAP and AdaRank.NDCG over BM25, Ranking SVM, and RankBoost. Some of the improvements are not statistically significant. This is because we have only 50 queries used in the experiments, and the number of queries is too small. 4.5 Discussions We investigated the reasons that AdaRank outperforms the baseline methods, using the results of the OHSUMED dataset as examples. First, we examined the reason that AdaRank has higher performances than Ranking SVM and RankBoost. Specifically we com0.58 0.60 0.62 0.64 0.66 0.68 d-n d-p p-n accuracy pair type Ranking SVM RankBoost AdaRank.MAP AdaRank.NDCG Figure 6: Accuracy on ranking document pairs with OHSUMED dataset. 0 2 4 6 8 10 12 numberofqueries number of document pairs per query Figure 7: Distribution of queries with different number of document pairs in training data of trial 1. pared the error rates between different rank pairs made by Ranking SVM, RankBoost, AdaRank.MAP, and AdaRank.NDCG on the test data. The results averaged over four trials in the 4-fold cross validation are shown in Figure 6. We use ‘d-n" to stand for the pairs between ‘definitely relevant" and ‘not relevant", ‘d-p" the pairs between ‘definitely relevant" and ‘partially relevant", and ‘p-n" the pairs between ‘partially relevant" and ‘not relevant". From Figure 6, we can see that AdaRank.MAP and AdaRank.NDCG make fewer errors for ‘d-n" and ‘d-p", which are related to the tops of rankings and are important. This is because AdaRank.MAP and AdaRank.NDCG can naturally focus upon the training on the tops by optimizing MAP and NDCG@5, respectively. We also made statistics on the number of document pairs per query in the training data (for trial 1). The queries are clustered into different groups based on the the number of their associated document pairs. Figure 7 shows the distribution of the query groups. In the figure, for example, ‘0-1k" is the group of queries whose number of document pairs are between 0 and 999. We can see that the numbers of document pairs really vary from query to query. Next we evaluated the accuracies of AdaRank.MAP and RankBoost in terms of MAP for each of the query group. The results are reported in Figure 8. We found that the average MAP of AdaRank.MAP over the groups is two points higher than RankBoost. Furthermore, it is interesting to see that AdaRank.MAP performs particularly better than RankBoost for queries with small numbers of document pairs (e.g., ‘0-1k", ‘1k-2k", and ‘2k-3k"). The results indicate that AdaRank.MAP can effectively avoid creating a model biased towards queries with more document pairs. For AdaRank.NDCG, similar results can be observed. 0.2 0.3 0.4 0.5 MAP query group RankBoost AdaRank.MAP Figure 8: Differences in MAP for different query groups. 0.30 0.31 0.32 0.33 0.34 trial 1 trial 2 trial 3 trial 4 MAP AdaRank.MAP AdaRank.NDCG Figure 9: MAP on training set when model is trained with MAP or NDCG@5. We further conducted an experiment to see whether AdaRank has the ability to improve the ranking accuracy in terms of a measure by using the measure in training. Specifically, we trained ranking models using AdaRank.MAP and AdaRank.NDCG and evaluated their accuracies on the training dataset in terms of both MAP and NDCG@5. The experiment was conducted for each trial. Figure 9 and Figure 10 show the results in terms of MAP and NDCG@5, respectively. We can see that, AdaRank.MAP trained with MAP performs better in terms of MAP while AdaRank.NDCG trained with NDCG@5 performs better in terms of NDCG@5. The results indicate that AdaRank can indeed enhance ranking performance in terms of a measure by using the measure in training. Finally, we tried to verify the correctness of Theorem 1. That is, the ranking accuracy in terms of the performance measure can be continuously improved, as long as e−δt min 1 − ϕ(t)2 < 1 holds. As an example, Figure 11 shows the learning curve of AdaRank.MAP in terms of MAP during the training phase in one trial of the cross validation. From the figure, we can see that the ranking accuracy of AdaRank.MAP steadily improves, as the training goes on, until it reaches to the peak. The result agrees well with Theorem 1. 5. CONCLUSION AND FUTURE WORK In this paper we have proposed a novel algorithm for learning ranking models in document retrieval, referred to as AdaRank. In contrast to existing methods, AdaRank optimizes a loss function that is directly defined on the performance measures. It employs a boosting technique in ranking model learning. AdaRank offers several advantages: ease of implementation, theoretical soundness, efficiency in training, and high accuracy in ranking. Experimental results based on four benchmark datasets show that AdaRank can significantly outperform the baseline methods of BM25, Ranking SVM, and RankBoost. 0.49 0.50 0.51 0.52 0.53 trial 1 trial 2 trial 3 trial 4 NDCG@5 AdaRank.MAP AdaRank.NDCG Figure 10: NDCG@5 on training set when model is trained with MAP or NDCG@5. 0.29 0.30 0.31 0.32 0 50 100 150 200 250 300 350 MAP number of rounds Figure 11: Learning curve of AdaRank. Future work includes theoretical analysis on the generalization error and other properties of the AdaRank algorithm, and further empirical evaluations of the algorithm including comparisons with other algorithms that can directly optimize performance measures. 6. ACKNOWLEDGMENTS We thank Harry Shum, Wei-Ying Ma, Tie-Yan Liu, Gu Xu, Bin Gao, Robert Schapire, and Andrew Arnold for their valuable comments and suggestions to this paper. 7. REFERENCES [1] R. Baeza-Yates and B. Ribeiro-Neto. Modern Information Retrieval. Addison Wesley, May 1999. [2] C. Burges, R. Ragno, and Q. Le. Learning to rank with nonsmooth cost functions. In Advances in Neural Information Processing Systems 18, pages 395-402. MIT Press, Cambridge, MA, 2006. [3] C. Burges, T. Shaked, E. Renshaw, A. Lazier, M. Deeds, N. Hamilton, and G. Hullender. Learning to rank using gradient descent. In ICML 22, pages 89-96, 2005. [4] Y. Cao, J. Xu, T.-Y. Liu, H. Li, Y. Huang, and H.-W. Hon. Adapting ranking SVM to document retrieval. In SIGIR 29, pages 186-193, 2006. [5] D. Cossock and T. Zhang. Subset ranking using regression. In COLT, pages 605-619, 2006. [6] N. Craswell, D. Hawking, R. Wilkinson, and M. Wu. Overview of the TREC 2003 web track. In TREC, pages 78-92, 2003. [7] N. Duffy and D. Helmbold. Boosting methods for regression. Mach. Learn., 47(2-3):153-200, 2002. [8] Y. Freund, R. D. Iyer, R. E. Schapire, and Y. Singer. An efficient boosting algorithm for combining preferences. Journal of Machine Learning Research, 4:933-969, 2003. [9] Y. Freund and R. E. Schapire. A decision-theoretic generalization of on-line learning and an application to boosting. J. Comput. Syst. Sci., 55(1):119-139, 1997. [10] J. Friedman, T. Hastie, and R. Tibshirani. Additive logistic regression: A statistical view of boosting. The Annals of Statistics, 28(2):337-374, 2000. [11] G. Fung, R. Rosales, and B. Krishnapuram. Learning rankings via convex hull separation. In Advances in Neural Information Processing Systems 18, pages 395-402. MIT Press, Cambridge, MA, 2006. [12] T. Hastie, R. Tibshirani, and J. H. Friedman. The Elements of Statistical Learning. Springer, August 2001. [13] R. Herbrich, T. Graepel, and K. Obermayer. Large Margin rank boundaries for ordinal regression. MIT Press, Cambridge, MA, 2000. [14] W. Hersh, C. Buckley, T. J. Leone, and D. Hickam. Ohsumed: an interactive retrieval evaluation and new large test collection for research. In SIGIR, pages 192-201, 1994. [15] K. Jarvelin and J. Kekalainen. IR evaluation methods for retrieving highly relevant documents. In SIGIR 23, pages 41-48, 2000. [16] T. Joachims. Optimizing search engines using clickthrough data. In SIGKDD 8, pages 133-142, 2002. [17] T. Joachims. A support vector method for multivariate performance measures. In ICML 22, pages 377-384, 2005. [18] J. Lafferty and C. Zhai. Document language models, query models, and risk minimization for information retrieval. In SIGIR 24, pages 111-119, 2001. [19] D. A. Metzler, W. B. Croft, and A. McCallum. Direct maximization of rank-based metrics for information retrieval. Technical report, CIIR, 2005. [20] R. Nallapati. Discriminative models for information retrieval. In SIGIR 27, pages 64-71, 2004. [21] L. Page, S. Brin, R. Motwani, and T. Winograd. The pagerank citation ranking: Bringing order to the web. Technical report, Stanford Digital Library Technologies Project, 1998. [22] J. M. Ponte and W. B. Croft. A language modeling approach to information retrieval. In SIGIR 21, pages 275-281, 1998. [23] T. Qin, T.-Y. Liu, X.-D. Zhang, Z. Chen, and W.-Y. Ma. A study of relevance propagation for web search. In SIGIR 28, pages 408-415, 2005. [24] S. E. Robertson and D. A. Hull. The TREC-9 filtering track final report. In TREC, pages 25-40, 2000. [25] R. E. Schapire, Y. Freund, P. Barlett, and W. S. Lee. Boosting the margin: A new explanation for the effectiveness of voting methods. In ICML 14, pages 322-330, 1997. [26] R. E. Schapire and Y. Singer. Improved boosting algorithms using confidence-rated predictions. Mach. Learn., 37(3):297-336, 1999. [27] R. Song, J. Wen, S. Shi, G. Xin, T. yan Liu, T. Qin, X. Zheng, J. Zhang, G. Xue, and W.-Y. Ma. Microsoft Research Asia at web track and terabyte track of TREC 2004. In TREC, 2004. [28] A. Trotman. Learning to rank. Inf. Retr., 8(3):359-381, 2005. [29] J. Xu, Y. Cao, H. Li, and Y. Huang. Cost-sensitive learning of SVM for ranking. In ECML, pages 833-840, 2006. [30] G.-R. Xue, Q. Yang, H.-J. Zeng, Y. Yu, and Z. Chen. Exploiting the hierarchical structure for link analysis. In SIGIR 28, pages 186-193, 2005. [31] H. Yu. SVM selective sampling for ranking with application to data retrieval. In SIGKDD 11, pages 354-363, 2005. APPENDIX Here we give the proof of Theorem 1. P. Set ZT = m i=1 exp {−E(π(qi, di, fT ), yi)} and φ(t) = 1 2 (1 + ϕ(t)). According to the definition of αt, we know that eαt = φ(t) 1−φ(t) . ZT = m i=1 exp {−E(π(qi, di, fT−1 + αT hT ), yi)} = m i=1 exp −E(π(qi, di, fT−1), yi) − αT E(π(qi, di, hT ), yi) − δT i ≤ m i=1 exp {−E(π(qi, di, fT−1), yi)} exp {−αT E(π(qi, di, hT ), yi)} e−δT min = e−δT min ZT−1 m i=1 exp {−E(π(qi, di, fT−1), yi)} ZT−1 exp{−αT E(π(qi, di, hT ), yi)} = e−δT min ZT−1 m i=1 PT (i) exp{−αT E(π(qi, di, hT ), yi)}. Moreover, if E(π(qi, di, hT ), yi) ∈ [−1, +1] then, ZT ≤ e−δT minZT−1 m i=1 PT (i) 1+E(π(qi, di, hT ), yi) 2 e−αT + 1−E(π(qi, di, hT ), yi) 2 eαT = e−δT min ZT−1  φ(T) 1 − φ(T) φ(T) + (1 − φ(T)) φ(T) 1 − φ(T)   = ZT−1e−δT min 4φ(T)(1 − φ(T)) ≤ ZT−2 T t=T−1 e−δt min 4φ(t)(1 − φ(t)) ≤ Z1 T t=2 e−δt min 4φ(t)(1 − φ(t)) = m m i=1 1 m exp{−E(π(qi, di, α1h1), yi)} T t=2 e−δt min 4φ(t)(1 − φ(t)) = m m i=1 1 m exp{−α1E(π(qi, di, h1), yi) − δ1 i } T t=2 e−δt min 4φ(t)(1 − φ(t)) ≤ me−δ1 min m i=1 1 m exp{−α1E(π(qi, di, h1), yi)} T t=2 e−δt min 4φ(t)(1 − φ(t)) ≤ m e−δ1 min 4φ(1)(1 − φ(1)) T t=2 e−δt min 4φ(t)(1 − φ(t)) = m T t=1 e−δt min 1 − ϕ(t)2. ∴ 1 m m i=1 E(π(qi, di, fT ), yi) ≥ 1 m m i=1 {1 − exp(−E(π(qi, di, fT ), yi))} ≥ 1 − T t=1 e−δt min 1 − ϕ(t)2.

support vector machine;trained ranking model;rankboost;weak ranker;ranking model tuning;ranking model;boost;novel learning algorithm;learn to rank;training process;machine learning;new learning algorithm;document retrieval;information retrieval;re-weighted training datum

train_H-37

Relaxed Online SVMs for Spam Filtering

Spam is a key problem in electronic communication, including large-scale email systems and the growing number of blogs. Content-based filtering is one reliable method of combating this threat in its various forms, but some academic researchers and industrial practitioners disagree on how best to filter spam. The former have advocated the use of Support Vector Machines (SVMs) for content-based filtering, as this machine learning methodology gives state-of-the-art performance for text classification. However, similar performance gains have yet to be demonstrated for online spam filtering. Additionally, practitioners cite the high cost of SVMs as reason to prefer faster (if less statistically robust) Bayesian methods. In this paper, we offer a resolution to this controversy. First, we show that online SVMs indeed give state-of-the-art classification performance on online spam filtering on large benchmark data sets. Second, we show that nearly equivalent performance may be achieved by a Relaxed Online SVM (ROSVM) at greatly reduced computational cost. Our results are experimentally verified on email spam, blog spam, and splog detection tasks.

1. INTRODUCTION Electronic communication is increasingly plagued by unwanted or harmful content known as spam. The most well known form of spam is email spam, which remains a major problem for large email systems. Other forms of spam are also becoming problematic, including blog spam, in which spammers post unwanted comments in blogs [21], and splogs, which are fake blogs constructed to enable link spam with the hope of boosting the measured importance of a given webpage in the eyes of automated search engines [17]. There are a variety of methods for identifying these many forms of spam, including compiling blacklists of known spammers, and conducting link analysis. The approach of content analysis has shown particular promise and generality for combating spam. In content analysis, the actual message text (often including hyper-text and meta-text, such as HTML and headers) is analyzed using machine learning techniques for text classification to determine if the given content is spam. Content analysis has been widely applied in detecting email spam [11], and has also been used for identifying blog spam [21] and splogs [17]. In this paper, we do not explore the related problem of link spam, which is currently best combated by link analysis [13]. 1.1 An Anti-Spam Controversy The anti-spam community has been divided on the choice of the best machine learning method for content-based spam detection. Academic researchers have tended to favor the use of Support Vector Machines (SVMs), a statistically robust machine learning method [7] which yields state-of-theart performance on general text classification [14]. However, SVMs typically require training time that is quadratic in the number of training examples, and are impractical for largescale email systems. Practitioners requiring content-based spam filtering have typically chosen to use the faster (if less statistically robust) machine learning method of Naive Bayes text classification [11, 12, 20]. This Bayesian method requires only linear training time, and is easily implemented in an online setting with incremental updates. This allows a deployed system to easily adapt to a changing environment over time. Other fast methods for spam filtering include compression models [1] and logistic regression [10]. It has not yet been empirically demonstrated that SVMs give improved performance over these methods in an online spam detection setting [4]. 1.2 Contributions In this paper, we address the anti-spam controversy and offer a potential resolution. We first demonstrate that online SVMs do indeed provide state-of-the-art spam detection through empirical tests on several large benchmark data sets of email spam. We then analyze the effect of the tradeoff parameter in the SVM objective function, which shows that the expensive SVM methodology may, in fact, be overkill for spam detection. We reduce the computational cost of SVM learning by relaxing this requirement on the maximum margin in online settings, and create a Relaxed Online SVM, ROSVM, appropriate for high performance content-based spam filtering in large-scale settings. 2. SPAM AND ONLINE SVMS The controversy between academics and practitioners in spam filtering centers on the use of SVMs. The former advocate their use, but have yet to demonstrate strong performance with SVMs on online spam filtering. Indeed, the results of [4] show that, when used with default parameters, SVMs actually perform worse than other methods. In this section, we review the basic workings of SVMs and describe a simple Online SVM algorithm. We then show that Online SVMs indeed achieve state-of-the-art performance on filtering email spam, blog comment spam, and splogs, so long as the tradeoff parameter C is set to a high value. However, the cost of Online SVMs turns out to be prohibitive for largescale applications. These findings motivate our proposal of Relaxed Online SVMs in the following section. 2.1 Background: SVMs SVMs are a robust machine learning methodology which has been shown to yield state-of-the-art performance on text classification [14]. by finding a hyperplane that separates two classes of data in data space while maximizing the margin between them. We use the following notation to describe SVMs, which draws from [23]. A data set X contains n labeled example vectors {(x1, y1) . . . (xn, yn)}, where each xi is a vector containing features describing example i, and each yi is the class label for that example. In spam detection, the classes spam and ham (i.e., not spam) are assigned the numerical class labels +1 and −1, respectively. The linear SVMs we employ in this paper use a hypothesis vector w and bias term b to classify a new example x, by generating a predicted class label f(x): f(x) = sign(< w, x > +b) SVMs find the hypothesis w, which defines the separating hyperplane, by minimizing the following objective function over all n training examples: τ(w, ξ) = 1 2 ||w||2 + C nX i=i ξi under the constraints that ∀i = {1..n} : yi(< w, xi > +b) ≥ 1 − ξi, ξi ≥ 0 In this objective function, each slack variable ξi shows the amount of error that the classifier makes on a given example xi. Minimizing the sum of the slack variables corresponds to minimizing the loss function on the training data, while minimizing the term 1 2 ||w||2 corresponds to maximizing the margin between the two classes [23]. These two optimization goals are often in conflict; the tradeoff parameter C determines how much importance to give each of these tasks. Linear SVMs exploit data sparsity to classify a new instance in O(s) time, where s is the number of non-zero features. This is the same classification time as other linear Given: data set X = (x1, y1), . . . , (xn, yn), C, m: Initialize w := 0, b := 0, seenData := { } For Each xi ∈ X do: Classify xi using f(xi) = sign(< w, xi > +b) IF yif(xi) < 1 Find w , b using SMO on seenData, using w, b as seed hypothesis. Add xi to seenData done Figure 1: Pseudo code for Online SVM. classifiers, and as Naive Bayesian classification. Training SVMs, however, typically takes O(n2 ) time, for n training examples. A variant for linear SVMs was recently proposed which trains in O(ns) time [15], but because this method has a high constant, we do not explore it here. 2.2 Online SVMs In many traditional machine learning applications, SVMs are applied in batch mode. That is, an SVM is trained on an entire set of training data, and is then tested on a separate set of testing data. Spam filtering is typically tested and deployed in an online setting, which proceeds incrementally. Here, the learner classifies a new example, is told if its prediction is correct, updates its hypothesis accordingly, and then awaits a new example. Online learning allows a deployed system to adapt itself in a changing environment. Re-training an SVM from scratch on the entire set of previously seen data for each new example is cost prohibitive. However, using an old hypothesis as the starting point for re-training reduces this cost considerably. One method of incremental and decremental SVM learning was proposed in [2]. Because we are only concerned with incremental learning, we apply a simpler algorithm for converting a batch SVM learner into an online SVM (see Figure 1 for pseudocode), which is similar to the approach of [16]. Each time the Online SVM encounters an example that was poorly classified, it retrains using the old hypothesis as a starting point. Note that due to the Karush-Kuhn-Tucker (KKT) conditions, it is not necessary to re-train on wellclassified examples that are outside the margins [23]. We used Platt"s SMO algorithm [22] as a core SVM solver, because it is an iterative method that is well suited to converge quickly from a good initial hypothesis. Because previous work (and our own initial testing) indicates that binary feature values give the best results for spam filtering [20, 9], we optimized our implementation of the Online SMO to exploit fast inner-products with binary vectors. 1 2.3 Feature Mapping Spam Content Extracting machine learning features from text may be done in a variety of ways, especially when that text may include hyper-content and meta-content such as HTML and header information. However, previous research has shown that simple methods from text classification, such as bag of words vectors, and overlapping character-level n-grams, can achieve strong results [9]. Formally, a bag of words vector is a vector x with a unique dimension for each possible 1 Our source code is freely available at www.cs.tufts.edu/∼dsculley/onlineSMO. 1 0.999 0.995 0.99 0.985 0.98 0.1 1 10 100 1000 ROCArea C 2-grams 3-grams 4-grams words Figure 2: Tuning the Tradeoff Parameter C. Tests were conducted with Online SMO, using binary feature vectors, on the spamassassin data set of 6034 examples. Graph plots C versus Area under the ROC curve. word, defined as a contiguous substring of non-whitespace characters. An n-gram vector is a vector x with a unique dimension for each possible substring of n total characters. Note that n-grams may include whitespace, and are overlapping. We use binary feature scoring, which has been shown to be most effective for a variety of spam detection methods [20, 9]. We normalize the vectors with the Euclidean norm. Furthermore, with email data, we reduce the impact of long messages (for example, with attachments) by considering only the first 3,000 characters of each string. For blog comments and splogs, we consider the whole text, including any meta-data such as HTML tags, as given. No other feature selection or domain knowledge was used. 2.4 Tuning the Tradeoff Parameter, C The SVM tradeoff parameter C must be tuned to balance the (potentially conflicting) goals of maximizing the margin and minimizing the training error. Early work on SVM based spam detection [9] showed that high values of C give best performance with binary features. Later work has not always followed this lead: a (low) default setting of C was used on splog detection [17], and also on email spam [4]. Following standard machine learning practice, we tuned C on separate tuning data not used for later testing. We used the publicly available spamassassin email spam data set, and created an online learning task by randomly interleaving all 6034 labeled messages to create a single ordered set. For tuning, we performed a coarse parameter search for C using powers of ten from .0001 to 10000. We used the Online SVM described above, and tested both binary bag of words vectors and n-gram vectors with n = {2, 3, 4}. We used the first 3000 characters of each message, which included header information, body of the email, and possibly attachments. Following the recommendation of [6], we use Area under the ROC curve as our evaluation measure. The results (see Figure 2) agree with [9]: there is a plateau of high performance achieved with all values of C ≥ 10, and performance degrades sharply with C < 1. For the remainder of our experiments with SVMs in this paper, we set C = 100. We will return to the observation that very high values of C do not degrade performance as support for the intuition that relaxed SVMs should perform well on spam. Table 1: Results for Email Spam filtering with Online SVM on benchmark data sets. Score reported is (1-ROCA)%, where 0 is optimal. trec05p-1 trec06p OnSVM: words 0.015 (.011-.022) 0.034 (.025-.046) 3-grams 0.011 (.009-.015) 0.025 (.017-.035) 4-grams 0.008 (.007-.011) 0.023 (.017-.032) SpamProbe 0.059 (.049-.071) 0.092 (.078-.110) BogoFilter 0.048 (.038-.062) 0.077 (.056-.105) TREC Winners 0.019 (.015-.023) 0.054 (.034-.085) 53-Ensemble 0.007 (.005-.008) 0.020 (.007-.050) Table 2: Results for Blog Comment Spam Detection using SVMs and Leave One Out Cross Validation. We report the same performance measures as in the prior work for meaningful comparison. accuracy precision recall SVM C = 100: words 0.931 0.946 0.954 3-grams 0.951 0.963 0.965 4-grams 0.949 0.967 0.956 Prior best method 0.83 0.874 0.874 2.5 Email Spam and Online SVMs With C tuned on a separate tuning set, we then tested the performance of Online SVMs in spam detection. We used two large benchmark data sets of email spam as our test corpora. These data sets are the 2005 TREC public data set trec05p-1 of 92,189 messages, and the 2006 TREC public data sets, trec06p, containing 37,822 messages in English. (We do not report our strong results on the trec06c corpus of Chinese messages as there have been questions raised over the validity of this test set.) We used the canonical ordering provided with each of these data sets for fair comparison. Results for these experiments, with bag of words vectors and and n-gram vectors appear in Table 1. To compare our results with previous scores on these data sets, we use the same (1-ROCA)% measure described in [6], which is one minus the area under the ROC curve, expressed as a percent. This measure shows the percent chance of error made by a classifier asserting that one message is more likely to be spam than another. These results show that Online SVMs do give state of the art performance on email spam. The only known system that out-performs the Online SVMs on the trec05p-1 data set is a recent ensemble classifier which combines the results of 53 unique spam filters [19]. To our knowledge, the Online SVM has out-performed every other single filter on these data sets, including those using Bayesian methods [5, 3], compression models [5, 3], logistic regression [10], and perceptron variants [3], the TREC competition winners [5, 3], and open source email spam filters BogoFilter v1.1.5 and SpamProbe v1.4d. 2.6 Blog Comment Spam and SVMs Blog comment spam is similar to email spam in many regards, and content-based methods have been proposed for detecting these spam comments [21]. However, large benchmark data sets of labeled blog comment spam do not yet exist. Thus, we run experiments on the only publicly available data set we know of, which was used in content-based blog Table 3: Results for Splog vs. Blog Detection using SVMs and Leave One Out Cross Validation. We report the same evaluation measures as in the prior work for meaningful comparison. features precision recall F1 SVM C = 100: words 0.921 0.870 0.895 3-grams 0.904 0.866 0.885 4-grams 0.928 0.876 0.901 Prior SVM with: words 0.887 0.864 0.875 4-grams 0.867 0.844 0.855 words+urls 0.893 0.869 0.881 comment spam detection experiments by [21]. Because of the small size of the data set, and because prior researchers did not conduct their experiments in an on-line setting, we test the performance of linear SVMs using leave-one-out cross validation, with SVM-Light, a standard open-source SVM implementation [14]. We use the parameter setting C = 100, with the same feature space mappings as above. We report accuracy, precision, and recall to compare these to the results given on the same data set by [21]. These results (see Table 2) show that SVMs give superior performance on this data set to the prior methodology. 2.7 Splogs and SVMs As with blog comment spam, there is not yet a large, publicly available benchmark corpus of labeled splog detection test data. However, the authors of [17] kindly provided us with the labeled data set of 1,389 blogs and splogs that they used to test content-based splog detection using SVMs. The only difference between our methodology and that of [17] is that they used default parameters for C, which SVM-Light sets to 1 avg||x||2 . (For normalized vectors, this default value sets C = 1.) They also tested several domain-informed feature mappings, such as giving special features to url tags. For our experiments, we used the same feature mappings as above, and tested the effect of setting C = 100. As with the methodology of [17], we performed leave one out cross validation for apples-to-apples comparison on this data. The results (see Table 3) show that a high value of C produces higher performance for the same feature space mappings, and even enables the simple 4-gram mapping to out-perform the previous best mapping which incorporated domain knowledge by using words and urls. 2.8 Computational Cost The results presented in this section demonstrate that linfeatures trec06p trec05p-1 words 12196s 66478s 3-grams 44605s 128924s 4-grams 87519s 242160s corpus size 32822 92189 Table 4: Execution time for Online SVMs with email spam detection, in CPU seconds. These times do not include the time spent mapping strings to feature vectors. The number of examples in each data set is given in the last row as corpus size. A B Figure 3: Visualizing the effect of C. Hyperplane A maximizes the margin while accepting a small amount of training error. This corresponds to setting C to a low value. Hyperplane B accepts a smaller margin in order to reduce training error. This corresponds to setting C to a high value. Content-based spam filtering appears to do best with high values of C. ear SVMs give state of the art performance on content-based spam filtering. However, this performance comes at a price. Although the blog comment spam and splog data sets are too small for the quadratic training time of SVMs to appear problematic, the email data sets are large enough to illustrate the problems of quadratic training cost. Table 4 shows computation time versus data set size for each of the online learning tasks (on same system). The training cost of SVMs are prohibitive for large-scale content based spam detection, or a large blog host. In the following section, we reduce this cost by relaxing the expensive requirements of SVMs. 3. RELAXED ONLINE SVMS (ROSVM) One of the main benefits of SVMs is that they find a decision hyperplane that maximizes the margin between classes in the data space. Maximizing the margin is expensive, typically requiring quadratic training time in the number of training examples. However, as we saw in the previous section, the task of content-based spam detection is best achieved by SVMs with a high value of C. Setting C to a high value for this domain implies that minimizing training loss is more important than maximizing the margin (see Figure 3). Thus, while SVMs do create high performance spam filters, applying them in practice is overkill. The full margin maximization feature that they provide is unnecessary, and relaxing this requirement can reduce computational cost. We propose three ways to relax Online SVMs: • Reduce the size of the optimization problem by only optimizing over the last p examples. • Reduce the number of training updates by only training on actual errors. • Reduce the number of iterations in the iterative SVM Given: dataset X = (x1, y1), . . . , (xn, yn), C, m, p: Initialize w := 0, b := 0, seenData := { } For Each xi ∈ X do: Classify xi using f(xi) = sign(< w, xi > +b) If yif(xi) < m Find w , b with SMO on seenData, using w, b as seed hypothesis. set (w, b) := (w",b") If size(seenData) > p remove oldest example from seenData Add xi to seenData done Figure 4: Pseudo-code for Relaxed Online SVM. solver by allowing an approximate solution to the optimization problem. As we describe in the remainder of this subsection, all of these methods trade statistical robustness for reduced computational cost. Experimental results reported in the following section show that they equal or approach the performance of full Online SVMs on content-based spam detection. 3.1 Reducing Problem Size In the full Online SVMs, we re-optimize over the full set of seen data on every update, which becomes expensive as the number of seen data points grows. We can bound this expense by only considering the p most recent examples for optimization (see Figure 4 for pseudo-code). Note that this is not equivalent to training a new SVM classifier from scratch on the p most recent examples, because each successive optimization problem is seeded with the previous hypothesis w [8]. This hypothesis may contain values for features that do not occur anywhere in the p most recent examples, and these will not be changed. This allows the hypothesis to remember rare (but informative) features that were learned further than p examples in the past. Formally, the optimization problem is now defined most clearly in the dual form [23]. In this case, the original softmargin SVM is computed by maximizing at example n: W (α) = nX i=1 αi − 1 2 nX i,j=1 αiαjyiyj < xi, xj >, subject to the previous constraints [23]: ∀i ∈ {1, . . . , n} : 0 ≤ αi ≤ C and nX i=1 αiyi = 0 To this, we add the additional lookback buffer constraint ∀j ∈ {1, . . . , (n − p)} : αj = cj where cj is a constant, fixed as the last value found for αj while j > (n − p). Thus, the margin found by an optimization is not guaranteed to be one that maximizes the margin for the global data set of examples {x1, . . . , xn)}, but rather one that satisfies a relaxed requirement that the margin be maximized over the examples { x(n−p+1), . . . , xn}, subject to the fixed constraints on the hyperplane that were found in previous optimizations over examples {x1, . . . , x(n−p)}. (For completeness, when p ≥ n, define (n − p) = 1.) This set of constraints reduces the number of free variables in the optimization problem, reducing computational cost. 3.2 Reducing Number of Updates As noted before, the KKT conditions show that a well classified example will not change the hypothesis; thus it is not necessary to re-train when we encounter such an example. Under the KKT conditions, an example xi is considered well-classified when yif(xi) > 1. If we re-train on every example that is not well-classified, our hyperplane will be guaranteed to be optimal at every step. The number of re-training updates can be reduced by relaxing the definition of well classified. An example xi is now considered well classified when yif(xi) > M, for some 0 ≤ M ≤ 1. Here, each update still produces an optimal hyperplane. The learner may encounter an example that lies within the margins, but farther from the margins than M. Such an example means the hypothesis is no longer globally optimal for the data set, but it is considered good enough for continued use without immediate retraining. This update procedure is similar to that used by variants of the Perceptron algorithm [18]. In the extreme case, we can set M = 0, which creates a mistake driven Online SVM. In the experimental section, we show that this version of Online SVMs, which updates only on actual errors, does not significantly degrade performance on content-based spam detection, but does significantly reduce cost. 3.3 Reducing Iterations As an iterative solver, SMO makes repeated passes over the data set to optimize the objective function. SMO has one main loop, which can alternate between passing over the entire data set, or the smaller active set of current support vectors [22]. Successive iterations of this loop bring the hyperplane closer to an optimal value. However, it is possible that these iterations provide less benefit than their expense justifies. That is, a close first approximation may be good enough. We introduce a parameter T to control the maximum number of iterations we allow. As we will see in the experimental section, this parameter can be set as low as 1 with little impact on the quality of results, providing computational savings. 4. EXPERIMENTS In Section 2, we argued that the strong performance on content-based spam detection with SVMs with a high value of C show that the maximum margin criteria is overkill, incurring unnecessary computational cost. In Section 3, we proposed ROSVM to address this issue, as both of these methods trade away guarantees on the maximum margin hyperplane in return for reduced computational cost. In this section, we test these methods on the same benchmark data sets to see if state of the art performance may be achieved by these less costly methods. We find that ROSVM is capable of achieving these high levels of performance with greatly reduced cost. Our main tests on content-based spam detection are performed on large benchmark sets of email data. We then apply these methods on the smaller data sets of blog comment spam and blogs, with similar performance. 4.1 ROSVM Tests In Section 3, we proposed three approaches for reducing the computational cost of Online SMO: reducing the prob0.005 0.01 0.025 0.05 0.1 10 100 1000 10000 100000 (1-ROCA)% Buffer Size trec05p-1 trec06p 0 50000 100000 150000 200000 250000 10 100 1000 10000 100000 CPUSec. Buffer Size trec05p-1 trec06p Figure 5: Reduced Size Tests. lem size, reducing the number of optimization iterations, and reducing the number of training updates. Each of these approaches relax the maximum margin criteria on the global set of previously seen data. Here we test the effect that each of these methods has on both effectiveness and efficiency. In each of these tests, we use the large benchmark email data sets, trec05p-1 and trec06p. 4.1.1 Testing Reduced Size For our first ROSVM test, we experiment on the effect of reducing the size of the optimization problem by only considering the p most recent examples, as described in the previous section. For this test, we use the same 4-gram mappings as for the reference experiments in Section 2, with the same value C = 100. We test a range of values p in a coarse grid search. Figure 5 reports the effect of the buffer size p in relationship to the (1-ROCA)% performance measure (top), and the number of CPU seconds required (bottom). The results show that values of p < 100 do result in degraded performance, although they evaluate very quickly. However, p values from 500 to 10,000 perform almost as well as the original Online SMO (represented here as p = 100, 000), at dramatically reduced computational cost. These results are important for making state of the art performance on large-scale content-based spam detection practical with online SVMs. Ordinarily, the training time would grow quadratically with the number of seen examples. However, fixing a value of p ensures that the training time is independent of the size of the data set. Furthermore, a lookback buffer allows the filter to adjust to concept drift. 0.005 0.01 0.025 0.05 0.1 10521 (1-ROCA)% Max Iters. trec06p trec05p-1 50000 100000 150000 200000 250000 10521 CPUSec. Max Iters. trec06p trec05p-1 Figure 6: Reduced Iterations Tests. 4.1.2 Testing Reduced Iterations In the second ROSVM test, we experiment with reducing the number of iterations. Our initial tests showed that the maximum number of iterations used by Online SMO was rarely much larger than 10 on content-based spam detection; thus we tested values of T = {1, 2, 5, ∞}. Other parameters were identical to the original Online SVM tests. The results on this test were surprisingly stable (see Figure 6). Reducing the maximum number of SMO iterations per update had essentially no impact on classification performance, but did result in a moderate increase in speed. This suggests that any additional iterations are spent attempting to find improvements to a hyperplane that is already very close to optimal. These results show that for content-based spam detection, we can reduce computational cost by allowing only a single SMO iteration (that is, T = 1) with effectively equivalent performance. 4.1.3 Testing Reduced Updates For our third ROSVM experiment, we evaluate the impact of adjusting the parameter M to reduce the total number of updates. As noted before, when M = 1, the hyperplane is globally optimal at every step. Reducing M allows a slightly inconsistent hyperplane to persist until it encounters an example for which it is too inconsistent. We tested values of M from 0 to 1, at increments of 0.1. (Note that we used p = 10000 to decrease the cost of evaluating these tests.) The results for these tests are appear in Figure 7, and show that there is a slight degradation in performance with reduced values of M, and that this degradation in performance is accompanied by an increase in efficiency. Values of 0.005 0.01 0.025 0.05 0.1 0 0.2 0.4 0.6 0.8 1 (1-ROCA)% M trec05p-1 trec06p 5000 10000 15000 20000 25000 30000 35000 40000 0 0.2 0.4 0.6 0.8 1 CPUSec. M trec05p-1 trec06p Figure 7: Reduced Updates Tests. M > 0.7 give effectively equivalent performance as M = 1, and still reduce cost. 4.2 Online SVMs and ROSVM We now compare ROSVM against Online SVMs on the email spam, blog comment spam, and splog detection tasks. These experiments show comparable performance on these tasks, at radically different costs. In the previous section, the effect of the different relaxation methods was tested separately. Here, we tested these methods together to create a full implementation of ROSVM. We chose the values p = 10000, T = 1, M = 0.8 for the email spam detection tasks. Note that these parameter values were selected as those allowing ROSVM to achieve comparable performance results with Online SVMs, in order to test total difference in computational cost. The splog and blog data sets were much smaller, so we set p = 100 for these tasks to allow meaningful comparisons between the reduced size and full size optimization problems. Because these values were not hand-tuned, both generalization performance and runtime results are meaningful in these experiments. 4.2.1 Experimental Setup We compared Online SVMs and ROSVM on email spam, blog comment spam, and splog detection. For the email spam, we used the two large benchmark corpora, trec05p-1 and trec06p, in the standard online ordering. We randomly ordered both the blog comment spam corpus and the splog corpus to create online learning tasks. Note that this is a different setting than the leave-one-out cross validation task presented on these corpora in Section 2 - the results are not directly comparable. However, this experimental design Table 5: Email Spam Benchmark Data. These results compare Online SVM and ROSVM on email spam detection, using binary 4-gram feature space. Score reported is (1-ROCA)%, where 0 is optimal. trec05p-1 trec05p-1 trec06p trec06p (1-ROC)% CPUs (1-ROC)% CPUs OnSVM 0.0084 242,160 0.0232 87,519 ROSVM 0.0090 24,720 0.0240 18,541 Table 6: Blog Comment Spam. These results comparing Online SVM and ROSVM on blog comment spam detection using binary 4-gram feature space. Acc. Prec. Recall F1 CPUs OnSVM 0.926 0.930 0.962 0.946 139 ROSVM 0.923 0.925 0.965 0.945 11 does allow meaningful comparison between our two online methods on these content-based spam detection tasks. We ran each method on each task, and report the results in Tables 5, 6, and 7. Note that the CPU time reported for each method was generated on the same computing system. This time reflects only the time needed to complete online learning on tokenized data. We do not report the time taken to tokenize the data into binary 4-grams, as this is the same additive constant for all methods on each task. In all cases, ROSVM was significantly less expensive computationally. 4.3 Discussion The comparison results shown in Tables 5, 6, and 7 are striking in two ways. First, they show that the performance of Online SVMs can be matched and even exceeded by relaxed margin methods. Second, they show a dramatic disparity in computational cost. ROSVM is an order of magnitude more efficient than the normal Online SVM, and gives comparable results. Furthermore, the fixed lookback buffer ensures that the cost of each update does not depend on the size of the data set already seen, unlike Online SVMs. Note the blog and splog data sets are relatively small, and results on these data sets must be considered preliminary. Overall, these results show that there is no need to pay the high cost of SVMs to achieve this level of performance on contentbased detection of spam. ROSVMs offer a far cheaper alternative with little or no performance loss. 5. CONCLUSIONS In the past, academic researchers and industrial practitioners have disagreed on the best method for online contentbased detection of spam on the web. We have presented one resolution to this debate. Online SVMs do, indeed, proTable 7: Splog Data Set. These results compare Online SVM and ROSVM on splog detection using binary 4-gram feature space. Acc. Prec. Recall F1 CPUs OnSVM 0.880 0.910 0.842 0.874 29353 ROSVM 0.878 0.902 0.849 0.875 1251 duce state-of-the-art performance on this task with proper adjustment of the tradeoff parameter C, but with cost that grows quadratically with the size of the data set. The high values of C required for best performance with SVMs show that the margin maximization of Online SVMs is overkill for this task. Thus, we have proposed a less expensive alternative, ROSVM, that relaxes this maximum margin requirement, and produces nearly equivalent results. These methods are efficient enough for large-scale filtering of contentbased spam in its many forms. It is natural to ask why the task of content-based spam detection gets strong performance from ROSVM. After all, not all data allows the relaxation of SVM requirements. We conjecture that email spam, blog comment spam, and splogs all share the characteristic that a subset of features are particularly indicative of content being either spam or not spam. These indicative features may be sparsely represented in the data set, because of spam methods such as word obfuscation, in which common spam words are intentionally misspelled in an attempt to reduce the effectiveness of word-based spam detection. Maximizing the margin may cause these sparsely represented features to be ignored, creating an overall reduction in performance. It appears that spam data is highly separable, allowing ROSVM to be successful with high values of C and little effort given to maximizing the margin. Future work will determine how applicable relaxed SVMs are to the general problem of text classification. Finally, we note that the success of relaxed SVM methods for content-based spam detection is a result that depends on the nature of spam data, which is potentially subject to change. Although it is currently true that ham and spam are linearly separable given an appropriate feature space, this assumption may be subject to attack. While our current methods appear robust against primitive attacks along these lines, such as the good word attack [24], we must explore the feasibility of more sophisticated attacks. 6. REFERENCES [1] A. Bratko and B. Filipic. Spam filtering using compression models. Technical Report IJS-DP-9227, Department of Intelligent Systems, Jozef Stefan Institute, L jubljana, Slovenia, 2005. [2] G. Cauwenberghs and T. Poggio. Incremental and decremental support vector machine learning. In NIPS, pages 409-415, 2000. [3] G. V. Cormack. TREC 2006 spam track overview. In To appear in: The Fifteenth Text REtrieval Conference (TREC 2006) Proceedings, 2006. [4] G. V. Cormack and A. Bratko. Batch and on-line spam filter comparison. In Proceedings of the Third Conference on Email and Anti-Spam (CEAS), 2006. [5] G. V. Cormack and T. R. Lynam. TREC 2005 spam track overview. In The Fourteenth Text REtrieval Conference (TREC 2005) Proceedings, 2005. [6] G. V. Cormack and T. R. Lynam. On-line supervised spam filter evaluation. Technical report, David R. Cheriton School of Computer Science, University of Waterloo, Canada, February 2006. [7] N. Cristianini and J. Shawe-Taylor. An introduction to support vector machines. Cambridge University Press, 2000. [8] D. DeCoste and K. Wagstaff. Alpha seeding for support vector machines. In KDD "00: Proceedings of the sixth ACM SIGKDD international conference on Knowledge discovery and data mining, pages 345-349, 2000. [9] H. Drucker, V. Vapnik, and D. Wu. Support vector machines for spam categorization. IEEE Transactions on Neural Networks, 10(5):1048-1054, 1999. [10] J. Goodman and W. Yin. Online discriminative spam filter training. In Proceedings of the Third Conference on Email and Anti-Spam (CEAS), 2006. [11] P. Graham. A plan for spam. 2002. [12] P. Graham. Better bayesian filtering. 2003. [13] Z. Gyongi and H. Garcia-Molina. Spam: It"s not just for inboxes anymore. Computer, 38(10):28-34, 2005. [14] T. Joachims. Text categorization with suport vector machines: Learning with many relevant features. In ECML "98: Proceedings of the 10th European Conference on Machine Learning, pages 137-142, 1998. [15] T. Joachims. Training linear svms in linear time. In KDD "06: Proceedings of the 12th ACM SIGKDD international conference on Knowledge discovery and data mining, pages 217-226, 2006. [16] J. Kivinen, A. Smola, and R. Williamson. Online learning with kernels. In Advances in Neural Information Processing Systems 14, pages 785-793. MIT Press, 2002. [17] P. Kolari, T. Finin, and A. Joshi. SVMs for the blogosphere: Blog identification and splog detection. AAAI Spring Symposium on Computational Approaches to Analyzing Weblogs, 2006. [18] W. Krauth and M. M´ezard. Learning algorithms with optimal stability in neural networks. Journal of Physics A, 20(11):745-752, 1987. [19] T. Lynam, G. Cormack, and D. Cheriton. On-line spam filter fusion. In SIGIR "06: Proceedings of the 29th annual international ACM SIGIR conference on Research and development in information retrieval, pages 123-130, 2006. [20] V. Metsis, I. Androutsopoulos, and G. Paliouras. Spam filtering with naive bayes - which naive bayes? Third Conference on Email and Anti-Spam (CEAS), 2006. [21] G. Mishne, D. Carmel, and R. Lempel. Blocking blog spam with language model disagreement. Proceedings of the 1st International Workshop on Adversarial Information Retrieval on the Web (AIRWeb), May 2005. [22] J. Platt. Sequenital minimal optimization: A fast algorithm for training support vector machines. In B. Scholkopf, C. Burges, and A. Smola, editors, Advances in Kernel Methods - Support Vector Learning. MIT Press, 1998. [23] B. Scholkopf and A. Smola. Learning with Kernels: Support Vector Machines, Regularization, Optimization, and Beyond. MIT Press, 2001. [24] G. L. Wittel and S. F. Wu. On attacking statistical spam filters. CEAS: First Conference on Email and Anti-Spam, 2004.

logistic regression;support vector machine;feature mapping;bayesian method;hyperplane;link spam;spam filter;incremental update;link analysis;machine learning technique;blog;spam filtering;content-based spam detection;splog;content-based filtering

train_H-38

DiffusionRank: A Possible Penicillin for Web Spamming

While the PageRank algorithm has proven to be very effective for ranking Web pages, the rank scores of Web pages can be manipulated. To handle the manipulation problem and to cast a new insight on the Web structure, we propose a ranking algorithm called DiffusionRank. DiffusionRank is motivated by the heat diffusion phenomena, which can be connected to Web ranking because the activities flow on the Web can be imagined as heat flow, the link from a page to another can be treated as the pipe of an air-conditioner, and heat flow can embody the structure of the underlying Web graph. Theoretically we show that DiffusionRank can serve as a generalization of PageRank when the heat diffusion coefficient γ tends to infinity. In such a case 1/γ = 0, DiffusionRank (PageRank) has low ability of anti-manipulation. When γ = 0, DiffusionRank obtains the highest ability of anti-manipulation, but in such a case, the web structure is completely ignored. Consequently, γ is an interesting factor that can control the balance between the ability of preserving the original Web and the ability of reducing the effect of manipulation. It is found empirically that, when γ = 1, DiffusionRank has a Penicillin-like effect on the link manipulation. Moreover, DiffusionRank can be employed to find group-to-group relations on the Web, to divide the Web graph into several parts, and to find link communities. Experimental results show that the DiffusionRank algorithm achieves the above mentioned advantages as expected.

1. INTRODUCTION While the PageRank algorithm [13] has proven to be very effective for ranking Web pages, inaccurate PageRank results are induced because of web page manipulations by people for commercial interests. The manipulation problem is also called the Web spam, which refers to hyperlinked pages on the World Wide Web that are created with the intention of misleading search engines [7]. It is reported that approximately 70% of all pages in the .biz domain and about 35% of the pages in the .us domain belong to the spam category [12]. The reason for the increasing amount of Web spam is explained in [12]: some web site operators try to influence the positioning of their pages within search results because of the large fraction of web traffic originating from searches and the high potential monetary value of this traffic. From the viewpoint of the Web site operators who want to increase the ranking value of a particular page for search engines, Keyword Stuffing and Link Stuffing are being used widely [7, 12]. From the viewpoint of the search engine managers, the Web spam is very harmful to the users" evaluations and thus their preference to choosing search engines because people believe that a good search engine should not return irrelevant or low-quality results. There are two methods being employed to combat the Web spam problem. Machine learning methods are employed to handle the keyword stuffing. To successfully apply machine learning methods, we need to dig out some useful textual features for Web pages, to mark part of the Web pages as either spam or non-spam, then to apply supervised learning techniques to mark other pages. For example, see [5, 12]. Link analysis methods are also employed to handle the link stuffing problem. One example is the TrustRank [7], a link-based method, in which the link structure is utilized so that human labelled trusted pages can propagate their trust scores trough their links. This paper focuses on the link-based method. The rest of the materials are organized as follows. In the next section, we give a brief literature review on various related ranking techniques. We establish the Heat Diffusion Model (HDM) on various cases in Section 3, and propose DiffusionRank in Section 4. In Section 5, we describe the data sets that we worked on and the experimental results. Finally, we draw conclusions in Section 6. 2. LITERATURE REVIEW The importance of a Web page is determined by either the textual content of pages or the hyperlink structure or both. As in previous work [7, 13], we focus on ranking methods solely determined by hyperlink structure of the Web graph. All the mentioned ranking algorithms are established on a graph. For our convenience, we first give some notations. Denote a static graph by G = (V, E), where V = {v1, v2, . . . , vn}, E = {(vi, vj) | there is an edge from vi to vj}. Ii and di denote the in-degree and the out-degree of page i respectively. 2.1 PageRank The importance of a Web page is an inherently subjective matter, which depends on the reader"s interests, knowledge and attitudes [13]. However, the average importance of all readers can be considered as an objective matter. PageRank tries to find such average importance based on the Web link structure, which is considered to contain a large amount of statistical data. The Web is modelled by a directed graph G in the PageRank algorithms, and the rank or importance xi for page vi ∈ V is defined recursively in terms of pages which point to it: xi = (j,i)∈E aijxj, where aij is assumed to be 1/dj if there is a link from j to i, and 0 otherwise. Or in matrix terms, x = Ax. When the concept of random jump is introduced, the matrix form is changed to x = [(1 − α)g1T + αA]x, (1) where α is the probability of following the actual link from a page, (1 − α) is the probability of taking a random jump, and g is a stochastic vector, i.e., 1T g = 1. Typically, α = 0.85, and g = 1 n 1 is one of the standard settings, where 1 is the vector of all ones [6, 13]. 2.2 TrustRank TrustRank [7] is composed of two parts. The first part is the seed selection algorithm, in which the inverse PageRank was proposed to help an expert of determining a good node. The second part is to utilize the biased PageRank, in which the stochastic distribution g is set to be shared by all the trusted pages found in the first part. Moreover, the initial input of x is also set to be g. The justification for the inverse PageRank and the solid experiments support its advantage in combating the Web spam. Although there are many variations of PageRank, e.g., a family of link-based ranking algorithms in [2], TrustRank is especially chosen for comparisons for three reasonss: (1) it is designed for combatting spamming; (2) its fixed parameters make a comparison easy; and (3) it has a strong theoretical relations with PageRank and DiffusionRank. 2.3 Manifold Ranking In [17], the idea of ranking on the data manifolds was proposed. The data points represented as vectors in Euclidean space are considered to be drawn from a manifold. From the data points on such a manifold, an undirected weighted graph is created, then the weight matrix is given by the Gaussian Kernel smoothing. While the manifold ranking algorithm achieves an impressive result on ranking images, the biased vector g and the parameter k in the general personalized PageRank in [17] are unknown in the Web graph setting; therefore we do not include it in the comparisons. 2.4 Heat Diffusion Heat diffusion is a physical phenomena. In a medium, heat always flow from position with high temperature to position with low temperature. Heat kernel is used to describe the amount of heat that one point receives from another point. Recently, the idea of heat kernel on a manifold is borrowed in applications such as dimension reduction [3] and classification [9, 10, 14]. In these work, the input data is considered to lie in a special structure. All the above topics are related to our work. The readers can find that our model is a generalization of PageRank in order to resist Web manipulation, that we inherit the first part of TrustRank, that we borrow the concept of ranking on the manifold to introduce our model, and that heat diffusion is a main scheme in this paper. 3. HEAT DIFFUSION MODEL Heat diffusion provides us with another perspective about how we can view the Web and also a way to calculate ranking values. In this paper, the Web pages are considered to be drawn from an unknown manifold, and the link structure forms a directed graph, which is considered as an approximation to the unknown manifold. The heat kernel established on the Web graph is considered as the representation of the relationship between Web pages. The temperature distribution after a fixed time period, induced by a special initial temperature distribution, is considered as the rank scores on the Web pages. Before establishing the proposed models, we first show our motivations. 3.1 Motivations There are two points to explain that PageRank is susceptible to web spam. • Over-democratic. There is a belief behind PageRank-all pages are born equal. This can be seen from the equal voting ability of one page: the sum of each column is equal to one. This equal voting ability of all pages gives the chance for a Web site operator to increase a manipulated page by creating a large number of new pages pointing to this page since all the newly created pages can obtain an equal voting right. • Input-independent. For any given non-zero initial input, the iteration will converge to the same stable distribution corresponding to the maximum eigenvalue 1 of the transition matrix. This input-independent property makes it impossible to set a special initial input (larger values for trusted pages and less values even negative values for spam pages) to avoid web spam. The input-independent feature of PageRank can be further explained as follows. P = [(1 − α)g1T + αA] is a positive stochastic matrix if g is set to be a positive stochastic vector (the uniform distribution is one of such settings), and so the largest eigenvalue is 1 and no other eigenvalue whose absolute value is equal to 1, which is guaranteed by the Perron Theorem [11]. Let y be the eigenvector corresponding to 1, then we have Py = y. Let {xk} be the sequence generated from the iterations xk+1 = Pxk, and x0 is the initial input. If {xk} converges to x, then xk+1 = Pxk implies that x must satisfy Px = x. Since the only maximum eigenvalue is 1, we have x = cy where c is a constant, and if both x and y are normalized by their sums, then c = 1. The above discussions show that PageRank is independent of the initial input x0. In our opinion, g and α are objective parameters determined by the users" behaviors and preferences. A, α and g are the true web structure. While A is obtained by a crawler and the setting α = 0.85 is accepted by the people, we think that g should be determined by a user behavior investigation, something like [1]. Without any prior knowledge, g has to be set as g = 1 n 1. TrustRank model does not follow the true web structure by setting a biased g, but the effects of combatting spamming are achieved in [7]; PageRank is on the contrary in some ways. We expect a ranking algorithm that has an effect of anti-manipulation as TrustRank while respecting the true web structure as PageRank. We observe that the heat diffusion model is a natural way to avoid the over-democratic and input-independent feature of PageRank. Since heat always flows from a position with higher temperatures to one with lower temperatures, points are not equal as some points are born with high temperatures while others are born with low temperatures. On the other hand, different initial temperature distributions will give rise to different temperature distributions after a fixed time period. Based on these considerations, we propose the novel DiffusionRank. This ranking algorithm is also motivated by the viewpoint for the Web structure. We view all the Web pages as points drawn from a highly complex geometric structure, like a manifold in a high dimensional space. On a manifold, heat can flow from one point to another through the underlying geometric structure in a given time period. Different geometric structures determine different heat diffusion behaviors, and conversely the diffusion behavior can reflect the geometric structure. More specifically, on the manifold, the heat flows from one point to another point, and in a given time period, if one point x receives a large amount of heat from another point y, we can say x and y are well connected, and thus x and y have a high similarity in the sense of a high mutual connection. We note that on a point with unit mass, the temperature and the heat of this point are equivalent, and these two terms are interchangeable in this paper. In the following, we first show the HDM on a manifold, which is the origin of HDM, but cannot be employed to the World Wide Web directly, and so is considered as the ideal case. To connect the ideal case and the practical case, we then establish HDM on a graph as an intermediate case. To model the real world problem, we further build HDM on a random graph as a practical case. Finally we demonstrate the DiffusionRank which is derived from the HDM on a random graph. 3.2 Heat Flow On a Known Manifold If the underlying manifold is known, the heat flow throughout a geometric manifold with initial conditions can be described by the following second order differential equation: ∂f(x,t) ∂t − ∆f(x, t) = 0, where f(x, t) is the heat at location x at time t, and ∆f is the Laplace-Beltrami operator on a function f. The heat diffusion kernel Kt(x, y) is a special solution to the heat equation with a special initial condition-a unit heat source at position y when there is no heat in other positions. Based on this, the heat kernel Kt(x, y) describes the heat distribution at time t diffusing from the initial unit heat source at position y, and thus describes the connectivity (which is considered as a kind of similarity) between x and y. However, it is very difficult to represent the World Wide Web as a regular geometry with a known dimension; even the underlying is known, it is very difficult to find the heat kernel Kt(x, y), which involves solving the heat equation with the delta function as the initial condition. This motivates us to investigate the heat flow on a graph. The graph is considered as an approximation to the underlying manifold, and so the heat flow on the graph is considered as an approximation to the heat flow on the manifold. 3.3 On an Undirected Graph On an undirected graph G, the edge (vi, vj) is considered as a pipe that connects nodes vi and vj. The value fi(t) describes the heat at node vi at time t, beginning from an initial distribution of heat given by fi(0) at time zero. f(t) (f(0)) denotes the vector consisting of fi(t) (fi(0)). We construct our model as follows. Suppose, at time t, each node i receives M(i, j, t, ∆t) amount of heat from its neighbor j during a period of ∆t. The heat M(i, j, t, ∆t) should be proportional to the time period ∆t and the heat difference fj(t) − fi(t). Moreover, the heat flows from node j to node i through the pipe that connects nodes i and j. Based on this consideration, we assume that M(i, j, t, ∆t) = γ(fj(t) − fi(t))∆t. As a result, the heat difference at node i between time t + ∆t and time t will be equal to the sum of the heat that it receives from all its neighbors. This is formulated as fi(t + ∆t) − fi(t) = j:(j,i)∈E γ(fj(t) − fi(t))∆t, (2) where E is the set of edges. To find a closed form solution to Eq. (2), we express it in a matrix form: (f(t + ∆t) − f(t))/∆t = γHf(t), where d(v) denotes the degree of the node v. In the limit ∆t → 0, it becomes d dt f(t) = γHf(t). Solving it, we obtain f(t) = eγtH f(0), especially we have f(1) = eγH f(0), Hij =    −d(vj), j = i, 1, (vj, vi) ∈ E, 0, otherwise, (3) where eγH is defined as eγH = I+γH+ γ2 2! H2 + γ3 3! H3 +· · · . 3.4 On a Directed Graph The above heat diffusion model must be modified to fit the situation where the links between Web pages are directed. On one Web page, when the page-maker creates a link (a, b) to another page b, he actually forces the energy flow, for example, people"s click-through activities, to that page, and so there is added energy imposed on the link. As a result, heat flows in a one-way manner, only from a to b, not from b to a. Based on such consideration, we modified the heat diffusion model on an undirected graph as follows. On a directed graph G, the pipe (vi, vj) is forced by added energy such that heat flows only from vi to vj. Suppose, at time t, each node vi receives RH = RH(i, j, t, ∆t) amount of heat from vj during a period of ∆t. We have three assumptions: (1) RH should be proportional to the time period ∆t; (2) RH should be proportional to the the heat at node vj; and (3) RH is zero if there is no link from vj to vi. As a result, vi will receive j:(vj ,vi)∈E σjfj(t)∆t amount of heat from all its neighbors that points to it. On the other hand, node vi diffuses DH(i, t, ∆t) amount of heat to its subsequent nodes. We assume that: (1) The heat DH(i, t, ∆t) should be proportional to the time period ∆t. (2) The heat DH(i, t, ∆t) should be proportional to the the heat at node vi. (3) Each node has the same ability of diffusing heat. This fits the intuition that a Web surfer only has one choice to find the next page that he wants to browse. (4) The heat DH(i, t, ∆t) should be uniformly distributed to its subsequent nodes. The real situation is more complex than what we assume, but we have to make this simple assumption in order to make our model concise. As a result, node vi will diffuse γfi(t)∆t/di amount of heat to any of its subsequent nodes, and each of its subsequent node should receive γfi(t)∆t/di amount of heat. Therefore σj = γ/dj. To sum up, the heat difference at node vi between time t+∆t and time t will be equal to the sum of the heat that it receives, deducted by what it diffuses. This is formulated as fi(t + ∆t) − fi(t) = −γfi(t)∆t + j:(vj ,vi)∈E γ/djfj(t)∆t. Similarly, we obtain f(1) = eγH f(0), Hij =    −1, j = i, 1/dj, (vj, vi) ∈ E, 0, otherwise. (4) 3.5 On a Random Directed Graph For real world applications, we have to consider random edges. This can be seen in two viewpoints. The first one is that in Eq. (1), the Web graph is actually modelled as a random graph, there is an edge from node vi to node vj with a probability of (1 − α)gj (see the item (1 − α)g1T ), and that the Web graph is predicted by a random graph [15, 16]. The second one is that the Web structure is a random graph in essence if we consider the content similarity between two pages, though this is not done in this paper. For these reasons, the model would become more flexible if we extend it to random graphs. The definition of a random graph is given below. Definition 1. A random graph RG = (V, P = (pij)) is defined as a graph with a vertex set V in which the edges are chosen independently, and for 1 ≤ i, j ≤ |V | the probability of (vi, vj) being an edge is exactly pij. The original definition of random graphs in [4], is changed slightly to consider the situation of directed graphs. Note that every static graph can be considered as a special random graph in the sense that pij can only be 0 or 1. On a random graph RG = (V, P), where P = (pij) is the probability of the edge (vi, vj) exists. In such a random graph, the expected heat difference at node i between time t + ∆t and time t will be equal to the sum of the expected heat that it receives from all its antecedents, deducted by the expected heat that it diffuses. Since the probability of the link (vj, vi) is pji, the expected heat flow from node j to node i should be multiplied by pji, and so we have fi(t + ∆t) − fi(t) = −γ fi(t) ∆t + j:(vj ,vi)∈E γpjifj(t)∆t/RD+ (vj), where RD+ (vi) is the expected out-degree of node vi, it is defined as k pik. Similarly we have f(1) = eγR f(0), Rij =    −1, j = i; pji/RD+ (vj), j = i. (5) When the graph is large, a direct computation of eγR is time-consuming, and we adopt its discrete approximation: f(1) = (I + γ N R)N f(0). (6) The matrix (I+ γ N R)N in Eq. (6) and matrix eγR in Eq. (5) are called Discrete Diffusion Kernel and the Continuous Diffusion Kernel respectively. Based on the Heat Diffusion Models and their solutions, DiffusionRank can be established on undirected graphs, directed graphs, and random graphs. In the next section, we mainly focus on DiffusionRank in the random graph setting. 4. DIFFUSIONRANK For a random graph, the matrix (I + γ N R)N or eγR can measure the similarity relationship between nodes. Let fi(0)= 1, fj(0) = 0 if j = i, then the vector f(0) represent the unit heat at node vi while all other nodes has zero heat. For such f(0) in a random graph, we can find the heat distribution at time 1 by using Eq. (5) or Eq. (6). The heat distribution is exactly the i−th row of the matrix of (I + γ N R)N or eγR . So the ith-row jth-column element hij in the matrix (I + γ∆tR)N or eγR means the amount of heat that vi can receive from vj from time 0 to 1. Thus the value hij can be used to measure the similarity from vj to vi. For a static graph, similarly the matrix (I + γ N H)N or eγH can measure the similarity relationship between nodes. The intuition behind is that the amount h(i, j) of heat that a page vi receives from a unit heat in a page vj in a unit time embodies the extent of the link connections from page vj to page vi. Roughly speaking, when there are more uncrossed paths from vj to vi, vi will receive more heat from vj; when the path length from vj to vi is shorter, vi will receive more heat from vj; and when the pipe connecting vj and vi is wide, the heat will flow quickly. The final heat that vi receives will depend on various paths from vj to vi, their length, and the width of the pipes. Algorithm 1 DiffusionRank Function Input: The transition matrix A; the inverse transition matrix U; the decay factor αI for the inverse PageRank; the decay factor αB for PageRank; number of iterations MI for the inverse PageRank; the number of trusted pages L; the thermal conductivity coefficient γ. Output: DiffusionRank score vector h. 1: s = 1 2: for i = 1 TO MI do 3: s = αI · U · s + (1 − αI ) · 1 n · 1 4: end for 5: Sort s in a decreasing order: π = Rank({1, . . . , n}, s) 6: d = 0, Count = 0, i = 0 7: while Count ≤ L do 8: if π(i) is evaluated as a trusted page then 9: d(π(i)) = 1, Count + + 10: end if 11: i + + 12: end while 13: d = d/|d| 14: h = d 15: Find the iteration number MB according to λ 16: for i = 1 TO MB do 17: h = (1 − γ MB )h + γ MB (αB · A · h + (1 − αB) · 1 n · 1) 18: end for 19: RETURN h 4.1 Algorithm For the ranking task, we adopt the heat kernel on a random graph. Formally the DiffusionRank is described in Algorithm 1, in which,the element Uij in the inverse transition matrix U is defined to be 1/Ij if there is a link from i to j, and 0 otherwise. This trusted pages selection procedure by inverse PageRank is completely borrowed from TrustRank [7] except for a fix number of the size of the trusted set. Although the inverse PageRank is not perfect in its ability of determining the maximum coverage, it is appealing because of its polynomial execution time and its reasonable intuition-we actually inverse the original link when we try to build the seed set from those pages that point to many pages that in turn point to many pages and so on. In the algorithm, the underlying random graph is set as P = αB · A + (1 − αB) · 1 n · 1n×n, which is induced by the Web graph. As a result, R = −I + P. In fact, the more general setting for DiffusionRank is P = αB ·A+(1−αB)· 1 n ·g·1T . By such a setting, DiffusionRank is a generalization of TrustRank when γ tends to infinity and when g is set in the same way as TrustRank. However, the second part of TrustRank is not adopted by us. In our model, g should be the true teleportation determined by the user"s browse habits, popularity distribution over all the Web pages, and so on; P should be the true model of the random nature of the World Wide Web. Setting g according to the trusted pages will not be consistent with the basic idea of Heat Diffusion on a random graph. We simply set g = 1 only because we cannot find it without any priori knowledge. Remark. In a social network interpretation, DiffusionRank first recognizes a group of trusted people, who may not be highly ranked, but they know many other people. The initially trusted people are endowed with the power to decide who can be further trusted, but cannot decide the final voting results, and so they are not dictators. 4.2 Advantages Next we show the four advantages for DiffusionRank. 4.2.1 Two closed forms First, its solutions have two forms, both of which are closed form. One takes the discrete form, and has the advantage of fast computing while the other takes the continuous form, and has the advantage of being easily analyzed in theoretical aspects. The theoretical advantage has been shown in the proof of theorem in the next section. (a) Group to Group Relations (b) An undirected graph Figure 1: Two graphs 4.2.2 Group-group relations Second, it can be naturally employed to detect the groupgroup relation. For example, let G2 and G1 denote two groups, containing pages (j1, j2, . . . , js) and (i1, i2, . . . , it), respectively. Then u,v hiu,jv is the total amounts of heat that G1 receives from G2, where hiu,jv is the iu−th row jv−th column element of the heat kernel. More specifically, we need to first set f(0) for such an application as follows. In f(0) = (f1(0), f2(0), . . . , fn(0))T , if i ∈ {j1, j2, . . . , js}, then fi(0) = 1, and 0 otherwise. Then we employ Eq. (5) to calculate f(1) = (f1(1), f2(1), . . . , fn(1))T , finally we sum those fj(1) where j ∈ {i1, i2, . . . , it}. Fig. 1 (a) shows the results generated by the DiffusionRank. We consider five groups-five departments in our Engineering Faculty: CSE, MAE, EE, IE, and SE. γ is set to be 1, the numbers in Fig. 1 (a) are the amount of heat that they diffuse to each other. These results are normalized by the total number of each group, and the edges are ignored if the values are less than 0.000001. The group-to-group relations are therefore detected, for example, we can see that the most strong overall tie is from EE to IE. While it is a natural application for DiffusionRank because of the easy interpretation by the amount heat from one group to another group, it is difficult to apply other ranking techniques to such an application because they lack such a physical meaning. 4.2.3 Graph cut Third, it can be used to partition the Web graph into several parts. A quick example is shown below. The graph in Fig. 1 (b) is an undirected graph, and so we employ the Eq. (3). If we know that node 1 belongs to one community and that node 12 belongs to another community, then we can put one unit positive heat source on node 1 and one unit negative heat source on node 12. After time 1, if we set γ = 0.5, the heat distribution is [0.25, 0.16, 0.17, 0.16, 0.15, 0.09, 0.01, -0.04, -0.18 -0.21, -0.21, -0.34], and if we set γ = 1, it will be [0.17, 0.16, 0.17, 0.16, 0.16, 0.12, 0.02, -0.07, -0.18, -0.22, -0.24, -0.24]. In both settings, we can easily divide the graph into two parts: {1, 2, 3, 4, 5, 6, 7} with positive temperatures and {8, 9, 10, 11, 12} with negative temperatures. For directed graphs and random graphs, similarly we can cut them by employing corresponding heat solution. 4.2.4 Anti-manipulation Fourth, it can be used to combat manipulation. Let G2 contain trusted Web pages (j1, j2, . . . , js), then for each page i, v hi,jv is the heat that page i receives from G2, and can be computed by the discrete approximation of Eq. (4) in the case of a static graph or Eq. (6) in the case of a random graph, in which f(0) is set to be a special initial heat distribution so that the trusted Web pages have unit heat while all the others have zero heat. In doing so, manipulated Web page will get a lower rank unless it has strong in-links from the trusted Web pages directly or indirectly. The situation is quite different for PageRank because PageRank is inputindependent as we have shown in Section 3.1. Based on the fact that the connection from a trusted page to a bad page should be weak-less uncross paths, longer distance and narrower pipe, we can say DiffusionRank can resist web spam if we can select trusted pages. It is fortunate that the trusted pages selection method in [7]-the first part of TrustRank can help us to fulfill this task. For such an application of DiffusionRank, the computation complexity for Discrete Diffusion Kernel is the same as that for PageRank in cases of both a static graph and a random graph. This can be seen in Eq. (6), by which we need N iterations and for each iteration we need a multiplication operation between a matrix and a vector, while in Eq. (1) we also need a multiplication operation between a matrix and a vector for each iteration. 4.3 The Physical Meaning of γ γ plays an important role in the anti-manipulation effect of DiffusionRank. γ is the thermal conductivity-the heat diffusion coefficient. If it has a high value, heat will diffuse very quickly. Conversely, if it is small, heat will diffuse slowly. In the extreme case, if it is infinitely large, then heat will diffuse from one node to other nodes immediately, and this is exactly the case corresponding to PageRank. Next, we will interpret it mathematically. Theorem 1. When γ tends to infinity and f(0) is not the zero vector, eγR f(0) is proportional to the stable distribution produced by PageRank. Let g = 1 n 1. By the Perron Theorem [11], we have shown that 1 is the largest eigenvalue of P = [(1 − α)g1T + αA], and that no other eigenvalue whose absolute value is equal to 1. Let x be the stable distribution, and so Px = x. x is the eigenvector corresponding to the eigenvalue 1. Assume the n − 1 other eigenvalues of P are |λ2| < 1, . . . , |λn| < 1, we can find an invertible matrix S = ( x S1 ) such that S−1 PS =      1 ∗ ∗ ∗ 0 λ2 ∗ ∗ 0 0 ... ∗ 0 0 0 λn      . (7) Since eγR = eγ(−I+P) = S−1      1 ∗ ∗ ∗ 0 eγ(λ2−1) ∗ ∗ 0 0 ... ∗ 0 0 0 eγ(λn−1)      S, (8) all eigenvalues of the matrix eγR are 1, eγ(λ2−1) , . . . , eγ(λn−1) . When γ → ∞, they become 1, 0, . . . , 0, which means that 1 is the only nonzero eigenvalue of eγR when γ → ∞. We can see that when γ → ∞, eγR eγR f(0) = eγR f(0), and so eγR f(0) is an eigenvector of eγR when γ → ∞. On the other hand, eγR x = (I+γR+γ2 2! R2 +γ3 3! R3 +. . .)x = Ix+γRx+γ2 2! R2 x+ γ3 3! R3 x + . . . = x since Rx = (−I + P)x = −x + x = 0, and hence x is the eigenvector of eγR for any γ. Therefore both x and eγR f(0) are the eigenvectors corresponding the unique eigenvalue 1 of eγR when γ → ∞, and consequently x = ceγR f(0). By this theorem, we see that DiffusionRank is a generalization of PageRank. When γ = 0, the ranking value is most robust to manipulation since no heat is diffused and the system is unchangeable, but the Web structure is completely ignored since eγR f(0) = e0R f(0) = If(0) = f(0); when γ = ∞, DiffusionRank becomes PageRank, it can be manipulated easily. We expect an appropriate setting of γ that can balance both. For this, we have no theoretical result, but in practice we find that γ = 1 works well in Section 5. Next we discuss how to determine the number of iterations if we employ the discrete heat kernel. 4.4 The Number of Iterations While we enjoy the advantage of the concise form of the exponential heat kernel, it is better for us to calculate DiffusionRank by employing Eq. (6) in an iterative way. Then the problem about determining N-the number of iterations arises: For a given threshold , find N such that ||((I + γ N R)N − eγR )f(0)|| < for any f(0) whose sum is one. Since it is difficult to solve this problem, we propose a heuristic motivated by the following observations. When R = −I+P, by Eq. (7), we have (I+ γ N R)N = (I+ γ N (−I+ P))N = S−1      1 ∗ ∗ ∗ 0 (1 + γ(λ2−1) N )N ∗ ∗ 0 0 ... ∗ 0 0 0 (1 + γ(λn−1) N )N      S. (9) Comparing Eq. (8) and Eq. (9), we observe that the eigenvalues of (I + γ N R)N − eγR are (1 + γ(λn−1) N )N − eγ(λn−1) . We propose a heuristic method to determine N so that the difference between the eigenvalues are less than a threshold for only positive λs. We also observe that if γ = 1, λ < 1, then |(1+ γ(λ−1) N )N − eγ(λ−1) | < 0.005 if N ≥ 100, and |(1+ γ(λ−1) N )N −eγ(λ−1) | < 0.01 if N ≥ 30. So we can set N = 30, or N = 100, or others according to different accuracy requirements. In this paper, we use the relatively accurate setting N = 100 to make the real eigenvalues in (I + γ N R)N − eγR less than 0.005. 5. EXPERIMENTS In this section, we show the experimental data, the methodology, the setting, and the results. 5.1 Data Preparation Our input data consist of a toy graph, a middle-size realworld graph, and a large-size real-world graph. The toy graph is shown in Fig. 2 (a). The graph below it shows node 1 is being manipulated by adding new nodes A, B, C, . . . such that they all point to node 1, and node 1 points to them all. The data of two real Web graph were obtained from the domain in our institute in October, 2004. The total number of pages found are 18,542 in the middle-size graph, and 607,170 in the large-size graph respectively. The middle-size graph is a subgraph of the large-size graph, and they were obtained by the same crawler: one is recorded by the crawler in its earlier time, and the other is obtained when the crawler stopped. 5.2 Methodology The algorithms we run include PageRank, TrustRank and DiffusionRank. All the rank values are multiplied by the number of nodes so that the sum of the rank values is equal to the number of nodes. By this normalization, we can compare the results on graphs with different sizes since the average rank value is one for any graph after such normalization. We will need value difference and pairwise order difference as comparison measures. Their definitions are listed as follows. Value Difference. The value difference between A = {Ai}n i=1 and B = {Bi}n i=1 is measured as n i=1 |Ai − Bi|. Pairwise Order Difference. The order difference between A and B is measured as the number of significant order differences between A and B. The pair (A[i], A[j]) and (B[i], B[j]) is considered as a significant order difference if one of the following cases happens: both A[i] > [ <]A[j]+0.1 and B[i] ≤ [ ≥]A[j]; both A[i] ≤ [ ≥]A[j] and B[i] > [ < ]A[j] + 0.1. A 1 B C ... 2 5 6 3 4 1 2 5 6 3 4 0 1 2 3 4 5 6 0 2 4 6 8 10 12 Gamma ValueDifference Trust set={1} Trust set={2} Trust set={3} Trust set={4} Trust set={5} Trust set={6} (a) (b) Figure 2: (a) The toy graph consisting of six nodes, and node 1 is being manipulated by adding new nodes A, B, C, . . . (b) The approximation tendency to PageRank by DiffusionRank 5.3 Experimental Set-up The experiments on the middle-size graph and the largesize graphs are conducted on the workstation, whose hardware model is Nix Dual Intel Xeon 2.2GHz with 1GB RAM and a Linux Kernel 2.4.18-27smp (RedHat7.3). In calculating DiffusionRank, we employ Eq. (6) and the discrete approximation of Eq. (4) for such graphs. The related tasks are implemented using C language. While in the toy graph, we employ the continuous diffusion kernel in Eq. (4) and Eq. (5), and implement related tasks using Matlab. For nodes that have zero out-degree (dangling nodes), we employ the method in the modified PageRank algorithm [8], in which dangling nodes of are considered to have random links uniformly to each node. We set α = αI = αB = 0.85 in all algorithms. We also set g to be the uniform distribution in both PageRank and DiffusionRank. For DiffusionRank, we set γ = 1. According to the discussions in Section 4.3 and Section 4.4, we set the iteration number to be MB = 100 in DiffusionRank, and for accuracy consideration, the iteration number in all the algorithms is set to be 100. 5.4 Approximation of PageRank We show that when γ tends to infinity, the value differences between DiffusionRank and PageRank tend to zero. Fig. 2 (b) shows the approximation property of DiffusionRank, as proved in Theorem 1, on the toy graph. The horizontal axis of Fig. 2 (b) marks the γ value, and vertical axis corresponds to the value difference between DiffusionRank and PageRank. All the possible trusted sets with L = 1 are considered. For L > 1, the results should be the linear combination of some of these curves because of the linearity of the solutions to heat equations. On other graphs, the situations are similar. 5.5 Results of Anti-manipulation In this section, we show how the rank values change as the intensity of manipulation increases. We measure the intensity of manipulation by the number of newly added points that point to the manipulated point. The horizontal axes of Fig. 3 stand for the numbers of newly added points, and vertical axes show the corresponding rank values of the manipulated nodes. To be clear, we consider all six situations. Every node in Fig. 2 (a) is manipulated respectively, and its 0 50 100 0 10 20 30 40 50 RankoftheManipulatdNode−1 DiffusionRank−Trust 4 PageRank TrustRanl−Trust 4 0 50 100 0 10 20 30 40 50 RankoftheManipulatdNode−2 DiffusionRank−Trust 4 PageRank TrustRanl−Trust 4 0 50 100 0 10 20 30 40 50 RankoftheManipulatdNode−3 DiffusionRank−Trust 4 PageRank TrustRanl−Trust 4 0 50 100 0 10 20 30 40 50 Number of New Added Nodes RankoftheManipulatdNode−4 DiffusionRank−Trust 3 PageRank TrustRanl−Trust 3 0 50 100 0 10 20 30 40 50 Number of New Added Nodes RankoftheManipulatdNode−5 DiffusionRank−Trust 4 PageRank TrustRanl−Trust 4 0 50 100 0 10 20 30 40 50 Number of New Added Nodes RankoftheManipulatdNode−6 DiffusionRank−Trust 4 PageRank TrustRanl−Trust 4 Figure 3: The rank values of the manipulated nodes on the toy graph 200040006000800010000 0 1000 2000 3000 4000 5000 6000 7000 8000 Number of New Added Points RankoftheManipulatdNode PageRank DiffusionRank−uniform DiffusionRank0 DiffusionRank1 DiffusionRank2 DiffusionRank3 TrustRank0 TrustRank1 TrustRank2 TrustRank3 2000 4000 6000 8000 10000 0 20 40 60 80 100 120 140 160 180 Number of New Added Points RankoftheManipulatdNode PageRank DiffusionRank TrustRank DiffusionRank−uniform (a) (b) Figure 4: (a) The rank values of the manipulated nodes on the middle-size graph; (b) The rank values of the manipulated nodes on the large-size graph corresponding values for PageRank, TrustRank (TR), DiffusionRank (DR) are shown in the one of six sub-figures in Fig. 3. The vertical axes show which node is being manipulated. In each sub-figure, the trusted sets are computed below. Since the inverse PageRank yields the results [1.26, 0.85, 1.31, 1.36, 0.51, 0.71]. Let L = 1. If the manipulated node is not 4, then the trusted set is {4}, and otherwise {3}. We observe that in all the cases, rank values of the manipulated node for DiffusionRank grow slowest as the number of the newly added nodes increases. On the middle-size graph and the large-size graph, this conclusion is also true, see Fig. 4. Note that, in Fig. 4 (a), we choose four trusted sets (L = 1), on which we test DiffusionRank and TrustRank, the results are denoted by DiffusionRanki and TrustRanki (i = 0, 1, 2, 3 denotes the four trusted set); in Fig. 4 (b), we choose one trusted set (L = 1). Moreover, in both Fig. 4 (a) and Fig. 4 (b), we show the results for DiffusionRank when we have no trusted set, and we trust all the pages before some of them are manipulated. We also test the order difference between the ranking order A before the page is manipulated and the ranking order PA after the page is manipulated. Because after manipulation, the number of pages changes, we only compare the common part of A and PA. This experiment is used to test the stability of all these algorithms. The less the order difference, the stabler the algorithm, in the sense that only a smaller part of the order relations is affected by the manipulation. Figure 5 (a) shows that the order difference values change when we add new nodes that point to the manipulated node. We give several γ settings. We find that when γ = 1, the least order difference is achieved by DiffusionRank. It is interesting to point out that as γ increases, the order difference will increase first; after reaching a maximum value, it will decrease, and finally it tends to the PageRank results. We show this tendency in Fig. 5 (b), in which we choose three different settings-the number of manipulated nodes are 2,000, 5,000, and 10,000 respectively. From these figures, we can see that when γ < 2, the values are less than those for PageRank, and that when γ > 20, the difference between PageRank and DiffusionRank is very small. After these investigations, we find that in all the graphs we tested, DiffusionRank (when γ = 1) is most robust to manipulation both in value difference and order difference. The trust set selection algorithm proposed in [7] is effective for both TrustRank and DiffusionRank. 0 1000 2000 3000 4000 5000 6000 7000 8000 9000 10000 0 0.5 1 1.5 2 2.5 3 x 10 5 Number of New Added Points PairwiseOrderDifference PageRank DiffusionRank−Gamma=1 DiffusionRank−Gamma=2 DiffusionRank−Gamma=3 DiffusionRank−Gamma=4 DiffusionRank−Gamma=5 DiffusionRank−Gamma=15 TrustRank 0 5 10 15 20 0 0.5 1 1.5 2 2.5 x 10 5 Gamma PairwiseOrderDifference DiffusionRank: when added 2000 nodes DiffusionRank: when added 5000 nodes DiffusionRank: when added 10000 nodes PageRank (a) (b) Figure 5: (a) Pairwise order difference on the middle-size graph, the least it is, the more stable the algorithm; (b) The tendency of varying γ 6. CONCLUSIONS We conclude that DiffusionRank is a generalization of PageRank, which is interesting in that the heat diffusion coefficient γ can balance the extent that we want to model the original Web graph and the extent that we want to reduce the effect of link manipulations. The experimental results show that we can actually achieve such a balance by setting γ = 1, although the best setting including varying γi is still under further investigation. This anti-manipulation feature enables DiffusionRank to be a candidate as a penicillin for Web spamming. Moreover, DiffusionRank can be employed to find group-group relations and to partition Web graph into small communities. All these advantages can be achieved in the same computational complexity as PageRank. For the special application of anti-manipulation, DiffusionRank performs best both in reduction effects and in its stability among all the three algorithms. 7. ACKNOWLEDGMENTS We thank Patrick Lau, Zhenjiang Lin and Zenglin Xu for their help. This work is fully supported by two grants from the Research Grants Council of the Hong Kong Special administrative Region, China (Project No. CUHK4205/04E and Project No. CUHK4235/04E). 8. REFERENCES [1] E. Agichtein, E. Brill, and S. T. Dumais. Improving web search ranking by incorporating user behavior information. In E. N. Efthimiadis, S. T. Dumais, D. Hawking, and K. J¨arvelin, editors, Proceedings of the 29th Annual International ACM SIGIR Conference on Research and Development in Information Retrieval (SIGIR), pages 19-26, 2006. [2] R. A. Baeza-Yates, P. Boldi, and C. Castillo. Generalizing pagerank: damping functions for link-based ranking algorithms. In E. N. Efthimiadis, S. T. Dumais, D. Hawking, and K. J¨arvelin, editors, Proceedings of the 29th Annual International ACM SIGIR Conference on Research and Development in Information Retrieval (SIGIR), pages 308-315, 2006. [3] M. Belkin and P. Niyogi. Laplacian eigenmaps for dimensionality reduction and data representation. Neural Computation, 15(6):1373-1396, Jun 2003. [4] B. Bollob´as. Random Graphs. Academic Press Inc. (London), 1985. [5] C. Burges, T. Shaked, E. Renshaw, A. Lazier, M. Deeds, N. Hamilton, and G. Hullender. Learning to rank using gradient descent. In Proceedings of the 22nd international conference on Machine learning (ICML), pages 89-96, 2005. [6] N. Eiron, K. S. McCurley, and J. A. Tomlin. Ranking the web frontier. In Proceeding of the 13th World Wide Web Conference (WWW), pages 309-318, 2004. [7] Z. Gy¨ongyi, H. Garcia-Molina, and J. Pedersen. Combating web spam with trustrank. In M. A. Nascimento, M. T. ¨Ozsu, D. Kossmann, R. J. Miller, J. A. Blakeley, and K. B. Schiefer, editors, Proceedings of the Thirtieth International Conference on Very Large Data Bases (VLDB), pages 576-587, 2004. [8] S. D. Kamvar, T. H. Haveliwala, C. D. Manning, and G. H. Golub. Exploiting the block structure of the web for computing pagerank. Technical report, Stanford University, 2003. [9] R. I. Kondor and J. D. Lafferty. Diffusion kernels on graphs and other discrete input spaces. In C. Sammut and A. G. Hoffmann, editors, Proceedings of the Nineteenth International Conference on Machine Learning (ICML), pages 315-322, 2002. [10] J. Lafferty and G. Lebanon. Diffusion kernels on statistical manifolds. Journal of Machine Learning Research, 6:129-163, Jan 2005. [11] C. R. MacCluer. The many proofs and applications of perron"s theorem. SIAM Review, 42(3):487-498, 2000. [12] A. Ntoulas, M. Najork, M. Manasse, and D. Fetterly. Detecting spam web pages through content analysis. In Proceedings of the 15th international conference on World Wide Web (WWW), pages 83-92, 2006. [13] L. Page, S. Brin, R. Motwani, and T. Winograd. The pagerank citation ranking: Bringing order to the web. Technical Report Paper SIDL-WP-1999-0120 (version of 11/11/1999), Stanford Digital Library Technologies Project, 1999. [14] H. Yang, I. King, and M. R. Lyu. NHDC and PHDC: Non-propagating and propagating heat diffusion classifiers. In Proceedings of the 12th International Conference on Neural Information Processing (ICONIP), pages 394-399, 2005. [15] H. Yang, I. King, and M. R. Lyu. Predictive ranking: a novel page ranking approach by estimating the web structure. In Proceedings of the 14th international conference on World Wide Web (WWW) - Special interest tracks and posters, pages 944-945, 2005. [16] H. Yang, I. King, and M. R. Lyu. Predictive random graph ranking on the web. In Proceedings of the IEEE World Congress on Computational Intelligence (WCCI), pages 3491-3498, 2006. [17] D. Zhou, J. Weston, A. Gretton, O. Bousquet, and B. Sch¨olkopf. Ranking on data manifolds. In S. Thrun, L. Saul, and B. Sch¨olkopf, editors, Advances in Neural Information Processing Systems 16 (NIPS 2003), 2004.

link stuffing;web graph;group-to-group relation;keyword stuffing;gaussian kernel smoothing;ranking;diffusionrank;pagerank;equal voting ability;random graph;web spam;link community;machine learning;link analysis;seed selection algorithm

train_H-40

Cross-Lingual Query Suggestion Using Query Logs of Different Languages

Query suggestion aims to suggest relevant queries for a given query, which help users better specify their information needs. Previously, the suggested terms are mostly in the same language of the input query. In this paper, we extend it to cross-lingual query suggestion (CLQS): for a query in one language, we suggest similar or relevant queries in other languages. This is very important to scenarios of cross-language information retrieval (CLIR) and cross-lingual keyword bidding for search engine advertisement. Instead of relying on existing query translation technologies for CLQS, we present an effective means to map the input query of one language to queries of the other language in the query log. Important monolingual and cross-lingual information such as word translation relations and word co-occurrence statistics, etc. are used to estimate the cross-lingual query similarity with a discriminative model. Benchmarks show that the resulting CLQS system significantly outperforms a baseline system based on dictionary-based query translation. Besides, the resulting CLQS is tested with French to English CLIR tasks on TREC collections. The results demonstrate higher effectiveness than the traditional query translation methods.

1. INTRODUCTION Query suggestion is a functionality to help users of a search engine to better specify their information need, by narrowing down or expanding the scope of the search with synonymous queries and relevant queries, or by suggesting related queries that have been frequently used by other users. Search engines, such as Google, Yahoo!, MSN, Ask Jeeves, all have implemented query suggestion functionality as a valuable addition to their core search method. In addition, the same technology has been leveraged to recommend bidding terms to online advertiser in the pay-forperformance search market [12]. Query suggestion is closely related to query expansion which extends the original query with new search terms to narrow the scope of the search. But different from query expansion, query suggestion aims to suggest full queries that have been formulated by users so that the query integrity and coherence are preserved in the suggested queries. Typical methods for query suggestion exploit query logs and document collections, by assuming that in the same period of time, many users share the same or similar interests, which can be expressed in different manners [12, 14, 26]. By suggesting the related and frequently used formulations, it is hoped that the new query can cover more relevant documents. However, all of the existing studies dealt with monolingual query suggestion and to our knowledge, there is no published study on cross-lingual query suggestion (CLQS). CLQS aims to suggest related queries but in a different language. It has wide applications on World Wide Web: for cross-language search or for suggesting relevant bidding terms in a different language. 1 CLQS can be approached as a query translation problem, i.e., to suggest the queries that are translations of the original query. Dictionaries, large size of parallel corpora and existing commercial machine translation systems can be used for translation. However, these kinds of approaches usually rely on static knowledge and data. It cannot effectively reflect the quickly shifting interests of Web users. Moreover, there are some problems with translated queries in target language. For instance, the translated terms can be reasonable translations, but they are not popularly used in the target language. For example, the French query aliment biologique is translated into biologic food by Google translation tool2 , yet the correct formulation nowadays should be organic food. Therefore, there exist many mismatch cases between the translated terms and the really used terms in target language. This mismatch makes the suggested terms in the target language ineffective. A natural thinking of solving this mismatch is to map the queries in the source language and the queries in the target language, by using the query log of a search engine. We exploit the fact that the users of search engines in the same period of time have similar interests, and they submit queries on similar topics in different languages. As a result, a query written in a source language likely has an equivalent in a query log in the target language. In particular, if the user intends to perform CLIR, then original query is even more likely to have its correspondent included in the target language query log. Therefore, if a candidate for CLQS appears often in the query log, then it is more likely the appropriate one to be suggested. In this paper, we propose a method of calculating the similarity between source language query and the target language query by exploiting, in addition to the translation information, a wide spectrum of bilingual and monolingual information, such as term co-occurrences, query logs with click-through data, etc. A discriminative model is used to learn the cross-lingual query similarity based on a set of manually translated queries. The model is trained by optimizing the cross-lingual similarity to best fit the monolingual similarity between one query and the other query"s translation. Besides being benchmarked as an independent module, the resulting CLQS system is tested as a new means of query translation in CLIR task on TREC collections. The results show that this new translation method is more effective than the traditional query translation method. The remainder of this paper is organized as follows: Section 2 introduces the related work; Section 3 describes in detail the discriminative model for estimating cross-lingual query similarity; Section 4 presents a new CLIR approach using cross-lingual query suggestion as a bridge across language boundaries. Section 5 discusses the experiments and benchmarks; finally, the paper is concluded in Section 6. 2. RELATED WORK Most approaches to CLIR perform a query translation followed by a monolingual IR. Typically, queries are translated either using a bilingual dictionary [22], a machine translation software [9] or a parallel corpus [20]. Despite the various types of resources used, out-of-vocabulary (OOV) words and translation disambiguation are the two major bottlenecks for CLIR [20]. In [7, 27], OOV term translations are mined from the Web using a search engine. In [17], bilingual knowledge is acquired based on anchor text analysis. In addition, word co-occurrence statistics in the target language has been leveraged for translation disambiguation [3, 10, 11, 19]. 2 http://www.google.com/language_tools Nevertheless, it is arguable that accurate query translation may not be necessary for CLIR. Indeed, in many cases, it is helpful to introduce words even if they are not direct translations of any query word, but are closely related to the meaning of the query. This observation has led to the development of cross-lingual query expansion (CLQE) techniques [2, 16, 18]. [2] reports the enhancement on CLIR by post-translation expansion. [16] develops a cross-lingual relevancy model by leveraging the crosslingual co-occurrence statistics in parallel texts. [18] makes performance comparison on multiple CLQE techniques, including pre-translation expansion and post-translation expansion. However, there is lack of a unified framework to combine the wide spectrum of resources and recent advances of mining techniques for CLQE. CLQS is different from CLQE in that it aims to suggest full queries that have been formulated by users in another language. As CLQS exploits up-to-date query logs, it is expected that for most user queries, we can find common formulations on these topics in the query log in the target language. Therefore, CLQS also plays a role of adapting the original query formulation to the common formulations of similar topics in the target language. Query logs have been successfully used for monolingual IR [8, 12, 15, 26], especially in monolingual query suggestions [12] and relating the semantically relevant terms for query expansion [8, 15]. In [1], the target language query log has been exploited to help query translation in CLIR. 3. ESTIMATING CROSS-LINGUAL QUERY SIMILARITY A search engine has a query log containing user queries in different languages within a certain period of time. In addition to query terms, click-through information is also recorded. Therefore, we know which documents have been selected by users for each query. Given a query in the source language, our CLQS task is to determine one or several similar queries in the target language from the query log. The key problem with cross-lingual query suggestion is how to learn a similarity measure between two queries in different languages. Although various statistical similarity measures have been studied for monolingual terms [8, 26], most of them are based on term co-occurrence statistics, and can hardly be applied directly in cross-lingual settings. In order to define a similarity measure across languages, one has to use at least one translation tool or resource. So the measure is based on both translation relation and monolingual similarity. In this paper, as our purpose is to provide up-to-date query similarity measure, it may not be sufficient to use only a static translation resource. Therefore, we also integrate a method to mine possible translations on the Web. This method is particularly useful for dealing with OOV terms. Given a set of resources of different natures, the next question is how to integrate them in a principled manner. In this paper, we propose a discriminative model to learn the appropriate similarity measure. The principle is as follows: we assume that we have a reasonable monolingual query similarity measure. For any training query example for which a translation exists, its similarity measure (with any other query) is transposed to its translation. Therefore, we have the desired cross-language similarity value for this example. Then we use a discriminative model to learn the cross-language similarity function which fits the best these examples. In the following sections, let us first describe the detail of the discriminative model for cross-lingual query similarity estimation. Then we introduce all the features (monolingual and cross-lingual information) that we will use in the discriminative model. 3.1 Discriminative Model for Estimating Cross-Lingual Query Similarity In this section, we propose a discriminative model to learn crosslingual query similarities in a principled manner. The principle is as follows: for a reasonable monolingual query similarity between two queries, a cross-lingual correspondent can be deduced between one query and another query"s translation. In other words, for a pair of queries in different languages, their crosslingual similarity should fit the monolingual similarity between one query and the other query"s translation. For example, the similarity between French query pages jaunes (i.e., yellow page in English) and English query telephone directory should be equal to the monolingual similarity between the translation of the French query yellow page and telephone directory. There are many ways to obtain a monolingual similarity measure between terms, e.g., term co-occurrence based mutual information and 2 χ . Any of them can be used as the target for the cross-lingual similarity function to fit. In this way, cross-lingual query similarity estimation is formulated as a regression task as follows: Given a source language query fq , a target language query eq , and a monolingual query similarity MLsim , the corresponding cross-lingual query similarity CLsim is defined as follows: ),(),( eqMLefCL qTsimqqsim f = (1) where fqT is the translation of fq in the target language. Based on Equation (1), it would be relatively easy to create a training corpus. All it requires is a list of query translations. Then an existing monolingual query suggestion system can be used to automatically produce similar query to each translation, and create the training corpus for cross-lingual similarity estimation. Another advantage is that it is fairly easy to make use of arbitrary information sources within a discriminative modeling framework to achieve optimal performance. In this paper, support vector machine (SVM) regression algorithm [25] is used to learn the cross-lingual term similarity function. Given a vector of feature functions f between fq and eq , ),( efCL ttsim is represented as an inner product between a weight vector and the feature vector in a kernel space as follows: )),((),( efefCL ttfwttsim φ•= (2) where φ is the mapping from the input feature space onto the kernel space, and wis the weight vector in the kernel space which will be learned by the SVM regression training. Once the weight vector is learned, the Equation (2) can be used to estimate the similarity between queries of different languages. We want to point out that instead of regression, one can definitely simplify the task as a binary or ordinal classification, in which case CLQS can be categorized according to discontinuous class labels, e.g., relevant and irrelevant, or a series of levels of relevancies, e.g., strongly relevant, weakly relevant, and irrelevant. In either case, one can resort to discriminative classification approaches, such as an SVM or maximum entropy model, in a straightforward way. However, the regression formalism enables us to fully rank the suggested queries based on the similarity score given by Equation (1). The Equations (1) and (2) construct a regression model for cross-lingual query similarity estimation. In the following sections, the monolingual query similarity measure (see Section 3.2) and the feature functions used for SVM regression (see Section 3.3) will be presented. 3.2 Monolingual Query Similarity Measure Based on Click-through Information Any monolingual term similarity measure can be used as the regression target. In this paper, we select the monolingual query similarity measure presented in [26] which reports good performance by using search users" click-through information in query logs. The reason to choose this monolingual similarity is that it is defined in a similar context as ours − according to a user log that reflects users" intention and behavior. Therefore, we can expect that the cross-language term similarity learned from it can also reflect users" intention and expectation. Following [26], our monolingual query similarity is defined by combining both query content-based similarity and click-through commonality in the query log. First the content similarity between two queries p and q is defined as follows: ))(),(( ),( ),( qknpknMax qpKN qpsimilarity content = (3) where )( xkn is the number of keywords in a query x, ),( qpKN is the number of common keywords in the two queries. Secondly, the click-through based similarity is defined as follows, ))(),(( ),( ),( qrdprdMax qpRD qpsimilarity throughclick =− (4) where )(xrd is the number of clicked URLs for a query x, and ),( qpRD is the number of common URLs clicked for two queries. Finally, the similarity between two queries is a linear combination of the content-based and click-through-based similarities, and is presented as follows: ),(* ),(*),( qpsimilarity qpsimilarityqpsimilarity throughclick content − += β α (5) where α and β are the relative importance of the two similarity measures. In this paper, we set ,4.0=α and 6.0=β following the practice in [26]. Queries with similarity measure higher than a threshold with another query will be regarded as relevant monolingual query suggestions (MLQS) for the latter. In this paper, the threshold is set as 0.9 empirically. 3.3 Features Used for Learning Cross-Lingual Query Similarity Measure This section presents the extraction of candidate relevant queries from the log with the assistance of various monolingual and bilingual resources. Meanwhile, feature functions over source query and the cross-lingual relevant candidates are defined. Some of the resources being used here, such as bilingual lexicon and parallel corpora, were for query translation in previous work. But note that we employ them here as an assistant means for finding relevant candidates in the log rather than for acquiring accurate translations. 3.3.1 Bilingual Dictionary In this subsection, a built-in-house bilingual dictionary containing 120,000 unique entries is used to retrieve candidate queries. Since multiple translations may be associated with each source word, co-occurrence based translation disambiguation is performed [3, 10]. The process is presented as follows: Given an input query }{ ,2,1 fnfff wwwq K= in the source language, for each query term fiw , a set of unique translations are provided by the bilingual dictionary D : },,{)( ,2,1 imiifi tttwD K= . Then the cohesion between the translations of two query terms is measured using mutual information which is computed as follows: )()( ),( log),()( , klij klij klijklij tPtP ttP ttPttMI = (6) where . )( )(, ),( ),( N tC tP N ttC ttP klij klij == Here ),( yxC is the number of queries in the log containing both x and y , )(xC is the number of queries containing term x , and N is the total number of queries in the log. Based on the term-term cohesion defined in Equation (6), all the possible query translations are ranked using the summation of the term-term cohesion ∑≠ = kiki klijqdict ttMITS f ,, ),()( . The set of top-4 query translations is denoted as )( fqTS . For each possible query translation )( fqTST∈ , we retrieve all the queries containing the same keywords as T from the target language log. The retrieved queries are candidate target queries, and are assigned )(TSdict as the value of the feature Dictionary-based Translation Score. 3.3.2 Parallel Corpora Parallel corpora are precious resources for bilingual knowledge acquisition. Different from the bilingual dictionary, the bilingual knowledge learned from parallel corpora assigns probability for each translation candidate which is useful in acquiring dominant query translations. In this paper, the Europarl corpus (a set of parallel French and English texts from the proceedings of the European Parliament) is used. The corpus is first sentence aligned. Then word alignments are derived by training an IBM translation model 1 [4] using GIZA++ [21]. The learned bilingual knowledge is used to extract candidate queries from the query log. The process is presented as follows: Given a pair of queries, fq in the source language and eq in the target language, the Bi-Directional Translation Score is defined as follows: )|()|(),( 111 feIBMefIBMefIBM qqpqqpqqS = (7) where )|(1 xypIBM is the word sequence translation probability given by IBM model 1 which has the following form: ∏∑= =+ = || 1 || 0 ||1 )|( )1|(| 1 )|( y j x i ijyIBM xyp x xyp (8) where )|( ij xyp is the word to word translation probability derived from the word-aligned corpora. The reason to use bidirectional translation probability is to deal with the fact that common words can be considered as possible translations of many words. By using bidirectional translation, we test whether the translation words can be translated back to the source words. This is helpful to focus on the translation probability onto the most specific translation candidates. Now, given an input query fq , the top 10 queries }{ eq with the highest bidirectional translation scores with fq are retrieved from the query log, and ),(1 efIBM qqS in Equation (7) is assigned as the value for the feature Bi-Directional Translation Score. 3.3.3 Online Mining for Related Queries OOV word translation is a major knowledge bottleneck for query translation and CLIR. To overcome this knowledge bottleneck, web mining has been exploited in [7, 27] to acquire EnglishChinese term translations based on the observation that Chinese terms may co-occur with their English translations in the same web page. In this section, this web mining approach is adapted to acquire not only translations but semantically related queries in the target language. It is assumed that if a query in the target language co-occurs with the source query in many web pages, they are probably semantically related. Therefore, a simple method is to send the source query to a search engine (Google in our case) for Web pages in the target language in order to find related queries in the target language. For instance, by sending a French query pages jaunes to search for English pages, the English snippets containing the key words yellow pages or telephone directory will be returned. However, this simple approach may induce significant amount of noise due to the non-relevant returns from the search engine. In order to improve the relevancy of the bilingual snippets, we extend the simple approach by the following query modification: the original query is used to search with the dictionary-based query keyword translations, which are unified by the ∧ (and) ∨ (OR) operators into a single Boolean query. For example, for a given query abcq = where the set of translation entries in the dictionary of for a is },,{ 321 aaa , b is },{ 21 bb and c is }{ 1c , we issue 121321 )()( cbbaaaq ∧∨∧∨∨∧ as one web query. From the returned top 700 snippets, the most frequent 10 target queries are identified, and are associated with the feature Frequency in the Snippets. Furthermore, we use Co-Occurrence Double-Check (CODC) Measure to weight the association between the source and target queries. CODC Measure is proposed in [6] as an association measure based on snippet analysis, named Web Search with Double Checking (WSDC) model. In WSDC model, two objects a and b are considered to have an association if b can be found by using a as query (forward process), and a can be found by using b as query (backward process) by web search. The forward process counts the frequency of b in the top N snippets of query a, denoted as )@( abfreq . Similarly, the backward process count the frequency of a in the top N snippets of query b, denoted as )@( bafreq . Then the CODC association score is defined as follows: ⎪ ⎩ ⎪ ⎨ ⎧ =× = ⎥ ⎥ ⎦ ⎤ ⎢ ⎢ ⎣ ⎡ × otherwise, 0)@()@(if,0 ),( )( )@( )( )@( log α e ef f fe qfreq qqfreq qfreq qqfreq effe efCODC e qqfreqqqfreq qqS (9) CODC measures the association of two terms in the range between 0 and 1, where under the two extreme cases, eq and fq are of no association when 0)@( =fe qqfreq or 0)@( =ef qqfreq , and are of the strongest association when )()@( ffe qfreqqqfreq = and )()@( eef qfreqqqfreq = . In our experiment, α is set at 0.15 following the practice in [6]. Any query eq mined from the Web will be associated with a feature CODC Measure with ),( efCODC qqS as its value. 3.3.4 Monolingual Query Suggestion For all the candidate queries 0Q being retrieved using dictionary (see Section 3.3.1), parallel data (see Section 3.3.2) and web mining (see Section 3.3.3), monolingual query suggestion system (described in Section 3.1) is called to produce more related queries in the target language. For each target query eq , its monolingual source query )( eML qSQ is defined as the query in 0Q with the highest monolingual similarity with eq , i.e., ),(maxarg)( 0 eeMLQqeML qqsimqSQ e ′= ∈′ (10) Then the monolingual similarity between eq and )( eML qSQ is used as the value of the eq "s Monolingual Query Suggestion Feature. For any target query 0Qq∈ , its Monolingual Query Suggestion Feature is set as 1. For any query 0Qqe ∉ , its values of Dictionary-based Translation Score, Bi-Directional Translation Score, Frequency in the Snippet, and CODC Measure are set to be equal to the feature values of )( eML qSQ . 3.4 Estimating Cross-lingual Query Similarity In summary, four categories of features are used to learn the crosslingual query similarity. SVM regression algorithm [25] is used to learn the weights in Equation (2). In this paper, LibSVM toolkit [5] is used for the regression training. In the prediction stage, the candidate queries will be ranked using the cross-lingual query similarity score computed in terms of )),((),( efefCL ttfwttsim φ•= , and the queries with similarity score lower than a threshold will be regarded as nonrelevant. The threshold is learned using a development data set by fitting MLQS"s output. 4. CLIR BASED ON CROSS-LINGUAL QUERY SUGGESTION In Section 3, we presented a discriminative model for cross lingual query suggestion. However, objectively benchmarking a query suggestion system is not a trivial task. In this paper, we propose to use CLQS as an alternative to query translation, and test its effectiveness in CLIR tasks. The resulting good performance of CLIR corresponds to the high quality of the suggested queries. Given a source query fq , a set of relevant queries }{ eq in the target language are recommended using the cross-lingual query suggestion system. Then a monolingual IR system based on the BM25 model [23] is called using each }{ eqq∈ as queries to retrieve documents. Then the retrieved documents are re-ranked based on the sum of the BM25 scores associated with each monolingual retrieval. 5. PERFORMACNCE EVALUATION In this section, we will benchmark the cross-lingual query suggestion system, comparing its performance with monolingual query suggestion, studying the contribution of various information sources, and testing its effectiveness when being used in CLIR tasks. 5.1 Data Resources In our experiments, French and English are selected as the source and target language respectively. Such selection is due to the fact that large scale query logs are readily available for these two languages. A one-month English query log (containing 7 million unique English queries with occurrence frequency more than 5) of MSN search engine is used as the target language log. And a monolingual query suggestion system is built based on it. In addition, 5,000 French queries are selected randomly from a French query log (containing around 3 million queries), and are manually translated into English by professional French-English translators. Among the 5,000 French queries, 4,171 queries have their translations in the English query log, and are used for CLQS training and testing. Furthermore, among the 4,171 French queries, 70% are used for cross-lingual query similarity training, 10% are used as the development data to determine the relevancy threshold, and 20% are used for testing. To retrieve the crosslingual related queries, a built-in-house French-English bilingual lexicon (containing 120,000 unique entries) and the Europarl corpus are used. Besides benchmarking CLQS as an independent system, the CLQS is also tested as a query translation system for CLIR tasks. Based on the observation that the CLIR performance heavily relies on the quality of the suggested queries, this benchmark measures the quality of CLQS in terms of its effectiveness in helping CLIR. To perform such benchmark, we use the documents of TREC6 CLIR data (AP88-90 newswire, 750MB) with officially provided 25 short French-English queries pairs (CL1-CL25). The selection of this data set is due to the fact that the average length of the queries are 3.3 words long, which matches the web query logs we use to train CLQS. 5.2 Performance of Cross-lingual Query Suggestion Mean-square-error (MSE) is used to measure the regression error and it is defined as follows: ( )2 ),(),( 1 ∑ −= i eiqMLeifiCL qTsimqqsim l MSE fi where l is the total number of cross-lingual query pairs in the testing data. As described in Section 3.4, a relevancy threshold is learned using the development data, and only CLQS with similarity value above the threshold is regarded as truly relevant to the input query. In this way, CLQS can also be benchmarked as a classification task using precision (P) and recall (R) which are defined as follows: CLQS MLQSCLQS P S SS I = , MLQS MLQSCLQS R S SS I = where CLQSS is the set of relevant queries suggested by CLQS, MLQSS is the set of relevant queries suggested by MLQS (see Section 3.2). The benchmarking results with various feature configurations are shown in Table 1. Regression Classification Features MSE P R DD 0.274 0.723 0.098 DD+PC 0.224 0.713 0.125 DD+PC+ Web 0.115 0.808 0.192 DD+PC+ Web+ML QS 0.174 0.796 0.421 Table 1. CLQS performance with different feature settings (DD: dictionary only; DD+PC: dictionary and parallel corpora; DD+PC+Web: dictionary, parallel corpora, and web mining; DD+PC+Web+MLQS: dictionary, parallel corpora, web mining and monolingual query suggestion) Table 1 reports the performance comparison with various feature settings. The baseline system (DD) uses a conventional query translation approach, i.e., a bilingual dictionary with cooccurrence-based translation disambiguation. The baseline system only covers less than 10% of the suggestions made by MLQS. Using additional features obviously enables CLQS to generate more relevant queries. The most significant improvement on recall is achieved by exploiting MLQS. The final CLQS system is able to generate 42% of the queries suggested by MLQS. Among all the feature combinations, there is no significant change in precision. This indicates that our methods can improve the recall by effectively leveraging various information sources without losing the accuracy of the suggestions. Besides benchmarking CLQS by comparing its output with MLQS output, 200 French queries are randomly selected from the French query log. These queries are double-checked to make sure that they are not in the CLQS training corpus. Then CLQS system is used to suggest relevant English queries for them. On average, for each French query, 8.7 relevant English queries are suggested. Then the total 1,740 suggested English queries are manually checked by two professional English/French translators with cross-validation. Among the 1,747 suggested queries, 1,407 queries are recognized as relevant to the original ones, hence the accuracy is 80.9%. Figure 1 shows an example of CLQS of the French query terrorisme international (international terrorism in English). 5.3 CLIR Performance In this section, CLQS is tested with French to English CLIR tasks. We conduct CLIR experiments using the TREC 6 CLIR dataset described in Section 5.1. The CLIR is performed using a query translation system followed by a BM25-based [23] monolingual IR module. The following three different systems have been used to perform query translation: (1) CLQS: our CLQS system; (2) MT: Google French to English machine translation system; (3) DT: a dictionary based query translation system using cooccurrence statistics for translation disambiguation. The translation disambiguation algorithm is presented in Section 3.3.1. Besides, the monolingual IR performance is also reported as a reference. The average precision of the four IR systems are reported in Table 2, and the 11-point precision-recall curves are shown in Figure 2. Table 2. Average precision of CLIR on TREC 6 Dataset (Monolingual: monolingual IR system; MT: CLIR based on machine translation; DT: CLIR based on dictionary translation; CLQS: CLQS-based CLIR) IR System Average Precision % of Monolingual IR Monolingual 0.266 100% MT 0.217 81.6% DT 0.186 69.9% CLQS 0.233 87.6% Figure 1. An example of CLQS of the French query terrorisme international international terrorism (0.991); what is terrorism (0.943); counter terrorism (0.920); terrorist (0.911); terrorist attacks (0.898); international terrorist (0.853); world terrorism (0.845); global terrorism (0.833); transnational terrorism (0.821); human rights (0.811); terrorist groups (0. 777); patterns of global terrorism (0.762) september 11 (0.734) 11-point P-R curves (TREC6) 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 Recall Precison Monolingual MT DT CLQS The benchmark shows that using CLQS as a query translation tool outperforms CLIR based on machine translation by 7.4%, outperforms CLIR based on dictionary translation by 25.2%, and achieves 87.6% of the monolingual IR performance. The effectiveness of CLQS lies in its ability in suggesting closely related queries besides accurate translations. For example, for the query CL14 terrorisme international (international terrorism), although the machine translation tool translates the query correctly, CLQS system still achieves higher score by recommending many additional related terms such as global terrorism, world terrorism, etc. (as shown in Figure 1). Another example is the query La pollution causée par l'automobile (air pollution due to automobile) of CL6. The MT tool provides the translation the pollution caused by the car, while CLQS system enumerates all the possible synonyms of car, and suggest the following queries car pollution, auto pollution, automobile pollution. Besides, other related queries such as global warming are also suggested. For the query CL12 La culture écologique (organic farming), the MT tool fails to generate the correct translation. Although the correct translation is neither in our French-English dictionary, CLQS system generates organic farm as a relevant query due to successful web mining. The above experiment demonstrates the effectiveness of using CLQS to suggest relevant queries for CLIR enhancement. A related research is to perform query expansion to enhance CLIR [2, 18]. So it is very interesting to compare the CLQS approach with the conventional query expansion approaches. Following [18], post-translation expansion is performed based on pseudorelevance feedback (PRF) techniques. We first perform CLIR in the same way as before. Then we use the traditional PRF algorithm described in [24] to select expansion terms. In our experiments, the top 10 terms are selected to expand the original query, and the new query is used to search the collection for the second time. The new CLIR performance in terms of average precision is shown in Table 3. The 11-point P-R curves are drawn in Figure 3. Although being enhanced by pseudo-relevance feedback, the CLIR using either machine translation or dictionary-based query translation still does not perform as well as CLQS-based approach. Statistical t-test [13] is conducted to indicate whether the CLQS-based CLIR performs significantly better. Pair-wise pvalues are shown in Table 4. Clearly, CLQS significantly outperforms MT and DT without PRF as well as DT+PRF, but its superiority over MT+PRF is not significant. However, when combined with PRF, CLQS significant outperforms all the other methods. This indicates the higher effectiveness of CLQS in related term identification by leveraging a wide spectrum of resources. Furthermore, post-translation expansion is capable of improving CLQS-based CLIR. This is due to the fact that CLQS and pseudo-relevance feedback are leveraging different categories of resources, and both approaches can be complementary. IR System AP without PRF AP with PRF Monolingual 0.266 (100%) 0.288 (100%) MT 0.217 (81.6%) 0.222 (77.1%) DT 0.186 (69.9%) 0.220 (76.4%) CLQS 0.233 (87.6%) 0.259 (89.9%) 11-point P-R curves with pseudo relevance feedback (TREC6) 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 Recall Precison Monolingual MT DT CLQS MT DT MT+PRF DT+PRF CLQS 0.0298 3.84e-05 0.1472 0.0282 CLQS+PR F 0.0026 2.63e-05 0.0094 0.0016 6. CONCLUSIONS In this paper, we proposed a new approach to cross-lingual query suggestion by mining relevant queries in different languages from query logs. The key solution to this problem is to learn a crosslingual query similarity measure by a discriminative model exploiting multiple monolingual and bilingual resources. The model is trained based on the principle that cross-lingual similarity should best fit the monolingual similarity between one query and the other query"s translation. Figure 2. 11 points precision-recall on TREC6 CLIR data set Figure 3. 11 points precision-recall on TREC6 CLIR dataset with pseudo relevance feedback Table 3. Comparison of average precision (AP) on TREC 6 without and with post-translation expansion. (%) are the relative percentages over the monolingual IR performance Table 4. The results of pair-wise significance t-test. Here pvalue < 0.05 is considered statistically significant The baseline CLQS system applies a typical query translation approach, using a bilingual dictionary with co-occurrence-based translation disambiguation. This approach only covers 10% of the relevant queries suggested by an MLQS system (when the exact translation of the original query is given). By leveraging additional resources such as parallel corpora, web mining and logbased monolingual query expansion, the final system is able to cover 42% of the relevant queries suggested by an MLQS system with precision as high as 79.6%. To further test the quality of the suggested queries, CLQS system is used as a query translation system in CLIR tasks. Benchmarked using TREC 6 French to English CLIR task, CLQS demonstrates higher effectiveness than the traditional query translation methods using either bilingual dictionary or commercial machine translation tools. The improvement on TREC French to English CLIR task by using CLQS demonstrates the high quality of the suggested queries. This also shows the strong correspondence between the input French queries and English queries in the log. In the future, we will build CLQS system between languages which may be more loosely correlated, e.g., English and Chinese, and study the CLQS performance change due to the less strong correspondence among queries in such languages. 7. REFERENCES [1] Ambati, V. and Rohini., U. Using Monolingual Clickthrough Data to Build Cross-lingual Search Systems. In Proceedings of New Directions in Multilingual Information Access Workshop of SIGIR 2006. [2] Ballestors, L. A. and Croft, W. B. Phrasal Translation and Query Expansion Techniques for Cross-Language Information Retrieval. In Proc. SIGIR 1997, pp. 84-91. [3] Ballestors, L. A. and Croft, W. B. Resolving Ambiguity for Cross-Language Retrieval. In Proc. SIGIR 1998, pp. 64-71. [4] Brown, P. F., Pietra, D. S. A., Pietra, D. V. J., and Mercer, R. L. The Mathematics of Statistical Machine Translation: Parameter Estimation. Computational Linguistics, 19(2):263311, 1993. [5] Chang, C. C. and Lin, C. LIBSVM: a Library for Support Vector Machines (Version 2.3). 2001. http://citeseer.ist.psu.edu/chang01libsvm.html [6] Chen, H.-H., Lin, M.-S., and Wei, Y.-C. Novel Association Measures Using Web Search with Double Checking. In Proc. COLING/ACL 2006, pp. 1009-1016. [7] Cheng, P.-J., Teng, J.-W., Chen, R.-C., Wang, J.-H., Lu, W.H., and Chien, L.-F. Translating Unknown Queries with Web Corpora for Cross-Language Information Retrieval. In Proc. SIGIR 2004, pp. 146-153. [8] Cui, H., Wen, J. R., Nie, J.-Y., and Ma, W. Y. Query Expansion by Mining User Logs. IEEE Trans. on Knowledge and Data Engineering, 15(4):829-839, 2003. [9] Fujii A. and Ishikawa, T. Applying Machine Translation to Two-Stage Cross-Language Information Retrieval. In Proceedings of 4th Conference of the Association for Machine Translation in the Americas, pp. 13-24, 2000. [10] Gao, J. F., Nie, J.-Y., Xun, E., Zhang, J., Zhou, M., and Huang, C. Improving query translation for CLIR using statistical Models. In Proc. SIGIR 2001, pp. 96-104. [11] Gao, J. F., Nie, J.-Y., He, H., Chen, W., and Zhou, M. Resolving Query Translation Ambiguity using a Decaying Co-occurrence Model and Syntactic Dependence Relations. In Proc. SIGIR 2002, pp. 183-190. [12] Gleich, D., and Zhukov, L. SVD Subspace Projections for Term Suggestion Ranking and Clustering. In Technical Report, Yahoo! Research Labs, 2004. [13] Hull, D. Using Statistical Testing in the Evaluation of Retrieval Experiments. In Proc. SIGIR 1993, pp. 329-338. [14] Jeon, J., Croft, W. B., and Lee, J. Finding Similar Questions in Large Question and Answer Archives. In Proc. CIKM 2005, pp. 84-90. [15] Joachims, T. Optimizing Search Engines Using Clickthrough Data. In Proc. SIGKDD 2002, pp. 133-142. [16] Lavrenko, V., Choquette, M., and Croft, W. B. Cross-Lingual Relevance Models. In Proc. SIGIR 2002, pp. 175-182. [17] Lu, W.-H., Chien, L.-F., and Lee, H.-J. Anchor Text Mining for Translation Extraction of Query Terms. In Proc. SIGIR 2001, pp. 388-389. [18] McNamee, P. and Mayfield, J. Comparing Cross-Language Query Expansion Techniques by Degrading Translation Resources. In Proc. SIGIR 2002, pp. 159-166. [19] Monz, C. and Dorr, B. J. Iterative Translation Disambiguation for Cross-Language Information Retrieval. In Proc. SIGIR 2005, pp. 520-527. [20] Nie, J.-Y., Simard, M., Isabelle, P., and Durand, R. CrossLanguage Information Retrieval based on Parallel Text and Automatic Mining of Parallel Text from the Web. In Proc. SIGIR 1999, pp. 74-81. [21] Och, F. J. and Ney, H. A Systematic Comparison of Various Statistical Alignment Models. Computational Linguistics, 29(1):19-51, 2003. [22] Pirkola, A., Hedlund, T., Keshusalo, H., and Järvelin, K. Dictionary-Based Cross-Language Information Retrieval: Problems, Methods, and Research Findings. Information Retrieval, 4(3/4):209-230, 2001. [23] Robertson, S. E., Walker, S., Hancock-Beaulieu, M. M., and Gatford, M. OKAPI at TREC-3. In Proc.TREC-3, pp. 200225, 1995. [24] Robertson, S. E. and Jones, K. S. Relevance Weighting of Search Terms. Journal of the American Society of Information Science, 27(3):129-146, 1976. [25] Smola, A. J. and Schölkopf, B. A. Tutorial on Support Vector Regression. Statistics and Computing, 14(3):199-222, 2004. [26] Wen, J. R., Nie, J.-Y., and Zhang, H. J. Query Clustering Using User Logs. ACM Trans. Information Systems, 20(1):59-81, 2002. [27] Zhang, Y. and Vines, P. Using the Web for Automated Translation Extraction in Cross-Language Information Retrieval. In Proc. SIGIR 2004, pp. 162-169.

bidding term;benchmark;target language query log;map;query expansion;query suggestion;query log;cross-language information retrieval;query translation;search engine advertisement;search engine;keyword bidding;monolingual query suggestion

train_H-41

HITS on the Web: How does it Compare?

This paper describes a large-scale evaluation of the effectiveness of HITS in comparison with other link-based ranking algorithms, when used in combination with a state-ofthe-art text retrieval algorithm exploiting anchor text. We quantified their effectiveness using three common performance measures: the mean reciprocal rank, the mean average precision, and the normalized discounted cumulative gain measurements. The evaluation is based on two large data sets: a breadth-first search crawl of 463 million web pages containing 17.6 billion hyperlinks and referencing 2.9 billion distinct URLs; and a set of 28,043 queries sampled from a query log, each query having on average 2,383 results, about 17 of which were labeled by judges. We found that HITS outperforms PageRank, but is about as effective as web-page in-degree. The same holds true when any of the link-based features are combined with the text retrieval algorithm. Finally, we studied the relationship between query specificity and the effectiveness of selected features, and found that link-based features perform better for general queries, whereas BM25F performs better for specific queries.

1. INTRODUCTION Link graph features such as in-degree and PageRank have been shown to significantly improve the performance of text retrieval algorithms on the web. The HITS algorithm is also believed to be of interest for web search; to some degree, one may expect HITS to be more informative that other link-based features because it is query-dependent: it tries to measure the interest of pages with respect to a given query. However, it remains unclear today whether there are practical benefits of HITS over other link graph measures. This is even more true when we consider that modern retrieval algorithms used on the web use a document representation which incorporates the document"s anchor text, i.e. the text of incoming links. This, at least to some degree, takes the link graph into account, in a query-dependent manner. Comparing HITS to PageRank or in-degree empirically is no easy task. There are two main difficulties: scale and relevance. Scale is important because link-based features are known to improve in quality as the document graph grows. If we carry out a small experiment, our conclusions won"t carry over to large graphs such as the web. However, computing HITS efficiently on a graph the size of a realistic web crawl is extraordinarily difficult. Relevance is also crucial because we cannot measure the performance of a feature in the absence of human judgments: what is crucial is ranking at the top of the ten or so documents that a user will peruse. To our knowledge, this paper is the first attempt to evaluate HITS at a large scale and compare it to other link-based features with respect to human evaluated judgment. Our results confirm many of the intuitions we have about link-based features and their relationship to text retrieval methods exploiting anchor text. This is reassuring: in the absence of a theoretical model capable of tying these measures with relevance, the only way to validate our intuitions is to carry out realistic experiments. However, we were quite surprised to find that HITS, a query-dependent feature, is about as effective as web page in-degree, the most simpleminded query-independent link-based feature. This continues to be true when the link-based features are combined with a text retrieval algorithm exploiting anchor text. The remainder of this paper is structured as follows: Section 2 surveys related work. Section 3 describes the data sets we used in our study. Section 4 reviews the performance measures we used. Sections 5 and 6 describe the PageRank and HITS algorithms in more detail, and sketch the computational infrastructure we employed to carry out large scale experiments. Section 7 presents the results of our evaluations, and Section 8 offers concluding remarks. 2. RELATED WORK The idea of using hyperlink analysis for ranking web search results arose around 1997, and manifested itself in the HITS [16, 17] and PageRank [5, 21] algorithms. The popularity of these two algorithms and the phenomenal success of the Google search engine, which uses PageRank, have spawned a large amount of subsequent research. There are numerous attempts at improving the effectiveness of HITS and PageRank. Query-dependent link-based ranking algorithms inspired by HITS include SALSA [19], Randomized HITS [20], and PHITS [7], to name a few. Query-independent link-based ranking algorithms inspired by PageRank include TrafficRank [22], BlockRank [14], and TrustRank [11], and many others. Another line of research is concerned with analyzing the mathematical properties of HITS and PageRank. For example, Borodin et al. [3] investigated various theoretical properties of PageRank, HITS, SALSA, and PHITS, including their similarity and stability, while Bianchini et al. [2] studied the relationship between the structure of the web graph and the distribution of PageRank scores, and Langville and Meyer examined basic properties of PageRank such as existence and uniqueness of an eigenvector and convergence of power iteration [18]. Given the attention that has been paid to improving the effectiveness of PageRank and HITS, and the thorough studies of the mathematical properties of these algorithms, it is somewhat surprising that very few evaluations of their effectiveness have been published. We are aware of two studies that have attempted to formally evaluate the effectiveness of HITS and of PageRank. Amento et al. [1] employed quantitative measures, but based their experiments on the result sets of just 5 queries and the web-graph induced by topical crawls around the result set of each query. A more recent study by Borodin et al. [4] is based on 34 queries, result sets of 200 pages per query obtained from Google, and a neighborhood graph derived by retrieving 50 in-links per result from Google. By contrast, our study is based on over 28,000 queries and a web graph covering 2.9 billion URLs. 3. OUR DATA SETS Our evaluation is based on two data sets: a large web graph and a substantial set of queries with associated results, some of which were labeled by human judges. Our web graph is based on a web crawl that was conducted in a breadth-first-search fashion, and successfully retrieved 463,685,607 HTML pages. These pages contain 17,672,011,890 hyperlinks (after eliminating duplicate hyperlinks embedded in the same web page), which refer to a total of 2,897,671,002 URLs. Thus, at the end of the crawl there were 2,433,985,395 URLs in the frontier set of the crawler that had been discovered, but not yet downloaded. The mean out-degree of crawled web pages is 38.11; the mean in-degree of discovered pages (whether crawled or not) is 6.10. Also, it is worth pointing out that there is a lot more variance in in-degrees than in out-degrees; some popular pages have millions of incoming links. As we will see, this property affects the computational cost of HITS. Our query set was produced by sampling 28,043 queries from the MSN Search query log, and retrieving a total of 66,846,214 result URLs for these queries (using commercial search engine technology), or about 2,838 results per query on average. It is important to point out that our 2.9 billion URL web graph does not cover all these result URLs. In fact, only 9,525,566 of the result URLs (about 14.25%) were covered by the graph. 485,656 of the results in the query set (about 0.73% of all results, or about 17.3 results per query) were rated by human judges as to their relevance to the given query, and labeled on a six-point scale (the labels being definitive, excellent, good, fair, bad and detrimental). Results were selected for judgment based on their commercial search engine placement; in other words, the subset of labeled results is not random, but biased towards documents considered relevant by pre-existing ranking algorithms. Involving a human in the evaluation process is extremely cumbersome and expensive; however, human judgments are crucial for the evaluation of search engines. This is so because no document features have been found yet that can effectively estimate the relevance of a document to a user query. Since content-match features are very unreliable (and even more so link features, as we will see) we need to ask a human to evaluate the results in order to compare the quality of features. Evaluating the retrieval results from document scores and human judgments is not trivial and has been the subject of many investigations in the IR community. A good performance measure should correlate with user satisfaction, taking into account that users will dislike having to delve deep in the results to find relevant documents. For this reason, standard correlation measures (such as the correlation coefficient between the score and the judgment of a document), or order correlation measures (such as Kendall tau between the score and judgment induced orders) are not adequate. 4. MEASURING PERFORMANCE In this study, we quantify the effectiveness of various ranking algorithms using three measures: NDCG, MRR, and MAP. The normalized discounted cumulative gains (NDCG) measure [13] discounts the contribution of a document to the overall score as the document"s rank increases (assuming that the best document has the lowest rank). Such a measure is particularly appropriate for search engines, as studies have shown that search engine users rarely consider anything beyond the first few results [12]. NDCG values are normalized to be between 0 and 1, with 1 being the NDCG of a perfect ranking scheme that completely agrees with the assessment of the human judges. The discounted cumulative gain at a particular rank-threshold T (DCG@T) is defined to be PT j=1 1 log(1+j) 2r(j) − 1 , where r(j) is the rating (0=detrimental, 1=bad, 2=fair, 3=good, 4=excellent, and 5=definitive) at rank j. The NDCG is computed by dividing the DCG of a ranking by the highest possible DCG that can be obtained for that query. Finally, the NDGCs of all queries in the query set are averaged to produce a mean NDCG. The reciprocal rank (RR) of the ranked result set of a query is defined to be the reciprocal value of the rank of the highest-ranking relevant document in the result set. The RR at rank-threshold T is defined to be 0 if none of the highestranking T documents is relevant. The mean reciprocal rank (MRR) of a query set is the average reciprocal rank of all queries in the query set. Given a ranked set of n results, let rel(i) be 1 if the result at rank i is relevant and 0 otherwise. The precision P(j) at rank j is defined to be 1 j Pj i=1 rel(i), i.e. the fraction of the relevant results among the j highest-ranking results. The average precision (AP) at rank-threshold k is defined to be Pk i=1 P (i)rel(i) Pn i=1 rel(i) . The mean average precision (MAP) of a query set is the mean of the average precisions of all queries in the query set. The above definitions of MRR and MAP rely on the notion of a relevant result. We investigated two definitions of relevance: One where all documents rated fair or better were deemed relevant, and one were all documents rated good or better were deemed relevant. For reasons of space, we only report MAP and MRR values computed using the latter definition; using the former definition does not change the qualitative nature of our findings. Similarly, we computed NDCG, MAP, and MRR values for a wide range of rank-thresholds; we report results here at rank 10; again, changing the rank-threshold never led us to different conclusions. Recall that over 99% of documents are unlabeled. We chose to treat all these documents as irrelevant to the query. For some queries, however, not all relevant documents have been judged. This introduces a bias into our evaluation: features that bring new documents to the top of the rank may be penalized. This will be more acute for features less correlated to the pre-existing commercial ranking algorithms used to select documents for judgment. On the other hand, most queries have few perfect relevant documents (i.e. home page or item searches) and they will most often be within the judged set. 5. COMPUTING PAGERANK ON A LARGE WEB GRAPH PageRank is a query-independent measure of the importance of web pages, based on the notion of peer-endorsement: A hyperlink from page A to page B is interpreted as an endorsement of page B"s content by page A"s author. The following recursive definition captures this notion of endorsement: R(v) = X (u,v)∈E R(u) Out(u) where R(v) is the score (importance) of page v, (u, v) is an edge (hyperlink) from page u to page v contained in the edge set E of the web graph, and Out(u) is the out-degree (number of embedded hyperlinks) of page u. However, this definition suffers from a severe shortcoming: In the fixedpoint of this recursive equation, only edges that are part of a strongly-connected component receive a non-zero score. In order to overcome this deficiency, Page et al. grant each page a guaranteed minimum score, giving rise to the definition of standard PageRank: R(v) = d |V | + (1 − d) X (u,v)∈E R(u) Out(u) where |V | is the size of the vertex set (the number of known web pages), and d is a damping factor, typically set to be between 0.1 and 0.2. Assuming that scores are normalized to sum up to 1, PageRank can be viewed as the stationary probability distribution of a random walk on the web graph, where at each step of the walk, the walker with probability 1 − d moves from its current node u to a neighboring node v, and with probability d selects a node uniformly at random from all nodes in the graph and jumps to it. In the limit, the random walker is at node v with probability R(v). One issue that has to be addressed when implementing PageRank is how to deal with sink nodes, nodes that do not have any outgoing links. One possibility would be to select another node uniformly at random and transition to it; this is equivalent to adding edges from each sink nodes to all other nodes in the graph. We chose the alternative approach of introducing a single phantom node. Each sink node has an edge to the phantom node, and the phantom node has an edge to itself. In practice, PageRank scores can be computed using power iteration. Since PageRank is query-independent, the computation can be performed off-line ahead of query time. This property has been key to PageRank"s success, since it is a challenging engineering problem to build a system that can perform any non-trivial computation on the web graph at query time. In order to compute PageRank scores for all 2.9 billion nodes in our web graph, we implemented a distributed version of PageRank. The computation consists of two distinct phases. In the first phase, the link files produced by the web crawler, which contain page URLs and their associated link URLs in textual form, are partitioned among the machines in the cluster used to compute PageRank scores, and converted into a more compact format along the way. Specifically, URLs are partitioned across the machines in the cluster based on a hash of the URLs" host component, and each machine in the cluster maintains a table mapping the URL to a 32-bit integer. The integers are drawn from a densely packed space, so as to make suitable indices into the array that will later hold the PageRank scores. The system then translates our log of pages and their associated hyperlinks into a compact representation where both page URLs and link URLs are represented by their associated 32-bit integers. Hashing the host component of the URLs guarantees that all URLs from the same host are assigned to the same machine in our scoring cluster. Since over 80% of all hyperlinks on the web are relative (that is, are between two pages on the same host), this property greatly reduces the amount of network communication required by the second stage of the distributed scoring computation. The second phase performs the actual PageRank power iteration. Both the link data and the current PageRank vector reside on disk and are read in a streaming fashion; while the new PageRank vector is maintained in memory. We represent PageRank scores as 64-bit floating point numbers. PageRank contributions to pages assigned to remote machines are streamed to the remote machine via a TCP connection. We used a three-machine cluster, each machine equipped with 16 GB of RAM, to compute standard PageRank scores for all 2.9 billion URLs that were contained in our web graph. We used a damping factor of 0.15, and performed 200 power iterations. Starting at iteration 165, the L∞ norm of the change in the PageRank vector from one iteration to the next had stopped decreasing, indicating that we had reached as much of a fixed point as the limitations of 64-bit floating point arithmetic would allow. 0.07 0.08 0.09 0.10 0.11 1 10 100 Number of back-links sampled per result NDCG@10 hits-aut-all hits-aut-ih hits-aut-id 0.01 0.02 0.03 0.04 1 10 100 Number of back-links sampled per result MAP@10 hits-aut-all hits-aut-ih hits-aut-id 0.07 0.08 0.09 0.10 0.11 0.12 0.13 0.14 1 10 100 Number of back-links sampled per result MRR@10 hits-aut-all hits-aut-ih hits-aut-id Figure 1: Effectiveness of authority scores computed using different parameterizations of HITS. A post-processing phase uses the final PageRank vectors (one per machine) and the table mapping URLs to 32-bit integers (representing indices into each PageRank vector) to score the result URL in our query log. As mentioned above, our web graph covered 9,525,566 of the 66,846,214 result URLs. These URLs were annotated with their computed PageRank score; all other URLs received a score of 0. 6. HITS HITS, unlike PageRank, is a query-dependent ranking algorithm. HITS (which stands for Hypertext Induced Topic Search) is based on the following two intuitions: First, hyperlinks can be viewed as topical endorsements: A hyperlink from a page u devoted to topic T to another page v is likely to endorse the authority of v with respect to topic T. Second, the result set of a particular query is likely to have a certain amount of topical coherence. Therefore, it makes sense to perform link analysis not on the entire web graph, but rather on just the neighborhood of pages contained in the result set, since this neighborhood is more likely to contain topically relevant links. But while the set of nodes immediately reachable from the result set is manageable (given that most pages have only a limited number of hyperlinks embedded into them), the set of pages immediately leading to the result set can be enormous. For this reason, Kleinberg suggests sampling a fixed-size random subset of the pages linking to any high-indegree page in the result set. Moreover, Kleinberg suggests considering only links that cross host boundaries, the rationale being that links between pages on the same host (intrinsic links) are likely to be navigational or nepotistic and not topically relevant. Given a web graph (V, E) with vertex set V and edge set E ⊆ V × V , and the set of result URLs to a query (called the root set R ⊆ V ) as input, HITS computes a neighborhood graph consisting of a base set B ⊆ V (the root set and some of its neighboring vertices) and some of the edges in E induced by B. In order to formalize the definition of the neighborhood graph, it is helpful to first introduce a sampling operator and the concept of a linkselection predicate. Given a set A, the notation Sn[A] draws n elements uniformly at random from A; Sn[A] = A if |A| ≤ n. A link section predicate P takes an edge (u, v) ∈ E. In this study, we use the following three link section predicates: all(u, v) ⇔ true ih(u, v) ⇔ host(u) = host(v) id(u, v) ⇔ domain(u) = domain(v) where host(u) denotes the host of URL u, and domain(u) denotes the domain of URL u. So, all is true for all links, whereas ih is true only for inter-host links, and id is true only for inter-domain links. The outlinked-set OP of the root set R w.r.t. a linkselection predicate P is defined to be: OP = [ u∈R {v ∈ V : (u, v) ∈ E ∧ P(u, v)} The inlinking-set IP s of the root set R w.r.t. a link-selection predicate P and a sampling value s is defined to be: IP s = [ v∈R Ss[{u ∈ V : (u, v) ∈ E ∧ P(u, v)}] The base set BP s of the root set R w.r.t. P and s is defined to be: BP s = R ∪ IP s ∪ OP The neighborhood graph (BP s , NP s ) has the base set BP s as its vertex set and an edge set NP s containing those edges in E that are covered by BP s and permitted by P: NP s = {(u, v) ∈ E : u ∈ BP s ∧ v ∈ BP s ∧ P(u, v)} To simplify notation, we write B to denote BP s , and N to denote NP s . For each node u in the neighborhood graph, HITS computes two scores: an authority score A(u), estimating how authoritative u is on the topic induced by the query, and a hub score H(u), indicating whether u is a good reference to many authoritative pages. This is done using the following algorithm: 1. For all u ∈ B do H(u) := q 1 |B| , A(u) := q 1 |B| . 2. Repeat until H and A converge: (a) For all v ∈ B : A (v) := P (u,v)∈N H(u) (b) For all u ∈ B : H (u) := P (u,v)∈N A(v) (c) H := H 2, A := A 2 where X 2 normalizes the vector X to unit length in euclidean space, i.e. the squares of its elements sum up to 1. In practice, implementing a system that can compute HITS within the time constraints of a major search engine (where the peak query load is in the thousands of queries per second, and the desired query response time is well below one second) is a major engineering challenge. Among other things, the web graph cannot reasonably be stored on disk, since .221 .106 .105 .104 .102 .095 .092 .090 .038 .036 .035 .034 .032 .032 .011 0.00 0.05 0.10 0.15 0.20 0.25 bm25f degree-in-id degree-in-ih hits-aut-id-25 hits-aut-ih-100 degree-in-all pagerank hits-aut-all-100 hits-hub-all-100 hits-hub-ih-100 hits-hub-id-100 degree-out-all degree-out-ih degree-out-id random NDCG@10 .100 .035 .033 .033 .033 .029 .027 .027 .008 .007 .007 .006 .006 .006 .002 0.00 0.02 0.04 0.06 0.08 0.10 0.12 bm25f hits-aut-id-9 degree-in-id hits-aut-ih-15 degree-in-ih degree-in-all pagerank hits-aut-all-100 hits-hub-all-100 hits-hub-ih-100 hits-hub-id-100 degree-out-all degree-out-ih degree-out-id random MAP@10 .273 .132 .126 .117 .114 .101 .101 .097 .032 .032 .030 .028 .027 .027 .007 0.00 0.05 0.10 0.15 0.20 0.25 0.30 bm25f hits-aut-id-9 hits-aut-ih-15 degree-in-id degree-in-ih degree-in-all hits-aut-all-100 pagerank hits-hub-all-100 hits-hub-ih-100 hits-hub-id-100 degree-out-all degree-out-ih degree-out-id random MRR@10 Figure 2: Effectiveness of different features. seek times of modern hard disks are too slow to retrieve the links within the time constraints, and the graph does not fit into the main memory of a single machine, even when using the most aggressive compression techniques. In order to experiment with HITS and other query-dependent link-based ranking algorithms that require non-regular accesses to arbitrary nodes and edges in the web graph, we implemented a system called the Scalable Hyperlink Store, or SHS for short. SHS is a special-purpose database, distributed over an arbitrary number of machines that keeps a highly compressed version of the web graph in memory and allows very fast lookup of nodes and edges. On our hardware, it takes an average of 2 microseconds to map a URL to a 64-bit integer handle called a UID, 15 microseconds to look up all incoming or outgoing link UIDs associated with a page UID, and 5 microseconds to map a UID back to a URL (the last functionality not being required by HITS). The RPC overhead is about 100 microseconds, but the SHS API allows many lookups to be batched into a single RPC request. We implemented the HITS algorithm using the SHS infrastructure. We compiled three SHS databases, one containing all 17.6 billion links in our web graph (all), one containing only links between pages that are on different hosts (ih, for inter-host), and one containing only links between pages that are on different domains (id). We consider two URLs to belong to different hosts if the host portions of the URLs differ (in other words, we make no attempt to determine whether two distinct symbolic host names refer to the same computer), and we consider a domain to be the name purchased from a registrar (for example, we consider news.bbc.co.uk and www.bbc.co.uk to be different hosts belonging to the same domain). Using each of these databases, we computed HITS authority and hub scores for various parameterizations of the sampling operator S, sampling between 1 and 100 back-links of each page in the root set. Result URLs that were not covered by our web graph automatically received authority and hub scores of 0, since they were not connected to any other nodes in the neighborhood graph and therefore did not receive any endorsements. We performed forty-five different HITS computations, each combining one of the three link selection predicates (all, ih, and id) with a sampling value. For each combination, we loaded one of the three databases into an SHS system running on six machines (each equipped with 16 GB of RAM), and computed HITS authority and hub scores, one query at a time. The longest-running combination (using the all database and sampling 100 back-links of each root set vertex) required 30,456 seconds to process the entire query set, or about 1.1 seconds per query on average. 7. EXPERIMENTAL RESULTS For a given query Q, we need to rank the set of documents satisfying Q (the result set of Q). Our hypothesis is that good features should be able to rank relevant documents in this set higher than non-relevant ones, and this should result in an increase in each performance measure over the query set. We are specifically interested in evaluating the usefulness of HITS and other link-based features. In principle, we could do this by sorting the documents in each result set by their feature value, and compare the resulting NDCGs. We call this ranking with isolated features. Let us first examine the relative performance of the different parameterizations of the HITS algorithm we examined. Recall that we computed HITS for each combination of three link section schemes - all links (all), inter-host links only (ih), and inter-domain links only (id) - with back-link sampling values ranging from 1 to 100. Figure 1 shows the impact of the number of sampled back-links on the retrieval performance of HITS authority scores. Each graph is associated with one performance measure. The horizontal axis of each graph represents the number of sampled back-links, the vertical axis represents performance under the appropriate measure, and each curve depicts a link selection scheme. The id scheme slightly outperforms ih, and both vastly outperform the all scheme - eliminating nepotistic links pays off. The performance of the all scheme increases as more back-links of each root set vertex are sampled, while the performance of the id and ih schemes peaks at between 10 and 25 samples and then plateaus or even declines, depending on the performance measure. Having compared different parameterizations of HITS, we will now fix the number of sampled back-links at 100 and compare the three link selection schemes against other isolated features: PageRank, in-degree and out-degree counting links of all pages, of different hosts only and of different domains only (all, ih and id datasets respectively), and a text retrieval algorithm exploiting anchor text: BM25F[24]. BM25F is a state-of-the art ranking function solely based on textual content of the documents and their associated anchor texts. BM25F is a descendant of BM25 that combines the different textual fields of a document, namely title, body and anchor text. This model has been shown to be one of the best-performing web search scoring functions over the last few years [8, 24]. BM25F has a number of free parameters (2 per field, 6 in our case); we used the parameter values described in [24]. .341 .340 .339 .337 .336 .336 .334 .311 .311 .310 .310 .310 .310 .231 0.22 0.24 0.26 0.28 0.30 0.32 0.34 0.36 degree-in-id degree-in-ih degree-in-all hits-aut-ih-100 hits-aut-all-100 pagerank hits-aut-id-10 degree-out-all hits-hub-all-100 degree-out-ih hits-hub-ih-100 degree-out-id hits-hub-id-10 bm25f NDCG@10 .152 .152 .151 .150 .150 .149 .149 .137 .136 .136 .128 .127 .127 .100 0.09 0.10 0.11 0.12 0.13 0.14 0.15 0.16 degree-in-ih degree-in-id degree-in-all hits-aut-ih-100 hits-aut-all-100 hits-aut-id-10 pagerank hits-hub-all-100 degree-out-ih hits-hub-id-100 degree-out-all degree-out-id hits-hub-ih-100 bm25f MAP@10 .398 .397 .394 .394 .392 .392 .391 .356 .356 .356 .356 .356 .355 .273 0.25 0.30 0.35 0.40 degree-in-id degree-in-ih degree-in-all hits-aut-ih-100 hits-aut-all-100 pagerank hits-aut-id-10 degree-out-all hits-hub-all-100 degree-out-ih hits-hub-ih-100 degree-out-id hits-hub-id-10 bm25f MRR@10 Figure 3: Effectiveness measures for linear combinations of link-based features with BM25F. Figure 2 shows the NDCG, MRR, and MAP measures of these features. Again all performance measures (and for all rank-thresholds we explored) agree. As expected, BM25F outperforms all link-based features by a large margin. The link-based features are divided into two groups, with a noticeable performance drop between the groups. The better-performing group consists of the features that are based on the number and/or quality of incoming links (in-degree, PageRank, and HITS authority scores); and the worse-performing group consists of the features that are based on the number and/or quality of outgoing links (outdegree and HITS hub scores). In the group of features based on incoming links, features that ignore nepotistic links perform better than their counterparts using all links. Moreover, using only inter-domain (id) links seems to be marginally better than using inter-host (ih) links. The fact that features based on outgoing links underperform those based on incoming links matches our expectations; if anything, it is mildly surprising that outgoing links provide a useful signal for ranking at all. On the other hand, the fact that in-degree features outperform PageRank under all measures is quite surprising. A possible explanation is that link-spammers have been targeting the published PageRank algorithm for many years, and that this has led to anomalies in the web graph that affect PageRank, but not other link-based features that explore only a distance-1 neighborhood of the result set. Likewise, it is surprising that simple query-independent features such as in-degree, which might estimate global quality but cannot capture relevance to a query, would outperform query-dependent features such as HITS authority scores. However, we cannot investigate the effect of these features in isolation, without regard to the overall ranking function, for several reasons. First, features based on the textual content of documents (as opposed to link-based features) are the best predictors of relevance. Second, link-based features can be strongly correlated with textual features for several reasons, mainly the correlation between in-degree and numFeature Transform function bm25f T(s) = s pagerank T(s) = log(s + 3 · 10−12 ) degree-in-* T(s) = log(s + 3 · 10−2 ) degree-out-* T(s) = log(s + 3 · 103 ) hits-aut-* T(s) = log(s + 3 · 10−8 ) hits-hub-* T(s) = log(s + 3 · 10−1 ) Table 1: Near-optimal feature transform functions. ber of textual anchor matches. Therefore, one must consider the effect of link-based features in combination with textual features. Otherwise, we may find a link-based feature that is very good in isolation but is strongly correlated with textual features and results in no overall improvement; and vice versa, we may find a link-based feature that is weak in isolation but significantly improves overall performance. For this reason, we have studied the combination of the link-based features above with BM25F. All feature combinations were done by considering the linear combination of two features as a document score, using the formula score(d) =Pn i=1 wiTi(Fi(d)), where d is a document (or documentquery pair, in the case of BM25F), Fi(d) (for 1 ≤ i ≤ n) is a feature extracted from d, Ti is a transform, and wi is a free scalar weight that needs to be tuned. We chose transform functions that we empirically determined to be well-suited. Table 1 shows the chosen transform functions. This type of linear combination is appropriate if we assume features to be independent with respect to relevance and an exponential model for link features, as discussed in [8]. We tuned the weights by selecting a random subset of 5,000 queries from the query set, used an iterative refinement process to find weights that maximized a given performance measure on that training set, and used the remaining 23,043 queries to measure the performance of the thus derived scoring functions. We explored the pairwise combination of BM25F with every link-based scoring function. Figure 3 shows the NDCG, MRR, and MAP measures of these feature combinations, together with a baseline BM25F score (the right-most bar in each graph), which was computed using the same subset of 23,045 queries that were used as the test set for the feature combinations. Regardless of the performance measure applied, we can make the following general observations: 1. Combining any of the link-based features with BM25F results in a substantial performance improvement over BM25F in isolation. 2. The combination of BM25F with features based on incoming links (PageRank, in-degree, and HITS authority scores) performs substantially better than the combination with features based on outgoing links (HITS hub scores and out-degree). 3. The performance differences between the various combinations of BM25F with features based on incoming links is comparatively small, and the relative ordering of feature combinations is fairly stable across the 0.00 0.02 0.04 0.06 0.08 0.10 0.12 0.14 0 2 4 6 8 10 12 14 16 18 20 22 24 MAP@10 26 374 1640 2751 3768 4284 3944 3001 2617 1871 1367 771 1629 bm25fnorm pagerank degree-in-id hits-aut-id-100 Figure 4: Effectiveness measures for selected isolated features, broken down by query specificity. ferent performance measures used. However, the combination of BM25F with any in-degree variant, and in particular with id in-degree, consistently outperforms the combination of BM25F with PageRank or HITS authority scores, and can be computed much easier and faster. Finally, we investigated whether certain features are better for some queries than for others. Particularly, we are interested in the relationship between the specificity of a query and the performance of different ranking features. The most straightforward measure of the specificity of a query Q would be the number of documents in a search engine"s corpus that satisfy Q. Unfortunately, the query set available to us did not contain this information. Therefore, we chose to approximate the specificity of Q by summing up the inverse document frequencies of the individual query terms comprising Q. The inverse document frequency (IDF) of a term t with respect to a corpus C is defined to be logN/doc(t), where doc(t) is the number of documents in C containing t and N is the total number of documents in C. By summing up the IDFs of the query terms, we make the (flawed) assumption that the individual query terms are independent of each other. However, while not perfect, this approximation is at least directionally correct. We broke down our query set into 13 buckets, each bucket associated with an interval of query IDF values, and we computed performance metrics for all ranking functions applied (in isolation) to the queries in each bucket. In order to keep the graphs readable, we will not show the performance of all the features, but rather restrict ourselves to the four most interesting ones: PageRank, id HITS authority scores, id in-degree, and BM25F. Figure 4 shows the MAP@10 for all 13 query specificity buckets. Buckets on the far left of each graph represent very general queries; buckets on the far right represent very specific queries. The figures on the upper x axis of each graph show the number of queries in each bucket (e.g. the right-most bucket contains 1,629 queries). BM25F performs best for medium-specific queries, peaking at the buckets representing the IDF sum interval [12,14). By comparison, HITS peaks at the bucket representing the IDF sum interval [4,6), and PageRank and in-degree peak at the bucket representing the interval [6,8), i.e. more general queries. 8. CONCLUSIONS AND FUTURE WORK This paper describes a large-scale evaluation of the effectiveness of HITS in comparison with other link-based ranking algorithms, in particular PageRank and in-degree, when applied in isolation or in combination with a text retrieval algorithm exploiting anchor text (BM25F). Evaluation is carried out with respect to a large number of human evaluated queries, using three different measures of effectiveness: NDCG, MRR, and MAP. Evaluating link-based features in isolation, we found that web page in-degree outperforms PageRank, and is about as effwective as HITS authority scores. HITS hub scores and web page out-degree are much less effective ranking features, but still outperform a random ordering. A linear combination of any link-based features with BM25F produces a significant improvement in performance, and there is a clear difference between combining BM25F with a feature based on incoming links (indegree, PageRank, or HITS authority scores) and a feature based on outgoing links (HITS hub scores and out-degree), but within those two groups the precise choice of link-based feature matters relatively little. We believe that the measurements presented in this paper provide a solid evaluation of the best well-known link-based ranking schemes. There are many possible variants of these schemes, and many other link-based ranking algorithms have been proposed in the literature, hence we do not claim this work to be the last word on this subject, but rather the first step on a long road. Future work includes evaluation of different parameterizations of PageRank and HITS. In particular, we would like to study the impact of changes to the PageRank damping factor on effectiveness, the impact of various schemes meant to counteract the effects of link spam, and the effect of weighing hyperlinks differently depending on whether they are nepotistic or not. Going beyond PageRank and HITS, we would like to measure the effectiveness of other link-based ranking algorithms, such as SALSA. Finally, we are planning to experiment with more complex feature combinations. 9. REFERENCES [1] B. Amento, L. Terveen, and W. Hill. Does authority mean quality? Predicting expert quality ratings of web documents. In Proc. of the 23rd Annual International ACM SIGIR Conference on Research and Development in Information Retrieval, pages 296-303, 2000. [2] M. Bianchini, M. Gori, and F. Scarselli. Inside PageRank. ACM Transactions on Internet Technology, 5(1):92-128, 2005. [3] A. Borodin, G. O. Roberts, and J. S. Rosenthal. Finding authorities and hubs from link structures on the World Wide Web. In Proc. of the 10th International World Wide Web Conference, pages 415-429, 2001. [4] A. Borodin, G. O. Roberts, J. S. Rosenthal, and P. Tsaparas. Link analysis ranking: algorithms, theory, and experiments. ACM Transactions on Interet Technology, 5(1):231-297, 2005. [5] S. Brin and L. Page. The anatomy of a large-scale hypertextual Web search engine. Computer Networks and ISDN Systems, 30(1-7):107-117, 1998. [6] C. Burges, T. Shaked, E. Renshaw, A. Lazier, M. Deeds, N. Hamilton, and G. Hullender. Learning to rank using gradient descent. In Proc. of the 22nd International Conference on Machine Learning, pages 89-96, New York, NY, USA, 2005. ACM Press. [7] D. Cohn and H. Chang. Learning to probabilistically identify authoritative documents. In Proc. of the 17th International Conference on Machine Learning, pages 167-174, 2000. [8] N. Craswell, S. Robertson, H. Zaragoza, and M. Taylor. Relevance weighting for query independent evidence. In Proc. of the 28th Annual International ACM SIGIR Conference on Research and Development in Information Retrieval, pages 416-423, 2005. [9] E. Garfield. Citation analysis as a tool in journal evaluation. Science, 178(4060):471-479, 1972. [10] Z. Gy¨ongyi and H. Garcia-Molina. Web spam taxonomy. In 1st International Workshop on Adversarial Information Retrieval on the Web, 2005. [11] Z. Gy¨ongyi, H. Garcia-Molina, and J. Pedersen. Combating web spam with TrustRank. In Proc. of the 30th International Conference on Very Large Databases, pages 576-587, 2004. [12] B. J. Jansen, A. Spink, J. Bateman, and T. Saracevic. Real life information retrieval: a study of user queries on the web. ACM SIGIR Forum, 32(1):5-17, 1998. [13] K. J¨arvelin and J. Kek¨al¨ainen. Cumulated gain-based evaluation of IR techniques. ACM Transactions on Information Systems, 20(4):422-446, 2002. [14] S. D. Kamvar, T. H. Haveliwala, C. D. Manning, and G. H. Golub. Extrapolation methods for accelerating PageRank computations. In Proc. of the 12th International World Wide Web Conference, pages 261-270, 2003. [15] M. M. Kessler. Bibliographic coupling between scientific papers. American Documentation, 14(1):10-25, 1963. [16] J. M. Kleinberg. Authoritative sources in a hyperlinked environment. In Proc. of the 9th Annual ACM-SIAM Symposium on Discrete Algorithms, pages 668-677, 1998. [17] J. M. Kleinberg. Authoritative sources in a hyperlinked environment. Journal of the ACM, 46(5):604-632, 1999. [18] A. N. Langville and C. D. Meyer. Deeper inside PageRank. Internet Mathematics, 1(3):2005, 335-380. [19] R. Lempel and S. Moran. The stochastic approach for link-structure analysis (SALSA) and the TKC effect. Computer Networks and ISDN Systems, 33(1-6):387-401, 2000. [20] A. Y. Ng, A. X. Zheng, and M. I. Jordan. Stable algorithms for link analysis. In Proc. of the 24th Annual International ACM SIGIR Conference on Research and Development in Information Retrieval, pages 258-266, 2001. [21] L. Page, S. Brin, R. Motwani, and T. Winograd. The PageRank citation ranking: Bringing order to the web. Technical report, Stanford Digital Library Technologies Project, 1998. [22] J. A. Tomlin. A new paradigm for ranking pages on the World Wide Web. In Proc. of the 12th International World Wide Web Conference, pages 350-355, 2003. [23] T. Upstill, N. Craswell, and D. Hawking. Predicting fame and fortune: Pagerank or indegree? In Proc. of the Australasian Document Computing Symposium, pages 31-40, 2003. [24] H. Zaragoza, N. Craswell, M. Taylor, S. Saria, and S. Robertson. Microsoft Cambridge at TREC-13: Web and HARD tracks. In Proc. of the 13th Text Retrieval Conference, 2004.

mean reciprocal rank;scale and relevance;quantitative measure;hit;breadth-first search crawl;mean average precision;link-based feature;rank;ranking;map;link graph;pagerank;ndcg;mrr;normalized discounted cumulative gain measurement;feature selection;query specificity;bm25f;crawled web page;hyperlink analysis

train_H-42

HITS Hits TRECExploring IR Evaluation Results with Network Analysis

We propose a novel method of analysing data gathered from TREC or similar information retrieval evaluation experiments. We define two normalized versions of average precision, that we use to construct a weighted bipartite graph of TREC systems and topics. We analyze the meaning of well known - and somewhat generalized - indicators from social network analysis on the Systems-Topics graph. We apply this method to an analysis of TREC 8 data; among the results, we find that authority measures systems performance, that hubness of topics reveals that some topics are better than others at distinguishing more or less effective systems, that with current measures a system that wants to be effective in TREC needs to be effective on easy topics, and that by using different effectiveness measures this is no longer the case.

1. INTRODUCTION Evaluation is a primary concern in the Information Retrieval (IR) field. TREC (Text REtrieval Conference) [12, 15] is an annual benchmarking exercise that has become a de facto standard in IR evaluation: before the actual conference, TREC provides to participants a collection of documents and a set of topics (representations of information needs). Participants use their systems to retrieve, and submit to TREC, a list of documents for each topic. After the lists have been submitted and pooled, the TREC organizers employ human assessors to provide relevance judgements on the pooled set. This defines a set of relevant documents for each topic. System effectiveness is then measured by well established metrics (Mean Average Precision being the most used). Other conferences such as NTCIR, INEX, CLEF provide comparable data. Network analysis is a discipline that studies features and properties of (usually large) networks, or graphs. Of particular importance is Social Network Analysis [16], that studies networks made up by links among humans (friendship, acquaintance, co-authorship, bibliographic citation, etc.). Network analysis and IR fruitfully meet in Web Search Engine implementation, as is already described in textbooks [3,6]. Current search engines use link analysis techniques to help rank the retrieved documents. Some indicators (and the corresponding algorithms that compute them) have been found useful in this respect, and are nowadays well known: Inlinks (the number of links to a Web page), PageRank [7], and HITS (Hyperlink-Induced Topic Search) [5]. Several extensions to these algorithms have been and are being proposed. These indicators and algorithms might be quite general in nature, and can be used for applications which are very different from search result ranking. One example is using HITS for stemming, as described by Agosti et al. [1]. In this paper, we propose and demonstrate a method for constructing a network, specifically a weighted complete bidirectional directed bipartite graph, on a set of TREC topics and participating systems. Links represent effectiveness measurements on system-topic pairs. We then apply analysis methods originally developed for search applications to the resulting network. This reveals phenomena previously hidden in TREC data. In passing, we also provide a small generalization to Kleinberg"s HITS algorithm, as well as to Inlinks and PageRank. The paper is organized as follows: Sect. 2 gives some motivations for the work. Sect. 3 presents the basic ideas of normalizing average precision and of constructing a systemstopics graph, whose properties are analyzed in Sect. 4; Sect. 5 presents some experiments on TREC 8 data; Sect. 6 discusses some issues and Sect. 7 closes the paper. 2. MOTIVATIONS We are interested in the following hypotheses: 1. Some systems are more effective than others; t1 · · · tn MAP s1 AP(s1, t1) · · · AP(s1, tn) MAP(s1) ... ... ... sm AP(sm, t1) · · · AP(sm, tn) MAP(sm) AAP AAP(t1) · · · AAP(tn) (a) t1 t2 · · · MAP s1 0.5 0.4 · · · 0.6 s2 0.4 · · · · · · 0.3 ... ... ... ... AAP 0.6 0.3 · · · (b) Table 1: AP, MAP and AAP 2. Some topics are easier than others; 3. Some systems are better than others at distinguishing easy and difficult topics; 4. Some topics are better than others at distinguishing more or less effective systems. The first of these hypotheses needs no further justification - every reported significant difference between any two systems supports it. There is now also quite a lot of evidence for the second, centered on the TREC Robust Track [14]. Our primary interest is in the third and fourth. The third might be regarded as being of purely academic interest; however, the fourth has the potential for being of major practical importance in evaluation studies. If we could identify a relatively small number of topics which were really good at distinguishing effective and ineffective systems, we could save considerable effort in evaluating systems. One possible direction from this point would be to attempt direct identification of such small sets of topics. However, in the present paper, we seek instead to explore the relationships suggested by the hypotheses, between what different topics tell us about systems and what different systems tell us about topics. We seek methods of building and analysing a matrix of system-topic normalised performances, with a view to giving insight into the issue and confirming or refuting the third and fourth hypotheses. It turns out that the obvious symmetry implied by the above formulation of the hypotheses is a property worth investigating, and the investigation does indeed give us valuable insights. 3. THE IDEA 3.1 1st step: average precision table From TREC results, one can produce an Average Precision (AP) table (see Tab. 1a): each AP(si, tj) value measures the AP of system si on topic tj. Besides AP values, the table shows Mean Average Precision (MAP) values i.e., the mean of the AP values for a single system over all topics, and what we call Average Average Precision (AAP) values i.e., the average of the AP values for a single topic over all systems: MAP(si) = 1 n nX j=1 AP(si, tj), (1) AAP(tj) = 1 m mX i=1 AP(si, tj). (2) MAPs are indicators of systems performance: higher MAP means good system. AAP are indicators of the performance on a topic: higher AAP means easy topic - a topic on which all or most systems have good performance. 3.2 Critique of pure AP MAP is a standard, well known, and widely used IR effectiveness measure. Single AP values are used too (e.g., in AP histograms). Topic difficulty is often discussed (e.g., in TREC Robust track [14]), although AAP values are not used and, to the best of our knowledge, have never been proposed (the median, not the average, of AP on a topic is used to produce TREC AP histograms [11]). However, the AP values in Tab. 1 present two limitations, which are symmetric in some respect: • Problem 1. They are not reliable to compare the effectiveness of a system on different topics, relative to the other systems. If, for example, AP(s1, t1) > AP(s1, t2), can we infer that s1 is a good system (i.e., has a good performance) on t1 and a bad system on t2? The answer is no: t1 might be an easy topic (with high AAP) and t2 a difficult one (low AAP). See an example in Tab. 1b: s1 is outperformed (on average) by the other systems on t1, and it outperforms the other systems on t2. • Problem 2. Conversely, if, for example, AP(s1, t1) > AP(s2, t1), can we infer that t1 is considered easier by s1 than by s2? No, we cannot: s1 might be a good system (with high MAP) and s2 a bad one (low MAP); see an example in Tab. 1b. These two problems are a sort of breakdown of the well known high influence of topics on IR evaluation; again, our formulation makes explicit the topics / systems symmetry. 3.3 2nd step: normalizations To avoid these two problems, we can normalize the AP table in two ways. The first normalization removes the influence of the single topic ease on system performance. Each AP(si, tj) value in the table depends on both system goodness and topic ease (the value will increase if a system is good and/or the topic is easy). By subtracting from each AP(si, tj) the AAP(tj) value, we obtain normalized AP values (APA(si, tj), Normalized AP according to AAP): APA(si, tj) = AP(si, tj) − AAP(tj), (3) that depend on system performance only (the value will increase only if system performance is good). See Tab. 2a. The second normalization removes the influence of the single system effectiveness on topic ease: by subtracting from each AP(si, tj) the MAP(si) value, we obtain normalized AP values (APM(si, tj), Normalized AP according to MAP): APM(si, tj) = AP(si, tj) − MAP(si), (4) that depend on topic ease only (the value will increase only if the topic is easy, i.e., all systems perform well on that topic). See Tab. 2b. In other words, APA avoids Problem 1: APA(s, t) values measure the performance of system s on topic t normalized t1 · · · tn MAP s1 APA(s1, t1) · · · APA(s1, tn) MAP(s1) ... ... ... sm APA(sm, t1) · · · APA(sm, tn) MAP(sm) 0 · · · 0 0 (a) t1 · · · tn s1 APM(s1, t1) · · · APM(s1, tn) 0 ... ... ... sm APM(sm, t1) · · · APM(sm, tn) 0 AAP AAP(t1) · · · AAP(tn) 0 (b) t1 t2 · · · MAP s1 −0.1 0.1 · · · . . . s2 0.2 · · · · · · . . . ... ... ... 0 0 · · · t1 t2 · · · s1 −0.1 −0.2 · · · 0 s2 0.1 · · · · · · 0 ... ... ... AAP . . . . . . · · · (c) (d) Table 2: Normalizations: APA and MAP: normalized AP (APA) and MAP (MAP) (a); normalized AP (APM) and AAP (AAP) (b); a numeric example (c) and (d) according to the ease of the topic (easy topics will not have higher APA values). Now, if, for example, APA(s1, t2) > APA(s1, t1), we can infer that s1 is a good system on t2 and a bad system on t1 (see Tab. 2c). Vice versa, APM avoids Problem 2: APM(s, t) values measure the ease of topic t according to system s, normalized according to goodness of the system (good systems will not lead to higher APM values). If, for example, APM(s2, t1) > APM(s1, t1), we can infer that t1 is considered easier by s2 than by s1 (see Tab. 2d). On the basis of Tables 2a and 2b, we can also define two new measures of system effectiveness and topic ease, i.e., a Normalized MAP (MAP), obtained by averaging the APA values on one row in Tab. 2a, and a Normalized AAP (AAP), obtained by averaging the APM values on one column in Tab. 2b: MAP(si) = 1 n nX j=1 APA(si, tj) (5) AAP(tj) = 1 m mX i=1 APM(si, tj). (6) Thus, overall system performance can be measured, besides by means of MAP, also by means of MAP. Moreover, MAP is equivalent to MAP, as can be immediately proved by using Eqs. (5), (3), and (1): MAP(si) = 1 n nX j=1 (AP(si, tj) − AAP(tj)) = = MAP(si) − 1 n nX j=1 AAP(tj) (and 1 n Pn j=1 AAP(tj) is the same for all systems). And, conversely, overall topic ease can be measured, besides by t1 · · · tn s1 ... APM sm t1 · · · tn s1 ... APA sm s1 · · · sm t1 · · · tn s1 ... 0 APM 0 sm t1 ... APA T 0 0 tn MAP AAP 0 Figure 1: Construction of the adjacency matrix. APA T is the transpose of APA. means of AAP, also by means of AAP, and this is equivalent (the proof is analogous, and relies on Eqs. (6), (4), and (2)). The two Tables 2a and 2b are interesting per se, and can be analyzed in several different ways. In the following we propose an analysis based on network analysis techniques, mainly Kleinberg"s HITS algorithm. There is a little further discussion of these normalizations in Sect. 6. 3.4 3rd step: Systems-Topics Graph The two tables 2a and 2b can be merged into a single one with the procedure shown in Fig. 1. The obtained matrix can be interpreted as the adjacency matrix of a complete weighted bipartite graph, that we call Systems-Topics graph. Arcs and weights in the graph can be interpreted as follows: • (weight on) arc s → t: how much the system s thinks that the topic t is easy - assuming that a system has no knowledge of the other systems (or in other words, how easy we might think the topic is, knowing only the results for this one system). This corresponds to APM values, i.e., to normalized topic ease (Fig. 2a). • (weight on) arc s ← t: how much the topic t thinks that the system s is good - assuming that a topic has no knowledge of the other topics (or in other words, how good we might think the system is, knowing only the results for this one topic). This corresponds to APA (normalized system effectiveness, Fig. 2b). Figs. 2c and 2d show the Systems-Topics complete weighted bipartite graph, on a toy example with 4 systems and 2 topics; the graph is split in two parts to have an understandable graphical representation: arcs in Fig. 2c are labeled with APM values; arcs in Fig. 2d are labeled with APA values. 4. ANALYSIS OF THE GRAPH 4.1 Weighted Inlinks, Outlinks, PageRank The sum of weighted outlinks, i.e., the sum of the weights on the outgoing arcs from each node, is always zero: • The outlinks on each node corresponding to a system s (Fig. 2c) is the sum of all the corresponding APM values on one row of the matrix in Tab. 2b. • The outlinks on each node corresponding to a topic t (Fig. 2d) is the sum of all the corresponding APA (a) (b) (c) (d) Figure 2: The relationships between systems and topics (a) and (b); and the Systems-Topics graph for a toy example (c) and (d). Dashed arcs correspond to negative values. h (a) s1 ... sm t1 ... tn = s1 · · · sm t1 · · · tn s1 ... 0 APM (APA) sm t1 ... APA T 0 tn (APM T ) · a (h) s1 ... sm t1 ... tn Figure 3: Hub and Authority computation values on one row of the transpose of the matrix in Tab. 2a. The average1 of weighted inlinks is: • MAP for each node corresponding to a system s; this corresponds to the average of all the corresponding APA values on one column of the APA T part of the adjacency matrix (see Fig. 1). • AAP for each node corresponding to a topic t; this corresponds to the average of all the corresponding APM values on one column of the APM part of the adjacency matrix (see Fig. 1). Therefore, weighted inlinks measure either system effectiveness or topic ease; weighted outlinks are not meaningful. We could also apply the PageRank algorithm to the network; the meaning of the PageRank of a node is not quite so obvious as Inlinks and Outlinks, but it also seems a sensible measure of either system effectiveness or topic ease: if a system is effective, it will have several incoming links with high 1 Usually, the sum of the weights on the incoming arcs to each node is used in place of the average; since the graph is complete, it makes no difference. weights (APA); if a topic is easy it will have high weights (APM) on the incoming links too. We will see experimental confirmation in the following. 4.2 Hubs and Authorities Let us now turn to more sophisticated indicators. Kleinberg"s HITS algorithm defines, for a directed graph, two indicators: hubness and authority; we reiterate here some of the basic details of the HITS algorithm in order to emphasize both the nature of our generalization and the interpretation of the HITS concepts in this context. Usually, hubness and authority are defined as h(x) = P x→y a(y) and a(x) = P y→x h(y), and described intuitively as a good hub links many good authorities; a good authority is linked from many good hubs. As it is well known, an equivalent formulation in linear algebra terms is (see also Fig. 3): h = Aa and a = AT h (7) (where h is the hubness vector, with the hub values for all the nodes; a is the authority vector; A is the adjacency matrix of the graph; and AT its transpose). Usually, A contains 0s and 1s only, corresponding to presence and absence of unweighted directed arcs, but Eq. (7) can be immediately generalized to (in fact, it is already valid for) A containing any real value, i.e., to weighted graphs. Therefore we can have a generalized version (or rather a generalized interpretation, since the formulation is still the original one) of hubness and authority for all nodes in a graph. An intuitive formulation of this generalized HITS is still available, although slightly more complex: a good hub links, by means of arcs having high weights, many good authorities; a good authority is linked, by means of arcs having high weights, from many good hubs. Since arc weights can be, in general, negative, hub and authority values can be negative, and one could speak of unhubness and unauthority; the intuitive formulation could be completed by adding that a good hub links good unauthorities by means of links with highly negative weights; a good authority is linked by good unhubs by means of links with highly negative weights. And, also, a good unhub links positively good unauthorities and negatively good authorities; a good unauthority is linked positively from good unhubs and negatively from good hubs. Let us now apply generalized HITS to our Systems-Topics graph. We compute a(s), h(s), a(t), and h(t). Intuitively, we expect that a(s) is somehow similar to Inlinks, so it should be a measure of either systems effectiveness or topic ease. Similarly, hubness should be more similar to Outlinks, thus less meaningful, although the interplay between hub and authority might lead to the discovery of something different. Let us start by remarking that authority of topics and hubness of systems depend only on each other; similarly hubness of topics and authority of systems depend only on each other: see Figs. 2c, 2d and 3. Thus the two graphs in Figs. 2c and 2d can be analyzed independently. In fact the entire HITS analysis could be done in one direction only, with just APM(s, t) values or alternatively with just APA(s, t). As discussed below, probably most interest resides in the hubness of topics and the authority of systems, so the latter makes sense. However, in this paper, we pursue both analyses together, because the symmetry itself is interesting. Considering Fig. 2c we can state that: • Authority a(t) of a topic node t increases when: - if h(si) > 0, APM(si, t) increases (or if APM(si, t) > 0, h(si) increases); - if h(si) < 0, APM(si, t) decreases (or if APM(si, t) < 0, h(si) decreases). • Hubness h(s) of a system node s increases when: - if a(tj) > 0, APM(s, tj) increases (or, if APM(s, tj) > 0, a(tj) increases); - if a(tj) < 0, APM(s, tj) decreases (or, if APM(s, tj) < 0, a(tj) decreases). We can summarize this as: a(t) is high if APM(s, t) is high for those systems with high h(s); h(s) is high if APM(s, t) is high for those topics with high a(t). Intuitively, authority a(t) of a topic measures topic ease; hubness h(s) of a system measures system"s capability to recognize easy topics. A system with high unhubness (negative hubness) would tend to regard easy topics as hard and hard ones as easy. The situation for Fig. 2d, i.e., for a(s) and h(t), is analogous. Authority a(s) of a system node s measures system effectiveness: it increases with the weight on the arc (i.e., APA(s, tj)) and the hubness of the incoming topic nodes tj. Hubness h(t) of a topic node t measures topic capability to recognize effective systems: if h(t) > 0, it increases further if APA(s, tj) increases; if h(t) < 0, it increases if APA(s, tj) decreases. Intuitively, we can state that A system has a higher authority if it is more effective on topics with high hubness; and A topic has a higher hubness if it is easier for those systems which are more effective in general. Conversely for system hubness and topic authority: A topic has a higher authority if it is easier on systems with high hubness; and A system has a higher hubness if it is more effective for those topics which are easier in general. Therefore, for each system we have two indicators: authority (a(s)), measuring system effectiveness, and hubness (h(s)), measuring system capability to estimate topic ease. And for each topic, we have two indicators: authority (a(t)), measuring topic ease, and hubness (h(t)), measuring topic capability to estimate systems effectiveness. We can define them formally as a(s) = X t h(t) · APA(s, t), h(t) = X s a(s) · APA(s, t), a(t) = X s h(s) · APM(s, t), h(s) = X t a(t) · APM(s, t). We observe that the hubness of topics may be of particular interest for evaluation studies. It may be that we can evaluate the effectiveness of systems efficiently by using relatively few high-hubness topics. 5. EXPERIMENTS We now turn to discuss if these indicators are meaningful and useful in practice, and how they correlate with standard measures used in TREC. We have built the Systems-Topics graph for TREC 8 data (featuring 1282 systems - actually, 2 Actually, TREC 8 data features 129 systems; due to some bug in our scripts, we did not include one system (8manexT3D1N0), but the results should not be affected. 0 0.1 0.2 0.3 -1 -0.5 0 0.5 1 NAPM NAPA AP Figure 4: Distributions of AP, APA, and APM values in TREC 8 data MAP In PR H A MAP 1 1.0 1.0 .80 .99 Inlinks 1 1.0 .80 .99 PageRank 1 .80 .99 Hub 1 .87 (a) AAP In PR H A AAP 1 1.0 1.0 .92 1.0 Inlinks 1 1.0 .92 1.0 PageRank 1 .92 1.0 Hub 1 .93 (b) Table 3: Correlations between network analysis measures and MAP (a) and AAP (b) runs - on 50 topics). This section illustrates the results obtained mining these data according to the method presented in previous sections. Fig. 4 shows the distributions of AP, APA, and APM: whereas AP is very skewed, both APA and APM are much more symmetric (as it should be, since they are constructed by subtracting the mean). Tables 3a and 3b show the Pearson"s correlation values between Inlinks, PageRank, Hub, Authority and, respectively, MAP or AAP (Outlinks values are not shown since they are always zero, as seen in Sect. 4). As expected, Inlinks and PageRank have a perfect correlation with MAP and AAP. Authority has a very high correlation too with MAP and AAP; Hub assumes slightly lower values. Let us analyze the correlations more in detail. The correlations chart in Figs. 5a and 5b demonstrate the high correlation between Authority and MAP or AAP. Hubness presents interesting phenomena: both Fig. 5c (correlation with MAP) and Fig. 5d (correlation with AAP) show that correlation is not exact, but neither is it random. This, given the meaning of hubness (capability in estimating topic ease and system effectiveness), means two things: (i) more effective systems are better at estimating topic ease; and (ii) easier topics are better at estimating system effectiveness. Whereas the first statement is fine (there is nothing against it), the second is a bit worrying. It means that system effectiveness in TREC is affected more by easy topics than by difficult topics, which is rather undesirable for quite obvious reasons: a system capable of performing well on a difficult topic, i.e., on a topic on which the other systems perform badly, would be an important result for IR effectiveness; con-8E-5 -6E-5 -4E-5 -2E-5 0E+0 2E-5 4E-5 6E-5 0.0 0.1 0.2 0.3 0.4 0.5 -3E-1 -2E-1 -1E-1 0E+0 1E-1 2E-1 3E-1 4E-1 5E-1 0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 (a) (b) 0E+0 2E-2 4E-2 6E-2 8E-2 1E-1 1E-1 1E-1 0.0 0.1 0.2 0.3 0.4 0.5 0E+0 1E-5 2E-5 3E-5 4E-5 5E-5 6E-5 7E-5 0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 (c) (d) Figure 5: Correlations: MAP (x axis) and authority (y axis) of systems (a); AAP and authority of topics (b); MAP and hub of systems (c) and AAP and hub of topics (d) versely, a system capable of performing well on easy topics is just a confirmation of the state of the art. Indeed, the correlation between hubness and AAP (statement (i) above) is higher than the correlation between hubness and MAP (corresponding to statement (ii)): 0.92 vs. 0.80. However, this phenomenon is quite strong. This is also confirmed by the work being done on the TREC Robust Track [14]. In this respect, it is interesting to see what happens if we use a different measure from MAP (and AAP). The GMAP (Geometric MAP) metric is defined as the geometric mean of AP values, or equivalently as the arithmetic mean of the logarithms of AP values [8]. GMAP has the property of giving more weight to the low end of the AP scale (i.e., to low AP values), and this seems reasonable, since, intuitively, a performance increase in MAP values from 0.01 to 0.02 should be more important than an increase from 0.81 to 0.82. To use GMAP in place of MAP and AAP, we only need to take the logarithms of initial AP values, i.e., those in Tab. 1a (zero values are modified into ε = 0.00001). We then repeat the same normalization process (with GMAP and GAAP - Geometric AAP - replacing MAP and AAP): whereas authority values still perfectly correlate with GMAP (0.99) and GAAP (1.00), the correlation with hubness largely disappears (values are −0.16 and −0.09 - slightly negative but not enough to concern us). This is yet another confirmation that TREC effectiveness as measured by MAP depends mainly on easy topics; GMAP appears to be a more balanced measure. Note that, perhaps surprisingly, GMAP is indeed fairly well balanced, not biased in the opposite direction - that is, it does not overemphasize the difficult topics. In Sect. 6.3 below, we discuss another transformation, replacing the log function used in GMAP with logit. This has a similar effect: the correlation of mean logitAP and average logitAP with hubness are now small positive numbers (0.23 and 0.15 respectively), still comfortably away from the high correlations with regular MAP and AAP, i.e., not presenting the problematic phenomenon (ii) above (over-dependency on easy topics). We also observe that hub values are positive, whereas authority assumes, as predicted, both positive and negative values. An intuitive justification is that negative hubness would indicate a node which disagrees with the other nodes, e.g., a system which does better on difficult topics, or a topic on which bad systems do better; such systems and topics would be quite strange, and probably do not appear in TREC. Finally, although one might think that topics with several relevant documents are more important and difficult, this is not the case: there is no correlation between hub (or any other indicator) and the number of documents relevant to a topic. 6. DISCUSSION 6.1 Related work There has been considerable interest in recent years in questions of statistical significance of effectiveness comparisons between systems (e.g. [2, 9]), and related questions of how many topics might be needed to establish differences (e.g. [13]). We regard some results of the present study as in some way complementary to this work, in that we make a step towards answering the question Which topics are best for establishing differences?. The results on evaluation without relevance judgements such as [10] show that, to some extent, good systems agree on which are the good documents. We have not addressed the question of individual documents in the present analysis, but this effect is certainly analogous to our results. 6.2 Are normalizations necessary? At this point it is also worthwhile to analyze what would happen without the MAP- and AAP-normalizations defined in Sect. 3.3. Indeed, the process of graph construction (Sect. 3.4) is still valid: both the APM and APA matrices are replaced by the AP one, and then everything goes on as above. Therefore, one might think that the normalizations are unuseful in this setting. This is not the case. From the theoretical point of view, the AP-only graph does not present the interesting properties above discussed: since the AP-only graph is symmetrical (the weight on each incoming link is equal to the weight on the corresponding outgoing link), Inlinks and Outlinks assume the same values. There is symmetry also in computing hub and authority, that assume the same value for each node since the weights on the incoming and outgoing arcs are the same. This could be stated in more precise and formal terms, but one might still wonder if on the overall graph there are some sort of counterbalancing effects. It is therefore easier to look at experimental data, which confirm that the normalizations are needed: the correlations between AP, Inlinks, Outlinks, Hub, and/or Authority are all very close to one (none of them is below 0.98). 6.3 Are these normalizations sufficient? It might be argued that (in the case of APA, for example) the amount we have subtracted from each AP value is topicdependent, therefore the range of the resulting APA value is also topic-dependent (e.g. the maximum is 1 − AAP(tj) and the minimum is − AAP(tj)). This suggests that the cross-topic comparisons of these values suggested in Sect. 3.3 may not be reliable. A similar issue arises for APM and comparisons across systems. One possible way to overcome this would be to use an unconstrained measure whose range is the full real line. Note that in applying the method to GMAP by using log AP, we avoid the problem with the lower limit but retain it for the upper limit. One way to achieve an unconstrainted range would be to use the logit function rather than the log [4,8]. We have also run this variant (as already reported in Sect. 5 above), and it appears to provide very similar results to the GMAP results already given. This is not surprising, since in practice the two functions are very similar over most of the operative range. The normalizations thus seem reliable. 6.4 On AAT and AT A It is well known that h and a vectors are the principal left eigenvectors of AAT and AT A, respectively (this can be easily derived from Eqs. (7)), and that, in the case of citation graphs, AAT and AT A represent, respectively, bibliographic coupling and co-citations. What is the meaning, if any, of AAT and AT A in our Systems-Topics graph? It is easy to derive that: AAT [i, j] = 8 >< >: 0  if i ∈ S ∧ j ∈ T or i ∈ T ∧ j ∈ S P k A[i, k] · A[j, k] otherwise AT A[i, j] = 8 >< >: 0  if i ∈ S ∧ j ∈ T or i ∈ T ∧ j ∈ S P k A[k, i] · A[k, j] otherwise (where S is the set of indices corresponding to systems and T the set of indices corresponding to topics). Thus AAT and AT A are block diagonal matrices, with two blocks each, one relative to systems and one relative to topics: (a) if i, j ∈ S, then AAT [i, j] = P k∈T APM(i, k)·APM(j, k) measures how much the two systems i and j agree in estimating topics ease (APM): high values mean that the two systems agree on topics ease. (b) if i, j ∈ T, then AAT [i, j] = P k∈S APA(k, i)·APA(k, j) measures how much the two topics i and j agree in estimating systems effectiveness (APA): high values mean that the two topics agree on systems effectiveness (and that TREC results would not change by leaving out one of the two topics). (c) if i, j ∈ S, then AT A[i, j] = P k∈T APA(i, k) · APA(j, k) measures how much agreement on the effectiveness of two systems i and j there is over all topics: high values mean that many topics quite agree on the two systems effectiveness; low values single out systems that are somehow controversial, and that need several topics to have a correct effectiveness assessment. (d) if i, j ∈ T, then AT A[i, j] = P k∈S APM(k, i)·APM(k, j) measures how much agreement on the ease of the two topics i and j there is over all systems: high values mean that many systems quite agree on the two topics ease. Therefore, these matrices are meaningful and somehow interesting. For instance, the submatrix (b) corresponds to a weighted undirected complete graph, whose nodes are the topics and whose arc weights are a measure of how much two topics agree on systems effectiveness. Two topics that are very close on this graph give the same information, and therefore one of them could be discarded without changes in TREC results. It would be interesting to cluster the topics on this graph. Furthermore, the matrix/graph (a) could be useful in TREC pool formation: systems that do not agree on topic ease would probably find different relevant documents, and should therefore be complementary in pool formation. Note that no notion of single documents is involved in the above analysis. 6.5 Insights As indicated, the primary contribution of this paper has been a method of analysis. However, in the course of applying this method to one set of TREC results, we have achieved some insights relating to the hypotheses formulated in Sect. 2: • We confirm Hypothesis 2 above, that some topics are easier than others. • Differences in the hubness of systems reveal that some systems are better than others at distinguishing easy and difficult topics; thus we have some confirmation of Hypothesis 3. • There are some relatively idiosyncratic systems which do badly on some topics generally considered easy but well on some hard topics. However, on the whole, the more effective systems are better at distinguishing easy and difficult topics. This is to be expected: a really bad system will do badly on everything, while even a good system may have difficulty with some topics. • Differences in the hubness of topics reveal that some topics are better than others at distinguising more or less effective systems; thus we have some confirmation of Hypothesis 4. • If we use MAP as the measure of effectiveness, it is also true that the easiest topics are better at distinguishing more or less effective systems. As argued in Sect. 5, this is an undesirable property. GMAP is more balanced. Clearly these ideas need to be tested on other data sets. However, they reveal that the method of analysis proposed in this paper can provide valuable information. 6.6 Selecting topics The confirmation of Hypothesis 4 leads, as indicated, to the idea that we could do reliable system evaluation on a much smaller set of topics, provided we could select such an appropriate set. This selection may not be straightforward, however. It is possible that simply selecting the high hubness topics will achieve this end; however, it is also possible that there are significant interactions between topics which would render such a simple rule ineffective. This investigation would therefore require serious experimentation. For this reason we have not attempted in this paper to point to the specific high hubness topics as being good for evaluation. This is left for future work. 7. CONCLUSIONS AND FUTURE DEVELOPMENTS The contribution of this paper is threefold: • we propose a novel way of normalizing AP values; • we propose a novel method to analyse TREC data; • the method applied on TREC data does indeed reveal some hidden properties. More particularly, we propose Average Average Precision (AAP), a measure of topic ease, and a novel way of normalizing the average precision measure in TREC, on the basis of both MAP (Mean Average Precision) and AAP. The normalized measures (APM and APA) are used to build a bipartite weighted Systems-Topics graph, that is then analyzed by means of network analysis indicators widely known in the (social) network analysis field, but somewhat generalised. We note that no such approach to TREC data analysis has been proposed so far. The analysis shows that, with current measures, a system that wants to be effective in TREC needs to be effective on easy topics. Also, it is suggested that a cluster analysis on topic similarity can lead to relying on a lower number of topics. Our method of analysis, as described in this paper, can be applied only a posteriori, i.e., once we have all the topics and all the systems available. Adding (removing) a new system / topic would mean re-computing hubness and authority indicators. Moreover, we are not explicitly proposing a change to current TREC methodology, although this could be a by-product of these - and further - analyses. This is an initial work, and further analyses could be performed. For instance, other effectiveness metrics could be used, in place of AP. Other centrality indicators, widely used in social network analysis, could be computed, although probably with similar results to PageRank. It would be interesting to compute the higher-order eigenvectors of AT A and AAT . The same kind of analysis could be performed at the document level, measuring document ease. Hopefully, further analyses of the graph defined in this paper, according to the approach described, can be insightful for a better understanding of TREC or similar data. Acknowledgments We would like to thank Nick Craswell for insightful discussions and the anonymous referees for useful remarks. Part of this research has been carried on while the first author was visiting Microsoft Research Cambridge, whose financial support is acknowledged. 8. REFERENCES [1] M. Agosti, M. Bacchin, N. Ferro, and M. Melucci. Improving the automatic retrieval of text documents. In Proceedings of the 3rd CLEF Workshop, volume 2785 of LNCS, pages 279-290, 2003. [2] C. Buckley and E. Voorhees. Evaluating evaluation measure stability. In 23rd SIGIR, pages 33-40, 2000. [3] S. Chakrabarti. Mining the Web. Morgan Kaufmann, 2003. [4] G. V. Cormack and T. R. Lynam. Statistical precision of information retrieval evaluation. In 29th SIGIR, pages 533-540, 2006. [5] J. Kleinberg. Authoritative sources in a hyperlinked environment. J. of the ACM, 46(5):604-632, 1999. [6] M. Levene. An Introduction to Search Engines and Web Navigation. Addison Wesley, 2006. [7] L. Page, S. Brin, R. Motwani, and T. Winograd. The PageRank Citation Ranking: Bringing Order to the Web, 1998. http://dbpubs.stanford.edu:8090/pub/1999-66. [8] S. Robertson. On GMAP - and other transformations. In 13th CIKM, pages 78-83, 2006. [9] M. Sanderson and J. Zobel. Information retrieval system evaluation: effort, sensitivity, and reliability. In 28th SIGIR, pages 162-169, 2005. http://doi.acm.org/10.1145/1076034.1076064. [10] I. Soboroff, C. Nicholas, and P. Cahan. Ranking retrieval systems without relevance judgments. In 24th SIGIR, pages 66-73, 2001. [11] TREC Common Evaluation Measures, 2005. http://trec.nist.gov/pubs/trec14/appendices/ CE.MEASURES05.pdf (Last visit: Jan. 2007). [12] Text REtrieval Conference (TREC). http://trec.nist.gov/ (Last visit: Jan. 2007). [13] E. Voorhees and C. Buckley. The effect of topic set size on retrieval experiment error. In 25th SIGIR, pages 316-323, 2002. [14] E. M. Voorhees. Overview of the TREC 2005 Robust Retrieval Track. In TREC 2005 Proceedings, 2005. [15] E. M. Voorhees and D. K. Harman. TRECExperiment and Evaluation in Information Retrieval. MIT Press, 2005. [16] S. Wasserman and K. Faust. Social Network Analysis. Cambridge University Press, Cambridge, UK, 1994.

trec;systems-topic graph;hit;social network analysis;mean average precision;link analysis technique;web search engine implementation;pagerank;network analysis;information retrieval evaluation experiment;inlink;ir evaluation;hit algorithm;kleinberg' hit algorithm;human assessor;weighted bipartite graph;stemming

train_H-43

Combining Content and Link for Classification using Matrix Factorization

The world wide web contains rich textual contents that are interconnected via complex hyperlinks. This huge database violates the assumption held by most of conventional statistical methods that each web page is considered as an independent and identical sample. It is thus difficult to apply traditional mining or learning methods for solving web mining problems, e.g., web page classification, by exploiting both the content and the link structure. The research in this direction has recently received considerable attention but are still in an early stage. Though a few methods exploit both the link structure or the content information, some of them combine the only authority information with the content information, and the others first decompose the link structure into hub and authority features, then apply them as additional document features. Being practically attractive for its great simplicity, this paper aims to design an algorithm that exploits both the content and linkage information, by carrying out a joint factorization on both the linkage adjacency matrix and the document-term matrix, and derives a new representation for web pages in a low-dimensional factor space, without explicitly separating them as content, hub or authority factors. Further analysis can be performed based on the compact representation of web pages. In the experiments, the proposed method is compared with state-of-the-art methods and demonstrates an excellent accuracy in hypertext classification on the WebKB and Cora benchmarks.

1. INTRODUCTION With the advance of the World Wide Web, more and more hypertext documents become available on the Web. Some examples of such data include organizational and personal web pages (e.g, the WebKB benchmark data set, which contains university web pages), research papers (e.g., data in CiteSeer), online news articles, and customer-generated media (e.g., blogs). Comparing to data in traditional information management, in addition to content, these data on the Web also contain links: e.g., hyperlinks from a student"s homepage pointing to the homepage of her advisor, paper citations, sources of a news article, comments of one blogger on posts from another blogger, and so on. Performing information management tasks on such structured data raises many new research challenges. In the following discussion, we use the task of web page classification as an illustrating example, while the techniques we develop in later sections are applicable equally well to many other tasks in information retrieval and data mining. For the classification problem of web pages, a simple approach is to treat web pages as independent documents. The advantage of this approach is that many off-the-shelf classification tools can be directly applied to the problem. However, this approach relies only on the content of web pages and ignores the structure of links among them. Link structures provide invaluable information about properties of the documents as well as relationships among them. For example, in the WebKB dataset, the link structure provides additional insights about the relationship among documents (e.g., links often pointing from a student to her advisor or from a faculty member to his projects). Since some links among these documents imply the inter-dependence among the documents, the usual i.i.d. (independent and identical distributed) assumption of documents does not hold any more. From this point of view, the traditional classification methods that ignore the link structure may not be suitable. On the other hand, a few studies, for example [25], rely solely on link structures. It is however a very rare case that content information can be ignorable. For example, in the Cora dataset, the content of a research article abstract largely determines the category of the article. To improve the performance of web page classification, therefore, both link structure and content information should be taken into consideration. To achieve this goal, a simple approach is to convert one type of information to the other. For example, in spam blog classification, Kolari et al. [13] concatenate outlink features with the content features of the blog. In document classification, Kurland and Lee [14] convert content similarity among documents into weights of links. However, link and content information have different properties. For example, a link is an actual piece of evidence that represents an asymmetric relationship whereas the content similarity is usually defined conceptually for every pair of documents in a symmetric way. Therefore, directly converting one type of information to the other usually degrades the quality of information. On the other hand, there exist some studies, as we will discuss in detail in related work, that consider link information and content information separately and then combine them. We argue that such an approach ignores the inherent consistency between link and content information and therefore fails to combine the two seamlessly. Some work, such as [3], incorporates link information using cocitation similarity, but this may not fully capture the global link structure. In Figure 1, for example, web pages v6 and v7 co-cite web page v8, implying that v6 and v7 are similar to each other. In turns, v4 and v5 should be similar to each other, since v4 and v5 cite similar web pages v6 and v7, respectively. But using cocitation similarity, the similarity between v4 and v5 is zero without considering other information. v1 v2 v3 v4 v5 v6 v7 v8 Figure 1: An example of link structure In this paper, we propose a simple technique for analyzing inter-connected documents, such as web pages, using factor analysis[18]. In the proposed technique, both content information and link structures are seamlessly combined through a single set of latent factors. Our model contains two components. The first component captures the content information. This component has a form similar to that of the latent topics in the Latent Semantic Indexing (LSI) [8] in traditional information retrieval. That is, documents are decomposed into latent topics/factors, which in turn are represented as term vectors. The second component captures the information contained in the underlying link structure, such as links from homepages of students to those of faculty members. A factor can be loosely considered as a type of documents (e.g., those homepages belonging to students). It is worth noting that we do not explicitly define the semantic of a factor a priori. Instead, similar to LSI, the factors are learned from the data. Traditional factor analysis models the variables associated with entities through the factors. However, in analysis of link structures, we need to model the relationship of two ends of links, i.e., edges between vertex pairs. Therefore, the model should involve factors of both vertices of the edge. This is a key difference between traditional factor analysis and our model. In our model, we connect two components through a set of shared factors, that is, the latent factors in the second component (for contents) are tied to the factors in the first component (for links). By doing this, we search for a unified set of latent factors that best explains both content and link structures simultaneously and seamlessly. In the formulation, we perform factor analysis based on matrix factorization: solution to the first component is based on factorizing the term-document matrix derived from content features; solution to the second component is based on factorizing the adjacency matrix derived from links. Because the two factorizations share a common base, the discovered bases (latent factors) explain both content information and link structures, and are then used in further information management tasks such as classification. This paper is organized as follows. Section 2 reviews related work. Section 3 presents the proposed approach to analyze the web page based on the combined information of links and content. Section 4 extends the basic framework and a few variants for fine tune. Section 5 shows the experiment results. Section 6 discusses the details of this approach and Section 7 concludes. 2. RELATED WORK In the content analysis part, our approach is closely related to Latent Semantic Indexing (LSI) [8]. LSI maps documents into a lower dimensional latent space. The latent space implicitly captures a large portion of information of documents, therefore it is called the latent semantic space. The similarity between documents could be defined by the dot products of the corresponding vectors of documents in the latent space. Analysis tasks, such as classification, could be performed on the latent space. The commonly used singular value decomposition (SVD) method ensures that the data points in the latent space can optimally reconstruct the original documents. Though our approach also uses latent space to represent web pages (documents), we consider the link structure as well as the content of web pages. In the link analysis approach, the framework of hubs and authorities (HITS) [12] puts web page into two categories, hubs and authorities. Using recursive notion, a hub is a web page with many outgoing links to authorities, while an authority is a web page with many incoming links from hubs. Instead of using two categories, PageRank [17] uses a single category for the recursive notion, an authority is a web page with many incoming links from authorities. He et al. [9] propose a clustering algorithm for web document clustering. The algorithm incorporates link structure and the co-citation patterns. In the algorithm, all links are treated as undirected edge of the link graph. The content information is only used for weighing the links by the textual similarity of both ends of the links. Zhang et al. [23] uses the undirected graph regularization framework for document classification. Achlioptas et al[2] decompose the web into hub and authority attributes then combine them with content. Zhou et al. [25] and [24] propose a directed graph regularization framework for semi-supervised learning. The framework combines the hub and authority information of web pages. But it is difficult to combine the content information into that framework. Our approach consider the content and the directed linkage between topics of source and destination web pages in one step, which implies the topic combines the information of web page as authorities and as hubs in a single set of factors. Cohn and Hofmann [6] construct the latent space from both content and link information, using content analysis based on probabilistic LSI (PLSI) [10] and link analysis based on PHITS [5]. The major difference between the approach of [6] (PLSI+PHITS) and our approach is in the part of link analysis. In PLSI+PHITS, the link is constructed with the linkage from the topic of the source web page to the destination web page. In the model, the outgoing links of the destination web page have no effect on the source web page. In other words, the overall link structure is not utilized in PHITS. In our approach, the link is constructed with the linkage between the factor of the source web page and the factor of the destination web page, instead of the destination web page itself. The factor of the destination web page contains information of its outgoing links. In turn, such information is passed to the factor of the source web page. As the result of matrix factorization, the factor forms a factor graph, a miniature of the original graph, preserving the major structure of the original graph. Taskar et al. [19] propose relational Markov networks (RMNs) for entity classification, by describing a conditional distribution of entity classes given entity attributes and relationships. The model was applied to web page classification, where web pages are entities and hyperlinks are treated as relationships. RMNs apply conditional random fields to define a set of potential functions on cliques of random variables, where the link structure provides hints to form the cliques. However the model does not give an off-the-shelf solution, because the success highly depends on the arts of designing the potential functions. On the other hand, the inference for RMNs is intractable and requires belief propagation. The following are some work on combining documents and links, but the methods are loosely related to our approach. The experiments of [21] show that using terms from the linked document improves the classification accuracy. Chakrabarti et al.[3] use co-citation information in their classification model. Joachims et al.[11] combine text kernels and co-citation kernels for classification. Oh et al [16] use the Naive Bayesian frame to combine link information with content. 3. OUR APPROACH In this section we will first introduce a novel matrix factorization method, which is more suitable than conventional matrix factorization methods for link analysis. Then we will introduce our approach that jointly factorizes the document-term matrix and link matrix and obtains compact and highly indicative factors for representing documents or web pages. 3.1 Link Matrix Factorization Suppose we have a directed graph G = (V, E), where the vertex set V = {vi}n i=1 represents the web pages and the edge set E represents the hyperlinks between web pages. Let A = {asd} denotes the n×n adjacency matrix of G, which is also called the link matrix in this paper. For a pair of vertices, vs and vd, let asd = 1 when there is an edge from vs to vd, and asd = 0, otherwise. Note that A is an asymmetric matrix, because hyperlinks are directed. Most machine learning algorithms assume a feature-vector representation of instances. For web page classification, however, the link graph does not readily give such a vector representation for web pages. If one directly uses each row or column of A for the job, she will suffer a very high computational cost because the dimensionality equals to the number of web pages. On the other hand, it will produces a poor classification accuracy (see our experiments in Section 5), because A is extremely sparse1 . The idea of link matrix factorization is to derive a high-quality feature representation Z of web pages based on analyzing the link matrix A, where Z is an n × l matrix, with each row being the ldimensional feature vector of a web page. The new representation of web pages captures the principal factors of the link structure and makes further processing more efficient. One may use a method similar to LSI, to apply the well-known principal component analysis (PCA) for deriving Z from A. The corresponding optimization problem 2 is min Z,U A − ZU 2 F + γ U 2 F (1) where γ is a small positive number, U is an l ×n matrix, and · F is the Frobenius norm. The optimization aims to approximate A by ZU , a product of two low-rank matrices, with a regularization on U. In the end, the i-th row vector of Z can be thought as the hub feature vector of vertex vi, and the row vector of U can be thought as the authority features. A link generation model proposed in [2] is similar to the PCA approach. Since A is a nonnegative matrix here, one can also consider to put nonnegative constraints on U and Z, which produces an algorithm similar to PLSA [10] and NMF [20]. 1 Due to the sparsity of A, links from two similar pages may not share any common target pages, which makes them to appear dissimilar. However the two pages may be indirectly linked to many common pages via their neighbors. 2 Another equivalent form is minZ,U A − ZU 2 F , s. t. U U = I. The solution Z is identical subject to a scaling factor. However, despite its popularity in matrix analysis, PCA (or other similar methods like PLSA) is restrictive for link matrix factorization. The major problem is that, PCA ignores the fact that the rows and columns of A are indexed by exactly the same set of objects (i.e., web pages). The approximating matrix ˜A = ZU shows no evidence that links are within the same set of objects. To see the drawback, let"s consider a link transitivity situation vi → vs → vj, where page i is linked to page s which itself is linked to page j. Since ˜A = ZU treats A as links from web pages {vi} to a different set of objects, let it be denoted by {oi}, ˜A = ZU actually splits an linked object os from vs and breaks down the link path into two parts vi → os and vs → oj. This is obviously a miss interpretation to the original link path. To overcome the problem of PCA, in this paper we suggest to use a different factorization: min Z,U A − ZUZ 2 F + γ U 2 F (2) where U is an l × l full matrix. Note that U is not symmetric, thus ZUZ produces an asymmetric matrix, which is the case of A. Again, each row vector of Z corresponds to a feature vector of a web pages. The new approximating form ˜A = ZUZ puts a clear meaning that the links are between the same set of objects, represented by features Z. The factor model actually maps each vertex, vi, into a vector zi = {zi,k; 1 ≤ k ≤ l} in the Rl space. We call the Rl space the factor space. Then, {zi} encodes the information of incoming and outgoing connectivity of vertices {vi}. The factor loadings, U, explain how these observed connections happened based on {zi}. Once we have the vector zi, we can use many traditional classification methods (such as SVMs) or clustering tools (such as K-Means) to perform the analysis. Illustration Based on a Synthetic Problem To further illustrate the advantages of the proposed link matrix factorization Eq. (2), let us consider the graph in Figure 1. Given v1 v2 v3 v4 v5 v6 v7 v8 Figure 2: Summarize Figure 1 with a factor graph these observations, we can summarize the graph by grouping as factor graph depicted in Figure 2. In the next we preform the two factorization methods Eq. (2) and Eq. (1) on this link matrix. A good low-rank representation should reveal the structure of the factor graph. First we try PCA-like decomposition, solving Eq. (1) and obtaining Z = U = 2 6 6 6 6 6 6 6 6 6 6 6 6 6 6 6 6 4 1. 1. 0 0 0 0 0 −.6 −.7 .1 0 0 .0 .6 −.0 0 0 .8 −.4 .3 0 0 .2 −.2 −.9 .7 .7 0 0 0 .7 .7 0 0 0 0 0 0 0 0 3 7 7 7 7 7 7 7 7 7 7 7 7 7 7 7 7 5 2 6 6 6 6 6 6 6 6 6 6 6 6 6 6 6 6 4 0 0 0 0 0 .5 −.5 0 0 0 .5 −.5 0 0 0 0 0 −0.6 −.7 .1 0 0 .0 .6 −.0 0 0 .8 −.4 .3 0 0 .2 −.2 −.9 .7 .7 0 0 0 3 7 7 7 7 7 7 7 7 7 7 7 7 7 7 7 7 5 We can see that the row vectors of v6 and v7 are the same in Z, indicating that v6 and v7 have the same hub attributes. The row vectors of v2 and v3 are the same in U, indicating that v2 and v3 have the same authority attributes. It is not clear to see the similarity between v4 and v5, because their inlinks (and outlinks) are different. Then, we factorize A by ZUZ via solving Eq. (2), and obtain the results Z = U = 2 6 6 6 6 6 6 6 6 6 6 6 6 6 6 6 6 4 −.8 −.5 .3 −.1 −.0 −.0 .4 .6 −.1 −.4 −.0 .4 .6 −.1 −.4 .3 −.2 .3 −.4 .3 .3 −.2 .3 −.4 .3 −.4 .5 .0 −.2 .6 −.4 .5 .0 −.2 .6 −.1 .1 −.4 −.8 −.4 3 7 7 7 7 7 7 7 7 7 7 7 7 7 7 7 7 5 2 6 6 6 6 6 6 6 6 4 −.1 −.2 −.4 .6 .7 .2 −.5 −.5 −.5 .0 .1 .1 .4 −.4 .3 .1 −.2 −.0 .3 −.1 −.3 .3 −.5 −.4 −.2 3 7 7 7 7 7 7 7 7 5 The resultant Z is very consistent with the clustering structure of vertices: the row vectors of v2 and v3 are the same, those of v4 and v5 are the same, those of v6 and v7 are the same. Even interestingly, if we add constraints to ensure Z and U be nonnegative, we have Z = U = 2 6 6 6 6 6 6 6 6 6 6 6 6 6 6 6 6 4 1. 0 0 0 0 0 .9 0 0 0 0 .9 0 0 0 0 0 .7 0 0 0 0 .7 0 0 0 0 0 .9 0 0 0 0 .9 0 0 0 0 0 1. 3 7 7 7 7 7 7 7 7 7 7 7 7 7 7 7 7 5 2 6 6 6 6 6 6 6 6 4 0 1. 0 0 0 0 0 .7 0 0 0 0 0 .7 0 0 0 0 0 1. 0 0 0 0 0 3 7 7 7 7 7 7 7 7 5 which clearly tells the assignment of vertices to clusters from Z and the links of factor graph from U. When the interpretability is not critical in some tasks, for example, classification, we found that it achieves better accuracies without the nonnegative constraints. Given our above analysis, it is clear that the factorization ZUZ is more expressive than ZU in representing the link matrix A. 3.2 Content Matrix Factorization Now let us consider the content information on the vertices. To combine the link information and content information, we want to use the same latent space to approximate the content as the latent space for the links. Using the bag-of-words approach, we denote the content of web pages by an n×m matrix C, each of whose rows represents a document, each column represents a keyword, where m is the number of keywords. Like the latent semantic indexing (LSI) [8], the l-dimensional latent space for words is denoted by an m × l matrix V . Therefore, we use ZV to approximate matrix C, min V,Z C − ZV 2 F + β V 2 F , (3) where β is a small positive number, β V 2 F serves as a regularization term to improve the robustness. 3.3 Joint Link-Content Matrix Factorization There are many ways to employ both the content and link information for web page classification. Our idea in this paper is not to simply combine them, but rather to fuse them into a single, consistent, and compact feature representation. To achieve this goal, we solve the following problem, min U,V,Z n J (U, V, Z) def = A − ZUZ 2 F + α C − ZV 2 F + γ U 2 F + β V 2 F o . (4) Eq. (4) is the joined matrix factorization of A and C with regularization. The new representation Z is ensured to capture both the structures of the link matrix A and the content matrix C. Once we find the optimal Z, we can apply the traditional classification or clustering methods on vectorial data Z. The relationship among these matrices can be depicted as Figure 3. A Y C U Z V Figure 3: Relationship among the matrices. Node Y is the target of classification. Eq. (4) can be solved using gradient methods, such as the conjugate gradient method and quasi-Newton methods. Then main computation of gradient methods is evaluating the object function J and its gradients against variables, ∂J ∂U = Z ZUZ Z − Z AZ + γU, ∂J ∂V =α V Z Z − C Z + βV, ∂J ∂Z = ZU Z ZU + ZUZ ZU − A ZU − AZU + α ZV V − CV . Because of the sparsity of A, the computational complexity of multiplication of A and Z is O(µAl), where µA is the number of nonzero entries in A. Similarly, the computational complexity of C Z and CV is O(µC l), where µC is the number of nonzero entries in C. The computational complexity of the rest multiplications in the gradient computation is O(nl2 ). Therefore, the total computational complexity in one iteration is O(µAl + µC l + nl2 ). The number of links and the number of words in a web page are relatively small comparing to the number of web pages, and are almost constant as the number of web pages/documents increases, i.e. µA = O(n) and µC = O(n). Therefore, theoretically the computation time is almost linear to the number of web pages/documents, n. 4. SUPERVISED MATRIX FACTORIZATION Consider a web page classification problem. We can solve Eq. (4) to obtain Z as Section 3, then use a traditional classifier to perform classification. However, this approach does not take data labels into account in the first step. Believing that using data labels improves the accuracy by obtaining a better Z for the classification, we consider to use the data labels to guide the matrix factorization, called supervised matrix factorization [22]. Because some data used in the matrix factorization have no label information, the supervised matrix factorization falls into the category of semi-supervised learning. Let C be the set of classes. For simplicity, we first consider binary class problem, i.e. C = {−1, 1}. Assume we know the labels {yi} for vertices in T ⊂ V. We want to find a hypothesis h : V → R, such that we assign vi to 1 when h(vi) ≥ 0, −1 otherwise. We assume a transform from the latent space to R is linear, i.e. h(vi) = w φ(vi) + b = w zi + b, (5) School course dept. faculty other project staff student total Cornell 44 1 34 581 18 21 128 827 Texas 36 1 46 561 20 2 148 814 Washington 77 1 30 907 18 10 123 1166 Wisconsin 85 0 38 894 25 12 156 1210 Table 1: Dataset of WebKB where w and b are parameters to estimate. Here, w is the norm of the decision boundary. Similar to Support Vector Machines (SVMs) [7], we can use the hinge loss to measure the loss, X i:vi∈T [1 − yih(vi)]+ , where [x]+ is x if x ≥ 0, 0 if x < 0. However, the hinge loss is not smooth at the hinge point, which makes it difficult to apply gradient methods on the problem. To overcome the difficulty, we use a smoothed version of hinge loss for each data point, g(yih(vi)), (6) where g(x) = 8 >< >: 0 when x ≥ 2, 1 − x when x ≤ 0, 1 4 (x − 2)2 when 0 < x < 2. We reduce a multiclass problem into multiple binary ones. One simple scheme of reduction is the one-against-rest coding scheme. In the one-against-rest scheme, we assign a label vector for each class label. The element of a label vector is 1 if the data point belongs the corresponding class, −1, if the data point does not belong the corresponding class, 0, if the data point is not labeled. Let Y be the label matrix, each column of which is a label vector. Therefore, Y is a matrix of n × c, where c is the number of classes, |C|. Then the values of Eq. (5) form a matrix H = ZW + 1b , (7) where 1 is a vector of size n, whose elements are all one, W is a c × l parameter matrix, and b is a parameter vector of size c. The total loss is proportional to the sum of Eq. (6) over all labeled data points and the classes, LY (W, b, Z) = λ X i:vi∈T ,j∈C g(YijHij), where λ is the parameter to scale the term. To derive a robust solution, we also use Tikhonov regularization for W, ΩW (W) = ν 2 W 2 F , where ν is the parameter to scale the term. Then the supervised matrix factorization problem becomes min U,V,Z,W,b Js(U, V, Z, W, b) (8) where Js(U, V, Z, W, b) = J (U, V, Z) + LY (W, b, Z) + ΩW (W). We can also use gradient methods to solve the problem of Eq. (8). The gradients are ∂Js ∂U = ∂J ∂U , ∂Js ∂V = ∂J ∂V , ∂Js ∂Z = ∂J ∂Z + λGW, ∂Js ∂W =λG Z + νW, ∂Js ∂b =λG 1, where G is an n×c matrix, whose ik-th element is Yikg (YikHik), and g (x) = 8 >< >: 0 when x ≥ 2, −1 when x ≤ 0, 1 2 (x − 2) when 0 < x < 2. Once we obtain w, b, and Z, we can apply h on the vertices with unknown class labels, or apply traditional classification algorithms on Z to get the classification results. 5. EXPERIMENTS 5.1 Data Description In this section, we perform classification on two datasets, to demonstrate the our approach. The two datasets are the WebKB data set[1] and the Cora data set [15]. The WebKB data set consists of about 6000 web pages from computer science departments of four schools (Cornell, Texas, Washington, and Wisconsin). The web pages are classified into seven categories. The numbers of pages in each category are shown in Table 1. The Cora data set consists of the abstracts and references of about 34,000 computer science research papers. We use part of them to categorize into one of subfields of data structure (DS), hardware and architecture (HA), machine learning (ML), and programing language (PL). We remove those articles without reference to other articles in the set. The number of papers and the number of subfields in each area are shown in Table 2. area # of papers # of subfields Data structure (DS) 751 9 Hardware and architecture (HA) 400 7 Machine learning (ML) 1617 7 Programing language (PL) 1575 9 Table 2: Dataset of Cora 5.2 Methods The task of the experiments is to classify the data based on their content information and/or link structure. We use the following methods: 65 70 75 80 85 90 95 100 WisconsinWashingtonTexasCornell accuracy(%) dataset SVM on content SVM on link SVM on link-content Directed graph reg. PLSI+PHITS link-content MF link-content sup. MF method Cornell Texas Washington Wisconsin SVM on content 81.00 ± 0.90 77.00 ± 0.60 85.20 ± 0.90 84.30 ± 0.80 SVM on links 70.10 ± 0.80 75.80 ± 1.20 80.50 ± 0.30 74.90 ± 1.00 SVM on link-content 80.00 ± 0.80 78.30 ± 1.00 85.20 ± 0.70 84.00 ± 0.90 Directed graph regularization 89.47 ± 1.41 91.28 ± 0.75 91.08 ± 0.51 89.26 ± 0.45 PLSI+PHITS 80.74 ± 0.88 76.15 ± 1.29 85.12 ± 0.37 83.75 ± 0.87 link-content MF 93.50 ± 0.80 96.20 ± 0.50 93.60 ± 0.50 92.60 ± 0.60 link-content sup. MF 93.80 ± 0.70 97.07 ± 1.11 93.70 ± 0.60 93.00 ± 0.30 Table 3: Classification accuracy (mean ± std-err %) on WebKB data set • SVM on content We apply support vector machines (SVM) on the content of documents. The features are the bag-ofwords and all word are stemmed. This method ignores link structure in the data. Linear SVM is used. The regularization parameter of SVM is selected using the cross-validation method. The implementation of SVM used in the experiments is libSVM[4]. • SVM on links We treat links as the features of each document, i.e. the i-th feature is link-to-pagei. We apply SVM on link features. This method uses link information, but not the link structure. • SVM on link-content We combine the features of the above two methods. We use different weights for these two set of features. The weights are also selected using crossvalidation. • Directed graph regularization This method is described in [25] and [24]. This method is solely based on link structure. • PLSI+PHITS This method is described in [6]. This method combines text content information and link structure for analysis. The PHITS algorithm is in spirit similar to Eq.1, with an additional nonnegative constraint. It models the outgoing and in-coming structures separately. • Link-content MF This is our approach of matrix factorization described in Section 3. We use 50 latent factors for Z. After we compute Z, we train a linear SVM using Z as the feature vectors, then apply SVM on testing portion of Z to obtain the final result, because of the multiclass output. • Link-content sup. MF This method is our approach of the supervised matrix factorization in Section 4. We use 50 latent factors for Z. After we compute Z, we train a linear SVM on the training portion of Z, then apply SVM on testing portion of Z to obtain the final result, because of the multiclass output. We randomly split data into five folds and repeat the experiment for five times, for each time we use one fold for test, four other folds for training. During the training process, we use the crossvalidation to select all model parameters. We measure the results by the classification accuracy, i.e., the percentage of the number of correct classified documents in the entire data set. The results are shown as the average classification accuracies and it standard deviation over the five repeats. 5.3 Results The average classification accuracies for the WebKB data set are shown in Table 3. For this task, the accuracies of SVM on links are worse than that of SVM on content. But the directed graph regularization, which is also based on link alone, achieves a much higher accuracy. This implies that the link structure plays an important role in the classification of this dataset, but individual links in a web page give little information. The combination of link and content using SVM achieves similar accuracy as that of SVM on content alone, which confirms individual links in a web page give little information. Since our approach consider the link structure as well as the content information, our two methods give results a highest accuracies among these approaches. The difference between the results of our two methods is not significant. However in the experiments below, we show the difference between them. The classification accuracies for the Cora data set are shown in Table 4. In this experiment, the accuracies of SVM on the combination of links and content are higher than either SVM on content or SVM on links. This indicates both content and links are infor45 50 55 60 65 70 75 80 PLMLHADS accuracy(%) dataset SVM on content SVM on link SVM on link-content Directed graph reg. PLSI+PHITS link-content MF link-content sup. MF method DS HA ML PL SVM on content 53.70 ± 0.50 67.50 ± 1.70 68.30 ± 1.60 56.40 ± 0.70 SVM on links 48.90 ± 1.70 65.80 ± 1.40 60.70 ± 1.10 58.20 ± 0.70 SVM on link-content 63.70 ± 1.50 70.50 ± 2.20 70.56 ± 0.80 62.35 ± 1.00 Directed graph regularization 46.07 ± 0.82 65.50 ± 2.30 59.37 ± 0.96 56.06 ± 0.84 PLSI+PHITS 53.60 ± 1.78 67.40 ± 1.48 67.51 ± 1.13 57.45 ± 0.68 link-content MF 61.00 ± 0.70 74.20 ± 1.20 77.50 ± 0.80 62.50 ± 0.80 link-content sup. MF 69.38 ± 1.80 74.20 ± 0.70 78.70 ± 0.90 68.76 ± 1.32 Table 4: Classification accuracy (mean ± std-err %) on Cora data set mative for classifying the articles into subfields. The method of directed graph regularization does not perform as good as SVM on link-content, which confirms the importance of the article content in this task. Though our method of link-content matrix factorization perform slightly better than other methods, our method of linkcontent supervised matrix factorization outperform significantly. 5.4 The Number of Factors As we discussed in Section 3, the computational complexity of each iteration for solving the optimization problem is quadratic to the number of factors. We perform experiments to study how the number of factors affects the accuracy of predication. We use different numbers of factors for the Cornell data of WebKB data set and the machine learning (ML) data of Cora data set. The result shown in Figure 4(a) and 4(b). The figures show that the accuracy 88 89 90 91 92 93 94 95 0 10 20 30 40 50 accuracy(%) number of factors link-content sup. MF link-content MF (a) Cornell data 62 64 66 68 70 72 74 76 78 80 0 10 20 30 40 50 accuracy(%) number of factors link-content sup. MF link-content MF (b) ML data Figure 4: Accuracy vs number of factors increases as the number of factors increases. It is a different concept from choosing the optimal number of clusters in clustering application. It is how much information to represent in the latent variables. We have considered the regularization over the factors, which avoids the overfit problem for a large number of factors. To choose of the number of factors, we need to consider the trade-off between the accuracy and the computation time, which is quadratic to the number of factors. The difference between the method of matrix factorization and that of supervised one decreases as the number of factors increases. This indicates that the usefulness of supervised matrix factorization at lower number of factors. 6. DISCUSSIONS The loss functions LA in Eq. (2) and LC in Eq. (3) use squared loss due to computationally convenience. Actually, squared loss does not precisely describe the underlying noise model, because the weights of adjacency matrix can only take nonnegative values, in our case, zero or one only, and the components of content matrix C can only take nonnegative integers. Therefore, we can apply other types of loss, such as hinge loss or smoothed hinge loss, e.g. LA(U, Z) = µh(A, ZUZ ), where h(A, B) =P i,j [1 − AijBij]+ . In our paper, we mainly discuss the application of classification. A entry of matrix Z means the relationship of a web page and a factor. The values of the entries are the weights of linear model, instead of the probabilities of web pages belonging to latent topics. Therefore, we allow the components take any possible real values. When we come to the clustering application, we can use this model to find Z, then apply K-means to partition the web pages into clusters. Actually, we can use the idea of nonnegative matrix factorization for clustering [20] to directly cluster web pages. As the example with nonnegative constraints shown in Section 3, we represent each cluster by a latent topic, i.e. the dimensionality of the latent space is set to the number of clusters we want. Then the problem of Eq. (4) becomes min U,V,Z J (U, V, Z), s.t.Z ≥ 0. (9) Solving Eq. (9), we can obtain more interpretable results, which could be used for clustering. 7. CONCLUSIONS In this paper, we study the problem of how to combine the information of content and links for web page analysis, mainly on classification application. We propose a simple approach using factors to model the text content and link structure of web pages/documents. The directed links are generated from the linear combination of linkage of between source and destination factors. By sharing factors between text content and link structure, it is easy to combine both the content information and link structure. Our experiments show our approach is effective for classification. We also discuss an extension for clustering application. Acknowledgment We would like to thank Dr. Dengyong Zhou for sharing his code of his algorithm. Also, thanks to the reviewers for constructive comments. 8. REFERENCES [1] CMU world wide knowledge base (WebKB) project. Available at http://www.cs.cmu.edu/∼WebKB/. [2] D. Achlioptas, A. Fiat, A. R. Karlin, and F. McSherry. Web search via hub synthesis. In IEEE Symposium on Foundations of Computer Science, pages 500-509, 2001. [3] S. Chakrabarti, B. E. Dom, and P. Indyk. Enhanced hypertext categorization using hyperlinks. In L. M. Haas and A. Tiwary, editors, Proceedings of SIGMOD-98, ACM International Conference on Management of Data, pages 307-318, Seattle, US, 1998. ACM Press, New York, US. [4] C.-C. Chang and C.-J. Lin. LIBSVM: a library for support vector machines, 2001. Software available at http://www.csie.ntu.edu.tw/∼cjlin/libsvm. [5] D. Cohn and H. Chang. Learning to probabilistically identify authoritative documents. Proc. ICML 2000. pp.167-174., 2000. [6] D. Cohn and T. Hofmann. The missing link - a probabilistic model of document content and hypertext connectivity. In T. K. Leen, T. G. Dietterich, and V. Tresp, editors, Advances in Neural Information Processing Systems 13, pages 430-436. MIT Press, 2001. [7] C. Cortes and V. Vapnik. Support-vector networks. Machine Learning, 20:273, 1995. [8] S. C. Deerwester, S. T. Dumais, T. K. Landauer, G. W. Furnas, and R. A. Harshman. Indexing by latent semantic analysis. Journal of the American Society of Information Science, 41(6):391-407, 1990. [9] X. He, H. Zha, C. Ding, and H. Simon. Web document clustering using hyperlink structures. Computational Statistics and Data Analysis, 41(1):19-45, 2002. [10] T. Hofmann. Probabilistic latent semantic indexing. In Proceedings of the Twenty-Second Annual International SIGIR Conference, 1999. [11] T. Joachims, N. Cristianini, and J. Shawe-Taylor. Composite kernels for hypertext categorisation. In C. Brodley and A. Danyluk, editors, Proceedings of ICML-01, 18th International Conference on Machine Learning, pages 250-257, Williams College, US, 2001. Morgan Kaufmann Publishers, San Francisco, US. [12] J. M. Kleinberg. Authoritative sources in a hyperlinked environment. J. ACM, 48:604-632, 1999. [13] P. Kolari, T. Finin, and A. Joshi. SVMs for the Blogosphere: Blog Identification and Splog Detection. In AAAI Spring Symposium on Computational Approaches to Analysing Weblogs, March 2006. [14] O. Kurland and L. Lee. Pagerank without hyperlinks: structural re-ranking using links induced by language models. In SIGIR "05: Proceedings of the 28th annual international ACM SIGIR conference on Research and development in information retrieval, pages 306-313, New York, NY, USA, 2005. ACM Press. [15] A. McCallum, K. Nigam, J. Rennie, and K. Seymore. Automating the contruction of internet portals with machine learning. Information Retrieval Journal, 3(127-163), 2000. [16] H.-J. Oh, S. H. Myaeng, and M.-H. Lee. A practical hypertext catergorization method using links and incrementally available class information. In SIGIR "00: Proceedings of the 23rd annual international ACM SIGIR conference on Research and development in information retrieval, pages 264-271, New York, NY, USA, 2000. ACM Press. [17] L. Page, S. Brin, R. Motowani, and T. Winograd. PageRank citation ranking: bring order to the web. Stanford Digital Library working paper 1997-0072, 1997. [18] C. Spearman. General Intelligence, objectively determined and measured. The American Journal of Psychology, 15(2):201-292, Apr 1904. [19] B. Taskar, P. Abbeel, and D. Koller. Discriminative probabilistic models for relational data. In Proceedings of 18th International UAI Conference, 2002. [20] W. Xu, X. Liu, and Y. Gong. Document clustering based on non-negative matrix factorization. In SIGIR "03: Proceedings of the 26th annual international ACM SIGIR conference on Research and development in informaion retrieval, pages 267-273. ACM Press, 2003. [21] Y. Yang, S. Slattery, and R. Ghani. A study of approaches to hypertext categorization. Journal of Intelligent Information Systems, 18(2-3):219-241, 2002. [22] K. Yu, S. Yu, and V. Tresp. Multi-label informed latent semantic indexing. In SIGIR "05: Proceedings of the 28th annual international ACM SIGIR conference on Research and development in information retrieval, pages 258-265, New York, NY, USA, 2005. ACM Press. [23] T. Zhang, A. Popescul, and B. Dom. Linear prediction models with graph regularization for web-page categorization. In KDD "06: Proceedings of the 12th ACM SIGKDD international conference on Knowledge discovery and data mining, pages 821-826, New York, NY, USA, 2006. ACM Press. [24] D. Zhou, J. Huang, and B. Sch¨olkopf. Learning from labeled and unlabeled data on a directed graph. In Proceedings of the 22nd International Conference on Machine Learning, Bonn, Germany, 2005. [25] D. Zhou, B. Sch¨olkopf, and T. Hofmann. Semi-supervised learning on directed graphs. Proc. Neural Info. Processing Systems, 2004.

low-dimensional factor space;web mining problem;classification;combining content and link;linkage adjacency matrix;content information;authority information;joint factorization;relationship;link structure;factor analysis;asymmetric relationship;document-term matrix;cocitation similarity;webkb and cora benchmark;text content;matrix factorization

train_H-44

A Time Machine for Text Search

Text search over temporally versioned document collections such as web archives has received little attention as a research problem. As a consequence, there is no scalable and principled solution to search such a collection as of a specified time t. In this work, we address this shortcoming and propose an efficient solution for time-travel text search by extending the inverted file index to make it ready for temporal search. We introduce approximate temporal coalescing as a tunable method to reduce the index size without significantly affecting the quality of results. In order to further improve the performance of time-travel queries, we introduce two principled techniques to trade off index size for its performance. These techniques can be formulated as optimization problems that can be solved to near-optimality. Finally, our approach is evaluated in a comprehensive series of experiments on two large-scale real-world datasets. Results unequivocally show that our methods make it possible to build an efficient time machine scalable to large versioned text collections.

1. INTRODUCTION In this work we address time-travel text search over temporally versioned document collections. Given a keyword query q and a time t our goal is to identify and rank relevant documents as if the collection was in its state as of time t. An increasing number of such versioned document collections is available today including web archives, collaborative authoring environments like Wikis, or timestamped information feeds. Text search on these collections, however, is mostly time-ignorant: while the searched collection changes over time, often only the most recent version of a documents is indexed, or, versions are indexed independently and treated as separate documents. Even worse, for some collections, in particular web archives like the Internet Archive [18], a comprehensive text-search functionality is often completely missing. Time-travel text search, as we develop it in this paper, is a crucial tool to explore these collections and to unfold their full potential as the following example demonstrates. For a documentary about a past political scandal, a journalist needs to research early opinions and statements made by the involved politicians. Sending an appropriate query to a major web search-engine, the majority of returned results contains only recent coverage, since many of the early web pages have disappeared and are only preserved in web archives. If the query could be enriched with a time point, say August 20th 2003 as the day after the scandal got revealed, and be issued against a web archive, only pages that existed specifically at that time could be retrieved thus better satisfying the journalist"s information need. Document collections like the Web or Wikipedia [32], as we target them here, are already large if only a single snapshot is considered. Looking at their evolutionary history, we are faced with even larger data volumes. As a consequence, na¨ıve approaches to time-travel text search fail, and viable approaches must scale-up well to such large data volumes. This paper presents an efficient solution to time-travel text search by making the following key contributions: 1. The popular well-studied inverted file index [35] is transparently extended to enable time-travel text search. 2. Temporal coalescing is introduced to avoid an indexsize explosion while keeping results highly accurate. 3. We develop two sublist materialization techniques to improve index performance that allow trading off space vs. performance. 4. In a comprehensive experimental evaluation our approach is evaluated on the English Wikipedia and parts of the Internet Archive as two large-scale real-world datasets with versioned documents. The remainder of this paper is organized as follows. The presented work is put in context with related work in Section 2. We delineate our model of a temporally versioned document collection in Section 3. We present our time-travel inverted index in Section 4. Building on it, temporal coalescing is described in Section 5. In Section 6 we describe principled techniques to improve index performance, before presenting the results of our experimental evaluation in Section 7. 2. RELATED WORK We can classify the related work mainly into the following two categories: (i) methods that deal explicitly with collections of versioned documents or temporal databases, and (ii) methods for reducing the index size by exploiting either the document-content overlap or by pruning portions of the index. We briefly review work under these categories here. To the best of our knowledge, there is very little prior work dealing with historical search over temporally versioned documents. Anick and Flynn [3], while pioneering this research, describe a help-desk system that supports historical queries. Access costs are optimized for accesses to the most recent versions and increase as one moves farther into the past. Burrows and Hisgen [10], in a patent description, delineate a method for indexing range-based values and mention its potential use for searching based on dates associated with documents. Recent work by Nørv˚ag and Nybø [25] and their earlier proposals concentrate on the relatively simpler problem of supporting text-containment queries only and neglect the relevance scoring of results. Stack [29] reports practical experiences made when adapting the open source search-engine Nutch to search web archives. This adaptation, however, does not provide the intended time-travel text search functionality. In contrast, research in temporal databases has produced several index structures tailored for time-evolving databases; a comprehensive overview of the state-of-art is available in [28]. Unlike the inverted file index, their applicability to text search is not well understood. Moving on to the second category of related work, Broder et al. [8] describe a technique that exploits large content overlaps between documents to achieve a reduction in index size. Their technique makes strong assumptions about the structure of document overlaps rendering it inapplicable to our context. More recent approaches by Hersovici et al. [17] and Zhang and Suel [34] exploit arbitrary content overlaps between documents to reduce index size. None of the approaches, however, considers time explicitly or provides the desired time-travel text search functionality. Static indexpruning techniques [11, 12] aim to reduce the effective index size, by removing portions of the index that are expected to have low impact on the query result. They also do not consider temporal aspects of documents, and thus are technically quite different from our proposal despite having a shared goal of index-size reduction. It should be noted that index-pruning techniques can be adapted to work along with the temporal text index we propose here. 3. MODEL In the present work, we deal with a temporally versioned document collection D that is modeled as described in the following. Each document d ∈ D is a sequence of its versions d = dt1 , dt2 , . . . . Each version dti has an associated timestamp ti reflecting when the version was created. Each version is a vector of searchable terms or features. Any modification to a document version results in the insertion of a new version with corresponding timestamp. We employ a discrete definition of time, so that timestamps are non-negative integers. The deletion of a document at time ti, i.e., its disappearance from the current state of the collection, is modeled as the insertion of a special tombstone version ⊥. The validity time-interval val(dti ) of a version dti is [ti, ti+1), if a newer version with associated timestamp ti+1 exists, and [ti, now) otherwise where now points to the greatest possible value of a timestamp (i.e., ∀t : t < now). Putting all this together, we define the state Dt of the collection at time t (i.e., the set of versions valid at t that are not deletions) as Dt = [ d∈D {dti ∈ d | t ∈ val(dti ) ∧ dti = ⊥} . As mentioned earlier, we want to enrich a keyword query q with a timestamp t, so that q be evaluated over Dt , i.e., the state of the collection at time t. The enriched time-travel query is written as q t for brevity. As a retrieval model in this work we adopt Okapi BM25 [27], but note that the proposed techniques are not dependent on this choice and are applicable to other retrieval models like tf-idf [4] or language models [26] as well. For our considered setting, we slightly adapt Okapi BM25 as w(q t , dti ) = X v∈q wtf (v, dti ) · widf (v, t) . In the above formula, the relevance w(q t , dti ) of a document version dti to the time-travel query q t is defined. We reiterate that q t is evaluated over Dt so that only the version dti valid at time t is considered. The first factor wtf (v, dti ) in the summation, further referred to as the tfscore is defined as wtf (v, dti ) = (k1 + 1) · tf(v, dti ) k1 · ((1 − b) + b · dl(d ti ) avdl(ti) ) + tf(v, dti ) . It considers the plain term frequency tf(v, dti ) of term v in version dti normalizing it, taking into account both the length dl(dti ) of the version and the average document length avdl(ti) in the collection at time ti. The length-normalization parameter b and the tf-saturation parameter k1 are inherited from the original Okapi BM25 and are commonly set to values 1.2 and 0.75 respectively. The second factor widf (v, t), which we refer to as the idf-score in the remainder, conveys the inverse document frequency of term v in the collection at time t and is defined as widf (v, t) = log N(t) − df(v, t) + 0.5 df(v, t) + 0.5 where N(t) = |Dt | is the collection size at time t and df(v, t) gives the number of documents in the collection that contain the term v at time t. While the idf-score depends on the whole corpus as of the query time t, the tf-score is specific to each version. 4. TIME-TRAVELINVERTEDFILEINDEX The inverted file index is a standard technique for text indexing, deployed in many systems. In this section, we briefly review this technique and present our extensions to the inverted file index that make it ready for time-travel text search. 4.1 Inverted File Index An inverted file index consists of a vocabulary, commonly organized as a B+-Tree, that maps each term to its idfscore and inverted list. The index list Lv belonging to term v contains postings of the form ( d, p ) where d is a document-identifier and p is the so-called payload. The payload p contains information about the term frequency of v in d, but may also include positional information about where the term appears in the document. The sort-order of index lists depends on which queries are to be supported efficiently. For Boolean queries it is favorable to sort index lists in document-order. Frequencyorder and impact-order sorted index lists are beneficial for ranked queries and enable optimized query processing that stops early after having identified the k most relevant documents [1, 2, 9, 15, 31]. A variety of compression techniques, such as encoding document identifiers more compactly, have been proposed [33, 35] to reduce the size of index lists. For an excellent recent survey about inverted file indexes we refer to [35]. 4.2 Time-Travel Inverted File Index In order to prepare an inverted file index for time travel we extend both inverted lists and the vocabulary structure by explicitly incorporating temporal information. The main idea for inverted lists is that we include a validity timeinterval [tb, te) in postings to denote when the payload information was valid. The postings in our time-travel inverted file index are thus of the form ( d, p, [tb, te) ) where d and p are defined as in the standard inverted file index above and [tb, te) is the validity time-interval. As a concrete example, in our implementation, for a version dti having the Okapi BM25 tf-score wtf (v, dti ) for term v, the index list Lv contains the posting ( d, wtf (v, dti ), [ti, ti+1) ) . Similarly, the extended vocabulary structure maintains for each term a time-series of idf-scores organized as a B+Tree. Unlike the tf-score, the idf-score of every term could vary with every change in the corpus. Therefore, we take a simplified approach to idf-score maintenance, by computing idf-scores for all terms in the corpus at specific (possibly periodic) times. 4.3 Query Processing During processing of a time-travel query q t , for each query term the corresponding idf-score valid at time t is retrieved from the extended vocabulary. Then, index lists are sequentially read from disk, thereby accumulating the information contained in the postings. We transparently extend the sequential reading, which is - to the best of our knowledgecommon to all query processing techniques on inverted file indexes, thus making them suitable for time-travel queryprocessing. To this end, sequential reading is extended by skipping all postings whose validity time-interval does not contain t (i.e., t ∈ [tb, te)). Whether a posting can be skipped can only be decided after the posting has been transferred from disk into memory and therefore still incurs significant I/O cost. As a remedy, we propose index organization techniques in Section 6 that aim to reduce the I/O overhead significantly. We note that our proposed extension of the inverted file index makes no assumptions about the sort-order of index lists. As a consequence, existing query-processing techniques and most optimizations (e.g., compression techniques) remain equally applicable. 5. TEMPORAL COALESCING If we employ the time-travel inverted index, as described in the previous section, to a versioned document collection, we obtain one posting per term per document version. For frequent terms and large highly-dynamic collections, this time score non-coalesced coalesced Figure 1: Approximate Temporal Coalescing leads to extremely long index lists with very poor queryprocessing performance. The approximate temporal coalescing technique that we propose in this section counters this blowup in index-list size. It builds on the observation that most changes in a versioned document collection are minor, leaving large parts of the document untouched. As a consequence, the payload of many postings belonging to temporally adjacent versions will differ only slightly or not at all. Approximate temporal coalescing reduces the number of postings in an index list by merging such a sequence of postings that have almost equal payloads, while keeping the maximal error bounded. This idea is illustrated in Figure 1, which plots non-coalesced and coalesced scores of postings belonging to a single document. Approximate temporal coalescing is greatly effective given such fluctuating payloads and reduces the number of postings from 9 to 3 in the example. The notion of temporal coalescing was originally introduced in temporal database research by B¨ohlen et al. [6], where the simpler problem of coalescing only equal information was considered. We next formally state the problem dealt with in approximate temporal coalescing, and discuss the computation of optimal and approximate solutions. Note that the technique is applied to each index list separately, so that the following explanations assume a fixed term v and index list Lv. As an input we are given a sequence of temporally adjacent postings I = ( d, pi, [ti, ti+1) ), . . . , ( d, pn−1, [tn−1, tn)) ) . Each sequence represents a contiguous time period during which the term was present in a single document d. If a term disappears from d but reappears later, we obtain multiple input sequences that are dealt with separately. We seek to generate the minimal length output sequence of postings O = ( d, pj, [tj, tj+1) ), . . . , ( d, pm−1, [tm−1, tm)) ) , that adheres to the following constraints: First, O and I must cover the same time-range, i.e., ti = tj and tn = tm. Second, when coalescing a subsequence of postings of the input into a single posting of the output, we want the approximation error to be below a threshold . In other words, if (d, pi, [ti, ti+1)) and (d, pj, [tj, tj+1)) are postings of I and O respectively, then the following must hold for a chosen error function and a threshold : tj ≤ ti ∧ ti+1 ≤ tj+1 ⇒ error(pi, pj) ≤ . In this paper, as an error function we employ the relative error between payloads (i.e., tf-scores) of a document in I and O, defined as: errrel(pi, pj) = |pi − pj| / |pi| . Finding an optimal output sequence of postings can be cast into finding a piecewise-constant representation for the points (ti, pi) that uses a minimal number of segments while retaining the above approximation guarantee. Similar problems occur in time-series segmentation [21, 30] and histogram construction [19, 20]. Typically dynamic programming is applied to obtain an optimal solution in O(n2 m∗ ) [20, 30] time with m∗ being the number of segments in an optimal sequence. In our setting, as a key difference, only a guarantee on the local error is retained - in contrast to a guarantee on the global error in the aforementioned settings. Exploiting this fact, an optimal solution is computable by means of induction [24] in O(n2 ) time. Details of the optimal algorithm are omitted here but can be found in the accompanying technical report [5]. The quadratic complexity of the optimal algorithm makes it inappropriate for the large datasets encountered in this work. As an alternative, we introduce a linear-time approximate algorithm that is based on the sliding-window algorithm given in [21]. This algorithm produces nearly-optimal output sequences that retain the bound on the relative error, but possibly require a few additional segments more than an optimal solution. Algorithm 1 Temporal Coalescing (Approximate) 1: I = ( d, pi, [ti, ti+1) ), . . . O = 2: pmin = pi pmax = pi p = pi tb = ti te = ti+1 3: for ( d, pj, [tj, tj+1) ) ∈ I do 4: pmin = min( pmin, pj ) pmax = max( pmax, pj ) 5: p = optrep(pmin, pmax) 6: if errrel(pmin, p ) ≤ ∧ errrel(pmax, p ) ≤ then 7: pmin = pmin pmax = pmax p = p te = tj+1 8: else 9: O = O ∪ ( d, p, [tb, te) ) 10: pmin = pj pmax = pj p = pj tb = tj te = tj+1 11: end if 12: end for 13: O = O ∪ ( d, p, [tb, te) ) Algorithm 1 makes one pass over the input sequence I. While doing so, it coalesces sequences of postings having maximal length. The optimal representative for a sequence of postings depends only on their minimal and maximal payload (pmin and pmax) and can be looked up using optrep in O(1) (see [16] for details). When reading the next posting, the algorithm tries to add it to the current sequence of postings. It computes the hypothetical new representative p and checks whether it would retain the approximation guarantee. If this test fails, a coalesced posting bearing the old representative is added to the output sequence O and, following that, the bookkeeping is reinitialized. The time complexity of the algorithm is in O(n). Note that, since we make no assumptions about the sort order of index lists, temporal-coalescing algorithms have an additional preprocessing cost in O(|Lv| log |Lv|) for sorting the index list and chopping it up into subsequences for each document. 6. SUBLIST MATERIALIZATION Efficiency of processing a query q t on our time-travel inverted index is influenced adversely by the wasted I/O due to read but skipped postings. Temporal coalescing implicitly addresses this problem by reducing the overall index list size, but still a significant overhead remains. In this section, we tackle this problem by proposing the idea of materializing sublists each of which corresponds to a contiguous subinterval of time spanned by the full index. Each of these sublists contains all coalesced postings that overlap with the corresponding time interval of the sublist. Note that all those postings whose validity time-interval spans across the temporal boundaries of several sublists are replicated in each of the spanned sublists. Thus, in order to process the query q t time t1 t2 t3 t4 t5 t6 t7 t8 t9 t10 d1 d2 d3 document 1 2 3 4 5 6 7 8 9 10 Figure 2: Sublist Materialization it is sufficient to scan any materialized sublist whose timeinterval contains t. We illustrate the idea of sublist materialization using an example shown in Figure 2. The index list Lv visualized in the figure contains a total of 10 postings from three documents d1, d2, and d3. For ease of description, we have numbered boundaries of validity time-intervals, in increasing time-order, as t1, . . . , t10 and numbered the postings themselves as 1, . . . , 10. Now, consider the processing of a query q t with t ∈ [t1, t2) using this inverted list. Although only three postings (postings 1, 5 and 8) are valid at time t, the whole inverted list has to be read in the worst case. Suppose that we split the time axis of the list at time t2, forming two sublists with postings {1, 5, 8} and {2, 3, 4, 5, 6, 7, 8, 9, 10} respectively. Then, we can process the above query with optimal cost by reading only those postings that existed at this t. At a first glance, it may seem counterintuitive to reduce index size in the first step (using temporal coalescing), and then to increase it again using the sublist materialization techniques presented in this section. However, we reiterate that our main objective is to improve the efficiency of processing queries, not to reduce the index size alone. The use of temporal coalescing improves the performance by reducing the index size, while the sublist materialization improves performance by judiciously replicating entries. Further, the two techniques, can be applied separately and are independent. If applied in conjunction, though, there is a synergetic effect - sublists that are materialized from a temporally coalesced index are generally smaller. We employ the notation Lv : [ti, tj) to refer to the materialized sublist for the time interval [ti, tj), that is formally defined as, Lv : [ti, tj) = {( d, p, [tb, te) ) ∈ Lv | tb < tj ∧ te > ti} . To aid the presentation in the rest of the paper, we first provide some definitions. Let T = t1 . . . tn be the sorted sequence of all unique time-interval boundaries of an inverted list Lv. Then we define E = { [ti, ti+1) | 1 ≤ i < n} to be the set of elementary time intervals. We refer to the set of time intervals for which sublists are materialized as M ⊆ { [ti, tj) | 1 ≤ i < j ≤ n } , and demand ∀ t ∈ [t1, tn) ∃ m ∈ M : t ∈ m , i.e., the time intervals in M must completely cover the time interval [t1, tn), so that time-travel queries q t for all t ∈ [t1, tn) can be processed. We also assume that intervals in M are disjoint. We can make this assumption without ruling out any optimal solution with regard to space or performance defined below. The space required for the materialization of sublists in a set M is defined as S( M ) = X m∈M |Lv : m| , i.e., the total length of all lists in M. Given a set M, we let π( [ti, ti+1) ) = [tj, tk) ∈ M : [ti, ti+1) ⊆ [tj, tk) denote the time interval that is used to process queries q t with t ∈ [ti, ti+1). The performance of processing queries q t for t ∈ [ti, ti+1) inversely depends on its processing cost PC( [ti, ti+1) ) = |Lv : π( [ti, ti+1) )| , which is assumed to be proportional to the length of the list Lv : π( [ti, ti+1) ). Thus, in order to optimize the performance of processing queries we minimize their processing costs. 6.1 Performance/Space-Optimal Approaches One strategy to eliminate the problem of skipped postings is to eagerly materialize sublists for all elementary time intervals, i.e., to choose M = E. In doing so, for every query q t only postings valid at time t are read and thus the best possible performance is achieved. Therefore, we will refer to this approach as Popt in the remainder. The initial approach described above that keeps only the full list Lv and thus picks M = { [t1, tn) } is referred to as Sopt in the remainder. This approach requires minimal space, since it keeps each posting exactly once. Popt and Sopt are extremes: the former provides the best possible performance but is not space-efficient, the latter requires minimal space but does not provide good performance. The two approaches presented in the rest of this section allow mutually trading off space and performance and can thus be thought of as means to explore the configuration spectrum between the Popt and the Sopt approach. 6.2 Performance-Guarantee Approach The Popt approach clearly wastes a lot of space materializing many nearly-identical sublists. In the example illustrated in Figure 2 materialized sublists for [t1, t2) and [t2, t3) differ only by one posting. If the sublist for [t1, t3) was materialized instead, one could save significant space while incurring only an overhead of one skipped posting for all t ∈ [t1, t3). The technique presented next is driven by the idea that significant space savings over Popt are achievable, if an upper-bounded loss on the performance can be tolerated, or to put it differently, if a performance guarantee relative to the optimum is to be retained. In detail, the technique, which we refer to as PG (Performance Guarantee) in the remainder, finds a set M that has minimal required space, but guarantees for any elementary time interval [ti, ti+1) (and thus for any query q t with t ∈ [ti, ti+1)) that performance is worse than optimal by at most a factor of γ ≥ 1. Formally, this problem can be stated as argmin M S( M ) s.t. ∀ [ti, ti+1) ∈ E : PC( [ti, ti+1) ) ≤ γ · |Lv : [ti, ti+1)| . An optimal solution to the problem can be computed by means of induction using the recurrence C( [t1, tk+1) ) = min {C( [t1, tj) ) + |Lv : [tj, tk+1)| | 1 ≤ j ≤ k ∧ condition} , where C( [t1, tj) ) is the optimal cost (i.e., the space required) for the prefix subproblem { [ti, ti+1) ∈ E | [ti, ti+1) ⊆ [t1, tj) } and condition stands for ∀ [ti, ti+1) ∈ E : [ti, ti+1) ⊆ [tj, tk+1) ⇒ |Lv : [tj, tk+1)| ≤ γ · |Lv : [ti, ti+1)| . Intuitively, the recurrence states that an optimal solution for [t1, tk+1) be combined from an optimal solution to a prefix subproblem C( [t1, tj) ) and a time interval [tj, tk+1) that can be materialized without violating the performance guarantee. Pseudocode of the algorithm is omitted for space reasons, but can be found in the accompanying technical report [5]. The time complexity of the algorithm is in O(n2 ) - for each prefix subproblem the above recurrence must be evaluated, which is possible in linear time if list sizes |L : [ti, tj)| are precomputed. The space complexity is in O(n2 ) - the cost of keeping the precomputed sublist lengths and memoizing optimal solutions to prefix subproblems. 6.3 Space-Bound Approach So far we considered the problem of materializing sublists that give a guarantee on performance while requiring minimal space. In many situations, though, the storage space is at a premium and the aim would be to materialize a set of sublists that optimizes expected performance while not exceeding a given space limit. The technique presented next, which is named SB, tackles this very problem. The space restriction is modeled by means of a user-specified parameter κ ≥ 1 that limits the maximum allowed blowup in index size from the space-optimal solution provided by Sopt. The SB technique seeks to find a set M that adheres to this space limit but minimizes the expected processing cost (and thus optimizes the expected performance). In the definition of the expected processing cost, P( [ti, ti+1) ) denotes the probability of a query time-point being in [ti, ti+1). Formally, this space-bound sublist-materialization problem can be stated as argmin M X [ti, ti+1) ∈ E P( [ti, ti+1) ) · PC( [ti, ti+1) ) s.t. X m∈M |Lv : m| ≤ κ |Lv| . The problem can be solved by using dynamic programming over an increasing number of time intervals: At each time interval in E the algorithms decides whether to start a new materialization time-interval, using the known best materialization decision from the previous time intervals, and keeping track of the required space consumption for materialization. A detailed description of the algorithm is omitted here, but can be found in the accompanying technical report [5]. Unfortunately, the algorithm has time complexity in O(n3 |Lv|) and its space complexity is in O(n2 |Lv|), which is not practical for large data sets. We obtain an approximate solution to the problem using simulated annealing [22, 23]. Simulated annealing takes a fixed number R of rounds to explore the solution space. In each round a random successor of the current solution is looked at. If the successor does not adhere to the space limit, it is always rejected (i.e., the current solution is kept). A successor adhering to the space limit is always accepted if it achieves lower expected processing cost than the current solution. If it achieves higher expected processing cost, it is randomly accepted with probability e−∆/r where ∆ is the increase in expected processing cost and R ≥ r ≥ 1 denotes the number of remaining rounds. In addition, throughout all rounds, the method keeps track of the best solution seen so far. The solution space for the problem at hand can be efficiently explored. As we argued above, we solely have to look at sets M that completely cover the time interval [t1, tn) and do not contain overlapping time intervals. We represent such a set M as an array of n boolean variables b1 . . . bn that convey the boundaries of time intervals in the set. Note that b1 and bn are always set to true. Initially, all n − 2 intermediate variables assume false, which corresponds to the set M = { [t1, tn) }. A random successor can now be easily generated by switching the value of one of the n − 2 intermediate variables. The time complexity of the method is in O(n2 ) - the expected processing cost must be computed in each round. Its space complexity is in O(n) - for keeping the n boolean variables. As a side remark note that for κ = 1.0 the SB method does not necessarily produce the solution that is obtained from Sopt, but may produce a solution that requires the same amount of space while achieving better expected performance. 7. EXPERIMENTAL EVALUATION We conducted a comprehensive series of experiments on two real-world datasets to evaluate the techniques proposed in this paper. 7.1 Setup and Datasets The techniques described in this paper were implemented in a prototype system using Java JDK 1.5. All experiments described below were run on a single SUN V40z machine having four AMD Opteron CPUs, 16GB RAM, a large network-attached RAID-5 disk array, and running Microsoft Windows Server 2003. All data and indexes are kept in an Oracle 10g database that runs on the same machine. For our experiments we used two different datasets. The English Wikipedia revision history (referred to as WIKI in the remainder) is available for free download as a single XML file. This large dataset, totaling 0.7 TBytes, contains the full editing history of the English Wikipedia from January 2001 to December 2005 (the time of our download). We indexed all encyclopedia articles excluding versions that were marked as the result of a minor edit (e.g., the correction of spelling errors etc.). This yielded a total of 892,255 documents with 13,976,915 versions having a mean (µ) of 15.67 versions per document at standard deviation (σ) of 59.18. We built a time-travel query workload using the query log temporarily made available recently by AOL Research as follows - we first extracted the 300 most frequent keyword queries that yielded a result click on a Wikipedia article (for e.g., french revolution, hurricane season 2005, da vinci code etc.). The thus extracted queries contained a total of 422 distinct terms. For each extracted query, we randomly picked a time point for each month covered by the dataset. This resulted in a total of 18, 000 (= 300 × 60) time-travel queries. The second dataset used in our experiments was based on a subset of the European Archive [13], containing weekly crawls of 11 .gov.uk websites throughout the years 2004 and 2005 amounting close to 2 TBytes of raw data. We filtered out documents not belonging to MIME-types text/plain and text/html, to obtain a dataset that totals 0.4 TBytes and which we refer to as UKGOV in rest of the paper. This included a total of 502,617 documents with 8,687,108 versions (µ = 17.28 and σ = 13.79). We built a corresponding query workload as mentioned before, this time choosing keyword queries that led to a site in the .gov.uk domain (e.g., minimum wage, inheritance tax , citizenship ceremony dates etc.), and randomly sampling a time point for every month within the two year period spanned by the dataset. Thus, we obtained a total of 7,200 (= 300 × 24) time-travel queries for the UKGOV dataset. In total 522 terms appear in the extracted queries. The collection statistics (i.e., N and avdl) and term statistics (i.e., DF) were computed at monthly granularity for both datasets. 7.2 Impact of Temporal Coalescing Our first set of experiments is aimed at evaluating the approximate temporal coalescing technique, described in Section 5, in terms of index-size reduction and its effect on the result quality. For both the WIKI and UKGOV datasets, we compare temporally coalesced indexes for different values of the error threshold computed using Algorithm 1 with the non-coalesced index as a baseline. WIKI UKGOV # Postings Ratio # Postings Ratio - 8,647,996,223 100.00% 7,888,560,482 100.00% 0.00 7,769,776,831 89.84% 2,926,731,708 37.10% 0.01 1,616,014,825 18.69% 744,438,831 9.44% 0.05 556,204,068 6.43% 259,947,199 3.30% 0.10 379,962,802 4.39% 187,387,342 2.38% 0.25 252,581,230 2.92% 158,107,198 2.00% 0.50 203,269,464 2.35% 155,434,617 1.97% Table 1: Index sizes for non-coalesced index (-) and coalesced indexes for different values of Table 1 summarizes the index sizes measured as the total number of postings. As these results demonstrate, approximate temporal coalescing is highly effective in reducing index size. Even a small threshold value, e.g. = 0.01, has a considerable effect by reducing the index size almost by an order of magnitude. Note that on the UKGOV dataset, even accurate coalescing ( = 0) manages to reduce the index size to less than 38% of the original size. Index size continues to reduce on both datasets, as we increase the value of . How does the reduction in index size affect the query results? In order to evaluate this aspect, we compared the top-k results computed using a coalesced index against the ground-truth result obtained from the original index, for different cutoff levels k. Let Gk and Ck be the top-k documents from the ground-truth result and from the coalesced index respectively. We used the following two measures for comparison: (i) Relative Recall at cutoff level k (RR@k), that measures the overlap between Gk and Ck, which ranges in [0, 1] and is defined as RR@k = |Gk ∩ Ck|/k . (ii) Kendall"s τ (see [7, 14] for a detailed definition) at cutoff level k (KT@k), measuring the agreement between two results in the relative order of items in Gk ∩ Ck, with value 1 (or -1) indicating total agreement (or disagreement). Figure 3 plots, for cutoff levels 10 and 100, the mean of RR@k and KT@k along with 5% and 95% percentiles, for different values of the threshold starting from 0.01. Note that for = 0, results coincide with those obtained by the original index, and hence are omitted from the graph. It is reassuring to see from these results that approximate temporal coalescing induces minimal disruption to the query results, since RR@k and KT@k are within reasonable limits. For = 0.01, the smallest value of in our experiments, RR@100 for WIKI is 0.98 indicating that the results are -1 -0.5 0 0.5 1 ε 0.01 0.05 0.10 0.25 0.50 Relative Recall @ 10 (WIKI) Kendall's τ @ 10 (WIKI) Relative Recall @ 10 (UKGOV) Kendall's τ @ 10 (UKGOV) (a) @10 -1 -0.5 0 0.5 1 ε 0.01 0.05 0.10 0.25 0.50 Relative Recall @ 100 (WIKI) Kendall's τ @ 100 (WIKI) Relative Recall @ 100 (UKGOV) Kendall's τ @ 100 (UKGOV) (b) @100 Figure 3: Relative recall and Kendall"s τ observed on coalesced indexes for different values of almost indistinguishable from those obtained through the original index. Even the relative order of these common results is quite high, as the mean KT@100 is close to 0.95. For the extreme value of = 0.5, which results in an index size of just 2.35% of the original, the RR@100 and KT@100 are about 0.8 and 0.6 respectively. On the relatively less dynamic UKGOV dataset (as can be seen from the σ values above), results were even better, with high values of RR and KT seen throughout the spectrum of values for both cutoff values. 7.3 Sublist Materialization We now turn our attention towards evaluating the sublist materialization techniques introduced in Section 6. For both datasets, we started with the coalesced index produced by a moderate threshold setting of = 0.10. In order to reduce the computational effort, boundaries of elementary time intervals were rounded to day granularity before computing the sublist materializations. However, note that the postings in the materialized sublists still retain their original timestamps. For a comparative evaluation of the four approaches - Popt, Sopt, PG, and SB - we measure space and performance as follows. The required space S(M), as defined earlier, is equal to the total number of postings in the materialized sublists. To assess performance we compute the expected processing cost (EPC) for all terms in the respective query workload assuming a uniform probability distribution among query time-points. We report the mean EPC, as well as the 5%- and 95%-percentile. In other words, the mean EPC reflects the expected length of the index list (in terms of index postings) that needs to be scanned for a random time point and a random term from the query workload. The Sopt and Popt approaches are, by their definition, parameter-free. For the PG approach, we varied its parameter γ, which limits the maximal performance degradation, between 1.0 and 3.0. Analogously, for the SB approach the parameter κ, as an upper-bound on the allowed space blowup, was varied between 1.0 and 3.0. Solutions for the SB approach were obtained running simulated annealing for R = 50, 000 rounds. Table 2 lists the obtained space and performance figures. Note that EPC values are smaller on WIKI than on UKGOV, since terms in the query workload employed for WIKI are relatively rarer in the corpus. Based on the depicted results, we make the following key observations. i) As expected, Popt achieves optimal performance at the cost of an enormous space consumption. Sopt, to the contrary, while consuming an optimal amount of space, provides only poor expected processing cost. The PG and SB methods, for different values of their respective parameter, produce solutions whose space and performance lie in between the extremes that Popt and Sopt represent. ii) For the PG method we see that for an acceptable performance degradation of only 10% (i.e., γ = 1.10) the required space drops by more than one order of magnitude in comparison to Popt on both datasets. iii) The SB approach achieves close-to-optimal performance on both datasets, if allowed to consume at most three times the optimal amount of space (i.e., κ = 3.0), which on our datasets still corresponds to a space reduction over Popt by more than one order of magnitude. We also measured wall-clock times on a sample of the queries with results indicating improvements in execution time by up to a factor of 12. 8. CONCLUSIONS In this work we have developed an efficient solution for time-travel text search over temporally versioned document collections. Experiments on two real-world datasets showed that a combination of the proposed techniques can reduce index size by up to an order of magnitude while achieving nearly optimal performance and highly accurate results. The present work opens up many interesting questions for future research, e.g.: How can we even further improve performance by applying (and possibly extending) encoding, compression, and skipping techniques [35]?. How can we extend the approach for queries q [tb, te] specifying a time interval instead of a time point? How can the described time-travel text search functionality enable or speed up text mining along the time axis (e.g., tracking sentiment changes in customer opinions)? 9. ACKNOWLEDGMENTS We are grateful to the anonymous reviewers for their valuable comments - in particular to the reviewer who pointed out the opportunity for algorithmic improvements in Section 5 and Section 6.2. 10. REFERENCES [1] V. N. Anh and A. Moffat. Pruned Query Evaluation Using Pre-Computed Impacts. In SIGIR, 2006. [2] V. N. Anh and A. Moffat. Pruning Strategies for Mixed-Mode Querying. In CIKM, 2006. WIKI UKGOV S(M) EPC S(M) EPC 5% Mean 95% 5% Mean 95% Popt 54,821,634,137 11.22 3,132.29 15,658.42 21,372,607,052 39.93 15,593.60 66,938.86 Sopt 379,962,802 114.05 30,186.52 149,820.1 187,387,342 63.15 22,852.67 102,923.85 PG γ = 1.10 3,814,444,654 11.30 3,306.71 16,512.88 1,155,833,516 40.66 16,105.61 71,134.99 PG γ = 1.25 1,827,163,576 12.37 3,629.05 18,120.86 649,884,260 43.62 17,059.47 75,749.00 PG γ = 1.50 1,121,661,751 13.96 4,128.03 20,558.60 436,578,665 46.68 18,379.69 78,115.89 PG γ = 1.75 878,959,582 15.48 4,560.99 22,476.77 345,422,898 51.26 19,150.06 82,028.48 PG γ = 2.00 744,381,287 16.79 4,992.53 24,637.62 306,944,062 51.48 19,499.78 87,136.31 PG γ = 2.50 614,258,576 18.28 5,801.66 28,849.02 269,178,107 53.36 20,279.62 87,897.95 PG γ = 3.00 552,796,130 21.04 6,485.44 32,361.93 247,666,812 55.95 20,800.35 89,591.94 SB κ = 1.10 412,383,387 38.97 12,723.68 60,350.60 194,287,671 63.09 22,574.54 102,208.58 SB κ = 1.25 467,537,173 26.87 9,011.81 45,119.08 204,454,800 57.42 22,036.39 95,337.33 SB κ = 1.50 557,341,140 19.84 6,699.36 32,810.85 246,323,383 53.24 20,566.68 91,691.38 SB κ = 1.75 647,187,522 16.59 5,769.40 28,272.89 296,345,976 49.56 19,065.99 84,377.44 SB κ = 2.00 737,819,354 15.86 5,358.99 27,112.01 336,445,773 47.58 18,569.08 81,386.02 SB κ = 2.50 916,308,766 13.99 4,639.77 23,037.59 427,122,038 44.89 17,153.94 74,449.28 SB κ = 3.00 1,094,973,140 13.01 4,343.72 22,708.37 511,470,192 42.15 16,772.65 72,307.43 Table 2: Required space and expected processing cost (in # postings) observed on coalesced indexes ( = 0.10) [3] P. G. Anick and R. A. Flynn. Versioning a Full-Text Information Retrieval System. In SIGIR, 1992. [4] R. A. Baeza-Yates and B. Ribeiro-Neto. Modern Information Retrieval. Addison-Wesley, 1999. [5] K. Berberich, S. Bedathur, T. Neumann, and G. Weikum. A Time Machine for Text search. Technical Report MPI-I-2007-5-002, Max-Planck Institute for Informatics, 2007. [6] M. H. B¨ohlen, R. T. Snodgrass, and M. D. Soo. Coalescing in Temporal Databases. In VLDB, 1996. [7] P. Boldi, M. Santini, and S. Vigna. Do Your Worst to Make the Best: Paradoxical Effects in PageRank Incremental Computations. In WAW, 2004. [8] A. Z. Broder, N. Eiron, M. Fontoura, M. Herscovici, R. Lempel, J. McPherson, R. Qi, and E. J. Shekita. Indexing Shared Content in Information Retrieval Systems. In EDBT, 2006. [9] C. Buckley and A. F. Lewit. Optimization of Inverted Vector Searches. In SIGIR, 1985. [10] M. Burrows and A. L. Hisgen. Method and Apparatus for Generating and Searching Range-Based Index of Word Locations. U.S. Patent 5,915,251, 1999. [11] S. B¨uttcher and C. L. A. Clarke. A Document-Centric Approach to Static Index Pruning in Text Retrieval Systems. In CIKM, 2006. [12] D. Carmel, D. Cohen, R. Fagin, E. Farchi, M. Herscovici, Y. S. Maarek, and A. Soffer. Static Index Pruning for Information Retrieval Systems. In SIGIR, 2001. [13] http://www.europarchive.org. [14] R. Fagin, R. Kumar, and D. Sivakumar. Comparing Top k Lists. SIAM J. Discrete Math., 17(1):134-160, 2003. [15] R. Fagin, A. Lotem, and M. Naor. Optimal Aggregation Algorithms for Middleware. J. Comput. Syst. Sci., 66(4):614-656, 2003. [16] S. Guha, K. Shim, and J. Woo. REHIST: Relative Error Histogram Construction Algorithms. In VLDB, 2004. [17] M. Hersovici, R. Lempel, and S. Yogev. Efficient Indexing of Versioned Document Sequences. In ECIR, 2007. [18] http://www.archive.org. [19] Y. E. Ioannidis and V. Poosala. Balancing Histogram Optimality and Practicality for Query Result Size Estimation. In SIGMOD, 1995. [20] H. V. Jagadish, N. Koudas, S. Muthukrishnan, V. Poosala, K. C. Sevcik, and T. Suel. Optimal Histograms with Quality Guarantees. In VLDB, 1998. [21] E. J. Keogh, S. Chu, D. Hart, and M. J. Pazzani. An Online Algorithm for Segmenting Time Series. In ICDM, 2001. [22] S. Kirkpatrick, D. G. Jr., and M. P. Vecchi. Optimization by Simulated Annealing. Science, 220(4598):671-680, 1983. [23] J. Kleinberg and E. Tardos. Algorithm Design. Addison-Wesley, 2005. [24] U. Manber. Introduction to Algorithms: A Creative Approach. Addison-Wesley, 1989. [25] K. Nørv˚ag and A. O. N. Nybø. DyST: Dynamic and Scalable Temporal Text Indexing. In TIME, 2006. [26] J. M. Ponte and W. B. Croft. A Language Modeling Approach to Information Retrieval. In SIGIR, 1998. [27] S. E. Robertson and S. Walker. Okapi/Keenbow at TREC-8. In TREC, 1999. [28] B. Salzberg and V. J. Tsotras. Comparison of Access Methods for Time-Evolving Data. ACM Comput. Surv., 31(2):158-221, 1999. [29] M. Stack. Full Text Search of Web Archive Collections. In IWAW, 2006. [30] E. Terzi and P. Tsaparas. Efficient Algorithms for Sequence Segmentation. In SIAM-DM, 2006. [31] M. Theobald, G. Weikum, and R. Schenkel. Top-k Query Evaluation with Probabilistic Guarantees. In VLDB, 2004. [32] http://www.wikipedia.org. [33] I. H. Witten, A. Moffat, and T. C. Bell. Managing Gigabytes: Compressing and Indexing Documents and Images. Morgan Kaufmann publishers Inc., 1999. [34] J. Zhang and T. Suel. Efficient Search in Large Textual Collections with Redundancy. In WWW, 2007. [35] J. Zobel and A. Moffat. Inverted Files for Text Search Engines. ACM Comput. Surv., 38(2):6, 2006.

text search;inverted file index;sublist materialization;temporal text index;time-travel text search;timestamped information feed;open source search-engine nutch;collaborative authoring environment;validity time-interval;web archive;approximate temporal coalescing;temporal search;versioned document collection;indexing range-based value;document-content overlap;time machine;static indexpruning technique

train_H-45

Query Performance Prediction in Web Search Environments

Current prediction techniques, which are generally designed for content-based queries and are typically evaluated on relatively homogenous test collections of small sizes, face serious challenges in web search environments where collections are significantly more heterogeneous and different types of retrieval tasks exist. In this paper, we present three techniques to address these challenges. We focus on performance prediction for two types of queries in web search environments: content-based and Named-Page finding. Our evaluation is mainly performed on the GOV2 collection. In addition to evaluating our models for the two types of queries separately, we consider a more challenging and realistic situation that the two types of queries are mixed together without prior information on query types. To assist prediction under the mixed-query situation, a novel query classifier is adopted. Results show that our prediction of web query performance is substantially more accurate than the current stateof-the-art prediction techniques. Consequently, our paper provides a practical approach to performance prediction in realworld web settings.

1. INTRODUCTION Query performance prediction has many applications in a variety of information retrieval (IR) areas such as improving retrieval consistency, query refinement, and distributed IR. The importance of this problem has been recognized by IR researchers and a number of new methods have been proposed for prediction recently [1, 2, 17]. Most work on prediction has focused on the traditional ad-hoc retrieval task where query performance is measured according to topical relevance. These prediction models are evaluated on TREC document collections which typically consist of no more than one million relatively homogenous newswire articles. With the popularity and influence of the Web, prediction techniques that will work well for web-style queries are highly preferable. However, web search environments pose significant challenges to current prediction models that are mainly designed for traditional TREC settings. Here we outline some of these challenges. First, web collections, which are much larger than conventional TREC collections, include a variety of documents that are different in many aspects such as quality and style. Current prediction techniques can be vulnerable to these characteristics of web collections. For example, the reported prediction accuracy of the ranking robustness technique and the clarity technique on the GOV2 collection (a large web collection) is significantly worse compared to the other TREC collections [1]. Similar prediction accuracy on the GOV2 collection using another technique is reported in [2], confirming the difficult of predicting query performance on a large web collection. Furthermore, web search goes beyond the scope of the ad-hoc retrieval task based on topical relevance. For example, the Named-Page (NP) finding task, which is a navigational task, is also popular in web retrieval. Query performance prediction for the NP task is still necessary since NP retrieval performance is far from perfect. In fact, according to the report on the NP task of the 2005 Terabyte Track [3], about 40% of the test queries perform poorly (no correct answer in the first 10 search results) even in the best run from the top group. To our knowledge, little research has explicitly addressed the problem of NP-query performance prediction. Current prediction models devised for content-based queries will be less effective for NP queries considering the fundamental differences between the two. Third, in real-world web search environments, user queries are usually a mixture of different types and prior knowledge about the type of each query is generally unavailable. The mixed-query situation raises new problems for query performance prediction. For instance, we may need to incorporate a query classifier into prediction models. Despite these problems, the ability to handle this situation is a crucial step towards turning query performance prediction from an interesting research topic into a practical tool for web retrieval. In this paper, we present three techniques to address the above challenges that current prediction models face in Web search environments. Our work focuses on query performance prediction for the content-based (ad-hoc) retrieval task and the name-page finding task in the context of web retrieval. Our first technique, called weighted information gain (WIG), makes use of both single term and term proximity features to estimate the quality of top retrieved documents for prediction. We find that WIG offers consistent prediction accuracy across various test collections and query types. Moreover, we demonstrate that good prediction accuracy can be achieved for the mixed-query situation by using WIG with the help of a query type classifier. Query feedback and first rank change, which are our second and third prediction techniques, perform well for content-based queries and NP queries respectively. Our main contributions include: (1) considerably improved prediction accuracy for web content-based queries over several state-of-the-art techniques. (2) new techniques for successfully predicting NP-query performance. (3) a practical and fully automatic solution to predicting mixed-query performance. In addition, one minor contribution is that we find that the robustness score [1], which was originally proposed for performance prediction, is helpful for query classification. 2. RELATED WORK As we mentioned in the introduction, a number of prediction techniques have been proposed recently that focus on contentbased queries in the topical relevance (ad-hoc) task. We know of no published work that addresses other types of queries such as NP queries, let alone a mixture of query types. Next we review some representative models. The major difficulty of performance prediction comes from the fact that many factors have an impact on retrieval performance. Each factor affects performance to a different degree and the overall effect is hard to predict accurately. Therefore, it is not surprising to notice that simple features, such as the frequency of query terms in the collection [4] and the average IDF of query terms [5], do not predict well. In fact, most of the successful techniques are based on measuring some characteristics of the retrieved document set to estimate topic difficulty. For example, the clarity score [6] measures the coherence of a list of documents by the KL-divergence between the query model and the collection model. The robustness score [1] quantifies another property of a ranked list: the robustness of the ranking in the presence of uncertainty. Carmel et al. [2] found that the distance measured by the Jensen-Shannon divergence between the retrieved document set and the collection is significantly correlated to average precision. Vinay et al.[7] proposed four measures to capture the geometry of the top retrieved documents for prediction. The most effective measure is the sensitivity to document perturbation, an idea somewhat similar to the robustness score. Unfortunately, their way of measuring the sensitivity does not perform equally well for short queries and prediction accuracy drops considerably when a state-of-the-art retrieval technique (like Okapi or a language modeling approach) is adopted for retrieval instead of the tf-idf weighting used in their paper [16]. The difficulties of applying these models in web search environments have already been mentioned. In this paper, we mainly adopt the clarity score and the robustness score as our baselines. We experimentally show that the baselines, even after being carefully tuned, are inadequate for the web environment. One of our prediction models, WIG, is related to the Markov random field (MRF) model for information retrieval [8]. The MRF model directly models term dependence and is found be to highly effective across a variety of test collections (particularly web collections) and retrieval tasks. This model is used to estimate the joint probability distribution over documents and queries, an important part of WIG. The superiority of WIG over other prediction techniques based on unigram features, which will be demonstrated later in our paper, coincides with that of MRF for retrieval. In other word, it is interesting to note that term dependence, when being modeled appropriately, can be helpful for both improving and predicting retrieval performance. 3. PREDICTION MODELS 3.1 Weighted Information Gain (WIG) This section introduces a weighted information gain approach that incorporates both single term and proximity features for predicting performance for both content-based and Named-Page (NP) finding queries. Given a set of queries Q={Qs} (s=1,2,..N) which includes all possible user queries and a set of documents D={Dt} (t=1,2…M), we assume that each query-document pair (Qs,Dt) is manually judged and will be put in a relevance list if Qs is found to be relevant to Dt. The joint probability P(Qs,Dt) over queries Q and documents D denotes the probability that pair (Qs,Dt) will be in the relevance list. Such assumptions are similar to those used in [8]. Assuming that the user issues query Qi ∈Q and the retrieval results in response to Qi is a ranked list L of documents, we calculate the amount of information contained in P(Qs,Dt) with respect to Qi and L by Eq.1 which is a variant of entropy called the weighted entropy[13]. The weights in Eq.1 are solely determined by Qi and L. )1(),(log),(),( , , ∑−= ts tststsLQ DQPDQweightDQH i In this paper, we choose the weights as follows: LindocumentsKtopthecontainsLTwhere otherwise LTDandisifK DQweight K Kt ts )( )2( ,0 )(,/1 ),( ⎩ ⎨ ⎧ ∈= = The cutoff rank K is a parameter in our model that will be discussed later. Accordingly, Eq.1 can be simplified as follows: )3(),(log 1 ),( )( , ∑∈ −= LTD titsLQ Kt i DQP K DQH Unfortunately, weighted entropy ),(, tsLQ DQH i computed by Eq.3, which represents the amount of information about how likely the top ranked documents in L would be relevant to query Qi on average, cannot be compared across different queries, making it inappropriate for directly predicting query performance. To mitigate this problem, we come up with a background distribution P(Qs,C) over Q and D by imagining that every document in D is replaced by the same special document C which represents average language usage. In this paper, C is created by concatenating every document in D. Roughly speaking, C is the collection (the document set) {Dt} without document boundaries. Similarly, weighted entropy ),(, CQH sLQi calculated by Eq.3 represents the amount of information about how likely an average document (represented by the whole collection) would be relevant to query Qi. Now we introduce our performance predictor WIG which is the weighted information gain [13] computed as the difference between ),(, tsLQ DQH i and ),(, CQH sLQi .Specifically, given query Qi, collection C and ranked list L of documents, WIG is calculated as follows: )4( ),( ),( log 1 ),( ),( log),( ),(),(),,( )(, ,, ∑∑ ∈ == −= LTD i ti ts s ts ts tsLQsLQi Kt ii CQP DQP KCQP DQP DQweight DQHCQHLCQWIG WIG computed by Eq.4 measures the change in information about the quality of retrieval (in response to query Qi) from an imaginary state that only an average document is retrieved to a posterior state that the actual search results are observed. We hypothesize that WIG is positively correlated with retrieval effectiveness because high quality retrieval should be much more effective than just returning the average document. The heart of this technique is how to estimate the joint distribution P(Qs,Dt). In the language modeling approach to IR, a variety of models can be applied readily to estimate this distribution. Although most of these models are based on the bagof-words assumption, recent work on modeling term dependence under the language modeling framework have shown consistent and significant improvements in retrieval effectiveness over bagof-words models. Inspired by the success of incorporating term proximity features into language models, we decide to adopt a good dependence model to estimate the probability P(Qs,Dt). The model we chose for this paper is Metzler and Croft"s Markov Random Field (MRF) model, which has already demonstrated superiority over a number of collections and different retrieval tasks [8,9]. According to the MRF model, log P(Qi, Dt) can be written as )5()|(loglog),(log )( 1 ∑∈ +−= iQF tti DPZDQP ξ ξ ξλ where Z1 is a constant that ensures that P(Qi, Dt) sums up to 1. F(Qi) consists of a set of features expanded from the original query Qi . For example, assuming that query Qi is talented student program, F(Qi) includes features like program and talented student. We consider two kinds of features: single term features T and proximity features P. Proximity features include exact phrase (#1) and unordered window (#uwN) features as described in [8]. Note that F(Qi) is the union of T(Qi) and P(Qi). For more details on F(Qi) such as how to expand the original query Qi to F(Qi), we refer the reader to [8] and [9]. P(ξ|Dt) denotes the probability that feature ξ will occur in Dt. More details on P(ξ|Dt) will be provided later in this section. The choice of λξ is somewhat different from that used in [8] since λξ plays a dual role in our model. The first role, which is the same as in [8], is to weight between single term and proximity features. The other role, which is specific to our prediction task, is to normalize the size of F(Qi).We found that the following weight strategy for λξ satisfies the above two roles and generalizes well on a variety of collections and query types. )6( )(, |)(| 1 )(, |)(| ⎪ ⎪ ⎩ ⎪ ⎪ ⎨ ⎧ ∈ − ∈ = i i T i i T QP QP QT QT ξ λ ξ λ λξ where |T(Qi)| and |P(Qi)| denote the number of single term and proximity features in F(Qi) respectively. The reason for choosing the square root function in the denominator of λξ is to penalize a feature set of large size appropriately, making WIG more comparable across queries of various lengths. λT is a fixed parameter and set to 0.8 according to [8] throughout this paper. Similarly, log P(Qi,C) can be written as: )7()|(loglog),(log )( 2 ∑∈ +−= iQF i CPZCQP ξ ξ ξλ When constant Z1 and Z2 are dropped, WIG computed in Eq.4 can be rewritten as follows by plugging in Eq.5 and Eq.7 : )8( )|( )|( log 1 ),,( )( )( ∑ ∑∈ ∈ = LTD QF t i Kt i CP DP K LCQWIG ξ ξ ξ ξ λ One of the advantages of WIG over other techniques is that it can handle well both content-based and NP queries. Based on the type (or the predicted type) of Qi, the calculation of WIG in Eq. 8 differs in two aspects: (1) how to estimate P(ξ|Dt) and P(ξ|C), and (2) how to choose K. For content-based queries, P(ξ|C) is estimated by the relative frequency of feature ξ in collection C as a whole. The estimation of P(ξ|Dt) is the same as in [8]. Namely, we estimate P(ξ|Dt) by the relative frequency of feature ξ in Dt linearly smoothed with collection frequency P(ξ|C). K in Eq.8 is treated as a free parameter. Note that K is the only free parameter in the computation of WIG for content-based queries because all parameters involved in P(ξ|Dt) are assumed to be fixed by taking the suggested values in [8]. Regarding NP queries, we make use of document structure to estimate P(ξ|Dt) and P(ξ|C) by the so-called mixture of language models proposed in [10] and incorporated into the MRF model for Named-Page finding retrieval in [9]. The basic idea is that a document (collection) is divided into several fields such as the title field, the main-body field and the heading field. P(ξ|Dt) and P(ξ|C) are estimated by a linear combination of the language models from each field. Due to space constraints, we refer the reader to [9] for details. We adopt the exact same set of parameters as used in [9] for estimation. With regard to K in Eq.8, we set K to 1 because the Named-Page finding task heavily focuses on the first ranked document. Consequently, there are no free parameters in the computation of WIG for NP queries. 3.2 Query Feedback In this section, we introduce another technique called query feedback (QF) for prediction. Suppose that a user issues query Q to a retrieval system and a ranked list L of documents is returned. We view the retrieval system as a noisy channel. Specifically, we assume that the output of the channel is L and the input is Q. After going through the channel, Q becomes corrupted and is transformed to ranked list L. By thinking about the retrieval process this way, the problem of predicting retrieval effectiveness turns to the task of evaluating the quality of the channel. In other words, prediction becomes finding a way to measure the degree of corruption that arises when Q is transformed to L. As directly computing the degree of the corruption is difficult, we tackle this problem by approximation. Our main idea is that we measure to what extent information on Q can be recovered from L on the assumption that only L is observed. Specifically, we design a decoder that can accurately translate L back into new query Q" and the similarity S between the original query Q and the new query Q" is adopted as a performance predictor. This is a sketch of how the QF technique predicts query performance. Before filling in more details, we briefly discuss why this method would work. There is a relation between the similarity S defined above and retrieval performance. On the one hand, if the retrieval has strayed from the original sense of the query Q, the new query Q" extracted from ranked list L in response to Q would be very different from the original query Q. On the other hand, a query distilled from a ranked list containing many relevant documents is likely to be similar to the original query. Further examples in support of the relation will be provided later. Next we detail how to build the decoder and how to measure the similarity S. In essence, the goal of the decoder is to compress ranked list L into a few informative terms that should represent the content of the top ranked documents in L. Our approach to this goal is to represent ranked list L by a language model (distribution over terms). Then terms are ranked by their contribution to the language model"s KL (Kullback-Leibler) divergence from the background collection model. Top ranked terms will be chosen to form the new query Q". This approach is similar to that used in Section 4.1 of [11]. Specifically, we take three steps to compress ranked list L into query Q" without referring to the original query. 1. We adopt the ranked list language model [14], to estimate a language model based on ranked list L. The model can be written as: )9()|()|()|( ∑∈ = LD LDPDwPLwP where w is any term, D is a document. P(D|L) is estimated by a linearly decreasing function of the rank of document D. 2. Each term in P(w|L) is ranked by the following KL-divergence contribution: )10( )|( )|( log)|( CwP LwP LwP where P(w|C) is the collection model estimated by the relative frequency of term w in collection C as a whole. 3. The top N ranked terms by Eq.10 form a weighted query Q"={(wi,ti)} i=1,N. where wi denotes the i-th ranked term and weight ti is the KL-divergence contribution of wi in Eq. 10. Term cruise ship vessel sea passenger KL contribution 0.050 0.040 0.012 0.010 0.009 Table 1: top 5 terms compressed from the ranked list in response to query Cruise ship damage sea life Two representative examples, one for a poorly performing query Cruise ship damage sea life (TREC topic 719; average precision: 0.08) and the other for a high performing query prostate cancer treatments( TREC topic 710; average precision: 0.49), are shown in Table 1 and 2 respectively. These examples indicate how the similarity between the original and the new query correlates with retrieval performance. The parameter N in step 3 is set to 20 empirically and choosing a larger value of N is unnecessary since the weights after the top 20 are usually too small to make any difference. Term prostate cancer treatment men therapy KL contribution 0.177 0.140 0.028 0.025 0.020 Table 2: top 5 terms compressed from the ranked list in response to query prostate cancer treatments To measure the similarity between original query Q and new query Q", we first use Q" to do retrieval on the same collection. A variant of the query likelihood model [15] is adopted for retrieval. Namely, documents are ranked by: )11()|()|'( '),( ∑∈ = Qtw t i ii i DwPDQP where wi is a term in Q" and ti is the associated weight. D is a document. Let L" denote the new ranked list returned from the above retrieval. The similarity is measured by the overlap of documents in L and L". Specifically, the percentage of documents in the top K documents of L that are also present in the top K documents in L". the cutoff K is treated as a free parameter. We summarize here how the QF technique predicts performance given a query Q and the associated ranked list L. We first obtain a weighted query Q" compressed from L by the above three steps. Then we use Q" to perform retrieval and the new ranked list is L". The overlap of documents in L and L" is used for prediction. 3.3 First Rank Change (FRC) In this section, we propose a method called the first rank change (FRC) for performance prediction for NP queries. This method is derived from the ranking robustness technique [1] that is mainly designed for content-based queries. When directly applied to NP queries, the robustness technique will be less effective because it takes the top ranked documents as a whole into account while NP queries usually have only one single relevant document. Instead, our technique focuses on the first rank document while the main idea of the robustness method remains. Specifically, the pseudocode for computing FRC is shown in figure 1. Input: (1) ranked list L={Di} where i=1,100. Di denotes the i-th ranked document. (2) query Q 1 initialize: (1) set the number of trials J=100000 (2) counter c=0; 2 for i=1 to J 3 Perturb every document in L, let the outcome be a set F={Di"} where Di" denotes the perturbed version of Di. 4 Do retrieval with query Q on set F 5 c=c+1 if and only if D1" is ranked first in step 4 6 end of for 7 return the ratio c/J Figure 1: pseudo-code for computing FRC FRC approximates the probability that the first ranked document in the original list L will remain ranked first even after the documents are perturbed. The higher the probability is, the more confidence we have in the first ranked document. On the other hand, in the extreme case of a random ranking, the probability would be as low as 0.5. We expect that FRC has a positive association with NP query performance. We adopt [1] to implement the document perturbation step (step 4 in Fig.1) using Poisson distributions. For more details, we refer the reader to [1]. 4. EVALUATION We now present the results of predicting query performance by our models. Three state-of-the-art techniques are adopted as our baselines. We evaluate our techniques across a variety of Web retrieval settings. As mentioned before, we consider two types of queries, that is, content-based (CB) queries and Named-Page(NP) finding queries. First, suppose that the query types are known. We investigate the correlation between the predicted retrieval performance and the actual performance for both types of queries separately. Results show that our methods yield considerable improvements over the baselines. We then consider a more challenging scenario where no prior information on query types is available. Two sub-cases are considered. In the first one, there exists only one type of query but the actual type is unknown. We assume a mixture of the two query types in the second case. We demonstrate that our models achieve good accuracy under this demanding scenario, making prediction practical in a real-world Web search environment. 4.1 Experimental Setup Our evaluation focuses on the GOV2 collection which contains about 25 million documents crawled from web sites in the .gov domain during 2004 [3]. We create two kinds of data set for CB queries and NP queries respectively. For the CB type, we use the ad-hoc topics of the Terabyte Tracks of 2004, 2005 and 2006 and name them TB04-adhoc, TB05-adhoc and TB06-adhoc respectively. In addition, we also use the ad-hoc topics of the 2004 Robust Track (RT04) to test the adaptability of our techniques to a non-Web environment. For NP queries, we use the Named-Page finding topics of the Terabyte Tracks of 2005 and 2006 and we name them TB05-NP and TB06-NP respectively. All queries used in our experiments are titles of TREC topics as we center on web retrieval. Table 3 summarizes the above data sets. Name Collection Topic Number Query Type TB04-adhoc GOV2 701-750 CB TB05-adhoc GOV2 751-800 CB TB06-adhoc GOV2 801-850 CB RT04 Disk 4+5 (minus CR) 301-450;601700 CB TB05-NP GOV2 NP601-NP872 NP TB06-NP GOV2 NP901-NP1081 NP Table 3: Summary of test collections and topics Retrieval performance of individual content-based and NP queries is measured by the average precision and reciprocal rank of the first correct answer respectively. We make use of the Markov Random field model for both ad-hoc and Named-Page finding retrieval. We adopt the same setting of retrieval parameters used in [8,9]. The Indri search engine [12] is used for all of our experiments. Though not reported here, we also tried the query likelihood model for ad-hoc retrieval and found that the results change little because of the very high correlation between the query performances obtained by the two retrieval models (0.96 measured by Pearson"s coefficient). 4.2 Known Query Types Suppose that query types are known. We treat each type of query separately and measure the correlation with average precision (or the reciprocal rank in the case of NP queries). We adopt the Pearson"s correlation test which reflects the degree of linear relationship between the predicted and the actual retrieval performance. 4.2.1 Content-based Queries Methods Clarity Robust JSD WIG QF WIG +QF TB04+0 5 adhoc 0.333 0.317 0.362 0.574 0.480 0.637 TB06 adhoc 0.076 0.294 N/A 0.464 0.422 0.511 Table 4: Pearson"s correlation coefficients for correlation with average precision on the Terabyte Tracks (ad-hoc) for clarity score, robustness score, the JSD-based method(we directly cites the score reported in [2]), WIG, query feedback(QF) and a linear combination of WIG and QF. Bold cases mean the results are statistically significant at the 0.01 level. Table 4 shows the correlation with average precision on two data sets: one is a combination of TB04-adhoc and TB05-adhoc(100 topics in total) and the other is TB06-adhoc (50 topics). The reason that we put TB04-adhoc and TB05-adhoc together is to make our results comparable to [2]. Our baselines are the clarity score (clarity) [6],the robustness score (robust)[1] and the JSDbased method (JSD) [2]. For the clarity and robustness score, we have tried different parameter settings and report the highest correlation coefficients we have found. We directly cite the result of the JSD-based method reported in [2]. The table also shows the results for the Weighted Information Gain (WIG) method and the Query Feedback (QF) method for predicting content-based queries. As we described in the previous section, both WIG and QF have one free parameter to set, that is, the cutoff rank K. We train the parameter on one dataset and test on the other. When combining WIG and QF, a simple linear combination is used and the combination weight is learned from the training data set. From these results, we can see that our methods are considerably more accurate compared to the baselines. We also observe that further improvements are obtained from the combination of WIG and QF, suggesting that they measure different properties of the retrieval process that relate to performance. We discover that our methods generalize well on TB06-adhoc while the correlation for the clarity score with retrieval performance on this data set is considerably worse. Further investigation shows that the mean average precision of TB06-adhoc is 0.342 and is about 10% better than that of the first data set. While the other three methods typically consider the top 100 or less documents given a ranked list, the clarity method usually needs the top 500 or more documents to adequately measure the coherence of a ranked list. Higher mean average precision makes ranked lists retrieved by different queries more similar in terms of coherence at the level of top 500 documents. We believe that this is the main reason for the low accuracy of the clarity score on the second data set. Though this paper focuses on a Web search environment, it is desirable that our techniques will work consistently well in other situations. To this end, we examine the effectiveness of our techniques on the Robust 2004 Track. For our methods, we evenly divide all of the test queries into five groups and perform five-fold cross validation. Each time we use one group for training and the remaining four groups for testing. We make use of all of the queries for our two baselines, that is, the clarity score and the robustness score. The parameters for our baselines are the same as those used in [1].The results shown in Table 5 demonstrate that the prediction accuracy of our methods is on a par with that of the two strong baselines. Clarity Robust WIG QF 0.464 0.539 0.468 0.464 Table 5: Comparison of Pearson"s correlation coefficients on the 2004 Robust Track for clarity score, robustness score, WIG and query feedback (QF). Bold cases mean the results are statistically significant at the 0.01 level. Furthermore, we examine the prediction sensitivity of our methods to the cutoff rank K. With respect to WIG, it is quite robust to K on the Terabyte Tracks (2004-2006) while it prefers a small value of K like 5 on the 2004 Robust Track. In other words, a small value of K is a nearly-optimal choice for both kinds of tracks. Considering the fact that all other parameters involved in WIG are fixed and consequently the same for the two cases, this means WIG can achieve nearly-optimal prediction accuracy in two considerably different situations with exactly the same parameter setting. Regarding QF, it prefers a larger value of K such as 100 on the Terabyte Tracks and a smaller value of K such as 25 on the 2004 Robust Track. 4.2.2 NP Queries We adopt WIG and first rank change (FRC) for predicting NPquery performance. We also try a linear combination of the two as in the previous section. The combination weight is obtained from the other data set. We use the correlation with the reciprocal ranks measured by the Pearson"s correlation test to evaluate prediction quality. The results are presented in Table 6. Again, our baselines are the clarity score and the robustness score. To make a fair comparison, we tune the clarity score in different ways. We found that using the first ranked document to build the query model yields the best prediction accuracy. We also attempted to utilize document structure by using the mixture of language models mentioned in section 3.1. Little improvement was obtained. The correlation coefficients for the clarity score reported in Table 6 are the best we have found. As we can see, our methods considerably outperform the clarity score technique on both of the runs. This confirms our intuition that the use of a coherence-based measure like the clarity score is inappropriate for NP queries. Methods Clarity Robust. WIG FRC WIG+FRC TB05-NP 0.150 -0.370 0.458 0.440 0.525 TB06-NP 0.112 -0.160 0.478 0.386 0.515 Table 6: Pearson"s correlation coefficients for correlation with reciprocal ranks on the Terabyte Tracks (NP) for clarity score, robustness score, WIG, the first rank change (FRC) and a linear combination of WIG and FRC. Bold cases mean the results are statistically significant at the 0.01 level. Regarding the robustness score, we also tune the parameters and report the best we have found. We observe an interesting and surprising negative correlation with reciprocal ranks. We explain this finding briefly. A high robustness score means that a number of top ranked documents in the original ranked list are still highly ranked after perturbing the documents. The existence of such documents is a good sign of high performance for content-based queries as these queries usually contain a number of relevant documents [1]. However, with regard to NP queries, one fundamental difference is that there is only one relevant document for each query. The existence of such documents can confuse the ranking function and lead to low retrieval performance. Although the negative correlation with retrieval performance exists, the strength of the correlation is weaker and less consistent compared to our methods as shown in Table 6. Based on the above analysis, we can see that current prediction techniques like clarity score and robustness score that are mainly designed for content-based queries face significant challenges and are inadequate to deal with NP queries. Our two techniques proposed for NP queries consistently demonstrate good prediction accuracy, displaying initial success in solving the problem of predicting performance for NP queries. Another point we want to stress is that the WIG method works well for both types of queries, a desirable property that most prediction techniques lack. 4.3 Unknown Query Types In this section, we run two kinds of experiments without access to query type labels. First, we assume that only one type of query exists but the type is unknown. Second, we experiment on a mixture of content-based and NP queries. The following two subsections will report results for the two conditions respectively. 4.3.1 Only One Type exists We assume that all queries are of the same type, that is, they are either NP queries or content-based queries. We choose WIG to deal with this case because it shows good prediction accuracy for both types of queries in the previous section. We consider two cases: (1) CB: all 150 title queries from the ad-hoc task of the Terabyte Tracks 2004-2006 (2)NP: all 433 NP queries from the named page finding task of the Terabyte Tracks 2005 and 2006. We take a simple strategy by labeling all of the queries in each case as the same type (either NP or CB) regardless of their actual type. The computation of WIG will be based on the labeled query type instead of the actual type. There are four possibilities with respect to the relation between the actual type and the labeled type. The correlation with retrieval performance under the four possibilities is presented in Table 7. For example, the value 0.445 at the intersection between the second row and the third column shows the Pearson"s correlation coefficient for correlation with average precision when the content-based queries are incorrectly labeled as the NP type. Based on these results, we recommend treating all queries as the NP type when only one query type exists and accurate query classification is not feasible, considering the risk that a large loss of accuracy will occur if NP queries are incorrectly labeled as content-based queries. These results also demonstrate the strong adaptability of WIG to different query types. CB (labeled) NP (labeled) CB (actual) 0.536 0.445 NP (actual) 0.174 0.467 Table 7: Comparison of Pearson"s correlation coefficients for correlation with retrieval performance under four possibilities on the Terabyte Tracks (NP). Bold cases mean the results are statistically significant at the 0.01 level. 4.3.2 A mixture of contented-based and NP queries A mixture of the two types of queries is a more realistic situation that a Web search engine will meet. We evaluate prediction accuracy by how accurately poorly-performing queries can be identified by the prediction method assuming that actual query types are unknown (but we can predict query types). This is a challenging task because both the predicted and actual performance for one type of query can be incomparable to that for the other type. Next we discuss how to implement our evaluation. We create a query pool which consists of all of the 150 ad-hoc title queries from Terabyte Track 2004-2006 and all of the 433 NP queries from Terabyte Track 2005&2006. We divide the queries in the pool into classes: good (better than 50% of the queries of the same type in terms of retrieval performance) and bad (otherwise). According to these standards, a NP query with the reciprocal rank above 0.2 or a content-based query with the average precision above 0.315 will be considered as good. Then, each time we randomly select one query Q from the pool with probability p that Q is contented-based. The remaining queries are used as training data. We first decide the type of query Q according to a query classifier. Namely, the query classifier tells us whether query Q is NP or content-based. Based on the predicted query type and the score computed for query Q by a prediction technique, a binary decision is made about whether query Q is good or bad by comparing to the score threshold of the predicted query type obtained from the training data. Prediction accuracy is measured by the accuracy of the binary decision. In our implementation, we repeatedly take a test query from the query pool and prediction accuracy is computed as the percentage of correct decisions, that is, a good(bad) query is predicted to be good (bad). It is obvious that random guessing will lead to 50% accuracy. Let us take the WIG method for example to illustrate the process. Two WIG thresholds (one for NP queries and the other for content-based queries) are trained by maximizing the prediction accuracy on the training data. When a test query is labeled as the NP (CB) type by the query type classifier, it will be predicted to be good if and only if the WIG score for this query is above the NP (CB) threshold. Similar procedures will be taken for other prediction techniques. Now we briefly introduce the automatic query type classifier used in this paper. We find that the robustness score, though originally proposed for performance prediction, is a good indicator of query types. We find that on average content-based queries have a much higher robustness score than NP queries. For example, Figure 2 shows the distributions of robustness scores for NP and content-based queries. According to this finding, the robustness score classifier will attach a NP (CB) label to the query if the robustness score for the query is below (above) a threshold trained from the training data. 0 0.5 1 1.5 2 2.5 -1 -0.8 -0.6 -0.4 -0.2 0 0.2 0.4 0.6 0.8 1 NP Content-based Figure 2: Distribution of robustness scores for NP and CB queries. The NP queries are the 252 NP topics from the 2005 Terabyte Track. The content-based queries are the 150 ad-hoc title from the Terabyte Tracks 2004-2006. The probability distributions are estimated by the Kernel density estimation method. Strategies Robust WIG-1 WIG-2 WIG-3 Optimal p=0.6 0.565 0.624 0.665 0.684 0.701 P=0.4 0.567 0.633 0.654 0.673 0.696 Table 8: Comparison of prediction accuracy for five strategies in the mixed-query situation. Two ways to sample a query from the pool: (1) the sampled query is content-based with the probability p=0.6. (that is, the query is NP with probability 0.4 ) (2) set the probability p=0.4. We consider five strategies in our experiments. In the first strategy (denoted by robust), we use the robustness score for query performance prediction with the help of a perfect query classifier that always correctly map a query into one of the two categories (that is, NP or CB). This strategy represents the level of prediction accuracy that current prediction techniques can achieve in an ideal condition that query types are known. In the next following three strategies, the WIG method is adopted for performance prediction. The difference among the three is that three different query classifiers are used for each strategy: (1) the classifier always classifies a query into the NP type. (2) the Robustness Score ProbabilityDensity classifier is the robust score classifier mentioned above. (3) the classifier is a perfect one. These three strategies are denoted by WIG-1, WIG-2 and WIG-3 respectively. The reason we are interested in WIG-1 is based on the results from section 4.3.1. In the last strategy (denoted by Optimal) which serves as an upper bound on how well we can do so far, we fully make use of our prediction techniques for each query type assuming a perfect query classifier is available. Specifically, we linearly combine WIG and QF for content-based queries and WIG and FRC for NP queries. The results for the five strategies are shown in Table 8. For each strategy, we try two ways to sample a query from the pool: (1) the sampled query is CB with probability p=0.6. (the query is NP with probability 0.4) (2) set the probability p=0.4. From Table 8 We can see that in terms of prediction accuracy WIG-2 (the WIG method with the automatic query classifier) is not only better than the first two cases, but also is close to WIG-3 where a perfect classifier is assumed. Some further improvements over WIG-3 are observed when combined with other prediction techniques. The merit of WIG-2 is that it provides a practical solution to automatically identifying poorly performing queries in a Web search environment with mixed query types, which poses considerable obstacles to traditional prediction techniques. 5. CONCLUSIONS AND FUTURE WORK To our knowledge, our paper is the first to thoroughly explore prediction of query performance in web search environments. We demonstrated that our models resulted in higher prediction accuracy than previously published techniques not specially devised for web search scenarios. In this paper, we focus on two types of queries in web search: content-based and Named-Page (NP) finding queries, corresponding to the ad-hoc retrieval task and the Named-Page finding task respectively. For both types of web queries, our prediction models were shown to be substantially more accurate than the current state-of-the-art techniques. Furthermore, we considered a more realistic case that no prior information on query types is available. We demonstrated that the WIG method is particularly suitable for this situation. Considering the adaptability of WIG to a range of collections and query types, one of our future plans is to apply this method to predict user preference of search results on realistic data collected from a commercial search engine. Other than accuracy, another major issue that prediction techniques have to deal with in a Web environment is efficiency. Fortunately, since the WIG score is computed just over the terms and the phrases that appear in the query, this calculation can be made very efficient with the support of index. On the other hand, the computation of QF and FRC is relatively less efficient since QF needs to retrieve the whole collection twice and FRC needs to repeatedly rank the perturbed documents. How to improve the efficiency of QF and FRC is our future work. In addition, the prediction techniques proposed in this paper have the potential of improving retrieval performance by combining with other IR techniques. For example, our techniques can be incorporated to popular query modification techniques such as query expansion and query relaxation. Guided by performance prediction, we can make a better decision on when to or how to modify queries to enhance retrieval effectiveness. We would like to carry out research in this direction in the future. 6. ACKNOWLEGMENTS This work was supported in part by the Center for Intelligent Information Retrieval, in part by the Defense Advanced Research Projects Agency (DARPA) under contract number HR0011-06-C0023, and in part by an award from Google. Any opinions, findings and conclusions or recommendations expressed in this material are those of the author and do not necessarily reflect those of the sponsor. In addition, we thank Donald Metzler for his valuable comments on this work. 7. REFERENCES [1] Y. Zhou ,W. B. Croft ,Ranking Robustness: A Novel Framework to Predict Query Performance, in Proceedings of CIKM 2006. [2] D.Carmel, E.Yom-Tov, A.Darlow,D.Pelleg, What Makes a Query Difficult?, in Proceedings of SIGIR 2006. [3] C.L.A. Clarke, F. Scholer, I.Soboroff, The TREC 2005 Terabyte Track, In the Online Proceedings of 2005 TREC. [4] B. He and I.Ounis. Inferring query performance using preretrieval predictors. In proceedings of the SPIRE 2004. [5] S. Tomlinson. Robust, Web and Terabyte Retrieval with Hummingbird SearchServer at TREC 2004. In the Online Proceedings of 2004 TREC. [6] S. Cronen-Townsend, Y. Zhou, W. B. Croft, Predicting Query Performance, in Proceedings of SIGIR 2002. [7] V.Vinay, I.J.Cox, N.Mill-Frayling,K.Wood, On Ranking the Effectiveness of Searcher, in Proceedings of SIGIR 2006. [8] D.Metzler, W.B.Croft, A Markov Random Filed Model for Term Dependencies, in Proceedings of SIGIR 2005. [9] D.Metzler, T.Strohman,Y.Zhou,W.B.Croft, Indri at TREC 2005: Terabyte Track, In the Online Proceedings of 2004 TREC. [10] P. Ogilvie and J. Callan, Combining document representations for known-item search, in Proceedings of SIGIR 2003. [11] A.Berger, J.Lafferty, Information retrieval as statistical translation, in Proceedings of SIGIR 1999. [12] Indri search engine : http://www.lemurproject.org/indri/ [13] I.J. Taneja: On Generalized Information Measures and Their Applications, Advances in Electronics and Electron Physics, Academic Press (USA), 76, 1989, 327-413. [14] S.Cronen-Townsend, Y. Zhou and Croft, W. B. , "A Framework for Selective Query Expansion," in Proceedings of CIKM 2004. [15] F.Song, W.B.Croft, A general language model for information retrieval, in Proceedings of SIGIR 1999. [16] Personal email contact with Vishwa Vinay and our own experiments [17] E.Yom-Tov, S.Fine, D.Carmel, A.Darlow, Learning to Estimate Query Difficulty Including Applications to Missing Content Detection and Distributed Information retrieval, in Proceedings of SIGIR 2005

homogenous test collection;web search environment;kl-divergence;content-based and named-page finding;jensen-shannon divergence;weighted information gain;trec document collection;query performance prediction;content-based query;robustness score probabilitydensity classifier;web search;gov2 collection;wig;ranking robustness technique;query classification;mixed-query situation;named-page finding task

train_H-46

Broad Expertise Retrieval in Sparse Data Environments

Expertise retrieval has been largely unexplored on data other than the W3C collection. At the same time, many intranets of universities and other knowledge-intensive organisations offer examples of relatively small but clean multilingual expertise data, covering broad ranges of expertise areas. We first present two main expertise retrieval tasks, along with a set of baseline approaches based on generative language modeling, aimed at finding expertise relations between topics and people. For our experimental evaluation, we introduce (and release) a new test set based on a crawl of a university site. Using this test set, we conduct two series of experiments. The first is aimed at determining the effectiveness of baseline expertise retrieval methods applied to the new test set. The second is aimed at assessing refined models that exploit characteristic features of the new test set, such as the organizational structure of the university, and the hierarchical structure of the topics in the test set. Expertise retrieval models are shown to be robust with respect to environments smaller than the W3C collection, and current techniques appear to be generalizable to other settings.

1. INTRODUCTION An organization"s intranet provides a means for exchanging information between employees and for facilitating employee collaborations. To efficiently and effectively achieve this, it is necessary to provide search facilities that enable employees not only to access documents, but also to identify expert colleagues. At the TREC Enterprise Track [22] the need to study and understand expertise retrieval has been recognized through the introduction of Expert Finding tasks. The goal of expert finding is to identify a list of people who are knowledgeable about a given topic. This task is usually addressed by uncovering associations between people and topics [10]; commonly, a co-occurrence of the name of a person with topics in the same context is assumed to be evidence of expertise. An alternative task, which using the same idea of people-topic associations, is expert profiling, where the task is to return a list of topics that a person is knowledgeable about [3]. The launch of the Expert Finding task at TREC has generated a lot of interest in expertise retrieval, with rapid progress being made in terms of modeling, algorithms, and evaluation aspects. However, nearly all of the expert finding or profiling work performed has been validated experimentally using the W3C collection [24] from the Enterprise Track. While this collection is currently the only publicly available test collection for expertise retrieval tasks, it only represents one type of intranet. With only one test collection it is not possible to generalize conclusions to other realistic settings. In this paper we focus on expertise retrieval in a realistic setting that differs from the W3C setting-one in which relatively small amounts of clean, multilingual data are available, that cover a broad range of expertise areas, as can be found on the intranets of universities and other knowledge-intensive organizations. Typically, this setting features several additional types of structure: topical structure (e.g., topic hierarchies as employed by the organization), organizational structure (faculty, department, ...), as well as multiple types of documents (research and course descriptions, publications, and academic homepages). This setting is quite different from the W3C setting in ways that might impact upon the performance of expertise retrieval tasks. We focus on a number of research questions in this paper: Does the relatively small amount of data available on an intranet affect the quality of the topic-person associations that lie at the heart of expertise retrieval algorithms? How do state-of-the-art algorithms developed on the W3C data set perform in the alternative scenario of the type described above? More generally, do the lessons from the Expert Finding task at TREC carry over to this setting? How does the inclusion or exclusion of different documents affect expertise retrieval tasks? In addition to, how can the topical and organizational structure be used for retrieval purposes? To answer our research questions, we first present a set of baseline approaches, based on generative language modeling, aimed at finding associations between topics and people. This allows us to formulate the expert finding and expert profiling tasks in a uniform way, and has the added benefit of allowing us to understand the relations between the two tasks. For our experimental evaluation, we introduce a new data set (the UvT Expert Collection) which is representative of the type of intranet that we described above. Our collection is based on publicly available data, crawled from the website of Tilburg University (UvT). This type of data is particularly interesting, since (1) it is clean, heterogeneous, structured, and focused, but comprises a limited number of documents; (2) contains information on the organizational hierarchy; (3) it is bilingual (English and Dutch); and (4) the list of expertise areas of an individual are provided by the employees themselves. Using the UvT Expert collection, we conduct two sets of experiments. The first is aimed at determining the effectiveness of baseline expertise finding and profiling methods in this new setting. A second group of experiments is aimed at extensions of the baseline methods that exploit characteristic features of the UvT Expert Collection; specifically, we propose and evaluate refined expert finding and profiling methods that incorporate topicality and organizational structure. Apart from the research questions and data set that we contribute, our main contributions are as follows. The baseline models developed for expertise finding perform well on the new data set. While on the W3C setting the expert finding task appears to be more difficult than profiling, for the UvT data the opposite is the case. We find that profiling on the UvT data set is considerably more difficult than on the W3C set, which we believe is due to the large (but realistic) number of topical areas that we used for profiling: about 1,500 for the UvT set, versus 50 in the W3C case. Taking the similarity between topics into account can significantly improve retrieval performance. The best performing similarity measures are content-based, therefore they can be applied on the W3C (and other) settings as well. Finally, we demonstrate that the organizational structure can be exploited in the form of a context model, improving MAP scores for certain models by up to 70%. The remainder of this paper is organized as follows. In the next section we review related work. Then, in Section 3 we provide detailed descriptions of the expertise retrieval tasks that we address in this paper: expert finding and expert profiling. In Section 4 we present our baseline models, of which the performance is then assessed in Section 6 using the UvT data set that we introduce in Section 5. Advanced models exploiting specific features of our data are presented in Section 7 and evaluated in Section 8. We formulate our conclusions in Section 9. 2. RELATED WORK Initial approaches to expertise finding often employed databases containing information on the skills and knowledge of each individual in the organization [11]. Most of these tools (usually called yellow pages or people-finding systems) rely on people to self-assess their skills against a predefined set of keywords. For updating profiles in these systems in an automatic fashion there is a need for intelligent technologies [5]. More recent approaches use specific document sets (such as email [6] or software [18]) to find expertise. In contrast with focusing on particular document types, there is also an increased interest in the development of systems that index and mine published intranet documents as sources of evidence for expertise. One such published approach is the P@noptic system [9], which builds a representation of each person by concatenating all documents associated with that person-this is similar to Model 1 of Balog et al. [4], who formalize and compare two methods. Balog et al."s Model 1 directly models the knowledge of an expert from associated documents, while their Model 2 first locates documents on the topic and then finds the associated experts. In the reported experiments the second method performs significantly better when there are sufficiently many associated documents per candidate. Most systems that took part in the 2005 and 2006 editions of the Expert Finding task at TREC implemented (variations on) one of these two models; see [10, 20]. Macdonald and Ounis [16] propose a different approach for ranking candidate expertise with respect to a topic based on data fusion techniques, without using collectionspecific heuristics; they find that applying field-based weighting models improves the ranking of candidates. Petkova and Croft [19] propose yet another approach, based on a combination of the above Model 1 and 2, explicitly modeling topics. Turning to other expert retrieval tasks that can also be addressed using topic-people associations, Balog and de Rijke [3] addressed the task of determining topical expert profiles. While their methods proved to be efficient on the W3C corpus, they require an amount of data that may not be available in the typical knowledge-intensive organization. Balog and de Rijke [2] study the related task of finding experts that are similar to a small set of experts given as input. As an aside, creating a textual summary of a person shows some similarities to biography finding, which has received a considerable amount of attention recently; see e.g., [13]. We use generative language modeling to find associations between topics and people. In our modeling of expert finding and profiling we collect evidence for expertise from multiple sources, in a heterogeneous collection, and integrate it with the co-occurrence of candidates" names and query terms-the language modeling setting allows us to do this in a transparent manner. Our modeling proceeds in two steps. In the first step, we consider three baseline models, two taken from [4] (the Models 1 and 2 mentioned above), and one a refined version of a model introduced in [3] (which we refer to as Model 3 below); this third model is also similar to the model described by Petkova and Croft [19]. The models we consider in our second round of experiments are mixture models similar to contextual language models [1] and to the expanded documents of Tao et al. [21]; however, the features that we use for definining our expansions-including topical structure and organizational structure-have not been used in this way before. 3. TASKS In the expertise retrieval scenario that we envisage, users seeking expertise within an organization have access to an interface that combines a search box (where they can search for experts or topics) with navigational structures (of experts and of topics) that allows them to click their way to an expert page (providing the profile of a person) or a topic page (providing a list of experts on the topic). To feed the above interface, we face two expertise retrieval tasks, expert finding and expert profiling, that we first define and then formalize using generative language models. In order to model either task, the probability of the query topic being associated to a candidate expert plays a key role in the final estimates for searching and profiling. By using language models, both the candidates and the query are characterized by distributions of terms in the vocabulary (used in the documents made available by the organization whose expertise retrieval needs we are addressing). 3.1 Expert finding Expert finding involves the task of finding the right person with the appropriate skills and knowledge: Who are the experts on topic X?. E.g., an employee wants to ascertain who worked on a particular project to find out why particular decisions were made without having to trawl through documentation (if there is any). Or, they may be in need a trained specialist for consultancy on a specific problem. Within an organization there are usually many possible candidates who could be experts for given topic. We can state this problem as follows: What is the probability of a candidate ca being an expert given the query topic q? That is, we determine p(ca|q), and rank candidates ca according to this probability. The candidates with the highest probability given the query are deemed the most likely experts for that topic. The challenge is how to estimate this probability accurately. Since the query is likely to consist of only a few terms to describe the expertise required, we should be able to obtain a more accurate estimate by invoking Bayes" Theorem, and estimating: p(ca|q) = p(q|ca)p(ca) p(q) , (1) where p(ca) is the probability of a candidate and p(q) is the probability of a query. Since p(q) is a constant, it can be ignored for ranking purposes. Thus, the probability of a candidate ca being an expert given the query q is proportional to the probability of a query given the candidate p(q|ca), weighted by the a priori belief p(ca) that candidate ca is an expert. p(ca|q) ∝ p(q|ca)p(ca) (2) In this paper our main focus is on estimating the probability of a query given the candidate p(q|ca), because this probability captures the extent to which the candidate knows about the query topic. Whereas the candidate priors are generally assumed to be uniformand thus will not influence the ranking-it has been demonstrated that a sensible choice of priors may improve the performance [20]. 3.2 Expert profiling While the task of expert searching was concerned with finding experts given a particular topic, the task of expert profiling seeks to answer a related question: What topics does a candidate know about? Essentially, this turns the questions of expert finding around. The profiling of an individual candidate involves the identification of areas of skills and knowledge that they have expertise about and an evaluation of the level of proficiency in each of these areas. This is the candidate"s topical profile. Generally, topical profiles within organizations consist of tabular structures which explicitly catalogue the skills and knowledge of each individual in the organization. However, such practice is limited by the resources available for defining, creating, maintaining, and updating these profiles over time. By focusing on automatic methods which draw upon the available evidence within the document repositories of an organization, our aim is to reduce the human effort associated with the maintenance of topical profiles1 . A topical profile of a candidate, then, is defined as a vector where each element i of the vector corresponds to the candidate ca"s expertise on a given topic ki, (i.e., s(ca, ki)). Each topic ki defines a particular knowledge area or skill that the organization uses to define the candidate"s topical profile. Thus, it is assumed that a list of topics, {k1, . . . , kn}, where n is the number of pre-defined topics, is given: profile(ca) = s(ca, k1), s(ca, k2), . . . , s(ca, kn) . (3) 1 Context and evidence are needed to help users of expertise finding systems to decide whom to contact when seeking expertise in a particular area. Examples of such context are: Who does she work with? What are her contact details? Is she well-connected, just in case she is not able to help us herself? What is her role in the organization? Who is her superior? Collaborators, and affiliations, etc. are all part of the candidate"s social profile, and can serve as a background against which the system"s recommendations should be interpreted. In this paper we only address the problem of determining topical profiles, and leave social profiling to further work. We state the problem of quantifying the competence of a person on a certain knowledge area as follows: What is the probability of a knowledge area (ki) being part of the candidate"s (expertise) profile? where s(ca, ki) is defined by p(ki|ca). Our task, then, is to estimate p(ki|ca), which is equivalent to the problem of obtaining p(q|ca), where the topic ki is represented as a query topic q, i.e., a sequence of keywords representing the expertise required. Both the expert finding and profiling tasks rely on the accurate estimation of p(q|ca). The only difference derives from the prior probability that a person is an expert (p(ca)), which can be incorporated into the expert finding task. This prior does not apply to the profiling task since the candidate (individual) is fixed. 4. BASELINE MODELS In this section we describe our baseline models for estimating p(q|ca), i.e., associations between topics and people. Both expert finding and expert profiling boil down to this estimation. We employ three models for calculating this probability. 4.1 From topics to candidates Using Candidate Models: Model 1 Model 1 [4] defines the probability of a query given a candidate (p(q|ca)) using standard language modeling techniques, based on a multinomial unigram language model. For each candidate ca, a candidate language model θca is inferred such that the probability of a term given θca is nonzero for all terms, i.e., p(t|θca) > 0. From the candidate model the query is generated with the following probability: p(q|θca) = Y t∈q p(t|θca)n(t,q) , where each term t in the query q is sampled identically and independently, and n(t, q) is the number of times t occurs in q. The candidate language model is inferred as follows: (1) an empirical model p(t|ca) is computed; (2) it is smoothed with background probabilities. Using the associations between a candidate and a document, the probability p(t|ca) can be approximated by: p(t|ca) = X d p(t|d)p(d|ca), where p(d|ca) is the probability that candidate ca generates a supporting document d, and p(t|d) is the probability of a term t occurring in the document d. We use the maximum-likelihood estimate of a term, that is, the normalised frequency of the term t in document d. The strength of the association between document d and candidate ca expressed by p(d|ca) reflects the degree to which the candidates expertise is described using this document. The estimation of this probability is presented later, in Section 4.2. The candidate model is then constructed as a linear interpolation of p(t|ca) and the background model p(t) to ensure there are no zero probabilities, which results in the final estimation: p(q|θca) = (4) Y t∈q ( (1 − λ) X d p(t|d)p(d|ca) ! + λp(t) )n(t,q) . Model 1 amasses all the term information from all the documents associated with the candidate, and uses this to represent that candidate. This model is used to predict how likely a candidate would produce a query q. This can can be intuitively interpreted as the probability of this candidate talking about the query topic, where we assume that this is indicative of their expertise. Using Document Models: Model 2 Model 2 [4] takes a different approach. Here, the process is broken into two parts. Given a candidate ca, (1) a document that is associated with a candidate is selected with probability p(d|ca), and (2) from this document a query q is generated with probability p(q|d). Then the sum over all documents is taken to obtain p(q|ca), such that: p(q|ca) = X d p(q|d)p(d|ca). (5) The probability of a query given a document is estimated by inferring a document language model θd for each document d in a similar manner as the candidate model was inferred: p(t|θd) = (1 − λ)p(t|d) + λp(t), (6) where p(t|d) is the probability of the term in the document. The probability of a query given the document model is: p(q|θd) = Y t∈q p(t|θd)n(t,q) . The final estimate of p(q|ca) is obtained by substituting p(q|d) for p(q|θd) into Eq. 5 (see [4] for full details). Conceptually, Model 2 differs from Model 1 because the candidate is not directly modeled. Instead, the document acts like a hidden variable in the process which separates the query from the candidate. This process is akin to how a user may search for candidates with a standard search engine: initially by finding the documents which are relevant, and then seeing who is associated with that document. By examining a number of documents the user can obtain an idea of which candidates are more likely to discuss the topic q. Using Topic Models: Model 3 We introduce a third model, Model 3. Instead of attempting to model the query generation process via candidate or document models, we represent the query as a topic language model and directly estimate the probability of the candidate p(ca|q). This approach is similar to the model presented in [3, 19]. As with the previous models, a language model is inferred, but this time for the query. We adapt the work of Lavrenko and Croft [14] to estimate a topic model from the query. The procedure is as follows. Given a collection of documents and a query topic q, it is assumed that there exists an unknown topic model θk that assigns probabilities p(t|θk) to the term occurrences in the topic documents. Both the query and the documents are samples from θk (as opposed to the previous approaches, where a query is assumed to be sampled from a specific document or candidate model). The main task is to estimate p(t|θk), the probability of a term given the topic model. Since the query q is very sparse, and as there are no examples of documents on the topic, this distribution needs to be approximated. Lavrenko and Croft [14] suggest a reasonable way of obtaining such an approximation, by assuming that p(t|θk) can be approximated by the probability of term t given the query q. We can then estimate p(t|q) using the joint probability of observing the term t together with the query terms, q1, . . . , qm, and dividing by the joint probability of the query terms: p(t|θk) ≈ p(t|q) = p(t, q1, . . . , qm) p(q1, . . . , qm) = p(t, q1, . . . , qm) P t ∈T p(t , q1, . . . , qm) , where p(q1, . . . , qm) = P t ∈T p(t , q1, . . . , qm), and T is the entire vocabulary of terms. In order to estimate the joint probability p(t, q1, . . . , qm), we follow [14, 15] and assume t and q1, . . . , qm are mutually independent, once we pick a source distribution from the set of underlying source distributions U. If we choose U to be a set of document models. then to construct this set, the query q would be issued against the collection, and the top n returned are assumed to be relevant to the topic, and thus treated as samples from the topic model. (Note that candidate models could be used instead.) With the document models forming U, the joint probability of term and query becomes: p(t, q1, . . . , qm) = X d∈U p(d) ˘ p(t|θd) mY i=1 p(qi|θd) ¯ . (7) Here, p(d) denotes the prior distribution over the set U, which reflects the relevance of the document to the topic. We assume that p(d) is uniform across U. In order to rank candidates according to the topic model defined, we use the Kullback-Leibler divergence metric (KL, [8]) to measure the difference between the candidate models and the topic model: KL(θk||θca) = X t p(t|θk) log p(t|θk) p(t|θca) . (8) Candidates with a smaller divergence from the topic model are considered to be more likely experts on that topic. The candidate model θca is defined in Eq. 4. By using KL divergence instead of the probability of a candidate given the topic model p(ca|θk), we avoid normalization problems. 4.2 Document-candidate associations For our models we need to be able to estimate the probability p(d|ca), which expresses the extent to which a document d characterizes the candidate ca. In [4], two methods are presented for estimating this probability, based on the number of person names recognized in a document. However, in our (intranet) setting it is reasonable to assume that authors of documents can unambiguously be identified (e.g., as the author of an article, the teacher assigned to a course, the owner of a web page, etc.) Hence, we set p(d|ca) to be 1 if candidate ca is author of document d, otherwise the probability is 0. In Section 6 we describe how authorship can be determined on different types of documents within the collection. 5. THE UVT EXPERT COLLECTION The UvT Expert collection used in the experiments in this paper fits the scenario outlined in Section 3. The collection is based on the Webwijs (Webwise) system developed at Tilburg University (UvT) in the Netherlands. Webwijs (http://www.uvt.nl/ webwijs/) is a publicly accessible database of UvT employees who are involved in research or teaching; currently, Webwijs contains information about 1168 experts, each of whom has a page with contact information and, if made available by the expert, a research description and publications list. In addition, each expert can select expertise areas from a list of 1491 topics and is encouraged to suggest new topics that need to be approved by the Webwijs editor. Each topic has a separate page that shows all experts associated with that topic and, if available, a list of related topics. Webwijs is available in Dutch and English, and this bilinguality has been preserved in the collection. Every Dutch Webwijs page has an English translation. Not all Dutch topics have an English translation, but the reverse is true: the 981 English topics all have a Dutch equivalent. About 42% of the experts teach courses at Tilburg University; these courses were also crawled and included in the profile. In addition, about 27% of the experts link to their academic homepage from their Webwijs page. These home pages were crawled and added to the collection. (This means that if experts put the full-text versions of their publications on their academic homepage, these were also available for indexing.) We also obtained 1880 full-text versions of publications from the UvT institutional repository and Dutch English no. of experts 1168 1168 no. of experts with ≥ 1 topic 743 727 no. of topics 1491 981 no. of expert-topic pairs 4318 3251 avg. no. of topics/expert 5.8 5.9 max. no. of topics/expert (no. of experts) 60 (1) 35 (1) min. no. of topics/expert (no. of experts) 1 (74) 1 (106) avg. no. of experts/topic 2.9 3.3 max. no. of experts/topic (no. of topics) 30 (1) 30 (1) min. no. of experts/topic (no. of topics) 1 (615) 1 (346) no. of experts with HP 318 318 no. of experts with CD 318 318 avg. no. of CDs per teaching expert 3.5 3.5 no. of experts with RD 329 313 no. of experts with PUB 734 734 avg. no. of PUBs per expert 27.0 27.0 avg. no. of PUB citations per expert 25.2 25.2 avg. no. of full-text PUBs per expert 1.8 1.8 Table 2: Descriptive statistics of the Dutch and English versions of the UvT Expert collection. converted them to plain text. We ran the TextCat [23] language identifier to classify the language of the home pages and the fulltext publications. We restricted ourselves to pages where the classifier was confident about the language used on the page. This resulted in four document types: research descriptions (RD), course descriptions (CD), publications (PUB; full-text and citationonly versions), and academic homepages (HP). Everything was bundled into the UvT Expert collection which is available at http: //ilk.uvt.nl/uvt-expert-collection/. The UvT Expert collection was extracted from a different organizational setting than the W3C collection and differs from it in a number of ways. The UvT setting is one with relatively small amounts of multilingual data. Document-author associations are clear and the data is structured and clean. The collection covers a broad range of expertise areas, as one can typically find on intranets of universities and other knowledge-intensive institutes. Additionally, our university setting features several types of structure (topical and organizational), as well as multiple document types. Another important difference between the two data sets is that the expertise areas in the UvT Expert collection are self-selected instead of being based on group membership or assignments by others. Size is another dimension along which the W3C and UvT Expert collections differ: the latter is the smaller of the two. Also realistic are the large differences in the amount of information available for each expert. Utilizing Webwijs is voluntary; 425 Dutch experts did not select any topics at all. This leaves us with 743 Dutch and 727 English usable expert profiles. Table 2 provides descriptive statistics for the UvT Expert collection. Universities tend to have a hierarchical structure that goes from the faculty level, to departments, research groups, down to the individual researchers. In the UvT Expert collection we have information about the affiliations of researchers with faculties and institutes, providing us with a two-level organizational hierarchy. Tilburg University has 22 organizational units at the faculty level (including the university office and several research institutes) and 71 departments, which amounts to 3.2 departments per faculty. As to the topical hierarchy used by Webwijs, 131 of the 1491 topics are top nodes in the hierarchy. This hierarchy has an average topic chain length of 2.65 and a maximum length of 7 topics. 6. EVALUATION Below, we evaluate Section 4"s models for expert finding and profiling onthe UvT Expert collection. We detail our research questions and experimental setup, and then present our results. 6.1 Research Questions We address the following research questions. Both expert finding and profiling rely on the estimations of p(q|ca). The question is how the models compare on the different tasks, and in the setting of the UvT Expert collection. In [4], Model 2 outperformed Model 1 on the W3C collection. How do they compare on our data set? And how does Model 3 compare to Model 1? What about performance differences between the two languages in our test collection? 6.2 Experimental Setup The output of our models was evaluated against the self-assigned topic labels, which were treated as relevance judgements. Results were evaluated separately for English and Dutch. For English we only used topics for which the Dutch translation was available; for Dutch all topics were considered. The results were averaged for the queries in the intersection of relevance judgements and results; missing queries do not contribute a value of 0 to the scores. We use standard information retrieval measures, such as Mean Average Precision (MAP) and Mean Reciprocal Rank (MRR). We also report the percentage of topics (%q) and candidates (%ca) covered, for the expert finding and profiling tasks, respectively. 6.3 Results Table 1 shows the performance of Model 1, 2, and 3 on the expert finding and profiling tasks. The rows of the table correspond to the various document types (RD, CD, PUB, and HP) and to their combinations. RD+CD+PUB+HP is equivalent to the full collection and will be referred as the BASELINE of our experiments. Looking at Table 1 we see that Model 2 performs the best across the board. However, when the data is clean and very focused (RD), Model 3 outperforms it in a number of cases. Model 1 has the best coverage of candidates (%ca) and topics (%q). The various document types differ in their characteristics and how they improve the finding and profiling tasks. Expert profiling benefits much from the clean data present in the RD and CD document types, while the publications contribute the most to the expert finding task. Adding the homepages does not prove to be particularly useful. When we compare the results across languages, we find that the coverage of English topics (%q) is higher than of the Dutch ones for expert finding. Apart from that, the scores fall in the same range for both languages. For the profiling task the coverage of the candidates (%ca) is very similar for both languages. However, the performance is substantially better for the English topics. While it is hard to compare scores across collections, we conclude with a brief comparison of the absolute scores in Table 1 to those reported in [3, 4] on the W3C test set (2005 edition). For expert finding the MAP scores for Model 2 reported here are about 50% higher than the corresponding figures in [4], while our MRR scores are slightly below those in [4]. For expert profiling, the differences are far more dramatic: the MAP scores for Model 2 reported here are around 50% below the scores in [3], while the (best) MRR scores are about the same as those in [3]. The cause for the latter differences seems to reside in the number of knowledge areas considered here-approx. 30 times more than in the W3C setting. 7. ADVANCED MODELS Now that we have developed and assessed basic language modeling techniques for expertise retrieval, we turn to refined models that exploit special features of our test collection. 7.1 Exploiting knowledge area similarity One way to improve the scoring of a query given a candidate is to consider what other requests the candidate would satisfy and use them as further evidence to support the original query, proportional Expert finding Expert profiling Document types Model 1 Model 2 Model 3 Model 1 Model 2 Model 3 %q MAP MRR %q MAP MRR %q MAP MRR %ca MAP MRR %ca MAP MRR %ca MAP MRR English RD 97.8 0.126 0.269 83.5 0.144 0.311 83.3 0.129 0.271 100 0.089 0.189 39.3 0.232 0.465 41.1 0.166 0.337 CD 97.8 0.118 0.227 91.7 0.123 0.248 91.7 0.118 0.226 32.8 0.188 0.381 32.4 0.195 0.385 32.7 0.203 0.370 PUB 97.8 0.200 0.330 98.0 0.216 0.372 98.0 0.145 0.257 78.9 0.167 0.364 74.5 0.212 0.442 78.9 0.135 0.299 HP 97.8 0.081 0.186 97.4 0.071 0.168 97.2 0.062 0.149 31.2 0.150 0.299 28.8 0.185 0.335 30.1 0.136 0.287 RD+CD 97.8 0.188 0.352 92.9 0.193 0.360 92.9 0.150 0.273 100 0.145 0.286 61.3 0.251 0.477 63.2 0.217 0.416 RD+CD+PUB 97.8 0.235 0.373 98.1 0.277 0.439 98.1 0.178 0.305 100 0.196 0.380 87.2 0.280 0.533 89.5 0.170 0.344 RD+CD+PUB+HP 97.8 0.237 0.372 98.6 0.280 0.441 98.5 0.166 0.293 100 0.199 0.387 88.7 0.281 0.525 90.9 0.169 0.329 Dutch RD 61.3 0.094 0.229 38.4 0.137 0.336 38.3 0.127 0.295 38.0 0.127 0.386 34.1 0.138 0.420 38.0 0.105 0.327 CD 61.3 0.107 0.212 49.7 0.128 0.256 49.7 0.136 0.261 32.5 0.151 0.389 31.8 0.158 0.396 32.5 0.170 0.380 PUB 61.3 0.193 0.319 59.5 0.218 0.368 59.4 0.173 0.291 78.8 0.126 0.364 76.0 0.150 0.424 78.8 0.103 0.294 HP 61.3 0.063 0.169 56.6 0.064 0.175 56.4 0.062 0.163 29.8 0.108 0.308 27.8 0.125 0.338 29.8 0.098 0.255 RD+CD 61.3 0.159 0.314 51.9 0.184 0.360 51.9 0.169 0.324 60.5 0.151 0.410 57.2 0.166 0.431 60.4 0.159 0.384 RD+CD+PUB 61.3 0.244 0.398 61.5 0.260 0.424 61.4 0.210 0.350 90.3 0.165 0.445 88.2 0.189 0.479 90.3 0.126 0.339 RD+CD+PUB+HP 61.3 0.249 0.401 62.6 0.265 0.436 62.6 0.195 0.344 91.9 0.164 0.426 90.1 0.195 0.488 91.9 0.125 0.328 Table 1: Performance of the models on the expert finding and profiling tasks, using different document types and their combinations. %q is the number of topics covered (applies to the expert finding task), %ca is the number of candidates covered (applies to the expert profiling task). The top and bottom blocks correspond to English and Dutch respectively. The best scores are in boldface. to how related the other requests are to the original query. This can be modeled by interpolating between the p(q|ca) and the further supporting evidence from all similar requests q , as follows: p (q|ca) = λp(q|ca) + (1 − λ) X q p(q|q )p(q |ca), (9) where p(q|q ) represents the similarity between the two topics q and q . To be able to work with similarity methods that are not necessarily probabilities, we set p(q|q ) = w(q,q ) γ , where γ is a normalizing constant, such that γ = P q w(q , q ). We consider four methods for calculating the similarity score between two topics. Three approaches are strictly content-based, and establish similarity by examining co-occurrence patterns of topics within the collection, while the last approach exploits the hierarchical structure of topical areas that may be present within an organization (see [7] for further examples of integrating word relationships into language models). The Kullback-Leibler (KL) divergence metric defined in Eq. 8 provides a measure of how different or similar two probability distributions are. A topic model is inferred for q and q using the method presented in Section 4.1 to describe the query across the entire vocabulary. Since a lower KL score means the queries are more similar, we let w(q, q ) = max(KL(θq||·) − KL(θq||θq )). Pointwise Mutual Information (PMI, [17]) is a measure of association used in information theory to determine the extent of independence between variables. The dependence between two queries is reflected by the SI(q, q ) score, where scores greater than zero indicate that it is likely that there is a dependence, which we take to mean that the queries are likely to be similar: SI(q, q ) = log p(q, q ) p(q)p(q ) (10) We estimate the probability of a topic p(q) using the number of documents relevant to query q within the collection. The joint probability p(q, q ) is estimated similarly, by using the concatenation of q and q as a query. To obtain p(q|q ), we then set w(q, q ) = SI(q, q ) when SI(q, q ) > 0 otherwise w(q, q ) = 0, because we are only interested in including queries that are similar. The log-likelihood statistic provides another measure of dependence, which is more reliable than the pointwise mutual information measure [17]. Let k1 be the number of co-occurrences of q and q , k2 the number of occurrences of q not co-occurring with q , n1 the total number of occurrences of q , and n2 the total number of topic tokens minus the number of occurrences of q . Then, let p1 = k1/n1, p2 = k2/n2, and p = (k1 + k2)/(n1 + n2), (q, q ) = 2( (p1, k1, n1) + (p2, k2, n2) − (p, k1, n1) − (p, k2, n2)), where (p, n, k) = k log p + (n − k) log(1 − p). The higher score indicate that queries are also likely to be similar, thus we set w(q, q ) = (q, q ). Finally, we also estimate the similarity of two topics based on their distance within the topic hierarchy. The topic hierarchy is viewed as a directed graph, and for all topic-pairs the shortest path SP(q, q ) is calculated. We set the similarity score to be the reciprocal of the shortest path: w(q, q ) = 1/SP(q, q ). 7.2 Contextual information Given the hierarchy of an organization, the units to which a person belong are regarded as a context so as to compensate for data sparseness. We model it as follows: p (q|ca) = 1 − P ou∈OU(ca) λou · p(q|ca) + P ou∈OU(ca) λou · p(q|ou), where OU(ca) is the set of organizational units of which candidate ca is a member of, and p(q|o) expresses the strength of the association between query q and the unit ou. The latter probability can be estimated using either of the three basic models, by simply replacing ca with ou in the corresponding equations. An organizational unit is associated with all the documents that its members have authored. That is, p(d|ou) = maxca∈ou p(d|ca). 7.3 A simple multilingual model For knowledge institutes in Europe, academic or otherwise, a multilingual (or at least bilingual) setting is typical. The following model builds on a kind of independence assumption: there is no spill-over of expertise/profiles across language boundaries. While a simplification, this is a sensible first approach. That is: p (q|ca) =P l∈L λl · p(ql|ca), where L is the set of languages used in the collection, ql is the translation of the query q to language l, and λl is a language specific smoothing parameter, such that P l∈L λl = 1. 8. ADVANCED MODELS: EVALUATION In this section we present an experimental evaluation of our advanced models. Expert finding Expert profiling Language Model 1 Model 2 Model 3 Model 1 Model 2 Model 3 %q MAP MRR %q MAP MRR %q MAP MRR %ca MAP MRR %ca MAP MRR %ca MAP MRR English only 97.8 0.237 0.372 98.6 0.280 0.441 98.5 0.166 0.293 100 0.199 0.387 88.7 0.281 0.525 90.9 0.169 0.329 Dutch only 61.3 0.249 0.401 62.6 0.265 0.436 62.6 0.195 0.344 91.9 0.164 0.426 90.1 0.195 0.488 91.9 0.125 0.328 Combination 99.4 0.297 0.444 99.7 0.324 0.491 99.7 0.223 0.388 100 0.241 0.445 92.1 0.313 0.564 93.2 0.224 0.411 Table 3: Performance of the combination of languages on the expert finding and profiling tasks (on candidates). Best scores for each model are in italic, absolute best scores for the expert finding and profiling tasks are in boldface. Method Model 1 Model 2 Model 3 MAP MRR MAP MRR MAP MRR English BASELINE 0.296 0.454 0.339 0.509 0.221 0.333 KLDIV 0.291 0.453 0.327 0.503 0.219 0.330 PMI 0.291 0.453 0.337 0.509 0.219 0.331 LL 0.319 0.490 0.360 0.524 0.233 0.368 HDIST 0.299 0.465 0.346 0.537 0.219 0.332 Dutch BASELINE 0.240 0.350 0.271 0.403 0.227 0.389 KLDIV 0.239 0.347 0.253 0.386 0.224 0.385 PMI 0.239 0.350 0.260 0.392 0.227 0.389 LL 0.255 0.372 0.281 0.425 0.231 0.389 HDIST 0.253 0.365 0.271 0.407 0.236 0.402 Method Model 1 Model 2 Model 3 MAP MRR MAP MRR MAP MRR English BASELINE 0.485 0.546 0.499 0.548 0.381 0.416 KLDIV 0.510 0.564 0.513 0.558 0.381 0.416 PMI 0.486 0.546 0.495 0.542 0.407 0.451 LL 0.558 0.589 0.586 0.617 0.408 0.453 HDIST 0.507 0.567 0.512 0.563 0.386 0.420 Dutch BASELINE 0.263 0.313 0.294 0.358 0.262 0.315 KLDIV 0.284 0.336 0.271 0.321 0.261 0.314 PMI 0.265 0.317 0.265 0.316 0.273 0.330 LL 0.312 0.351 0.330 0.377 0.284 0.331 HDIST 0.280 0.327 0.288 0.341 0.266 0.321 Table 4: Performance on the expert finding (top) and profiling (bottom) tasks, using knowledge area similarities. Runs were evaluated on the main topics set. Best scores are in boldface. 8.1 Research Questions Our questions follow the refinements presented in the preceding section: Does exploiting the knowledge area similarity improve effectiveness? Which of the various methods for capturing word relationships is most effective? Furthermore, is our way of bringing in contextual information useful? For which tasks? And finally, is our simple way of combining the monolingual scores sufficient for obtaining significant improvements? 8.2 Experimental setup Given that the self-assessments are also sparse in our collection, in order to be able to measure differences between the various models, we selected a subset of topics, and evaluated (some of the) runs only on this subset. This set is referred as main topics, and consists of topics that are located at the top level of the topical hierarchy. (A main topic has subtopics, but is not a subtopic of any other topic.) This main set consists of 132 Dutch and 119 English topics. The relevance judgements were restricted to the main topic set, but were not expanded with subtopics. 8.3 Exploiting knowledge area similarity Table 4 presents the results. The four methods used for estimating knowledge-area similarity are KL divergence (KLDIV), PointLang. Topics Model 1 Model 2 Model 3 MAP MRR MAP MRR MAP MRR Expert finding UK ALL 0.423 0.545 0.654 0.799 0.494 0.629 UK MAIN 0.500 0.621 0.704 0.834 0.587 0.699 NL ALL 0.439 0.560 0.672 0.826 0.480 0.630 NL MAIN 0.440 0.584 0.645 0.816 0.515 0.655 Expert profiling UK ALL 0.240 0.640 0.306 0.778 0.223 0.616 UK MAIN 0.523 0.677 0.519 0.648 0.461 0.587 NL ALL 0.203 0.716 0.254 0.770 0.183 0.627 NL MAIN 0.332 0.576 0.380 0.624 0.332 0.549 Table 5: Evaluating the context models on organizational units. wise mutual information (PMI), log-likelihood (LL), and distance within topic hierarchy (HDIST). We managed to improve upon the baseline in all cases, but the improvement is more noticeable for the profiling task. For both tasks, the LL method performed best. The content-based approaches performed consistently better than HDIST. 8.4 Contextual information A two level hierarchy of organizational units (faculties and institutes) is available in the UvT Expert collection. The unit a person belongs to is used as a context for that person. First, we evaluated the models of the organizational units, using all topics (ALL) and only the main topics (MAIN). An organizational unit is considered to be relevant for a given topic (or vice versa) if at least one member of the unit selected the given topic as an expertise area. Table 5 reports on the results. As far as expert finding goes, given a topic, the corresponding organizational unit can be identified with high precision. However, the expert profiling task shows a different picture: the scores are low, and the task seems hard. The explanation may be that general concepts (i.e., our main topics) may belong to several organizational units. Second, we performed another evaluation, where we combined the contextual models with the candidate models (to score candidates again). Table 6 reports on the results. We find a positive impact of the context models only for expert finding. Noticably, for expert finding (and Model 1), it improves over 50% (for English) and over 70% (for Dutch) on MAP. The poor performance on expert profiling may be due to the fact that context models alone did not perform very well on the profiling task to begin with. 8.5 Multilingual models In this subsection we evaluate the method for combining results across multiple languages that we described in Section 7.3. In our setting the set of languages consists of English and Dutch: L = {UK, NL}. The weights on these languages were set to be identical (λUK = λNL = 0.5). We performed experiments with various λ settings, but did not observe significant differences in performance. Table 3 reports on the multilingual results, where performance is evaluated on the full topic set. All three models significantly imLang. Method Model 1 Model 2 Model 3 MAP MRR MAP MRR MAP MRR Expert finding UK BL 0.296 0.454 0.339 0.509 0.221 0.333 UK CT 0.330 0.491 0.342 0.500 0.228 0.342 NL BL 0.240 0.350 0.271 0.403 0.227 0.389 NL CT 0.251 0.382 0.267 0.410 0.246 0.404 Expert profiling UK BL 0.485 0.546 0.499 0.548 0.381 0.416 UK CT 0.562 0.620 0.508 0.558 0.440 0.486 NL BL 0.263 0.313 0.294 0.358 0.262 0.315 NL CT 0.330 0.384 0.317 0.387 0.294 0.345 Table 6: Performance of the context models (CT) compared to the baseline (BL). Best scores are in boldface. proved over all measures for both tasks. The coverage of topics and candidates for the expert finding and profiling tasks, respectively, is close to 100% in all cases. The relative improvement of the precision scores ranges from 10% to 80%. These scores demonstrate that despite its simplicity, our method for combining results over multiple languages achieves substantial improvements over the baseline. 9. CONCLUSIONS In this paper we focused on expertise retrieval (expert finding and profiling) in a new setting of a typical knowledge-intensive organization in which the available data is of high quality, multilingual, and covering a broad range of expertise area. Typically, the amount of available data in such an organization (e.g., a university, a research institute, or a research lab) is limited when compared to the W3C collection that has mostly been used for the experimental evaluation of expertise retrieval so far. To examine expertise retrieval in this setting, we introduced (and released) the UvT Expert collection as a representative case of such knowledge intensive organizations. The new collection reflects the typical properties of knowledge-intensive institutes noted above and also includes several features which may are potentially useful for expertise retrieval, such as topical and organizational structure. We evaluated how current state-of-the-art models for expert finding and profiling performed in this new setting and then refined these models in order to try and exploit the different characteristics within the data environment (language, topicality, and organizational structure). We found that current models of expertise retrieval generalize well to this new environment; in addition we found that refining the models to account for the differences results in significant improvements, thus making up for problems caused by data sparseness issues. Future work includes setting up manual assessments of automatically generated profiles by the employees themselves, especially in cases where the employees have not provided a profile themselves. 10. ACKNOWLEDGMENTS Krisztian Balog was supported by the Netherlands Organisation for Scientific Research (NWO) under project number 220-80-001. Maarten de Rijke was also supported by NWO under project numbers 017.001.190, 220-80-001, 264-70-050, 354-20-005, 600.065.120, 612-13-001, 612.000.106, 612.066.302, 612.069.006, 640.001.501, 640.002.501, and by the E.U. IST programme of the 6th FP for RTD under project MultiMATCH contract IST-033104. The work of Toine Bogers and Antal van den Bosch was funded by the IOP-MMI-program of SenterNovem / The Dutch Ministry of Economic Affairs, as part of the `A Propos project. 11. REFERENCES [1] L. Azzopardi. Incorporating Context in the Language Modeling Framework for ad-hoc Information Retrieval. PhD thesis, University of Paisley, 2005. [2] K. Balog and M. de Rijke. Finding similar experts. In This volume, 2007. [3] K. Balog and M. de Rijke. Determining expert profiles (with an application to expert finding). In IJCAI "07: Proc. 20th Intern. Joint Conf. on Artificial Intelligence, pages 2657-2662, 2007. [4] K. Balog, L. Azzopardi, and M. de Rijke. Formal models for expert finding in enterprise corpora. In SIGIR "06: Proc. 29th annual intern. ACM SIGIR conf. on Research and development in information retrieval, pages 43-50, 2006. [5] I. Becerra-Fernandez. The role of artificial intelligence technologies in the implementation of people-finder knowledge management systems. In AAAI Workshop on Bringing Knowledge to Business Processes, March 2000. [6] C. S. Campbell, P. P. Maglio, A. Cozzi, and B. Dom. Expertise identification using email communications. In CIKM "03: Proc. twelfth intern. conf. on Information and knowledge management, pages 528531, 2003. [7] G. Cao, J.-Y. Nie, and J. Bai. Integrating word relationships into language models. In SIGIR "05: Proc. 28th annual intern. ACM SIGIR conf. on Research and development in information retrieval, pages 298-305, 2005. [8] T. M. Cover and J. A. Thomas. Elements of Information Theory. Wiley-Interscience, 1991. [9] N. Craswell, D. Hawking, A. M. Vercoustre, and P. Wilkins. P@noptic expert: Searching for experts not just for documents. In Ausweb, 2001. [10] N. Craswell, A. de Vries, and I. Soboroff. Overview of the TREC2005 Enterprise Track. In The Fourteenth Text REtrieval Conf. Proc. (TREC 2005), 2006. [11] T. H. Davenport and L. Prusak. Working Knowledge: How Organizations Manage What They Know. Harvard Business School Press, Boston, MA, 1998. [12] T. Dunning. Accurate methods for the statistics of surprise and coincidence. Computational Linguistics, 19(1):61-74, 1993. [13] E. Filatova and J. Prager. Tell me what you do and I"ll tell you what you are: Learning occupation-related activities for biographies. In HLT/EMNLP, 2005. [14] V. Lavrenko and W. B. Croft. Relevance based language models. In SIGIR "01: Proc. 24th annual intern. ACM SIGIR conf. on Research and development in information retrieval, pages 120-127, 2001. [15] V. Lavrenko, M. Choquette, and W. B. Croft. Cross-lingual relevance models. In SIGIR "02: Proc. 25th annual intern. ACM SIGIR conf. on Research and development in information retrieval, pages 175-182, 2002. [16] C. Macdonald and I. Ounis. Voting for candidates: adapting data fusion techniques for an expert search task. In CIKM "06: Proc. 15th ACM intern. conf. on Information and knowledge management, pages 387-396, 2006. [17] C. Manning and H. Sch¨utze. Foundations of Statistical Natural Language Processing. The MIT Press, 1999. [18] A. Mockus and J. D. Herbsleb. Expertise browser: a quantitative approach to identifying expertise. In ICSE "02: Proc. 24th Intern. Conf. on Software Engineering, pages 503-512, 2002. [19] D. Petkova and W. B. Croft. Hierarchical language models for expert finding in enterprise corpora. In Proc. ICTAI 2006, pages 599-608, 2006. [20] I. Soboroff, A. de Vries, and N. Craswell. Overview of the TREC 2006 Enterprise Track. In TREC 2006 Working Notes, 2006. [21] T. Tao, X. Wang, Q. Mei, and C. Zhai. Language model information retrieval with document expansion. In HLT-NAACL 2006, 2006. [22] TREC. Enterprise track, 2005. URL: http://www.ins.cwi. nl/projects/trec-ent/wiki/. [23] G. van Noord. TextCat Language Guesser. URL: http://www. let.rug.nl/˜vannoord/TextCat/. [24] W3C. The W3C test collection, 2005. URL: http://research. microsoft.com/users/nickcr/w3c-summary.html.

language model;broad expertise retrieval;expertise search;expert finding task;expert colleague;bayes' theorem;organizational structure;baseline model;intranet search;trec enterprise track;baseline expertise retrieval method;co-occurrence;expert find;generative language modeling;sparse datum environment;topicality and organizational structure

train_H-47

A Semantic Approach to Contextual Advertising

Contextual advertising or Context Match (CM) refers to the placement of commercial textual advertisements within the content of a generic web page, while Sponsored Search (SS) advertising consists in placing ads on result pages from a web search engine, with ads driven by the originating query. In CM there is usually an intermediary commercial ad-network entity in charge of optimizing the ad selection with the twin goal of increasing revenue (shared between the publisher and the ad-network) and improving the user experience. With these goals in mind it is preferable to have ads relevant to the page content, rather than generic ads. The SS market developed quicker than the CM market, and most textual ads are still characterized by bid phrases representing those queries where the advertisers would like to have their ad displayed. Hence, the first technologies for CM have relied on previous solutions for SS, by simply extracting one or more phrases from the given page content, and displaying ads corresponding to searches on these phrases, in a purely syntactic approach. However, due to the vagaries of phrase extraction, and the lack of context, this approach leads to many irrelevant ads. To overcome this problem, we propose a system for contextual ad matching based on a combination of semantic and syntactic features.

1. INTRODUCTION Web advertising supports a large swath of today"s Internet ecosystem. The total internet advertiser spend in US alone in 2006 is estimated at over 17 billion dollars with a growth rate of almost 20% year over year. A large part of this market consists of textual ads, that is, short text messages usually marked as sponsored links or similar. The main advertising channels used to distribute textual ads are: 1. Sponsored Search or Paid Search advertising which consists in placing ads on the result pages from a web search engine, with ads driven by the originating query. All major current web search engines (Google, Yahoo!, and Microsoft) support such ads and act simultaneously as a search engine and an ad agency. 2. Contextual advertising or Context Match which refers to the placement of commercial ads within the content of a generic web page. In contextual advertising usually there is a commercial intermediary, called an ad-network, in charge of optimizing the ad selection with the twin goal of increasing revenue (shared between publisher and ad-network) and improving user experience. Again, all major current web search engines (Google, Yahoo!, and Microsoft) provide such ad-networking services but there are also many smaller players. The SS market developed quicker than the CM market, and most textual ads are still characterized by bid phrases representing those queries where the advertisers would like to have their ad displayed. (See [5] for a brief history). However, today, almost all of the for-profit non-transactional web sites (that is, sites that do not sell anything directly) rely at least in part on revenue from context match. CM supports sites that range from individual bloggers and small niche communities to large publishers such as major newspapers. Without this model, the web would be a lot smaller! The prevalent pricing model for textual ads is that the advertisers pay a certain amount for every click on the advertisement (pay-per-click or PPC). There are also other models used: pay-per-impression, where the advertisers pay for the number of exposures of an ad and pay-per-action where the advertiser pays only if the ad leads to a sale or similar transaction. For simplicity, we only deal with the PPC model in this paper. Given a page, rather than placing generic ads, it seems preferable to have ads related to the content to provide a better user experience and thus to increase the probability of clicks. This intuition is supported by the analogy to conventional publishing where there are very successful magazines (e.g. Vogue) where a majority of the content is topical advertising (fashion in the case of Vogue) and by user studies that have confirmed that increased relevance increases the number of ad-clicks [4, 13]. Previous published approaches estimated the ad relevance based on co-occurrence of the same words or phrases within the ad and within the page (see [7, 8] and Section 3 for more details). However targeting mechanisms based solely on phrases found within the text of the page can lead to problems: For example, a page about a famous golfer named John Maytag might trigger an ad for Maytag dishwashers since Maytag is a popular brand. Another example could be a page describing the Chevy Tahoe truck (a popular vehicle in US) triggering an ad about Lake Tahoe vacations. Polysemy is not the only culprit: there is a (maybe apocryphal) story about a lurid news item about a headless body found in a suitcase triggering an ad for Samsonite luggage! In all these examples the mismatch arises from the fact that the ads are not appropriate for the context. In order to solve this problem we propose a matching mechanism that combines a semantic phase with the traditional keyword matching, that is, a syntactic phase. The semantic phase classifies the page and the ads into a taxonomy of topics and uses the proximity of the ad and page classes as a factor in the ad ranking formula. Hence we favor ads that are topically related to the page and thus avoid the pitfalls of the purely syntactic approach. Furthermore, by using a hierarchical taxonomy we allow for the gradual generalization of the ad search space in the case when there are no ads matching the precise topic of the page. For example if the page is about an event in curling, a rare winter sport, and contains the words Alpine Meadows, the system would still rank highly ads for skiing in Alpine Meadows as these ads belong to the class skiing which is a sibling of the class curling and both of these classes share the parent winter sports. In some sense, the taxonomy classes are used to select the set of applicable ads and the keywords are used to narrow down the search to concepts that are of too small granularity to be in the taxonomy. The taxonomy contains nodes for topics that do not change fast, for example, brands of digital cameras, say Canon. The keywords capture the specificity to a level that is more dynamic and granular. In the digital camera example this would correspond to the level of a particular model, say Canon SD450 whose advertising life might be just a few months. Updating the taxonomy with new nodes or even new vocabulary each time a new model comes to the market is prohibitively expensive when we are dealing with millions of manufacturers. In addition to increased click through rate (CTR) due to increased relevance, a significant but harder to quantify benefit of the semantic-syntactic matching is that the resulting page has a unified feel and improves the user experience. In the Chevy Tahoe example above, the classifier would establish that the page is about cars/automotive and only those ads will be considered. Even if there are no ads for this particular Chevy model, the chosen ads will still be within the automotive domain. To implement our approach we need to solve a challenging problem: classify both pages and ads within a large taxonomy (so that the topic granularity would be small enough) with high precision (to reduce the probability of mis-match). We evaluated several classifiers and taxonomies and in this paper we present results using a taxonomy with close to 6000 nodes using a variation of the Rocchio"s classifier [9]. This classifier gave the best results in both page and ad classification, and ultimately in ad relevance. The paper proceeds as follows. In the next section we review the basic principles of the contextual advertising. Section 3 overviews the related work. Section 4 describes the taxonomy and document classifier that were used for page and ad classification. Section 5 describes the semanticsyntactic method. In Section 6 we briefly discuss how to search efficiently the ad space in order to return the top-k ranked ads. Experimental evaluation is presented in Section 7. Finally, Section 8 presents the concluding remarks. 2. OVERVIEW OF CONTEXTUAL ADVERTISING Contextual advertising is an interplay of four players: • The publisher is the owner of the web pages on which the advertising is displayed. The publisher typically aims to maximize advertising revenue while providing a good user experience. • The advertiser provides the supply of ads. Usually the activity of the advertisers are organized around campaigns which are defined by a set of ads with a particular temporal and thematic goal (e.g. sale of digital cameras during the holiday season). As in traditional advertising, the goal of the advertisers can be broadly defined as the promotion of products or services. • The ad network is a mediator between the advertiser and the publisher and selects the ads that are put on the pages. The ad-network shares the advertisement revenue with the publisher. • Users visit the web pages of the publisher and interact with the ads. Contextual advertising usually falls into the category of direct marketing (as opposed to brand advertising), that is advertising whose aim is a direct response where the effect of an campaign is measured by the user reaction. One of the advantages of online advertising in general and contextual advertising in particular is that, compared to the traditional media, it is relatively easy to measure the user response. Usually the desired immediate reaction is for the user to follow the link in the ad and visit the advertiser"s web site and, as noted, the prevalent financial model is that the advertiser pays a certain amount for every click on the advertisement (PPC). The revenue is shared between the publisher and the network. Context match advertising has grown from Sponsored Search advertising, which consists in placing ads on the result pages from a web search engine, with ads driven by the originating query. In most networks, the amount paid by the advertiser for each SS click is determined by an auction process where the advertisers place bids on a search phrase, and their position in the tower of ads displayed in conjunction with the result is determined by their bid. Thus each ad is annotated with one or more bid phrases. The bid phrase has no direct bearing on the ad placement in CM. However, it is a concise description of target ad audience as determined by the advertiser and it has been shown to be an important feature for successful CM ad placement [8]. In addition to the bid phrase, an ad is also characterized by a title usually displayed in a bold font, and an abstract or creative, which is the few lines of text, usually less than 120 characters, displayed on the page. The ad-network model aligns the interests of the publishers, advertisers and the network. In general, clicks bring benefits to both the publisher and the ad network by providing revenue, and to the advertiser by bringing traffic to the target web site. The revenue of the network, given a page p, can be estimated as: R = X i=1..k P(click|p, ai)price(ai, i) where k is the number of ads displayed on page p and price(ai, i) is the click-price of the current ad ai at position i. The price in this model depends on the set of ads presented on the page. Several models have been proposed to determine the price, most of them based on generalizations of second price auctions. However, for simplicity we ignore the pricing model and concentrate on finding ads that will maximize the first term of the product, that is we search for arg max i P(click|p, ai) Furthermore we assume that the probability of click for a given ad and page is determined by its relevance score with respect to the page, thus ignoring the positional effect of the ad placement on the page. We assume that this is an orthogonal factor to the relevance component and could be easily incorporated in the model. 3. RELATED WORK Online advertising in general and contextual advertising in particular are emerging areas of research. The published literature is very sparse. A study presented in [13] confirms the intuition that ads need to be relevant to the user"s interest to avoid degrading the user"s experience and increase the probability of reaction. A recent work by Ribeiro-Neto et. al. [8] examines a number of strategies to match pages to ads based on extracted keywords. The ads and pages are represented as vectors in a vector space. The first five strategies proposed in that work match the pages and the ads based on the cosine of the angle between the ad vector and the page vector. To find out the important part of the ad, the authors explore using different ad sections (bid phrase, title, body) as a basis for the ad vector. The winning strategy out of the first five requires the bid phrase to appear on the page and then ranks all such ads by the cosine of the union of all the ad sections and the page vectors. While both pages and ads are mapped to the same space, there is a discrepancy (impendence mismatch) between the vocabulary used in the ads and in the pages. Furthermore, since in the vector model the dimensions are determined by the number of unique words, plain cosine similarity will not take into account synonyms. To solve this problem, Ribeiro-Neto et al expand the page vocabulary with terms from other similar pages weighted based on the overall similarity of the origin page to the matched page, and show improved matching precision. In a follow-up work [7] the authors propose a method to learn impact of individual features using genetic programming to produce a matching function. The function is represented as a tree composed of arithmetic operators and the log function as internal nodes, and different numerical features of the query and ad terms as leafs. The results show that genetic programming finds matching functions that significantly improve the matching compared to the best method (without page side expansion) reported in [8]. Another approach to contextual advertising is to reduce it to the problem of sponsored search advertising by extracting phrases from the page and matching them with the bid phrase of the ads. In [14] a system for phrase extraction is described that used a variety of features to determine the importance of page phrases for advertising purposes. The system is trained with pages that have been hand annotated with important phrases. The learning algorithm takes into account features based on tf-idf, html meta data and query logs to detect the most important phrases. During evaluation, each page phrase up to length 5 is considered as potential result and evaluated against a trained classifier. In our work we also experimented with a phrase extractor based on the work reported in [12]. While increasing slightly the precision, it did not change the relative performance of the explored algorithms. 4. PAGE AND AD CLASSIFICATION 4.1 Taxonomy Choice The semantic match of the pages and the ads is performed by classifying both into a common taxonomy. The matching process requires that the taxonomy provides sufficient differentiation between the common commercial topics. For example, classifying all medical related pages into one node will not result into a good classification since both sore foot and flu pages will end up in the same node. The ads suitable for these two concepts are, however, very different. To obtain sufficient resolution, we used a taxonomy of around 6000 nodes primarily built for classifying commercial interest queries, rather than pages or ads. This taxonomy has been commercially built by Yahoo! US. We will explain below how we can use the same taxonomy to classify pages and ads as well. Each node in our source taxonomy is represented as a collection of exemplary bid phrases or queries that correspond to that node concept. Each node has on average around 100 queries. The queries placed in the taxonomy are high volume queries and queries of high interest to advertisers, as indicated by an unusually high cost-per-click (CPC) price. The taxonomy has been populated by human editors using keyword suggestions tools similar to the ones used by ad networks to suggest keywords to advertisers. After initial seeding with a few queries, using the provided tools a human editor can add several hundreds queries to a given node. Nevertheless, it has been a significant effort to develop this 6000-nodes taxonomy and it has required several person-years of work. A similar-in-spirit process for building enterprise taxonomies via queries has been presented in [6]. However, the details and tools are completely different. Figure 1 provides some statistics about the taxonomy used in this work. 4.2 Classification Method As explained, the semantic phase of the matching relies on ads and pages being topically close. Thus we need to classify pages into the same taxonomy used to classify ads. In this section we overview the methods we used to build a page and an ad classifier pair. The detailed description and evaluation of this process is outside the scope of this paper. Given the taxonomy of queries (or bid-phrases - we use these terms interchangeably) described in the previous section, we tried three methods to build corresponding page and ad classifiers. For the first two methods we tried to find exemplary pages and ads for each concept as follows: Number of Categories By Level 0 200 400 600 800 1000 1200 1400 1600 1800 2000 1 2 3 4 5 6 7 8 9 Level NumberofCategories Number of Children per Nodes 0 50 100 150 200 250 300 350 400 0 2 4 6 8 10 12 14 16 18 20 22 24 29 31 33 35 52 Number of Children NumberofNodes Queries per Node 0 500 1000 1500 2000 2500 3000 1 50 80 120 160 200 240 280 320 360 400 440 480 Number Queries (up to 500+) NumberofNodes Figure 1: Taxonomy statistics: categories per level; fanout for non-leaf nodes; and queries per node We generated a page training set by running the queries in the taxonomy over a Web search index and using the top 10 results after some filtering as documents labeled with the query"s label. On the ad side we generated a training set for each class by selecting the ads that have a bid phrase assigned to this class. Using this training sets we then trained a hierarchical SVM [2] (one against all between every group of siblings) and a log-regression [11] classifier. (The second method differs from the first in the type of secondary filtering used. This filtering eliminates low content pages, pages deemed unsuitable for advertising, pages that lead to excessive class confusion, etc.) However, we obtained the best performance by using the third document classifier, based on the information in the source taxonomy queries only. For each taxonomy node we concatenated all the exemplary queries into a single metadocument. We then used the meta document as a centroid for a nearest-neighbor classifier based on Rocchio"s framework [9] with only positive examples and no relevance feedback. Each centroid is defined as a sum of the tf-idf values of each term, normalized by the number of queries in the class cj = 1 |Cj| X q∈Cj q q where cj is the centroid for class Cj; q iterates over the queries in a particular class. The classification is based on the cosine of the angle between the document d and the centroid meta-documents: Cmax = arg max Cj ∈C cj cj · d d = arg max Cj ∈C P i∈|F | ci j· di qP i∈|F |(ci j)2 qP i∈|F |(di)2 where F is the set of features. The score is normalized by the document and class length to produce comparable score. The terms ci and di represent the weight of the ith feature in the class centroid and the document respectively. These weights are based on the standard tf-idf formula. As the score of the max class is normalized with regard to document length, the scores for different documents are comparable. We conducted tests using professional editors to judge the quality of page and ad class assignments. The tests showed that for both ads and pages the Rocchio classifier returned the best results, especially on the page side. This is probably a result of the noise induced by using a search engine to machine generate training pages for the SVM and logregression classifiers. It is an area of current investigation how to improve the classification using a noisy training set. Based on the test results we decided to use the Rocchio"s classifier on both the ad and the page side. 5. SEMANTIC-SYNTACTIC MATCHING Contextual advertising systems process the content of the page, extract features, and then search the ad space to find the best matching ads. Given a page p and a set of ads A = {a1 . . . as} we estimate the relative probability of click P(click|p, a) with a score that captures the quality of the match between the page and the ad. To find the best ads for a page we rank the ads in A and select the top few for display. The problem can be formally defined as matching every page in the set of all pages P = {p1, . . . ppc} to one or more ads in the set of ads. Each page is represented as a set of page sections pi = {pi,1, pi,2 . . . pi,m}. The sections of the page represent different structural parts, such as title, metadata, body, headings, etc. In turn, each section is an unordered bag of terms (keywords). A page is represented by the union of the terms in each section: pi = {pws1 1 , pws1 2 . . . pwsi m} where pw stands for a page word and the superscript indicates the section of each term. A term can be a unigram or a phrase extracted by a phrase extractor [12]. Similarly, we represent each ad as a set of sections a = {a1, a2, . . . al}, each section in turn being an unordered set of terms: ai = {aws1 1 , aws1 2 . . . awsj l } where aw is an ad word. The ads in our experiments have 3 sections: title, body, and bid phrase. In this work, to produce the match score we use only the ad/page textual information, leaving user information and other data for future work. Next, each page and ad term is associated with a weight based on the tf-idf values. The tf value is determined based on the individual ad sections. There are several choices for the value of idf, based on different scopes. On the ad side, it has been shown in previous work that the set of ads of one campaign provide good scope for the estimation of idf that leads to improved matching results [8]. However, in this work for simplicity we do not take into account campaigns. To combine the impact of the term"s section and its tf-idf score, the ad/page term weight is defined as: tWeight(kwsi ) = weightSection(Si) · tf idf(kw) where tWeight stands for term weight and weightSection(Si) is the weight assigned to a page or ad section. This weight is a fixed parameter determined by experimentation. Each ad and page is classified into the topical taxonomy. We define these two mappings: Tax(pi) = {pci1, . . . pciu} Tax(aj) = {acj1 . . . acjv} where pc and ac are page and ad classes correspondingly. Each assignment is associated with a weight given by the classifier. The weights are normalized to sum to 1: X c∈T ax(xi) cWeight(c) = 1 where xi is either a page or an ad, and cWeights(c) is the class weight - normalized confidence assigned by the classifier. The number of classes can vary between different pages and ads. This corresponds to the number of topics a page/ad can be associated with and is almost always in the range 1-4. Now we define the relevance score of an ad ai and page pi as a convex combination of the keyword (syntactic) and classification (semantic) score: Score(pi, ai) = α · TaxScore(Tax(pi), Tax(ai)) +(1 − α) · KeywordScore(pi, ai) The parameter α determines the relative weight of the taxonomy score and the keyword score. To calculate the keyword score we use the vector space model [1] where both the pages and ads are represented in n-dimensional space - one dimension for each distinct term. The magnitude of each dimension is determined by the tWeight() formula. The keyword score is then defined as the cosine of the angle between the page and the ad vectors: KeywordScore(pi, ai) = P i∈|K| tWeight(pwi)· tWeight(awi) qP i∈|K|(tWeight(pwi))2 qP i∈|K|(tWeight(awi))2 where K is the set of all the keywords. The formula assumes independence between the words in the pages and ads. Furthermore, it ignores the order and the proximity of the terms in the scoring. We experimented with the impact of phrases and proximity on the keyword score and did not see a substantial impact of these two factors. We now turn to the definition of the TaxScore. This function indicates the topical match between a given ad and a page. As opposed to the keywords that are treated as independent dimensions, here the classes (topics) are organized into a hierarchy. One of the goals in the design of the TaxScore function is to be able to generalize within the taxonomy, that is accept topically related ads. Generalization can help to place ads in cases when there is no ad that matches both the category and the keywords of the page. The example in Figure 2 illustrates this situation. In this example, in the absence of ski ads, a page about skiing containing the word Atomic could be matched to the available snowboarding ad for the same brand. In general we would like the match to be stronger when both the ad and the page are classified into the same node Figure 2: Two generalization paths and weaker when the distance between the nodes in the taxonomy gets larger. There are multiple ways to specify the distance between two taxonomy nodes. Besides the above requirement, this function should lend itself to an efficient search of the ad space. Given a page we have to find the ad in a few milliseconds, as this impacts the presentation to a waiting user. This will be further discussed in the next section. To capture both the generalization and efficiency requirements we define the TaxScore function as follows: TaxScore(PC, AC) = X pc∈P C X ac∈AC idist(LCA(pc, ac), ac)·cWeight(pc)·cWeight(ac) In this function we consider every combination of page class and ad class. For each combination we multiply the product of the class weights with the inverse distance function between the least common ancestor of the two classes (LCA) and the ad class. The inverse distance function idist(c1, c2) takes two nodes on the same path in the class taxonomy and returns a number in the interval [0, 1] depending on the distance of the two class nodes. It returns 1 if the two nodes are the same, and declines toward 0 when LCA(pc, ac) or ac is the root of the taxonomy. The rate of decline determines the weight of the generalization versus the other terms in the scoring formula. To determine the rate of decline we consider the impact on the specificity of the match when we substitute a class with one of its ancestors. In general the impact is topic dependent. For example the node Hobby in our taxonomy has tens of children, each representing a different hobby, two examples being Sailing and Knitting. Placing an ad about Knitting on a page about Sailing does not make lots of sense. However, in the Winter Sports example above, in the absence of better alternative, skiing ads could be put on snowboarding pages as they might promote the same venues, equipment vendors etc. Such detailed analysis on a case by case basis is prohibitively expensive due to the size of the taxonomy. One option is to use the confidences of the ancestor classes as given by the classifier. However we found these numbers not suitable for this purpose as the magnitude of the confidences does not necessarily decrease when going up the tree. Another option is to use explore-exploit methods based on machine-learning from the click feedback as described in [10]. However for simplicity, in this work we chose a simple heuristic to determine the cost of generalization from a child to a parent. In this heuristic we look at the broadening of the scope when moving from a child to a parent. We estimate the broadening by the density of ads classified in the parent nodes vs the child node. The density is obtained by classifying a large set of ads in the taxonomy using the document classifier described above. Based on this, let nc be the number of document classified into the subtree rooted at c. Then we define: idist(c, p) = nc np where c represents the child node and p is the parent node. This fraction can be viewed as a probability of an ad belonging to the parent topic being suitable for the child topic. 6. SEARCHING THE AD SPACE In the previous section we discussed the choice of scoring function to estimate the match between an ad and a page. The top-k ads with the highest score are offered by the system for placement on the publisher"s page. The process of score calculation and ad selection is to be done in real time and therefore must be very efficient. As the ad collections are in the range of hundreds of millions of entries, there is a need for indexed access to the ads. Inverted indexes provide scalable and low latency solutions for searching documents. However, these have been traditionally used to search based on keywords. To be able to search the ads on a combination of keywords and classes we have mapped the classification match to term match and adapted the scoring function to be suitable for fast evaluation over inverted indexes. In this section we overview the ad indexing and the ranking function of our prototype ad search system for matching ads and pages. We used a standard inverted index framework where there is one posting list for each distinct term. The ads are parsed into terms and each term is associated with a weight based on the section in which it appears. Weights from distinct occurrences of a term in an ad are added together, so that the posting lists contain one entry per term/ad combination. The next challenge is how to index the ads so that the class information is preserved in the index? A simple method is to create unique meta-terms for the classes and annotate each ad with one meta-term for each assigned class. However this method does not allow for generalization, since only the ads matching an exact label of the page would be selected. Therefore we annotated each ad with one meta-term for each ancestor of the assigned class. The weights of meta-terms are assigned according to the value of the idist() function defined in the previous section. On the query side, given the keywords and the class of a page, we compose a keyword only query by inserting one class term for each ancestor of the classes assigned to the page. The scoring function is adapted to the two part scoreone for the class meta-terms and another for the text term. During the class score calculation, for each class path we use only the lowest class meta-term, ignoring the others. For example, if an ad belongs to the Skiing class and is annotated with both Skiing and its parent Winter Sports, the index will contain the special class meta-terms for both Skiing and Winter Sports (and all their ancestors) with the weights according to the product of the classifier confidence and the idist function. When matching with a page classified into Skiing, the query will contain terms for class Skiing and for each of its ancestors. However when scoring an ad classified into Skiing we will use the weight for the Skiing class meta-term. Ads classified into Snowboarding will be scored using the weight of the Winter Sports meta-term. To make this check efficiently we keep a sorted list of all the class paths and, at scoring time, we search the paths bottom up for a meta-term appearing in the ad. The first meta-term is used for scoring, the rest are ignored. At runtime, we evaluate the query using a variant of the WAND algorithm [3]. This is a document-at-a-time algorithm [1] that uses a branch-and-bound approach to derive efficient moves for the cursors associated to the postings lists. It finds the next cursor to be moved based on an upper bound of the score for the documents at which the cursors are currently positioned. The algorithm keeps a heap of current best candidates. Documents with an upper bound smaller than the current minimum score among the candidate documents can be eliminated from further considerations, and thus the cursors can skip over them. To find the upper bound for a document, the algorithm assumes that all cursors that are before it will hit this document (i.e. the document contains all those terms represented by cursors before or at that document). It has been shown that WAND can be used with any function that is monotonic with respect to the number of matching terms in the document. Our scoring function is monotonic since the score can never decrease when more terms are found in the document. In the special case when we add a cursor representing an ancestor of a class term already factored in the score, the score simply does not change (we add 0). Given these properties, we use an adaptation of the WAND algorithm where we change the calculation of the scoring function and the upper bound score calculation to reflect our scoring function. The rest of the algorithm remains unchanged. 7. EXPERIMENTAL EVALUATION 7.1 Data and Methodology We quantify the effect of the semantic-syntactic matching using a set of 105 pages. This set of pages was selected by a random sample of a larger set of around 20 million pages with contextual advertising. Ads for each of these pages have been selected from a larger pool of ads (tens of millions) by previous experiments conducted by Yahoo! US for other purposes. Each page-ad pair has been judged by three or more human judges on a 1 to 3 scale: 1. Relevant The ad is semantically directly related to the main subject of the page. For example if the page is about the National Football League and the ad is about tickets for NFL games, it would be scored as 1. 2. Somewhat relevant The ad is related to the secondary subject of the page, or is related to the main topic of the page in a general way. In the NFL page example, an ad about NFL branded products would be judged as 2. 3. Irrelevant The ad is unrelated to the page. For example a mention of the NFL player John Maytag triggers washing machine ads on a NFL page. pages 105 words per page 868 judgments 2946 judg. inter-editor agreement 84% unique ads 2680 unique ads per page 25.5 page classification precision 70% ad classification precision 86% Table 1: Dataset statistics To obtain a score for a page-ad pair we average all the scores and then round to the closest integer. We then used these judgments to evaluate how well our methods distinguish the positive and the negative ad assignments for each page. The statistics of the page dataset is given in Table 1. The original experiments that paired the pages and the ads are loosely related to the syntactic keyword based matching and classification based assignment but used different taxonomies and keyword extraction techniques. Therefore we could not use standard pooling as an evaluation method since we did not control the way the pairs were selected and could not precisely establish the set of ads from which the placed ads were selected. Instead, in our evaluation for each page we consider only those ads for which we have judgment. Each different method was applied to this set and the ads were ranked by the score. The relative effectiveness of the algorithms were judged by comparing how well the methods separated the ads with positive judgment from the ads with negative judgment. We present precision on various levels of recall within this set. As the set of ads per page is relatively small, the evaluation reports precision that is higher than it would be with a larger set of negative ads. However, these numbers still establish the relative performance of the algorithms and we can use it to evaluate performance at different score thresholds. In addition to the precision-recall over the judged ads, we also present Kendall"s τ rank correlation coefficient to establish how far from the perfect ordering are the orderings produced by our ranking algorithms. For this test we ranked the judged ads by the scores assigned by the judges and then compared this order to the order assigned by our algorithms. Finally we also examined the precision at position 1, 3 and 5. 7.2 Results Figure 3 shows the precision recall curves for the syntactic matching (keywords only used) vs. a syntactic-semantic matching with the optimal value of α = 0.8 (judged by the 11-point score [1]). In this figure, we assume that the adpage pairs judged with 1 or 2 are positive examples and the 3s are negative examples. We also examined counting only the pairs judged with 1 as positive examples and did not find a significant change in the relative performance of the tested methods. Overlaid are also results using phrases in the keyword match. We see that these additional features do not change the relative performance of the algorithm. The graphs show significant impact of the class information, especially in the mid range of recall values. In the low recall part of the chart the curves meet. This indicates that when the keyword match is really strong (i.e. when the ad is almost contained within the page) the precision 0.4 0.5 0.6 0.7 0.8 0.9 1 0 0.2 0.4 0.6 0.8 1 Recall Precision Alpha=.9, no phrase Alpha=0, no phrase Alpha=0, phrase Alpha=.9, phrase Figure 3: Data Set 2: Precision vs. Recall of syntactic match (α = 0) vs. syntactic-semantic match (α = 0.8) α Kendall"s τ α = 0 0.086 α = 0.25 0.155 α = 0.50 0.166 α = 0.75 0.158 α = 1 0.136 Table 2: Kendall"s τ for different α values of the syntactic keyword match is comparable with that of the semantic-syntactic match. Note however that the two methods might produce different ads and could be used as a complement at level of recall. The semantic components provides largest lift in precision at the mid range of recall where 25% improvement is achieved by using the class information for ad placement. This means that when there is somewhat of a match between the ad and the page, the restriction to the right classes provides a better scope for selecting the ads. Table 2 shows the Kendall"s τ values for different values of α. We calculated the values by ranking all the judged ads for each page and averaging the values over all the pages. The ads with tied judgment were given the rank of the middle position. The results show that when we take into account all the ad-page pairs, the optimal value of α is around 0.5. Note that purely syntactic match (α = 0) is by far the weakest method. Figure 4 shows the effect of the parameter α in the scoring. We see that in most cases precision grows or is flat when we increase α, except at the low level of recall where due to small number of data points there is a bit of jitter in the results. Table 3 shows the precision at positions 1, 3 and 5. Again, the purely syntactic method has clearly the lowest score by individual positions and the total number of correctly placed ads. The numbers are close due to the small number of the ads considered, but there are still some noticeable trends. For position 1 the optimal α is in the range of 0.25 to 0.75. For positions 3 and 5 the optimum is at α = 1. This also indicates that for those ads that have a very high keyword score, the semantic information is somewhat less important. If almost all the words in an ad appear in the page, this ad is Precision Vs Alpha for Different Levels of Recall (Data Set 2) 0.45 0.55 0.65 0.75 0.85 0.95 0 0.2 0.4 0.6 0.8 1 Alpha Precision 80% Recall 60% Recall 40% Recall 20% Recall Figure 4: Impact of α on precision for different levels of recall α #1 #3 #5 sum α = 0 80 70 68 218 α = 0.25 89 76 73 238 α = 0.5 89 74 73 236 α = 0.75 89 78 73 240 α = 1 86 79 74 239 Table 3: Precision at position 1, 3 and 5 likely to be relevant for this page. However when there is no such clear affinity, the class information becomes a dominant factor. 8. CONCLUSION Contextual advertising is the economic engine behind a large number of non-transactional sites on the Web. Studies have shown that one of the main success factors for contextual ads is their relevance to the surrounding content. All existing commercial contextual match solutions known to us evolved from search advertising solutions whereby a search query is matched to the bid phrase of the ads. A natural extension of search advertising is to extract phrases from the page and match them to the bid phrase of the ads. However, individual phrases and words might have multiple meanings and/or be unrelated to the overall topic of the page leading to miss-matched ads. In this paper we proposed a novel way of matching advertisements to web pages that rely on a topical (semantic) match as a major component of the relevance score. The semantic match relies on the classification of pages and ads into a 6000 nodes commercial advertising taxonomy to determine their topical distance. As the classification relies on the full content of the page, it is more robust than individual page phrases. The semantic match is complemented with a syntactic match and the final score is a convex combination of the two sub-scores with the relative weight of each determined by a parameter α. We evaluated the semantic-syntactic approach against a syntactic approach over a set of pages with different contextual advertising. As shown in our experimental evaluation, the optimal value of the parameter α depends on the precise objective of optimization (precision at particular position, precision at given recall). However in all cases the optimal value of α is between 0.25 and 0.9 indicating significant effect of the semantic score component. The effectiveness of the syntactic match depends on the quality of the pages used. In lower quality pages we are more likely to make classification errors that will then negatively impact the matching. We demonstrated that it is feasible to build a large scale classifier that has sufficient good precision for this application. We are currently examining how to employ machine learning algorithms to learn the optimal value of α based on a collection of features of the input pages. 9. REFERENCES [1] R. Baeza-Yates and B. Ribeiro-Neto. Modern Information Retrieval. ACM, 1999. [2] Bernhard E. Boser, Isabelle Guyon, and Vladimir Vapnik. A training algorithm for optimal margin classifiers. In Computational Learning Theory, pages 144-152, 1992. [3] A. Z. Broder, D. Carmel, M. Herscovici, A. Soffer, and J. Zien. Efficient query evaluation using a two-level retrieval process. In CIKM "03: Proc. of the twelfth intl. conf. on Information and knowledge management, pages 426-434, New York, NY, 2003. ACM. [4] P. Chatterjee, D. L. Hoffman, and T. P. Novak. Modeling the clickstream: Implications for web-based advertising efforts. Marketing Science, 22(4):520-541, 2003. [5] D. Fain and J. Pedersen. Sponsored search: A brief history. In In Proc. of the Second Workshop on Sponsored Search Auctions, 2006. Web publication, 2006. [6] S. C. Gates, W. Teiken, and K.-Shin F. Cheng. Taxonomies by the numbers: building high-performance taxonomies. In CIKM "05: Proc. of the 14th ACM intl. conf. on Information and knowledge management, pages 568-577, New York, NY, 2005. ACM. [7] A. Lacerda, M. Cristo, M. Andre; G., W. Fan, N. Ziviani, and B. Ribeiro-Neto. Learning to advertise. In SIGIR "06: Proc. of the 29th annual intl. ACM SIGIR conf., pages 549-556, New York, NY, 2006. ACM. [8] B. Ribeiro-Neto, M. Cristo, P. B. Golgher, and E. S. de Moura. Impedance coupling in content-targeted advertising. In SIGIR "05: Proc. of the 28th annual intl. ACM SIGIR conf., pages 496-503, New York, NY, 2005. ACM. [9] J. Rocchio. Relevance feedback in information retrieval. In The SMART Retrieval System: Experiments in Automatic Document Processing, pages 313-323. PrenticeHall, 1971. [10] P. Sandeep, D. Agarwal, D. Chakrabarti, and V. Josifovski. Bandits for taxonomies: A model-based approach. In In Proc. of the SIAM intl. conf. on Data Mining, 2007. [11] T. Santner and D. Duffy. The Statistical Analysis of Discrete Data. Springer-Verlag, 1989. [12] R. Stata, K. Bharat, and F. Maghoul. The term vector database: fast access to indexing terms for web pages. Computer Networks, 33(1-6):247-255, 2000. [13] C. Wang, P. Zhang, R. Choi, and M. D. Eredita. Understanding consumers attitude toward advertising. In Eighth Americas conf. on Information System, pages 1143-1148, 2002. [14] W. Yih, J. Goodman, and V. R. Carvalho. Finding advertising keywords on web pages. In WWW "06: Proc. of the 15th intl. conf. on World Wide Web, pages 213-222, New York, NY, 2006. ACM.

pay-per-click;semantics;contextual advertising;matching mechanism;semantic-syntactic matching;match;keyword matching;document classifier;ad relevance;contextual advertise;topical distance;top-k ad;hierarchical taxonomy class

train_H-48

A New Approach for Evaluating Query Expansion: Query-Document Term Mismatch

The effectiveness of information retrieval (IR) systems is influenced by the degree of term overlap between user queries and relevant documents. Query-document term mismatch, whether partial or total, is a fact that must be dealt with by IR systems. Query Expansion (QE) is one method for dealing with term mismatch. IR systems implementing query expansion are typically evaluated by executing each query twice, with and without query expansion, and then comparing the two result sets. While this measures an overall change in performance, it does not directly measure the effectiveness of IR systems in overcoming the inherent issue of term mismatch between the query and relevant documents, nor does it provide any insight into how such systems would behave in the presence of query-document term mismatch. In this paper, we propose a new approach for evaluating query expansion techniques. The proposed approach is attractive because it provides an estimate of system performance under varying degrees of query-document term mismatch, it makes use of readily available test collections, and it does not require any additional relevance judgments or any form of manual processing.

1. INTRODUCTION In our domain,1 and unlike web search, it is very important for attorneys to find all documents (e.g., cases) that are relevant to an issue. Missing relevant documents may have non-trivial consequences on the outcome of a court proceeding. Attorneys are especially concerned about missing relevant documents when researching a legal topic that is new to them, as they may not be aware of all language variations in such topics. Therefore, it is important to develop information retrieval systems that are robust with respect to language variations or term mismatch between queries and relevant documents. During our work on developing such systems, we concluded that current evaluation methods are not sufficient for this purpose. {Whooping cough, pertussis}, {heart attack, myocardial infarction}, {car wash, automobile cleaning}, {attorney, legal counsel, lawyer} are all examples of things that share the same meaning. Often, the terms chosen by users in their queries are different than those appearing in the documents relevant to their information needs. This query-document term mismatch arises from two sources: (1) the synonymy found in natural language, both at the term and the phrasal level, and (2) the degree to which the user is an expert at searching and/or has expert knowledge in the domain of the collection being searched. IR evaluations are comparative in nature (cf. TREC). Generally, IR evaluations show how System A did in relation to System B on the same test collection based on various precision- and recall-based metrics. Similarly, IR systems with QE capabilities are typically evaluated by executing each search twice, once with and once without query expansion, and then comparing the two result sets. While this approach shows which system may have performed better overall with respect to a particular test collection, it does not directly or systematically measure the effectiveness of IR systems in overcoming query-document term mismatch. If the goal of QE is to increase search performance by mitigating the effects of query-document term mismatch, then the degree to which a system does so should be measurable in evaluation. An effective evaluation method should measure the performance of IR systems under varying degrees of query-document term mismatch, not just in terms of overall performance on a collection relative to another system. 1 Thomson Corporation builds information based solutions to the professional markets including legal, financial, health care, scientific, and tax and accounting. In order to measure that a particular IR system is able to overcome query-document term mismatch by retrieving documents that are relevant to a user"s query, but that do not necessarily contain the query terms themselves, we systematically introduce term mismatch into the test collection by removing query terms from known relevant documents. Because we are purposely inducing term mismatch between the queries and known relevant documents in our test collections, the proposed evaluation framework is able to measure the effectiveness of QE in a way that testing on the whole collection is not. If a QE search method finds a document that is known to be relevant but that is nonetheless missing query terms, it shows that QE technique is indeed robust with respect to query-document term mismatch. 2. RELATED WORK Accounting for term mismatch between the terms in user queries and the documents relevant to users" information needs has been a fundamental issue in IR research for almost 40 years [38, 37, 47]. Query expansion (QE) is one technique used in IR to improve search performance by increasing the likelihood of term overlap (either explicitly or implicitly) between queries and documents that are relevant to users" information needs. Explicit query expansion occurs at run-time, based on the initial search results, as is the case with relevance feedback and pseudo relevance feedback [34, 37]. Implicit query expansion can be based on statistical properties of the document collection, or it may rely on external knowledge sources such as a thesaurus or an ontology [32, 17, 26, 50, 51, 2]. Regardless of method, QE algorithms that are capable of retrieving relevant documents despite partial or total term mismatch between queries and relevant documents should increase the recall of IR systems (by retrieving documents that would have previously been missed) as well as their precision (by retrieving more relevant documents). In practice, QE tends to improve the average overall retrieval performance, doing so by improving performance on some queries while making it worse on others. QE techniques are judged as effective in the case that they help more than they hurt overall on a particular collection [47, 45, 41, 27]. Often, the expansion terms added to a query in the query expansion phase end up hurting the overall retrieval performance because they introduce semantic noise, causing the meaning of the query to drift. As such, much work has been done with respect to different strategies for choosing semantically relevant QE terms to include in order to avoid query drift [34, 50, 51, 18, 24, 29, 30, 32, 3, 4, 5]. The evaluation of IR systems has received much attention in the research community, both in terms of developing test collections for the evaluation of different systems [11, 12, 13, 43] and in terms of the utility of evaluation metrics such as recall, precision, mean average precision, precision at rank, Bpref, etc. [7, 8, 44, 14]. In addition, there have been comparative evaluations of different QE techniques on various test collections [47, 45, 41]. In addition, the IR research community has given attention to differences between the performance of individual queries. Research efforts have been made to predict which queries will be improved by QE and then selectively applying it only to those queries [1, 5, 27, 29, 15, 48], to achieve optimal overall performance. In addition, related work on predicting query difficulty, or which queries are likely to perform poorly, has been done [1, 4, 5, 9]. There is general interest in the research community to improve the robustness of IR systems by improving retrieval performance on difficult queries, as is evidenced by the Robust track in the TREC competitions and new evaluation measures such as GMAP. GMAP (geometric mean average precision) gives more weight to the lower end of the average precision (as opposed to MAP), thereby emphasizing the degree to which difficult or poorly performing queries contribute to the score [33]. However, no attention is given to evaluating the robustness of IR systems implementing QE with respect to querydocument term mismatch in quantifiable terms. By purposely inducing mismatch between the terms in queries and relevant documents, our evaluation framework allows us a controlled manner in which to degrade the quality of the queries with respect to their relevant documents, and then to measure the both the degree of (induced) difficulty of the query and the degree to which QE improves the retrieval performance of the degraded query. The work most similar to our own in the literature consists of work in which document collections or queries are altered in a systematic way to measure differences query performance. [42] introduces into the document collection pseudowords that are ambiguous with respect to word sense, in order to measure the degree to which word sense disambiguation is useful in IR. [6] experiments with altering the document collection by adding semantically related expansion terms to documents at indexing time. In cross-language IR, [28] explores different query expansion techniques while purposely degrading their translation resources, in what amounts to expanding a query with only a controlled percentage of its translation terms. Although similar in introducing a controlled amount of variance into their test collections, these works differ from the work being presented in this paper in that the work being presented here explicitly and systematically measures query effectiveness in the presence of query-document term mismatch. 3. METHODOLOGY In order to accurately measure IR system performance in the presence of query-term mismatch, we need to be able to adjust the degree of term mismatch in a test corpus in a principled manner. Our approach is to introduce querydocument term mismatch into a corpus in a controlled manner and then measure the performance of IR systems as the degree of term mismatch changes. We systematically remove query terms from known relevant documents, creating alternate versions of a test collection that differ only in how many or which query terms have been removed from the documents relevant to a particular query. Introducing query-document term mismatch into the test collection in this manner allows us to manipulate the degree of term mismatch between relevant documents and queries in a controlled manner. This removal process affects only the relevant documents in the search collection. The queries themselves remain unaltered. Query terms are removed from documents one by one, so the differences in IR system performance can be measured with respect to missing terms. In the most extreme case (i.e., when the length of the query is less than or equal to the number of query terms removed from the relevant documents), there will be no term overlap between a query and its relevant documents. Notice that, for a given query, only relevant documents are modified. Non-relevant documents are left unchanged, even in the case that they contain query terms. Although, on the surface, we are changing the distribution of terms between the relevant and non-relevant documents sets by removing query terms from the relevant documents, doing so does not change the conceptual relevancy of these documents. Systematically removing query terms from known relevant documents introduces a controlled amount of query-document term mismatch by which we can evaluate the degree to which particular QE techniques are able to retrieve conceptually relevant documents, despite a lack of actual term overlap. Removing a query term from relevant documents simply masks the presence of that query term in those documents. It does not in any way change the conceptual relevancy of the documents. The evaluation framework presented in this paper consists of three elements: a test collection, C; a strategy for selecting which query terms to remove from the relevant documents in that collection, S; and a metric by which to compare performance of the IR systems, m. The test collection, C, consists of a document collection, queries, and relevance judgments. The strategy, S, determines the order and manner in which query terms are removed from the relevant documents in C. This evaluation framework is not metric-specific; any metric (MAP, P@10, recall, etc.) can be used to measure IR system performance. Although test collections are difficult to come by, it should be noted that this evaluation framework can be used on any available test collection. In fact, using this framework stretches the value of existing test collections in that one collection becomes several when query terms are removed from relevant documents, thereby increasing the amount of information that can be gained from evaluating on a particular collection. In other evaluations of QE effectiveness, the controlled variable is simply whether or not queries have been expanded or not, compared in terms of some metric. In contrast, the controlled variable in this framework is the query term that has been removed from the documents relevant to that query, as determined by the removal strategy, S. Query terms are removed one by one, in a manner and order determined by S, so that collections differ only with respect to the one term that has been removed (or masked) in the documents relevant to that query. It is in this way that we can explicitly measure the degree to which an IR system overcomes query-document term mismatch. The choice of a query term removal strategy is relatively flexible; the only restriction in choosing a strategy S is that query terms must be removed one at a time. Two decisions must be made when choosing a removal strategy S. The first is the order in which S removes terms from the relevant documents. Possible orders for removal could be based on metrics such as IDF or the global probability of a term in a document collection. Based on the purpose of the evaluation and the retrieval algorithm being used, it might make more sense to choose a removal order for S based on query term IDF or perhaps based on a measure of query term probability in the document collection. Once an order for removal has been decided, a manner for term removal/masking must be decided. It must be determined if S will remove the terms individually (i.e., remove just one different term each time) or additively (i.e., remove one term first, then that term in addition to another, and so on). The incremental additive removal of query terms from relevant documents allows the evaluation to show the degree to which IR system performance degrades as more and more query terms are missing, thereby increasing the degree of query-document term mismatch. Removing terms individually allows for a clear comparison of the contribution of QE in the absence of each individual query term. 4. EXPERIMENTAL SET-UP 4.1 IR Systems We used the proposed evaluation framework to evaluate four IR systems on two test collections. Of the four systems used in the evaluation, two implement query expansion techniques: Okapi (with pseudo-feedback for QE), and a proprietary concept search engine (we"ll call it TCS, for Thomson Concept Search). TCS is a language modeling based retrieval engine that utilizes a subject-appropriate external corpus (i.e., legal or news) as a knowledge source. This external knowledge source is a corpus separate from, but thematically related to, the document collection to be searched. Translation probabilities for QE [2] are calculated from these large external corpora. Okapi (without feedback) and a language model query likelihood (QL) model (implemented using Indri) are included as keyword-only baselines. Okapi without feedback is intended as an analogous baseline for Okapi with feedback, and the QL model is intended as an appropriate baseline for TCS, as they both implement language-modeling based retrieval algorithms. We choose these as baselines because they are dependent only on the words appearing in the queries and have no QE capabilities. As a result, we expect that when query terms are removed from relevant documents, the performance of these systems should degrade more dramatically than their counterparts that implement QE. The Okapi and QL model results were obtained using the Lemur Toolkit.2 Okapi was run with the parameters k1=1.2, b=0.75, and k3=7. When run with feedback, the feedback parameters used in Okapi were set at 10 documents and 25 terms. The QL model used Jelinek-Mercer smoothing, with λ = 0.6. 4.2 Test Collections We evaluated the performance of the four IR systems outlined above on two different test collections. The two test collections used were the TREC AP89 collection (TIPSTER disk 1) and the FSupp Collection. The FSupp Collection is a proprietary collection of 11,953 case law documents for which we have 44 queries (ranging from four to twenty-two words after stop word removal) with full relevance judgments.3 The average length of documents in the FSupp Collection is 3444 words. 2 www.lemurproject.org 3 Each of the 11,953 documents was evaluated by domain experts with respect to each of the 44 queries. The TREC AP89 test collection contains 84,678 documents, averaging 252 words in length. In our evaluation, we used both the title and the description fields of topics 151200 as queries, so we have two sets of results for the AP89 Collection. After stop word removal, the title queries range from two to eleven words and the description queries range from four to twenty-six terms. 4.3 Query Term Removal Strategy In our experiments, we chose to sequentially and additively remove query terms from highest-to-lowest inverse document frequency (IDF) with respect to the entire document collection. Terms with high IDF values tend to influence document ranking more than those with lower IDF values. Additionally, high IDF terms tend to be domainspecific terms that are less likely to be known to non-expert user, hence we start by removing these first. For the FSupp Collection, queries were evaluated incrementally with one, two, three, five, and seven terms removed from their corresponding relevant documents. The longer description queries from TREC topics 151-200 were likewise evaluated on the AP89 Collection with one, two, three, five, and seven query terms removed from their relevant documents. For the shorter TREC title queries, we removed one, two, three, and five terms from the relevant documents. 4.4 Metrics In this implementation of the evaluation framework, we chose three metrics by which to compare IR system performance: mean average precision (MAP), precision at 10 documents (P10), and recall at 1000 documents. Although these are the metrics we chose to demonstrate this framework, any appropriate IR metrics could be used within the framework. 5. RESULTS 5.1 FSupp Collection Figures 1, 2, and 3 show the performance (in terms of MAP, P10 and Recall, respectively) for the four search engines on the FSupp Collection. As expected, the performance of the keyword-only IR systems, QL and Okapi, drops quickly as query terms are removed from the relevant documents in the collection. The performance of Okapi with feedback (Okapi FB) is somewhat surprising in that on the original collection (i.e., prior to query term removal), its performance is worse than that of Okapi without feedback on all three measures. TCS outperforms the QL keyword baseline on every measure except for MAP on the original collection (i.e., prior to removing any query terms). Because TCS employs implicit query expansion using an external domain specific knowledge base, it is less sensitive to term removal (i.e., mismatch) than the Okapi FB, which relies on terms from the top-ranked documents retrieved by an initial keywordonly search. Because overall search engine performance is frequently measured in terms of MAP, and because other evaluations of QE often only consider performance on the entire collection (i.e., they do not consider term mismatch), the QE implemented in TCS would be considered (in an0 1 2 3 4 5 6 7 Number of Query Terms Removed from Relevant Documents 0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 MeanAveragePrecision(MAP) Okapi FB Okapi TCS QL FSupp: Mean Average Precision with Query Terms Removed Figure 1: The performance of the four retrieval systems on the FSupp collection in terms of Mean Average Precision (MAP) and as a function of the number of query terms removed (the horizontal axis). 0 1 2 3 4 5 6 7 Number of Query Terms Removed from Relevant Documents 0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5 0.55 0.6 Precisionat10Documents(P10) Okapi FB Okapi TCS QL FSupp: P10 with Query Terms Removed Figure 2: The performance of the four retrieval systems on the FSupp collection in terms of Precision at 10 and as a function of the number of query terms removed (the horizontal axis). other evaluation) to hurt performance on the FSupp Collection. However, when we look at the comparison of TCS to QL when query terms are removed from the relevant documents, we can see that the QE in TCS is indeed contributing positively to the search. 5.2 The AP89 Collection: using the description queries Figures 4, 5, and 6 show the performance of the four IR systems on the AP89 Collection, using the TREC topic descriptions as queries. The most interesting difference between the performance on the FSupp Collection and the AP89 collection is the reversal of Okapi FB and TCS. On FSupp, TCS outperformed the other engines consistently (see Figures 1, 2, and 3); on the AP89 Collection, Okapi FB is clearly the best performer (see Figures 4, 5, and 6). This is all the more interesting, based on the fact that QE in Okapi FB takes place after the first search iteration, which 0 1 2 3 4 5 6 7 Number of Query Terms Removed from Relevant Documents 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 Recall Okapi FB Okapi TCS Indri FSupp: Recall at 1000 documents with Query Terms Removed Figure 3: The Recall (at 1000) of the four retrieval systems on the FSupp collection as a function of the number of query terms removed (the horizontal axis). 0 1 2 3 4 5 6 7 Number of Query Terms Removed from Relevant Documents 0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 MeanAveragePrecision(MAP) Okapi FB Okapi TCS QL AP89: Mean Average Precision with Query Terms Removed (description queries) Figure 4: MAP of the four IR systems on the AP89 Collection, using TREC description queries. MAP is measured as a function of the number of query terms removed. 0 1 2 3 4 5 6 7 Number of Query Terms Removed from Relevant Documents 0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 Precisionat10Documents(P10) Okapi FB Okapi TCS QL AP89: P10 with Query Terms Removed (description queries) Figure 5: Precision at 10 of the four IR systems on the AP89 Collection, using TREC description queries. P at 10 is measured as a function of the number of query terms removed. 0 1 2 3 4 5 6 7 Number of Query Terms Removed from Relevant Documents 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 Recall Okapi FB Okapi TCS QL AP89: Recall at 1000 documents with Query Terms Removed (description queries) Figure 6: Recall (at 1000) of the four IR systems on the AP89 Collection, using TREC description queries, and as a function of the number of query terms removed. we would expect to be handicapped when query terms are removed. Looking at P10 in Figure 5, we can see that TCS and Okapi FB score similarly on P10, starting at the point where one query term is removed from relevant documents. At two query terms removed, TCS starts outperforming Okapi FB. If modeling this in terms of expert versus non-expert users, we could conclude that TCS might be a better search engine for non-experts to use on the AP89 Collection, while Okapi FB would be best for an expert searcher. It is interesting to note that on each metric for the AP89 description queries, TCS performs more poorly than all the other systems on the original collection, but quickly surpasses the baseline systems and approaches Okapi FB"s performance as terms are removed. This is again a case where the performance of a system on the entire collection is not necessarily indicative of how it handles query-document term mismatch. 5.3 The AP89 Collection: using the title queries Figures 7, 8, and 9 show the performance of the four IR systems on the AP89 Collection, using the TREC topic titles as queries. As with the AP89 description queries, Okapi FB is again the best performer of the four systems in the evaluation. As before, the performance of the Okapi and QL systems, the non-QE baseline systems, sharply degrades as query terms are removed. On the shorter queries, TCS seems to have a harder time catching up to the performance of Okapi FB as terms are removed. Perhaps the most interesting result from our evaluation is that although the keyword-only baselines performed consistently and as expected on both collections with respect to query term removal from relevant documents, the performances of the engines implementing QE techniques differed dramatically between collections. 0 1 2 3 4 5 Number of Query Terms Removed from Relevant Documents 0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 MeanAveragePrecision(MAP) Okapi FB Okapi TCS QL AP89: Mean Average Precision with Query Terms Removed (title queries) Figure 7: MAP of the four IR systems on the AP89 Collection, using TREC title queries and as a function of the number of query terms removed. 0 1 2 3 4 5 Number of Query Terms Removed from Relevant Documents 0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5 Precisionat10Documents(P10) Okapi FB Okapi TCS QL AP89: P10 with Query Terms Removed (title queries) Figure 8: Precision at 10 of the four IR systems on the AP89 Collection, using TREC title queries, and as a function of the number of query terms removed. 0 1 2 3 4 5 Number of Query Terms Removed from Relevant Documents 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 Recall Okapi FB Okapi TCS QL AP89: Recall at 1000 documents with Query Terms Removed (title queries) Figure 9: Recall (at 1000) of the four IR systems on the AP89 Collection, using TREC title queries and as a function of the number of query terms removed. 6. DISCUSSION The intuition behind this evaluation framework is to measure the degree to which various QE techniques overcome term mismatch between queries and relevant documents. In general, it is easy to evaluate the overall performance of different techniques for QE in comparison to each other or against a non-QE variant on any complete test collection. Such an approach does tell us which systems perform better on a complete test collection, but it does not measure the ability of a particular QE technique to retrieve relevant documents despite partial or complete term mismatch between queries and relevant documents. A systematic evaluation of IR systems as outlined in this paper is useful not only with respect to measuring the general success or failure of particular QE techniques in the presence of query-document term mismatch, but it also provides insight into how a particular IR system will perform when used by expert versus non-expert users on a particular collection. The less a user knows about the domain of the document collection on which they are searching, the more prevalent query-document term mismatch is likely to be. This distinction is especially relevant in the case that the test collection is domain-specific (i.e., medical or legal, as opposed to a more general domain, such as news), where the distinction between experts and non-experts may be more marked. For example, a non-expert in the medical domain might search for whooping cough, but relevant documents might instead contain the medical term pertussis. Since query terms are masked only the in relevant documents, this evaluation framework is actually biased against retrieving relevant documents. This is because non-relevant documents may also contain query terms, which can cause a retrieval system to rank such documents higher than it would have before terms were masked in relevant documents. Still, we think this is a more realistic scenario than removing terms from all documents regardless of relevance. The degree to which a QE technique is well-suited to a particular collection can be evaluated in terms of its ability to still find the relevant documents, even when they are missing query terms, despite the bias of this approach against relevant documents. However, given that Okapi FB and TCS outperformed each other on two different collection sets, further investigation into the degree of compatibility between QE expansion approach and target collection is probably warranted. Furthermore, the investigation of other term removal strategies could provide insight into the behavior of different QE techniques and their overall impact on the user experience. As mentioned earlier, our choice of the term removal strategy was motivated by (1) our desire to see the highest impact on system performance as terms are removed and (2) because high IDF terms, in our domain context, are more likely to be domain specific, which allows us to better understand the performance of an IR system as experienced by expert and non-expert users. Although not attempted in our experiments, another application of this evaluation framework would be to remove query terms individually, rather than incrementally, to analyze which terms (or possibly which types of terms) are being helped most by a QE technique on a particular test collection. This could lead to insight as to when QE should and should not be applied. This evaluation framework allows us to see how IR systems perform in the presence of query-document term mismatch. In other evaluations, the performance of a system is measured only on the entire collection, in which the degree of query-term document mismatch is not known. By systematically introducing this mismatch, we can see that even if an IR system is not the best performer on the entire collection, its performance may nonetheless be more robust to query-document term mismatch than other systems. Such robustness makes a system more user-friendly, especially to non-expert users. This paper presents a novel framework for IR system evaluation, the applications of which are numerous. The results presented in this paper are not by any means meant to be exhaustive or entirely representative of the ways in which this evaluation could be applied. To be sure, there is much future work that could be done using this framework. In addition to looking at average performance of IR systems, the results of individual queries could be examined and compared more closely, perhaps giving more insight into the classification and prediction of difficult queries, or perhaps showing which QE techniques improve (or degrade) individual query performance under differing degrees of querydocument term mismatch. Indeed, this framework would also benefit from further testing on a larger collection. 7. CONCLUSION The proposed evaluation framework allows us to measure the degree to which different IR systems overcome (or don"t overcome) term mismatch between queries and relevant documents. Evaluations of IR systems employing QE performed only on the entire collection do not take into account that the purpose of QE is to mitigate the effects of term mismatch in retrieval. By systematically removing query terms from relevant documents, we can measure the degree to which QE contributes to a search by showing the difference between the performances of a QE system and its keywordonly baseline when query terms have been removed from known relevant documents. Further, we can model the behavior of expert versus non-expert users by manipulating the amount of query-document term mismatch introduced into the collection. The evaluation framework proposed in this paper is attractive for several reasons. Most importantly, it provides a controlled manner in which to measure the performance of QE with respect to query-document term mismatch. In addition, this framework takes advantage and stretches the amount of information we can get from existing test collections. Further, this evaluation framework is not metricspecific: information in terms of any metric (MAP, P@10, etc.) can be gained from evaluating an IR system this way. It should also be noted that this framework is generalizable to any IR system, in that it evaluates how well IR systems evaluate users" information needs as represented by their queries. An IR system that is easy to use should be good at retrieving documents that are relevant to users" information needs, even if the queries provided by the users do not contain the same keywords as the relevant documents. 8. REFERENCES [1] Amati, G., C. Carpineto, and G. Romano. Query difficulty, robustness and selective application of query expansion. In Proceedings of the 25th European Conference on Information Retrieval (ECIR 2004), pp. 127-137. [2] Berger, A. and J.D. Lafferty. 1999. Information retrieval as statistical translation. In Research and Development in Information Retrieval, pages 222-229. [3] Billerbeck, B., F. Scholer, H. E. Williams, and J. Zobel. 2003. Query expansion using associated queries. In Proceedings of CIKM 2003, pp. 2-9. [4] Billerbeck, B., and J. Zobel. 2003. When Query Expansion Fails. In Proceedings of SIGIR 2003, pp. 387-388. [5] Billerbeck, B. and J. Zobel. 2004. Questioning Query Expansion: An Examination of Behaviour and Parameters. In Proceedings of the 15th Australasian Database Conference (ADC2004), pp. 69-76. [6] Billerbeck, B. and J. Zobel. 2005. Document Expansion versus Query Expansion for ad-hoc Retrieval. In Proceedings of the 10th Australasian Document Computing Symposium. [7] Buckley, C. and E.M. Voorhees. 2000. Evaluating Evaluation Measure Stability. In Proceedings of SIGIR 2000, pp. 33-40. [8] Buckley, C. and E.M. Voorhees. 2004. Retrieval evaluation with incomplete information. In Proceedings of SIGIR 2004, pp. 25-32. [9] Carmel, D., E. Yom-Tov, A. Darlow, D. Pelleg. 2006. What Makes A Query Difficult? In Proceedings of SIGIR 2006, pp. 390-397. [10] Carpineto, C., R. Mori and G. Romano. 1998. Informative Term Selection for Automatic Query Expansion. In The 7th Text REtrieval Conference, pp.363:369. [11] Carterette, B. and J. Allan. 2005. Incremental Test Collections. In Proceedings of CIKM 2005, pp. 680-687. [12] Carterette, B., J. Allan, and R. Sitaraman. 2006. Minimal Test Collections for Retrieval Evaluation. In Proceedings of SIGIR 2006, pp. 268-275. [13] Cormack, G.V., C. R. Palmer, and C.L. Clarke. 1998. Efficient Construction of Large Test Collections. In Proceedings of SIGIR 1998, pp. 282-289. [14] Cormack, G. and T.R. Lynam. 2006. Statistical Precision of Information Retrieval Evaluation. In Proceedings of SIGIR 2006, pp. 533-540. [15] Cronen-Townsend, S., Y. Zhou, and W.B. Croft. 2004. A Language Modeling Framework for Selective Query Expansion, CIIR Technical Report. [16] Efthimiadis, E.N. Query Expansion. 1996. In Martha E. Williams (ed.), Annual Review of Information Systems and Technology (ARIST), v31, pp 121- 187. [17] Evans, D.A. and Lefferts, R.G. 1995. CLARIT-TREC Experiments. Information Processing & Management. 31(3): 385-295. [18] Fang, H. and C.X. Zhai. 2006. Semantic Term Matching in Axiomatic Approaches to Information Retrieval. In Proceedings of SIGIR 2006, pp. 115-122. [19] Gao, J., J. Nie, G. Wu and G. Cao. 2004. Dependence language model for information retrieval. In Proceedings of SIGIR 2004, pp. 170-177. [20] Harman, D.K. 1992. Relevance feedback revisited. In Proceedings of ACM SIGIR 1992, pp. 1-10. [21] Harman, D.K., ed. 1993. The First Text REtrieval Conference (TREC-1): 1992. [22] Harman, D.K., ed. 1994. The Second Text REtrieval Conference (TREC-2): 1993. [23] Harman, D.K., ed. 1995. The Third Text REtrieval Conference (TREC-3): 1994. [24] Harman, D.K., 1998. Towards Interactive Query Expansion. In Proceedings of SIGIR 1998, pp. 321-331. [25] Hofmann, T. 1999. Probabilistic latent semantic indexing. In Proceedings of SIGIR 1999, pp 50-57. [26] Jing, Y. and W.B. Croft. 1994. The Association Thesaurus for Information Retrieval. In Proceedings of RIAO 1994, pp. 146-160 [27] Lu, X.A. and R.B. Keefer. Query expansion/reduction and its impact on retrieval effectiveness. In: D.K. Harman, ed. The Third Text REtrieval Conference (TREC-3). Gaithersburg, MD: National Institute of Standards and Technology, 1995,231-239. [28] McNamee, P. and J. Mayfield. 2002. Comparing Cross-Language Query Expansion Techniques by Degrading Translation Resources. In Proceedings of SIGIR 2002, pp. 159-166. [29] Mitra, M., A. Singhal, and C. Buckley. 1998. Improving Automatic Query Expansion. In Proceedings of SIGIR 1998, pp. 206-214. [30] Peat, H. J. and P. Willett. 1991. The limitations of term co-occurrence data for query expansion in document retrieval systems. Journal of the American Society for Information Science, 42(5): 378-383. [31] Ponte, J.M. and W.B. Croft. 1998. A language modeling approach to information retrieval. In Proceedings of SIGIR 1998, pp.275-281. [32] Qiu Y., and Frei H. 1993. Concept based query expansion. In Proceedings of SIGIR 1993, pp. 160-169. [33] Robertson, S. 2006. On GMAP - and other transformations. In Proceedings of CIKM 2006, pp. 78-83. [34] Robertson, S.E. and K. Sparck Jones. 1976. Relevance Weighting of Search Terms. Journal of the American Society for Information Science, 27(3): 129-146. [35] Robertson, S.E., S. Walker, S. Jones, M.M. Hancock-Beaulieu, and M. Gatford. 1994. Okapi at TREC-2. In D.K. Harman (ed). 1994. The Second Text REtrieval Conference (TREC-2): 1993, pp. 21-34. [36] Robertson, S.E., S. Walker, S. Jones, M.M. Hancock-Beaulieu, and M. Gatford. 1995. Okapi at TREC-3. In D.K. Harman (ed). 1995. The Third Text REtrieval Conference (TREC-2): 1993, pp. 109-126 [37] Rocchio, J.J. 1971. Relevance feedback in information retrieval. In G. Salton (Ed.), The SMART Retrieval System. Prentice-Hall, Inc., Englewood Cliffs, NJ, pp. 313-323. [38] Salton, G. 1968. Automatic Information Organization and Retrieval. McGraw-Hill. [39] Salton, G. 1971. The SMART Retrieval System: Experiments in Automatic Document Processing. Englewood Cliffs NJ; Prentice-Hall. [40] Salton,G. 1980. Automatic term class construction using relevance-a summary of work in automatic pseudoclassification. Information Processing & Management. 16(1): 1-15. [41] Salton, G., and C. Buckley. 1988. On the Use of Spreading Activation Methods in Automatic Information Retrieval. In Proceedings of SIGIR 1998, pp. 147-160. [42] Sanderson, M. 1994. Word sense disambiguation and information retrieval. In Proceedings of SIGIR 1994, pp. 161-175. [43] Sanderson, M. and H. Joho. 2004. Forming test collections with no system pooling. In Proceedings of SIGIR 2004, pp. 186-193. [44] Sanderson, M. and Zobel, J. 2005. Information Retrieval System Evaluation: Effort, Sensitivity, and Reliability. In Proceedings of SIGIR 2005, pp. 162-169. [45] Smeaton, A.F. and C.J. Van Rijsbergen. 1983. The Retrieval Effects of Query Expansion on a Feedback Document Retrieval System. Computer Journal. 26(3):239-246. [46] Song, F. and W.B. Croft. 1999. A general language model for information retrieval. In Proceedings of the Eighth International Conference on Information and Knowledge Management, pages 316-321. [47] Sparck Jones, K. 1971. Automatic Keyword Classification for Information Retrieval. London: Butterworths. [48] Terra, E. and C. L. Clarke. 2004. Scoring missing terms in information retrieval tasks. In Proceedings of CIKM 2004, pp. 50-58. [49] Turtle, Howard. 1994. Natural Language vs. Boolean Query Evaluation: A Comparison of Retrieval Performance. In Proceedings of SIGIR 1994, pp. 212-220. [50] Voorhees, E.M. 1994a. On Expanding Query Vectors with Lexically Related Words. In Harman, D. K., ed. Text REtrieval Conference (TREC-1): 1992. [51] Voorhees, E.M. 1994b. Query Expansion Using Lexical-Semantic Relations. In Proceedings of SIGIR 1994, pp. 61-69.

evaluation;document processing;query expansion;relevant document;query-document term mismatch;information retrieval;information search;document expansion

train_H-49

Performance Prediction Using Spatial Autocorrelation

Evaluation of information retrieval systems is one of the core tasks in information retrieval. Problems include the inability to exhaustively label all documents for a topic, nongeneralizability from a small number of topics, and incorporating the variability of retrieval systems. Previous work addresses the evaluation of systems, the ranking of queries by difficulty, and the ranking of individual retrievals by performance. Approaches exist for the case of few and even no relevance judgments. Our focus is on zero-judgment performance prediction of individual retrievals. One common shortcoming of previous techniques is the assumption of uncorrelated document scores and judgments. If documents are embedded in a high-dimensional space (as they often are), we can apply techniques from spatial data analysis to detect correlations between document scores. We find that the low correlation between scores of topically close documents often implies a poor retrieval performance. When compared to a state of the art baseline, we demonstrate that the spatial analysis of retrieval scores provides significantly better prediction performance. These new predictors can also be incorporated with classic predictors to improve performance further. We also describe the first large-scale experiment to evaluate zero-judgment performance prediction for a massive number of retrieval systems over a variety of collections in several languages.

1. INTRODUCTION In information retrieval, a user poses a query to a system. The system retrieves n documents each receiving a realvalued score indicating the predicted degree of relevance. If we randomly select pairs of documents from this set, we expect some pairs to share the same topic and other pairs to not share the same topic. Take two topically-related documents from the set and call them a and b. If the scores of a and b are very different, we may be concerned about the performance of our system. That is, if a and b are both on the topic of the query, we would like them both to receive a high score; if a and b are not on the topic of the query, we would like them both to receive a low score. We might become more worried as we find more differences between scores of related documents. We would be more comfortable with a retrieval where scores are consistent between related documents. Our paper studies the quantification of this inconsistency in a retrieval from a spatial perspective. Spatial analysis is appropriate since many retrieval models embed documents in some vector space. If documents are embedded in a space, proximity correlates with topical relationships. Score consistency can be measured by the spatial version of autocorrelation known as the Moran coefficient or IM [5, 10]. In this paper, we demonstrate a strong correlation between IM and retrieval performance. The discussion up to this point is reminiscent of the cluster hypothesis. The cluster hypothesis states: closely-related documents tend to be relevant to the same request [12]. As we shall see, a retrieval function"s spatial autocorrelation measures the degree to which closely-related documents receive similar scores. Because of this, we interpret autocorrelation as measuring the degree to which a retrieval function satisfies the clustering hypothesis. If this connection is reasonable, in Section 6, we present evidence that failure to satisfy the cluster hypothesis correlates strongly with poor performance. In this work, we provide the following contributions, 1. A general, robust method for predicting the performance of retrievals with zero relevance judgments (Section 3). 2. A theoretical treatment of the similarities and motivations behind several state-of-the-art performance prediction techniques (Section 4). 3. The first large-scale experiments of zero-judgment, single run performance prediction (Sections 5 and 6). 2. PROBLEM DEFINITION Given a query, an information retrieval system produces a ranking of documents in the collection encoded as a set of scores associated with documents. We refer to the set of scores for a particular query-system combination as a retrieval. We would like to predict the performance of this retrieval with respect to some evaluation measure (eg, mean average precision). In this paper, we present results for ranking retrievals from arbitrary systems. We would like this ranking to approximate the ranking of retrievals by the evaluation measure. This is different from ranking queries by the average performance on each query. It is also different from ranking systems by the average performance on a set of queries. Scores are often only computed for the top n documents from the collection. We place these scores in the length n vector, y, where yi refers to the score of the ith-ranked document. We adjust scores to have zero mean and unit variance. We use this method because of its simplicity and its success in previous work [15]. 3. SPATIAL CORRELATION In information retrieval, we often assume that the representations of documents exist in some high-dimensional vector space. For example, given a vocabulary, V, this vector space may be an arbitrary |V|-dimensional space with cosine inner-product or a multinomial simplex with a distributionbased distance measure. An embedding space is often selected to respect topical proximity; if two documents are near, they are more likely to share a topic. Because of the prevalence and success of spatial models of information retrieval, we believe that the application of spatial data analysis techniques are appropriate. Whereas in information retrieval, we are concerned with the score at a point in a space, in spatial data analysis, we are concerned with the value of a function at a point or location in a space. We use the term function here to mean a mapping from a location to a real value. For example, we might be interested in the prevalence of a disease in the neighborhood of some city. The function would map the location of a neighborhood to an infection rate. If we want to quantify the spatial dependencies of a function, we would employ a measure referred to as the spatial autocorrelation [5, 10]. High spatial autocorrelation suggests that knowing the value of a function at location a will tell us a great deal about the value at a neighboring location b. There is a high spatial autocorrelation for a function representing the temperature of a location since knowing the temperature at a location a will tell us a lot about the temperature at a neighboring location b. Low spatial autocorrelation suggests that knowing the value of a function at location a tells us little about the value at a neighboring location b. There is low spatial autocorrelation in a function measuring the outcome of a coin toss at a and b. In this section, we will begin by describing what we mean by spatial proximity for documents and then define a measure of spatial autocorrelation. We conclude by extending this model to include information from multiple retrievals from multiple systems for a single query. 3.1 Spatial Representation of Documents Our work does not focus on improving a specific similarity measure or defining a novel vector space. Instead, we choose an inner product known to be effective at detecting interdocument topical relationships. Specifically, we adopt tf.idf document vectors, ˜di = di log „ (n + 0.5) − ci 0.5 + ci « (1) where d is a vector of term frequencies, c is the length-|V| document frequency vector. We use this weighting scheme due to its success for topical link detection in the context of Topic Detection and Tracking (TDT) evaluations [6]. Assuming vectors are scaled by their L2 norm, we use the inner product, ˜di, ˜dj , to define similarity. Given documents and some similarity measure, we can construct a matrix which encodes the similarity between pairs of documents. Recall that we are given the top n documents retrieved in y. We can compute an n × n similarity matrix, W. An element of this matrix, Wij represents the similarity between documents ranked i and j. In practice, we only include the affinities for a document"s k-nearest neighbors. In all of our experiments, we have fixed k to 5. We leave exploration of parameter sensitivity to future work. We also row normalize the matrix so that Pn j=1 Wij = 1 for all i. 3.2 Spatial Autocorrelation of a Retrieval Recall that we are interested in measuring the similarity between the scores of spatially-close documents. One such suitable measure is the Moran coefficient of spatial autocorrelation. Assuming the function y over n locations, this is defined as ˜IM = n eTWe P i,j Wijyiyj P i y2 i = n eTWe yT Wy yTy (2) where eT We = P ij Wij. We would like to compare autocorrelation values for different retrievals. Unfortunately, the bound for Equation 2 is not consistent for different W and y. Therefore, we use the Cauchy-Schwartz inequality to establish a bound, ˜IM ≤ n eTWe s yTWTWy yTy And we define the normalized spatial autocorrelation as IM = yT Wy p yTy × yTWTWy Notice that if we let ˜y = Wy, then we can write this formula as, IM = yT ˜y y 2 ˜y 2 (3) which can be interpreted as the correlation between the original retrieval scores and a set of retrieval scores diffused in the space. We present some examples of autocorrelations of functions on a grid in Figure 1. 3.3 Correlation with Other Retrievals Sometimes we are interested in the performance of a single retrieval but have access to scores from multiple systems for (a) IM = 0.006 (b) IM = 0.241 (c) IM = 0.487 Figure 1: The Moran coefficient, IM for a several binary functions on a grid. The Moran coefficient is a local measure of function consistency. From the perspective of information retrieval, each of these grid spaces would represent a document and documents would be organized so that they lay next to topically-related documents. Binary retrieval scores would define a pattern on this grid. Notice that, as the Moran coefficient increases, neighboring cells tend to have similar values. the same query. In this situation, we can use combined information from these scores to construct a surrogate for a high-quality ranking [17]. We can treat the correlation between the retrieval we are interested in and the combined scores as a predictor of performance. Assume that we are given m score functions, yi, for the same n documents. We will represent the mean of these vectors as yµ = Pm i=1 yi. We use the mean vector as an approximation to relevance. Since we use zero mean and unit variance normalization, work in metasearch suggests that this assumption is justified [15]. Because yµ represents a very good retrieval, we hypothesize that a strong similarity between yµ and y will correlate positively with system performance. We use Pearson"s product-moment correlation to measure the similarity between these vectors, ρ(y, yµ) = yT yµ y 2 yµ 2 (4) We will comment on the similarity between Equation 3 and 4 in Section 7. Of course, we can combine ρ(y, ˜y) and ρ(y, yµ) if we assume that they capture different factors in the prediction. One way to accomplish this is to combine these predictors as independent variables in a linear regression. An alternative means of combination is suggested by the mathematical form of our predictors. Since ˜y encodes the spatial dependencies in y and yµ encodes the spatial properties of the multiple runs, we can compute a third correlation between these two vectors, ρ(˜y, yµ) = ˜yT yµ ˜y 2 yµ 2 (5) We can interpret Equation 5 as measuring the correlation between a high quality ranking (yµ) and a spatially smoothed version of the retrieval (˜y). 4. RELATIONSHIP WITH OTHER PREDICTORS One way to predict the effectiveness of a retrieval is to look at the shared vocabulary of the top n retrieved documents. If we computed the most frequent content words in this set, we would hope that they would be consistent with our topic. In fact, we might believe that a bad retrieval would include documents on many disparate topics, resulting in an overlap of terminological noise. The Clarity of a query attempts to quantify exactly this [7]. Specifically, Clarity measures the similarity of the words most frequently used in retrieved documents to those most frequently used in the whole corpus. The conjecture is that a good retrieval will use language distinct from general text; the overlapping language in a bad retrieval will tend to be more similar to general text. Mathematically, we can compute a representation of the language used in the initial retrieval as a weighted combination of document language models, P(w|θQ) = nX i=1 P(w|θi) P(Q|θi) Z (6) where θi is the language model of the ith-ranked document, P(Q|θi) is the query likelihood score of the ith-ranked document and Z = Pn i=1 P(Q|θi) is a normalization constant. The similarity between the multinomial P(w|θQ) and a model of general text can be computed using the Kullback-Leibler divergence, DV KL(θQ θC ). Here, the distribution P(w|θC ) is our model of general text which can be computed using term frequencies in the corpus. In Figure 2a, we present Clarity as measuring the distance between the weighted center of mass of the retrieval (labeled y) and the unweighted center of mass of the collection (labeled O). Clarity reaches a minimum when a retrieval assigns every document the same score. Let"s again assume we have a set of n documents retrieved for our query. Another way to quantify the dispersion of a set of documents is to look at how clustered they are. We may hypothesize that a good retrieval will return a single, tight cluster. A poorly performing retrieval will return a loosely related set of documents covering many topics. One proposed method of quantifying this dispersion is to measure the distance from a random document a to it"s nearest neighbor, b. A retrieval which is tightly clustered will, on average, have a low distance between a and b; a retrieval which is less tightly-closed will, on average have high distances between a and b. This average corresponds to using the Cox-Lewis statistic to measure the randomness of the top n documents retrieved from a system [18]. In Figure 2a, this is roughly equivalent to measuring the area of the set n. Notice that we are throwing away information about the retrieval function y. Therefore the Cox-Lewis statistic is highly dependent on selecting the top n documents.1 Remember that we have n documents and a set of scores. Let"s assume that we have access to the system which provided the original scores and that we can also request scores for new documents. This suggests a third method for predicting performance. Take some document, a, from the retrieved set and arbitrarily add or remove words at random to create a new document ˜a. Now, we can ask our system to score ˜a with respect to our query. If, on average over the n documents, the scores of a and ˜a tend to be very different, we might suspect that the system is failing on this query. So, an alternative approach is to measure the simi1 The authors have suggested coupling the query with the distance measure [18]. The information introduced by the query, though, is retrieval-independent so that, if two retrievals return the same set of documents, the approximate Cox-Lewis statistic will be the same regardless of the retrieval scores. yOy (a) Global Divergence µ(y)˜y y (b) Score Perturbation µ(y) y (c) Multirun Averaging Figure 2: Representation of several performance predictors on a grid. In Figure 2a, we depict predictors which measure the divergence between the center of mass of a retrieval and the center of the embedding space. In Figure 2b, we depict predictors which compare the original retrieval, y, to a perturbed version of the retrieval, ˜y. Our approach uses a particular type of perturbation based on score diffusion. Finally, in Figure 2c, we depict prediction when given retrievals from several other systems on the same query. Here, we can consider the fusion of these retrieval as a surrogate for relevance. larity between the retrieval and a perturbed version of that retrieval [18, 19]. This can be accomplished by either perturbing the documents or queries. The similarity between the two retrievals can be measured using some correlation measure. This is depicted in Figure 2b. The upper grid represents the original retrieval, y, while the lower grid represents the function after having been perturbed, ˜y. The nature of the perturbation process requires additional scorings or retrievals. Our predictor does not require access to the original scoring function or additional retrievals. So, although our method is similar to other perturbation methods in spirit, it can be applied in situations when the retrieval system is inaccessible or costly to access. Finally, assume that we have, in addition to the retrieval we want to evaluate, m retrievals from a variety of different systems. In this case, we might take a document a, compare its rank in the retrieval to its average rank in the m retrievals. If we believe that the m retrievals provide a satisfactory approximation to relevance, then a very large difference in rank would suggest that our retrieval is misranking a. If this difference is large on average over all n documents, then we might predict that the retrieval is bad. If, on the other hand, the retrieval is very consistent with the m retrievals, then we might predict that the retrieval is good. The similarity between the retrieval and the combined retrieval may be computed using some correlation measure. This is depicted in Figure 2c. In previous work, the Kullback-Leibler divergence between the normalized scores of the retrieval and the normalized scores of the combined retrieval provides the similarity [1]. 5. EXPERIMENTS Our experiments focus on testing the predictive power of each of our predictors: ρ(y, ˜y), ρ(y, yµ), and ρ(˜y, yµ). As stated in Section 2, we are interested in predicting the performance of the retrieval generated by an arbitrary system. Our methodology is consistent with previous research in that we predict the relative performance of a retrieval by comparing a ranking based on our predictor to a ranking based on average precision. We present results for two sets of experiments. The first set of experiments presents detailed comparisons of our predictors to previously-proposed predictors using identical data sets. Our second set of experiments demonstrates the generalizability of our approach to arbitrary retrieval methods, corpus types, and corpus languages. 5.1 Detailed Experiments In these experiments, we will predict the performance of language modeling scores using our autocorrelation predictor, ρ(y, ˜y); we do not consider ρ(y, yµ) or ρ(˜y, yµ) because, in these detailed experiments, we focus on ranking the retrievals from a single system. We use retrievals, values for baseline predictors, and evaluation measures reported in previous work [19]. 5.1.1 Topics and Collections These performance prediction experiments use language model retrievals performed for queries associated with collections in the TREC corpora. Using TREC collections allows us to confidently associate an average precision with a retrieval. In these experiments, we use the following topic collections: TREC 4 ad-hoc, TREC 5 ad-hoc, Robust 2004, Terabyte 2004, and Terabyte 2005. 5.1.2 Baselines We provide two baselines. Our first baseline is the classic Clarity predictor presented in Equation 6. Clarity is designed to be used with language modeling systems. Our second baseline is Zhou and Croft"s ranking robustness predictor. This predictor corrupts the top k documents from retrieval and re-computes the language model scores for these corrupted documents. The value of the predictor is the Spearman rank correlation between the original ranking and the corrupted ranking. In our tables, we will label results for Clarity using DV KL and the ranking robustness predictor using P. 5.2 Generalizability Experiments Our predictors do not require a particular baseline retrieval system; the predictors can be computed for an arbitrary retrieval, regardless of how scores were generated. We believe that that is one of the most attractive aspects of our algorithm. Therefore, in a second set of experiments, we demonstrate the ability of our techniques to generalize to a variety of collections, topics, and retrieval systems. 5.2.1 Topics and Collections We gathered a diverse set of collections from all possible TREC corpora. We cast a wide net in order to locate collections where our predictors might fail. Our hypothesis is that documents with high topical similarity should have correlated scores. Therefore, we avoided collections where scores were unlikely to be correlated (eg, question-answering) or were likely to be negatively correlated (eg, novelty). Nevertheless, our collections include corpora where correlations are weakly justified (eg, non-English corpora) or not justified at all (eg, expert search). We use the ad-hoc tracks from TREC3-8, TREC Robust 2003-2005, TREC Terabyte 20042005, TREC4-5 Spanish, TREC5-6 Chinese, and TREC Enterprise Expert Search 2005. In all cases, we use only the automatic runs for ad-hoc tracks submitted to NIST. For all English and Spanish corpora, we construct the matrix W according to the process described in Section 3.1. For Chinese corpora, we use na¨ıve character-based tf.idf vectors. For entities, entries in W are proportional to the number of documents in which two entities cooccur. 5.2.2 Baselines In our detailed experiments, we used the Clarity measure as a baseline. Since we are predicting the performance of retrievals which are not based on language modeling, we use a version of Clarity referred to as ranked-list Clarity [7]. Ranked-list clarity converts document ranks to P(Q|θi) values. This conversion begins by replacing all of the scores in y with the respective ranks. Our estimation of P(Q|θi) from the ranks, then is, P(Q|θi) = ( 2(c+1−yi) c(c+1) if yi ≤ c 0 otherwise (7) where c is a cutoff parameter. As suggested by the authors, we fix the algorithm parameters c and λ2 so that c = 60 and λ2 = 0.10. We use Equation 6 to estimate P(w|θQ) and DV KL(θQ θC ) to compute the value of the predictor. We will refer to this predictor as DV KL, superscripted by V to indicate that the Kullback-Leibler divergence is with respect to the term embedding space. When information from multiple runs on the same query is available, we use Aslam and Pavlu"s document-space multinomial divergence as a baseline [1]. This rank-based method first normalizes the scores in a retrieval as an n-dimensional multinomial. As with ranked-list Clarity, we begin by replacing all of the scores in y with their respective ranks. Then, we adjust the elements of y in the following way, ˆyi = 1 2n 0 @1 + nX k=yi 1 k 1 A (8) In our multirun experiments, we only use the top 75 documents from each retrieval (n = 75); this is within the range of parameter values suggested by the authors. However, we admit not tuning this parameter for either our system or the baseline. The predictor is the divergence between the candidate distribution, y, and the mean distribution, yµ . With the uniform linear combination of these m retrievals represented as yµ, we can compute the divergence as Dn KL(ˆy ˆyµ) where we use the superscript n to indicate that the summation is over the set of n documents. This baseline was developed in the context of predicting query difficulty but we adopt it as a reasonable baseline for predicting retrieval performance. 5.2.3 Parameter Settings When given multiple retrievals, we use documents in the union of the top k = 75 documents from each of the m retrievals for that query. If the size of this union is ˜n, then yµ and each yi is of length ˜n. In some cases, a system did not score a document in the union. Since we are making a Gaussian assumption about our scores, we can sample scores for these unseen documents from the negative tail of the distribution. Specifically, we sample from the part of the distribution lower than the minimum value of in the normalized retrieval. This introduces randomness into our algorithm but we believe it is more appropriate than assigning an arbitrary fixed value. We optimized the linear regression using the square root of each predictor. We found that this substantially improved fits for all predictors, including the baselines. We considered linear combinations of pairs of predictors (labeled by the components) and all predictors (labeled as β). 5.3 Evaluation Given a set of retrievals, potentially from a combination of queries and systems, we measure the correlation of the rank ordering of this set by the predictor and by the performance metric. In order to ensure comparability with previous results, we present Kendall"s τ correlation between the predictor"s ranking and ranking based on average precision of the retrieval. Unless explicitly noted, all correlations are significant with p < 0.05. Predictors can sometimes perform better when linearly combined [9, 11]. Although previous work has presented the coefficient of determination (R2 ) to measure the quality of the regression, this measure cannot be reliably used when comparing slight improvements from combining predictors. Therefore, we adopt the adjusted coefficient of determination which penalizes models with more variables. The adjusted R2 allows us to evaluate the improvement in prediction achieved by adding a parameter but loses the statistical interpretation of R2 . We will use Kendall"s τ to evaluate the magnitude of the correlation and the adjusted R2 to evaluate the combination of variables. 6. RESULTS We present results for our detailed experiments comparing the prediction of language model scores in Table 1. Although the Clarity measure is theoretically designed for language model scores, it consistently underperforms our system-agnostic predictor. Ranking robustness was presented as an improvement to Clarity for web collections (represented in our experiments by the terabyte04 and terabyte05 collections), shifting the τ correlation from 0.139 to 0.150 for terabyte04 and 0.171 to 0.208 for terabyte05. However, these improvements are slight compared to the performance of autocorrelation on these collections. Our predictor achieves a τ correlation of 0.454 for terabyte04 and 0.383 for terabyte05. Though not always the strongest, autocorrelation achieves correlations competitive with baseline predictors. When examining the performance of linear combinations of predictors, we note that in every case, autocorrelation factors as a necessary component of a strong predictor. We also note that the adjusted R2 for individual baselines are always significantly improved by incorporating autocorrelation. We present our generalizability results in Table 2. We begin by examining the situation in column (a) where we are presented with a single retrieval and no information from additional retrievals. For every collection except one, we achieve significantly better correlations than ranked-list Clarity. Surprisingly, we achieve relatively strong correlations for Spanish and Chinese collections despite our na¨ıve processing. We do not have a ranked-list clarity correlation for ent05 because entity modeling is itself an open research question. However, our autocorrelation measure does not achieve high correlations perhaps because relevance for entity retrieval does not propagate according to the cooccurrence links we use. As noted above, the poor Clarity performance on web data is consistent with our findings in the detailed experiments. Clarity also notably underperforms for several news corpora (trec5, trec7, and robust04). On the other hand, autocorrelation seems robust to the changes between different corpora. Next, we turn to the introduction of information from multiple retrievals. We compare the correlations between those predictors which do not use this information in column (a) and those which do in column (b). For every collection, the predictors in column (b) outperform the predictors in column (a), indicating that the information from additional runs can be critical to making good predictions. Inspecting the predictors in column (b), we only draw weak conclusions. Our new predictors tend to perform better on news corpora. And between our new predictors, the hybrid ρ(˜y, yµ) predictor tends to perform better. Recall that our ρ(˜y, yµ) measure incorporates both spatial and multiple retrieval information. Therefore, we believe that the improvement in correlation is the result of incorporating information from spatial behavior. In column (c), we can investigate the utility of incorporating spatial information with information from multiple retrievals. Notice that in the cases where autocorrelation, ρ(y, ˜y), alone performs well (trec3, trec5-spanish, and trec6-chinese), it is substantially improved by incorporating multiple-retrieval information from ρ(y, yµ) in the linear regression, β. In the cases where ρ(y, yµ) performs well, incorporating autocorrelation rarely results in a significant improvement in performance. In fact, in every case where our predictor outperforms the baseline, it includes information from multiple runs. 7. DISCUSSION The most important result from our experiments involves prediction when no information is available from multiple runs (Tables 1 and 2a). This situation arises often in system design. For example, a system may need to, at retrieval time, assess its performance before deciding to conduct more intensive processing such as pseudo-relevance feedback or interaction. Assuming the presence of multiple retrievals is unrealistic in this case. We believe that autocorrelation is, like multiple-retrieval algorithms, approximating a good ranking; in this case by diffusing scores. Why is ˜y a reasonable surrogate? We know that diffusion of scores on the web graph and language model graphs improves performance [14, 16]. Therefore, if score diffusion tends to, in general, improve performance, then diffused scores will, in general, provide a good surrogate for relevance. Our results demonstrate that this approximation is not as powerful as information from multiple retrievals. Nevertheless, in situations where this information is lacking, autocorrelation provides substantial information. The success of autocorrelation as a predictor may also have roots in the clustering hypothesis. Recall that we regard autocorrelation as the degree to which a retrieval satisfies the clustering hypothesis. Our experiments, then, demonstrate that a failure to respect the clustering hypothesis correlates with poor performance. Why might systems fail to conform to the cluster hypothesis? Query-based information retrieval systems often score documents independently. The score of document a may be computed by examining query term or phrase matches, the document length, and perhaps global collection statistics. Once computed, a system rarely compares the score of a to the score of a topically-related document b. With some exceptions, the correlation of document scores has largely been ignored. We should make it clear that we have selected tasks where topical autocorrelation is appropriate. There are certainly cases where there is no reason to believe that retrieval scores will have topical autocorrelation. For example, ranked lists which incorporate document novelty should not exhibit spatial autocorrelation; if anything autocorrelation should be negative for this task. Similarly, answer candidates in a question-answering task may or may not exhibit autocorrelation; in this case, the semantics of links is questionable too. It is important before applying this measure to confirm that, given the semantics for some link between two retrieved items, we should expect a correlation between scores. 8. RELATED WORK In this section we draw more general comparisons to other work in performance prediction and spatial data analysis. There is a growing body of work which attempts to predict the performance of individual retrievals [7, 3, 11, 9, 19]. We have attempted to place our work in the context of much of this work in Section 4. However, a complete comparison is beyond the scope of this paper. We note, though, that our experiments cover a larger and more diverse set of retrievals, collections, and topics than previously examined. Much previous work-particularly in the context of TRECfocuses on predicting the performance of systems. Here, each system generates k retrievals. The task is, given these retrievals, to predict the ranking of systems according to some performance measure. Several papers attempt to address this task under the constraint of few judgments [2, 4]. Some work even attempts to use zero judgments by leveraging multiple retrievals for the same query [17]. Our task differs because we focus on ranking retrievals independent of the generating system. The task here is not to test the hypothesis system A is superior to system B but to test the hypothesis retrieval A is superior to retrieval B. Autocorrelation manifests itself in many classification tasks. Neville and Jensen define relational autocorrelation for relational learning problems and demonstrate that many classification tasks manifest autocorrelation [13]. Temporal autocorrelation of initial retrievals has also been used to predict performance [9]. However, temporal autocorrelation is performed by projecting the retrieval function into the temporal embedding space. In our work, we focus on the behavior of the function over the relationships between documents. τ adjusted R2 DV KL P ρ(y, ˜y) DV KL P ρ(y, ˜y) DV KL, P DV KL, ρ(y, ˜y) Pρ(y, ˜y) β trec4 0.353 0.548 0.513 0.168 0.363 0.422 0.466 0.420 0.557 0.553 trec5 0.311 0.329 0.357 0.116 0.190 0.236 0.238 0.244 0.266 0.269 robust04 0.418 0.398 0.373 0.256 0.304 0.278 0.403 0.373 0.402 0.442 terabyte04 0.139 0.150 0.454 0.059 0.045 0.292 0.076 0.293 0.289 0.284 terabyte05 0.171 0.208 0.383 0.022 0.072 0.193 0.120 0.225 0.218 0.257 Table 1: Comparison to Robustness and Clarity measures for language model scores. Evaluation replicates experiments from [19]. We present correlations between the classic Clarity measure (DV KL), the ranking robustness measure (P), and autocorrelation (ρ(y, ˜y)) each with mean average precision in terms of Kendall"s τ. The adjusted coefficient of determination is presented to measure the effectiveness of combining predictors. Measures in bold represent the strongest correlation for that test/collection pair. multiple run (a) (b) (c) τ τ adjusted R2 DKL ρ(y, ˜y) Dn KL ρ(y, yµ) ρ(˜y, yµ) Dn KL ρ(y, ˜y) ρ(y, yµ) ρ(˜y, yµ) β trec3 0.201 0.461 0.461 0.439 0.456 0.444 0.395 0.394 0.386 0.498 trec4 0.252 0.396 0.455 0.482 0.489 0.379 0.263 0.429 0.482 0.483 trec5 0.016 0.277 0.433 0.459 0.393 0.280 0.157 0.375 0.323 0.386 trec6 0.230 0.227 0.352 0.428 0.418 0.203 0.089 0.323 0.325 0.325 trec7 0.083 0.326 0.341 0.430 0.483 0.264 0.182 0.363 0.442 0.400 trec8 0.235 0.396 0.454 0.508 0.567 0.402 0.272 0.490 0.580 0.523 robust03 0.302 0.354 0.377 0.385 0.447 0.269 0.206 0.274 0.392 0.303 robust04 0.183 0.308 0.301 0.384 0.453 0.200 0.182 0.301 0.393 0.335 robust05 0.224 0.249 0.371 0.377 0.404 0.341 0.108 0.313 0.328 0.336 terabyte04 0.043 0.245 0.544 0.420 0.392 0.516 0.105 0.357 0.343 0.365 terabyte05 0.068 0.306 0.480 0.434 0.390 0.491 0.168 0.384 0.309 0.403 trec4-spanish 0.307 0.388 0.488 0.398 0.395 0.423 0.299 0.282 0.299 0.388 trec5-spanish 0.220 0.458 0.446 0.484 0.475 0.411 0.398 0.428 0.437 0.529 trec5-chinese 0.092 0.199 0.367 0.379 0.384 0.379 0.199 0.273 0.276 0.310 trec6-chinese 0.144 0.276 0.265 0.353 0.376 0.115 0.128 0.188 0.223 0.199 ent05 - 0.181 0.324 0.305 0.282 0.211 0.043 0.158 0.155 0.179 Table 2: Large scale prediction experiments. We predict the ranking of large sets of retrievals for various collections and retrieval systems. Kendall"s τ correlations are computed between the predicted ranking and a ranking based on the retrieval"s average precision. In column (a), we have predictors which do not use information from other retrievals for the same query. In columns (b) and (c) we present performance for predictors which incorporate information from multiple retrievals. The adjusted coefficient of determination is computed to determine effectiveness of combining predictors. Measures in bold represent the strongest correlation for that test/collection pair. Finally, regularization-based re-ranking processes are also closely-related to our work [8]. These techniques seek to maximize the agreement between scores of related documents by solving a constrained optimization problem. The maximization of consistency is equivalent to maximizing the Moran autocorrelation. Therefore, we believe that our work provides explanation for why regularization-based re-ranking works. 9. CONCLUSION We have presented a new method for predicting the performance of a retrieval ranking without any relevance judgments. We consider two cases. First, when making predictions in the absence of retrievals from other systems, our predictors demonstrate robust, strong correlations with average precision. This performance, combined with a simple implementation, makes our predictors, in particular, very attractive. We have demonstrated this improvement for many, diverse settings. To our knowledge, this is the first large scale examination of zero-judgment, single-retrieval performance prediction. Second, when provided retrievals from other systems, our extended methods demonstrate competitive performance with state of the art baselines. Our experiments also demonstrate the limits of the usefulness of our predictors when information from multiple runs is provided. Our results suggest two conclusions. First, our results could affect retrieval algorithm design. Retrieval algorithms designed to consider spatial autocorrelation will conform to the cluster hypothesis and improve performance. Second, our results could affect the design of minimal test collection algorithms. Much of the recent work in ranking systems sometimes ignores correlations between document labels and scores. We believe that these two directions could be rewarding given the theoretical and experimental evidence in this paper. 10. ACKNOWLEDGMENTS This work was supported in part by the Center for Intelligent Information Retrieval and in part by the Defense Advanced Research Projects Agency (DARPA) under contract number HR0011-06-C-0023. Any opinions, findings and conclusions or recommendations expressed in this material are the author"s and do not necessarily reflect those of the sponsor. We thank Yun Zhou and Desislava Petkova for providing data and Andre Gauthier for technical assistance. 11. REFERENCES [1] J. Aslam and V. Pavlu. Query hardness estimation using jensen-shannon divergence among multiple scoring functions. In ECIR 2007: Proceedings of the 29th European Conference on Information Retrieval, 2007. [2] J. A. Aslam, V. Pavlu, and E. Yilmaz. A statistical method for system evaluation using incomplete judgments. In S. Dumais, E. N. Efthimiadis, D. Hawking, and K. Jarvelin, editors, Proceedings of the 29th Annual International ACM SIGIR Conference on Research and Development in Information Retrieval, pages 541-548. ACM Press, August 2006. [3] D. Carmel, E. Yom-Tov, A. Darlow, and D. Pelleg. What makes a query difficult? In SIGIR "06: Proceedings of the 29th annual international ACM SIGIR conference on Research and development in information retrieval, pages 390-397, New York, NY, USA, 2006. ACM Press. [4] B. Carterette, J. Allan, and R. Sitaraman. Minimal test collections for retrieval evaluation. In SIGIR "06: Proceedings of the 29th annual international ACM SIGIR conference on Research and development in information retrieval, pages 268-275, New York, NY, USA, 2006. ACM Press. [5] A. D. Cliff and J. K. Ord. Spatial Autocorrelation. Pion Ltd., 1973. [6] M. Connell, A. Feng, G. Kumaran, H. Raghavan, C. Shah, and J. Allan. Umass at tdt 2004. Technical Report CIIR Technical Report IR - 357, Department of Computer Science, University of Massachusetts, 2004. [7] S. Cronen-Townsend, Y. Zhou, and W. B. Croft. Precision prediction based on ranked list coherence. Inf. Retr., 9(6):723-755, 2006. [8] F. Diaz. Regularizing ad-hoc retrieval scores. In CIKM "05: Proceedings of the 14th ACM international conference on Information and knowledge management, pages 672-679, New York, NY, USA, 2005. ACM Press. [9] F. Diaz and R. Jones. Using temporal profiles of queries for precision prediction. In SIGIR "04: Proceedings of the 27th annual international ACM SIGIR conference on Research and development in information retrieval, pages 18-24, New York, NY, USA, 2004. ACM Press. [10] D. A. Griffith. Spatial Autocorrelation and Spatial Filtering. Springer Verlag, 2003. [11] B. He and I. Ounis. Inferring Query Performance Using Pre-retrieval Predictors. In The Eleventh Symposium on String Processing and Information Retrieval (SPIRE), 2004. [12] N. Jardine and C. J. V. Rijsbergen. The use of hierarchic clustering in information retrieval. Information Storage and Retrieval, 7:217-240, 1971. [13] D. Jensen and J. Neville. Linkage and autocorrelation cause feature selection bias in relational learning. In ICML "02: Proceedings of the Nineteenth International Conference on Machine Learning, pages 259-266, San Francisco, CA, USA, 2002. Morgan Kaufmann Publishers Inc. [14] O. Kurland and L. Lee. Corpus structure, language models, and ad-hoc information retrieval. In SIGIR "04: Proceedings of the 27th annual international conference on Research and development in information retrieval, pages 194-201, New York, NY, USA, 2004. ACM Press. [15] M. Montague and J. A. Aslam. Relevance score normalization for metasearch. In CIKM "01: Proceedings of the tenth international conference on Information and knowledge management, pages 427-433, New York, NY, USA, 2001. ACM Press. [16] T. Qin, T.-Y. Liu, X.-D. Zhang, Z. Chen, and W.-Y. Ma. A study of relevance propagation for web search. In SIGIR "05: Proceedings of the 28th annual international ACM SIGIR conference on Research and development in information retrieval, pages 408-415, New York, NY, USA, 2005. ACM Press. [17] I. Soboroff, C. Nicholas, and P. Cahan. Ranking retrieval systems without relevance judgments. In SIGIR "01: Proceedings of the 24th annual international ACM SIGIR conference on Research and development in information retrieval, pages 66-73, New York, NY, USA, 2001. ACM Press. [18] V. Vinay, I. J. Cox, N. Milic-Frayling, and K. Wood. On ranking the effectiveness of searches. In SIGIR "06: Proceedings of the 29th annual international ACM SIGIR conference on Research and development in information retrieval, pages 398-404, New York, NY, USA, 2006. ACM Press. [19] Y. Zhou and W. B. Croft. Ranking robustness: a novel framework to predict query performance. In CIKM "06: Proceedings of the 15th ACM international conference on Information and knowledge management, pages 567-574, New York, NY, USA, 2006. ACM Press.

query ranking;spatial autocorrelation;predictor predictive power;predictor relationship;language model score;ranking of query;regularization;autocorrelation;performance prediction;information retrieval;zero relevance judgment;cluster hypothesis;predictive power of predictor;relationship of predictor

train_H-50

An Outranking Approach for Rank Aggregation in Information Retrieval

Research in Information Retrieval usually shows performance improvement when many sources of evidence are combined to produce a ranking of documents (e.g., texts, pictures, sounds, etc.). In this paper, we focus on the rank aggregation problem, also called data fusion problem, where rankings of documents, searched into the same collection and provided by multiple methods, are combined in order to produce a new ranking. In this context, we propose a rank aggregation method within a multiple criteria framework using aggregation mechanisms based on decision rules identifying positive and negative reasons for judging whether a document should get a better rank than another. We show that the proposed method deals well with the Information Retrieval distinctive features. Experimental results are reported showing that the suggested method performs better than the well-known CombSUM and CombMNZ operators.

1. INTRODUCTION A wide range of current Information Retrieval (IR) approaches are based on various search models (Boolean, Vector Space, Probabilistic, Language, etc. [2]) in order to retrieve relevant documents in response to a user request. The result lists produced by these approaches depend on the exact definition of the relevance concept. Rank aggregation approaches, also called data fusion approaches, consist in combining these result lists in order to produce a new and hopefully better ranking. Such approaches give rise to metasearch engines in the Web context. We consider, in the following, cases where only ranks are available and no other additional information is provided such as the relevance scores. This corresponds indeed to the reality, where only ordinal information is available. Data fusion is also relevant in other contexts, such as when the user writes several queries of his/her information need (e.g., a boolean query and a natural language query) [4], or when many document surrogates are available [16]. Several studies argued that rank aggregation has the potential of combining effectively all the various sources of evidence considered in various input methods. For instance, experiments carried out in [16], [30], [4] and [19] showed that documents which appear in the lists of the majority of the input methods are more likely to be relevant. Moreover, Lee [19] and Vogt and Cottrell [31] found that various retrieval approaches often return very different irrelevant documents, but many of the same relevant documents. Bartell et al. [3] also found that rank aggregation methods improve the performances w.r.t. those of the input methods, even when some of them have weak individual performances. These methods also tend to smooth out biases of the input methods according to Montague and Aslam [22]. Data fusion has recently been proved to improve performances for both the ad-hoc retrieval and categorization tasks within the TREC genomics track in 2005 [1]. The rank aggregation problem was addressed in various fields such as i) in social choice theory which studies voting algorithms which specify winners of elections or winners of competitions in tournaments [29], ii) in statistics when studying correlation between rankings, iii) in distributed databases when results from different databases must be combined [12], and iv) in collaborative filtering [23]. Most current rank aggregation methods consider each input ranking as a permutation over the same set of items. They also give rigid interpretation to the exact ranking of the items. Both of these assumptions are rather not valid in the IR context, as will be shown in the following sections. The remaining of the paper is organized as follows. We first review current rank aggregation methods in Section 2. Then we outline the specificities of the data fusion problem in the IR context (Section 3). In Section 4, we present a new aggregation method which is proven to best fit the IR context. Experimental results are presented in Section 5 and conclusions are provided in a final section. 2. RELATED WORK As pointed out by Riker [25], we can distinguish two families of rank aggregation methods: positional methods which assign scores to items to be ranked according to the ranks they receive and majoritarian methods which are based on pairwise comparisons of items to be ranked. These two families of methods find their roots in the pioneering works of Borda [5] and Condorcet [7], respectively, in the social choice literature. 2.1 Preliminaries We first introduce some basic notations to present the rank aggregation methods in a uniform way. Let D = {d1, d2, . . . , dnd } be a set of nd documents. A list or a ranking j is an ordering defined on Dj ⊆ D (j = 1, . . . , n). Thus, di j di means di ‘is ranked better than" di in j. When Dj = D, j is said to be a full list. Otherwise, it is a partial list. If di belongs to Dj, rj i denotes the rank or position of di in j. We assume that the best answer (document) is assigned the position 1 and the worst one is assigned the position |Dj|. Let D be the set of all permutations on D or all subsets of D. A profile is a n-tuple of rankings PR = ( 1, 2, . . . , n). Restricting PR to the rankings containing document di defines PRi. We also call the number of rankings which contain document di the rank hits of di [19]. The rank aggregation or data fusion problem consists of finding a ranking function or mechanism Ψ (also called a social welfare function in the social choice theory terminology) defined by: Ψ : n D → D PR = ( 1, 2, . . . , n) → σ = Ψ(PR) where σ is called a consensus ranking. 2.2 Positional Methods 2.2.1 Borda Count This method [5] first assigns a score n j=1 rj i to each document di. Documents are then ranked by increasing order of this score, breaking ties, if any, arbitrarily. 2.2.2 Linear Combination Methods This family of methods basically combine scores of documents. When used for the rank aggregation problem, ranks are assumed to be scores or performances to be combined using aggregation operators such as the weighted sum or some variation of it [3, 31, 17, 28]. For instance, Callan et al. [6] used the inference networks model [30] to combine rankings. Fox and Shaw [15] proposed several combination strategies which are CombSUM, CombMIN, CombMAX, CombANZ and CombMNZ. The first three operators correspond to the sum, min and max operators, respectively. CombANZ and CombMNZ respectively divides and multiplies the CombSUM score by the rank hits. It is shown in [19] that the CombSUM and CombMNZ operators perform better than the others. Metasearch engines such as SavvySearch and MetaCrawler use the CombSUM strategy to fuse rankings. 2.2.3 Footrule Optimal Aggregation In this method, a consensus ranking minimizes the Spearman footrule distance from the input rankings [21]. Formally, given two full lists j and j , this distance is given by F( j, j ) = nd i=1 |rj i − rj i |. It extends to several lists as follows. Given a profile PR and a consensus ranking σ, the Spearman footrule distance of σ to PR is given by F(σ, PR) = n j=1 F(σ, j). Cook and Kress [8] proposed a similar method which consists in optimizing the distance D( j, j ) = 1 2 nd i,i =1 |rj i,i − rj i,i |, where rj i,i = rj i −rj i . This formulation has the advantage that it considers the intensity of preferences. 2.2.4 Probabilistic Methods This kind of methods assume that the performance of the input methods on a number of training queries is indicative of their future performance. During the training process, probabilities of relevance are calculated. For subsequent queries, documents are ranked based on these probabilities. For instance, in [20], each input ranking j is divided into a number of segments, and the conditional probability of relevance (R) of each document di depending on the segment k it occurs in, is computed, i.e. prob(R|di, k, j). For subsequent queries, the score of each document di is given by n j=1 prob(R|di,k, j ) k . Le Calve and Savoy [18] suggest using a logistic regression approach for combining scores. Training data is needed to infer the model parameters. 2.3 Majoritarian Methods 2.3.1 Condorcet Procedure The original Condorcet rule [7] specifies that a winner of the election is any item that beats or ties with every other item in a pairwise contest. Formally, let C(diσdi ) = { j∈ PR : di j di } be the coalition of rankings that are concordant with establishing diσdi , i.e. with the proposition di ‘should be ranked better than" di in the final ranking σ. di beats or ties with di iff |C(diσdi )| ≥ |C(di σdi)|. The repetitive application of the Condorcet algorithm can produce a ranking of items in a natural way: select the Condorcet winner, remove it from the lists, and repeat the previous two steps until there are no more documents to rank. Since there is not always Condorcet winners, variations of the Condorcet procedure have been developed within the multiple criteria decision aid theory, with methods such as ELECTRE [26]. 2.3.2 Kemeny Optimal Aggregation As in section 2.2.3, a consensus ranking minimizes a geometric distance from the input rankings, where the Kendall tau distance is used instead of the Spearman footrule distance. Formally, given two full lists j and j , the Kendall tau distance is given by K( j, j ) = |{(di, di ) : i < i , rj i < rj i , rj i > rj i }|, i.e. the number of pairwise disagreements between the two lists. It is easy to show that the consensus ranking corresponds to the geometric median of the input rankings and that the Kemeny optimal aggregation problem corresponds to the minimum feedback edge set problem. 2.3.3 Markov Chain Methods Markov chains (MCs) have been used by Dwork et al. [11] as a ‘natural" method to obtain a consensus ranking where states correspond to the documents to be ranked and the transition probabilities vary depending on the interpretation of the transition event. In the same reference, the authors proposed four specific MCs and experimental testing had shown that the following MC is the best performing one (see also [24]): • MC4: move from the current state di to the next state di by first choosing a document di uniformly from D. If for the majority of the rankings, we have rj i ≤ rj i , then move to di , else stay in di. The consensus ranking corresponds to the stationary distribution of MC4. 3. SPECIFICITIES OF THE RANK AGGREGATION PROBLEM IN THE IR CONTEXT 3.1 Limited Significance of the Rankings The exact positions of documents in one input ranking have limited significance and should not be overemphasized. For instance, having three relevant documents in the first three positions, any perturbation of these three items will have the same value. Indeed, in the IR context, the complete order provided by an input method may hide ties. In this case, we call such rankings semi orders. This was outlined in [13] as the problem of aggregation with ties. It is therefore important to build the consensus ranking based on robust information: • Documents with near positions in j are more likely to have similar interest or relevance. Thus a slight perturbation of the initial ranking is meaningless. • Assuming that document di is better ranked than document di in a ranking j, di is more likely to be definitively more relevant than di in j when the number of intermediate positions between di and di increases. 3.2 Partial Lists In real world applications, such as metasearch engines, rankings provided by the input methods are often partial lists. This was outlined in [14] as the problem of having to merge top-k results from various input lists. For instance, in the experiments carried out by Dwork et al. [11], authors found that among the top 100 best documents of 7 input search engines, 67% of the documents were present in only one search engine, whereas less than two documents were present in all the search engines. Rank aggregation of partial lists raises four major difficulties which we state hereafter, proposing for each of them various working assumptions: 1. Partial lists can have various lengths, which can favour long lists. We thus consider the following two working hypotheses: H1 k : We only consider the top k best documents from each input ranking. H1 all: We consider all the documents from each input ranking. 2. Since there are different documents in the input rankings, we must decide which documents should be kept in the consensus ranking. Two working hypotheses are therefore considered: H2 k : We only consider documents which are present in at least k input rankings (k > 1). H2 all: We consider all the documents which are ranked in at least one input ranking. Hereafter, we call documents which will be retained in the consensus ranking, candidate documents, and documents that will be excluded from the consensus ranking, excluded documents. We also call a candidate document which is missing in one or more rankings, a missing document. 3. Some candidate documents are missing documents in some input rankings. Main reasons for a missing document are that it was not indexed or it was indexed but deemed irrelevant ; usually this information is not available. We consider the following two working hypotheses: H3 yes: Each missing document in each j is assigned a position. H3 no: No assumption is made, that is each missing document is considered neither better nor worse than any other document. 4. When assumption H2 k holds, each input ranking may contain documents which will not be considered in the consensus ranking. Regarding the positions of the candidate documents, we can consider the following working hypotheses: H4 init: The initial positions of candidate documents are kept in each input ranking. H4 new: Candidate documents receive new positions in each input ranking, after discarding excluded ones. In the IR context, rank aggregation methods need to decide more or less explicitly which assumptions to retain w.r.t. the above-mentioned difficulties. 4. OUTRANKING APPROACH FOR RANK AGGREGATION 4.1 Presentation Positional methods consider implicitly that the positions of the documents in the input rankings are scores giving thus a cardinal meaning to an ordinal information. This constitutes a strong assumption that is questionable, especially when the input rankings have different lengths. Moreover, for positional methods, assumptions H3 and H4 , which are often arbitrary, have a strong impact on the results. For instance, let us consider an input ranking of 500 documents out of 1000 candidate documents. Whether we assign to each of the missing documents the position 1, 501, 750 or 1000 -corresponding to variations of H3 yes- will give rise to very contrasted results, especially regarding the top of the consensus ranking. Majoritarian methods do not suffer from the above-mentioned drawbacks of the positional methods since they build consensus rankings exploiting only ordinal information contained in the input rankings. Nevertheless, they suppose that such rankings are complete orders, ignoring that they may hide ties. Therefore, majoritarian methods base consensus rankings on illusory discriminant information rather than less discriminant but more robust information. Trying to overcome the limits of current rank aggregation methods, we found that outranking approaches, which were initially used for multiple criteria aggregation problems [26], can also be used for the rank aggregation purpose, where each ranking plays the role of a criterion. Therefore, in order to decide whether a document di should be ranked better than di in the consensus ranking σ, the two following conditions should be met: • a concordance condition which ensures that a majority of the input rankings are concordant with diσdi (majority principle). • a discordance condition which ensures that none of the discordant input rankings strongly refutes dσd (respect of minorities principle). Formally, the concordance coalition with diσdi is Csp (diσdi ) = { j∈ PR : rj i ≤ rj i − sp} where sp is a preference threshold which is the variation of document positions -whether it is absolute or relative to the ranking length- which draws the boundaries between an indifference and a preference situation between documents. The discordance coalition with diσdi is Dsv (diσdi ) = { j∈ PR : rj i ≥ rj i + sv} where sv is a veto threshold which is the variation of document positions -whether it is absolute or relative to the ranking length- which draws the boundaries between a weak and a strong opposition to diσdi . Depending on the exact definition of the preceding concordance and discordance coalitions leading to the definition of some decision rules, several outranking relations can be defined. They can be more or less demanding depending on i) the values of the thresholds sp and sv, ii) the importance or minimal size cmin required for the concordance coalition, and iii) the importance or maximum size dmax of the discordance coalition. A generic outranking relation can thus be defined as follows: diS(sp,sv,cmin,dmax)di ⇔ |Csp (diσdi )| ≥ cmin AND |Dsv (diσdi )| ≤ dmax This expression defines a family of nested outranking relations since S(sp,sv,cmin,dmax) ⊆ S(sp,sv,cmin,dmax) when cmin ≥ cmin and/or dmax ≤ dmax and/or sp ≥ sp and/or sv ≤ sv. This expression also generalizes the majority rule which corresponds to the particular relation S(0,∞, n 2 ,n). It also satisfies important properties of rank aggregation methods, called neutrality, Pareto-optimality, Condorcet property and Extended Condorcet property, in the social choice literature [29]. Outranking relations are not necessarily transitive and do not necessarily correspond to rankings since directed cycles may exist. Therefore, we need specific procedures in order to derive a consensus ranking. We propose the following procedure which finds its roots in [27]. It consists in partitioning the set of documents into r ranked classes. Each class Ch contains documents with the same relevance and results from the application of all relations (if possible) to the set of documents remaining after previous classes are computed. Documents within the same equivalence class are ranked arbitrarily. Formally, let • R be the set of candidate documents for a query, • S1 , S2 , . . . be a family of nested outranking relations, • Fk(di, E) = |{di ∈ E : diSk di }| be the number of documents in E(E ⊆ R) that could be considered ‘worse" than di according to relation Sk , • fk(di, E) = |{di ∈ E : di Sk di}| be the number of documents in E that could be considered ‘better" than di according to Sk , • sk(di, E) = Fk(di, E) − fk(di, E) be the qualification of di in E according to Sk . Each class Ch results from a distillation process. It corresponds to the last distillate of a series of sets E0 ⊇ E1 ⊇ . . . where E0 = R \ (C1 ∪ . . . ∪ Ch−1) and Ek is a reduced subset of Ek−1 resulting from the application of the following procedure: 1. compute for each di ∈ Ek−1 its qualification according to Sk , i.e. sk(di, Ek−1), 2. define smax = maxdi∈Ek−1 {sk(di, Ek−1)}, then 3. Ek = {di ∈ Ek−1 : sk(di, Ek−1) = smax} When one outranking relation is used, the distillation process stops after the first application of the previous procedure, i.e., Ch corresponds to distillate E1. When different outranking relations are used, the distillation process stops when all the pre-defined outranking relations have been used or when |Ek| = 1. 4.2 Illustrative Example This section illustrates the concepts and procedures of section 4.1. Let us consider a set of candidate documents R = {d1, d2, d3, d4, d5}. The following table gives a profile PR of different rankings of the documents of R: PR = ( 1 , 2, 3, 4). Table 1: Rankings of documents rj i 1 2 3 4 d1 1 3 1 5 d2 2 1 3 3 d3 3 2 2 1 d4 4 4 5 2 d5 5 5 4 4 Let us suppose that the preference and veto thresholds are set to values 1 and 4 respectively, and that the concordance and discordance thresholds are set to values 2 and 1 respectively. The following tables give the concordance, discordance and outranking matrices. Each entry csp (di, di ) (dsv (di, di )) in the concordance (discordance) matrix gives the number of rankings that are concordant (discordant) with diσdi , i.e. csp (di, di ) = |Csp (diσdi )| and dsv (di, di ) = |Dsv (diσdi )|. Table 2: Computation of the outranking relation d1 d2 d3 d4 d5 d1 - 2 2 3 3 d2 2 - 2 3 4 d3 2 2 - 4 4 d4 1 1 0 - 3 d5 1 0 0 1Concordance Matrix d1 d2 d3 d4 d5 d1 - 0 1 0 0 d2 0 - 0 0 0 d3 0 0 - 0 0 d4 1 0 0 - 0 d5 1 1 0 0Discordance Matrix d1 d2 d3 d4 d5 d1 - 1 1 1 1 d2 1 - 1 1 1 d3 1 1 - 1 1 d4 0 0 0 - 1 d5 0 0 0 0Outranking Matrix (S1) For instance, the concordance coalition for the assertion d1σd4 is C1(d1σd4) = { 1, 2, 3} and the discordance coalition for the same assertion is D4(d1σd4) = ∅. Therefore, c1(d1, d4) = 3, d4(d1, d4) = 0 and d1S1 d4 holds. Notice that Fk(di, R) (fk(di, R)) is given by summing the values of the ith row (column) of the outranking matrix. The consensus ranking is obtained as follows: to get the first class C1, we compute the qualifications of all the documents of E0 = R with respect to S1 . They are respectively 2, 2, 2, -2 and -4. Therefore smax equals 2 and C1 = E1 = {d1, d2, d3}. Observe that, if we had used a second outranking relation S2(⊇ S1), these three documents could have been possibly discriminated. At this stage, we remove documents of C1 from the outranking matrix and compute the next class C2: we compute the new qualifications of the documents of E0 = R \ C1 = {d4, d5}. They are respectively 1 and -1. So C3 = E1 = {d4}. The last document d5 is the only document of the last class C3. Thus, the consensus ranking is {d1, d2, d3} → {d4} → {d5}. 5. EXPERIMENTS AND RESULTS 5.1 Test Setting To facilitate empirical investigation of the proposed methodology, we developed a prototype metasearch engine that implements a version of our outranking approach for rank aggregation. In this paper, we apply our approach to the Topic Distillation (TD) task of TREC-2004 Web track [10]. In this task, there are 75 topics where only a short description of each is given. For each query, we retained the rankings of the 10 best runs of the TD task which are provided by TREC-2004 participating teams. The performances of these runs are reported in table 3. Table 3: Performances of the 10 best runs of the TD task of TREC-2004 Run Id MAP P@10 S@1 S@5 S@10 uogWebCAU150 17.9% 24.9% 50.7% 77.3% 89.3% MSRAmixed1 17.8% 25.1% 38.7% 72.0% 88.0% MSRC04C12 16.5% 23.1% 38.7% 74.7% 80.0% humW04rdpl 16.3% 23.1% 37.3% 78.7% 90.7% THUIRmix042 14.7% 20.5% 21.3% 58.7% 74.7% UAmsT04MWScb 14.6% 20.9% 36.0% 66.7% 76.0% ICT04CIIS1AT 14.1% 20.8% 33.3% 64.0% 78.7% SJTUINCMIX5 12.9% 18.9% 29.3% 57.3% 72.0% MU04web1 11.5% 19.9% 33.3% 64.0% 76.0% MeijiHILw3 11.5% 15.3% 30.7% 54.7% 64.0% Average 14.7% 21.2% 34.9% 66.8% 78.94% For each query, each run provides a ranking of about 1000 documents. The number of documents retrieved by all these runs ranges from 543 to 5769. Their average (median) number is 3340 (3386). It is worth noting that we found similar distributions of the documents among the rankings as in [11]. For evaluation, we used the ‘trec eval" standard tool which is used by the TREC community to calculate the standard measures of system effectiveness which are Mean Average Precision (MAP) and Success@n (S@n) for n=1, 5 and 10. Our approach effectiveness is compared against some high performing official results from TREC-2004 as well as against some standard rank aggregation algorithms. In the experiments, significance testing is mainly based on the t-student statistic which is computed on the basis of the MAP values of the compared runs. In the tables of the following section, statistically significant differences are marked with an asterisk. Values between brackets of the first column of each table, indicate the parameter value of the corresponding run. 5.2 Results We carried out several series of runs in order to i) study performance variations of the outranking approach when tuning the parameters and working assumptions, ii) compare performances of the outranking approach vs standard rank aggregation strategies , and iii) check whether rank aggregation performs better than the best input rankings. We set our basic run mcm with the following parameters. We considered that each input ranking is a complete order (sp = 0) and that an input ranking strongly refutes diσdi when the difference of both document positions is large enough (sv = 75%). Preference and veto thresholds are computed proportionally to the number of documents retained in each input ranking. They consequently may vary from one ranking to another. In addition, to accept the assertion diσdi , we supposed that the majority of the rankings must be concordant (cmin = 50%) and that every input ranking can impose its veto (dmax = 0). Concordance and discordance thresholds are computed for each tuple (di, di ) as the percentage of the input rankings of PRi ∩PRi . Thus, our choice of parameters leads to the definition of the outranking relation S(0,75%,50%,0). To test the run mcm, we had chosen the following assumptions. We retained the top 100 best documents from each input ranking (H1 100), only considered documents which are present in at least half of the input rankings (H2 5 ) and assumed H3 no and H4 new. In these conditions, the number of successful documents was about 100 on average, and the computation time per query was less than one second. Obviously, modifying the working assumptions should have deeper impact on the performances than tuning our model parameters. This was validated by preliminary experiments. Thus, we hereafter begin by studying performance variation when different sets of assumptions are considered. Afterwards, we study the impact of tuning parameters. Finally, we compare our model performances w.r.t. the input rankings as well as some standard data fusion algorithms. 5.2.1 Impact of the Working Assumptions Table 4 summarizes the performance variation of the outranking approach under different working hypotheses. In Table 4: Impact of the working assumptions Run Id MAP S@1 S@5 S@10 mcm 18.47% 41.33% 81.33% 86.67% mcm22 (H3 yes) 17.72% (-4.06%) 34.67% 81.33% 86.67% mcm23 (H4 init) 18.26% (-1.14%) 41.33% 81.33% 86.67% mcm24 (H1 all) 20.67% (+11.91%*) 38.66% 80.00% 86.66% mcm25 (H2 all) 21.68% (+17.38%*) 40.00% 78.66% 89.33% this table, we first show that run mcm22, in which missing documents are all put in the same last position of each input ranking, leads to performance drop w.r.t. run mcm. Moreover, S@1 moves from 41.33% to 34.67% (-16.11%). This shows that several relevant documents which were initially put at the first position of the consensus ranking in mcm, lose this first position but remain ranked in the top 5 documents since S@5 did not change. We also conclude that documents which have rather good positions in some input rankings are more likely to be relevant, even though they are missing in some other rankings. Consequently, when they are missing in some rankings, assigning worse ranks to these documents is harmful for performance. Also, from Table 4, we found that the performances of runs mcm and mcm23 are similar. Therefore, the outranking approach is not sensitive to keeping the initial positions of candidate documents or recomputing them by discarding excluded ones. From the same Table 4, performance of the outranking approach increases significantly for runs mcm24 and mcm25. Therefore, whether we consider all the documents which are present in half of the rankings (mcm24) or we consider all the documents which are ranked in the first 100 positions in one or more rankings (mcm25), increases performances. This result was predictable since in both cases we have more detailed information on the relative importance of documents. Tables 5 and 6 confirm this evidence. Table 5, where values between brackets of the first column give the number of documents which are retained from each input ranking, shows that selecting more documents from each input ranking leads to performance increase. It is worth mentioning that selecting more than 600 documents from each input ranking does not improve performance. Table 5: Impact of the number of retained documents Run Id MAP S@1 S@5 S@10 mcm (100) 18.47% 41.33% 81.33% 86.67% mcm24-1 (200) 19.32% (+4.60%) 42.67% 78.67% 88.00% mcm24-2 (400) 19.88% (+7.63%*) 37.33% 80.00% 88.00% mcm24-3 (600) 20.80% (+12.62%*) 40.00% 80.00% 88.00% mcm24-4 (800) 20.66% (+11.86%*) 40.00% 78.67% 86.67% mcm24 (1000) 20.67% (+11.91%*) 38.66% 80.00% 86.66% Table 6 reports runs corresponding to variations of H2 k . Values between brackets are rank hits. For instance, in the run mcm32, only documents which are present in 3 or more input rankings, were considered successful. This table shows that performance is significantly better when rare documents are considered, whereas it decreases significantly when these documents are discarded. Therefore, we conclude that many of the relevant documents are retrieved by a rather small set of IR models. Table 6: Performance considering different rank hits Run Id MAP S@1 S@5 S@10 mcm25 (1) 21.68% (+17.38%*) 40.00% 78.67% 89.33% mcm32 (3) 18.98% (+2.76%) 38.67% 80.00% 85.33% mcm (5) 18.47% 41.33% 81.33% 86.67% mcm33 (7) 15.83% (-14.29%*) 37.33% 78.67% 85.33% mcm34 (9) 10.96% (-40.66%*) 36.11% 66.67% 70.83% mcm35 (10) 7.42% (-59.83%*) 39.22% 62.75% 64.70% For both runs mcm24 and mcm25, the number of successful documents was about 1000 and therefore, the computation time per query increased and became around 5 seconds. 5.2.2 Impact of the Variation of the Parameters Table 7 shows performance variation of the outranking approach when different preference thresholds are considered. We found performance improvement up to threshold values of about 5%, then there is a decrease in the performance which becomes significant for threshold values greater than 10%. Moreover, S@1 improves from 41.33% to 46.67% when preference threshold changes from 0 to 5%. We can thus conclude that the input rankings are semi orders rather than complete orders. Table 8 shows the evolution of the performance measures w.r.t. the concordance threshold. We can conclude that in order to put document di before di in the consensus ranking, Table 7: Impact of the variation of the preference threshold from 0 to 12.5% Run Id MAP S@1 S@5 S@10 mcm (0%) 18.47% 41.33% 81.33% 86.67% mcm1 (1%) 18.57% (+0.54%) 41.33% 81.33% 86.67% mcm2 (2.5%) 18.63% (+0.87%) 42.67% 78.67% 86.67% mcm3 (5%) 18.69% (+1.19%) 46.67% 81.33% 86.67% mcm4 (7.5%) 18.24% (-1.25%) 46.67% 81.33% 86.67% mcm5 (10%) 17.93% (-2.92%) 40.00% 82.67% 86.67% mcm5b (12.5%) 17.51% (-5.20%*) 41.33% 80.00% 86.67% at least half of the input rankings of PRi ∩ PRi should be concordant. Performance drops significantly for very low and very high values of the concordance threshold. In fact, for such values, the concordance condition is either fulfilled rather always by too many document pairs or not fulfilled at all, respectively. Therefore, the outranking relation becomes either too weak or too strong respectively. Table 8: Impact of the variation of cmin Run Id MAP S@1 S@5 S@10 mcm11 (20%) 17.63% (-4.55%*) 41.33% 76.00% 85.33% mcm12 (40%) 18.37% (-0.54%) 42.67% 76.00% 86.67% mcm (50%) 18.47% 41.33% 81.33% 86.67% mcm13 (60%) 18.42% (-0.27%) 40.00% 78.67% 86.67% mcm14 (80%) 17.43% (-5.63%*) 40.00% 78.67% 86.67% mcm15 (100%) 16.12% (-12.72%*) 41.33% 70.67% 85.33% In the experiments, varying the veto threshold as well as the discordance threshold within reasonable intervals does not have significant impact on performance measures. In fact, runs with different veto thresholds (sv ∈ [50%; 100%]) had similar performances even though there is a slight advantage for runs with high threshold values which means that it is better not to allow the input rankings to put their veto easily. Also, tuning the discordance threshold was carried out for values 50% and 75% of the veto threshold. For these runs we did not get any noticeable performance variation, although for low discordance thresholds (dmax < 20%), performance slightly decreased. 5.2.3 Impact of the Variation of the Number of Input Rankings To study performance evolution when different sets of input rankings are considered, we carried three more runs where 2, 4, and 6 of the best performing sets of the input rankings are considered. Results reported in Table 9 are seemingly counter-intuitive and also do not support previous findings regarding rank aggregation research [3]. Nevertheless, this result shows that low performing rankings bring more noise than information to the establishment of the consensus ranking. Therefore, when they are considered, performance decreases. Table 9: Performance considering different best performing sets of input rankings Run Id MAP S@1 S@5 S@10 mcm (10) 18.47% 41.33% 81.33% 86.67% mcm27 (6) 18.60% (+0.70%) 41.33% 80.00% 85.33% mcm28 (4) 19.02% (+2.98%) 40.00% 86.67% 88.00% mcm29 (2) 18.33% (-0.76%) 44.00% 76.00% 88.00% 5.2.4 Comparison of the Performance of Different Rank Aggregation Methods In this set of runs, we compare the outranking approach with some standard rank aggregation methods which were proven to have acceptable performance in previous studies: we considered two positional methods which are the CombSUM and the CombMNZ strategies. We also examined the performance of one majoritarian method which is the Markov chain method (MC4). For the comparisons, we considered a specific outranking relation S∗ = S(5%,50%,50%,30%) which results in good overall performances when tuning all the parameters. The first row of Table 10 gives performances of the rank aggregation methods w.r.t. a basic assumption set A1 = (H1 100, H2 5 , H4 new): we only consider the 100 first documents from each ranking, then retain documents present in 5 or more rankings and update ranks of successful documents. For positional methods, we place missing documents at the queue of the ranking (H3 yes) whereas for our method as well as for MC4, we retained hypothesis H3 no. The three following rows of Table 10 report performances when changing one element from the basic assumption set: the second row corresponds to the assumption set A2 = (H1 1000, H2 5 , H4 new), i.e. changing the number of retained documents from 100 to 1000. The third row corresponds to the assumption set A3 = (H1 100, H2 all, H4 new), i.e. considering the documents present in at least one ranking. The fourth row corresponds to the assumption set A4 = (H1 100, H2 5 , H4 init), i.e. keeping the original ranks of successful documents. The fifth row of Table 10, labeled A5, gives performance when all the 225 queries of the Web track of TREC-2004 are considered. Obviously, performance level cannot be compared with previous lines since the additional queries are different from the TD queries and correspond to other tasks (Home Page and Named Page tasks [10]) of TREC-2004 Web track. This set of runs aims to show whether relative performance of the various methods is task-dependent. The last row of Table 10, labeled A6, reports performance of the various methods considering the TD task of TREC2002 instead of TREC-2004: we fused the results of input rankings of the 10 best official runs for each of the 50 TD queries [9] considering the set of assumptions A1 of the first row. This aims to show whether relative performance of the various methods changes from year to year. Values between brackets of Table 10 are variations of performance of each rank aggregation method w.r.t. performance of the outranking approach. Table 10: Performance (MAP) of different rank aggregation methods under 3 different test collections mcm combSUM combMNZ markov A1 18.79% 17.54% (-6.65%*) 17.08% (-9.10%*) 18.63% (-0.85%) A2 21.36% 19.18% (-10.21%*) 18.61% (-12.87%*) 21.33% (-0.14%) A3 21.92% 21.38% (-2.46%) 20.88% (-4.74%) 19.35% (-11.72%*) A4 18.64% 17.58% (-5.69%*) 17.18% (-7.83%*) 18.63% (-0.05%) A5 55.39% 52.16% (-5.83%*) 49.70% (-10.27%*) 53.30% (-3.77%) A6 16.95% 15.65% (-7.67%*) 14.57% (-14.04%*) 16.39% (-3.30%) From the analysis of table 10 the following can be established: • for all the runs, considering all the documents in each input ranking (A2) significantly improves performance (MAP increases by 11.62% on average). This is predictable since some initially unreported relevant documents would receive better positions in the consensus ranking. • for all the runs, considering documents even those present in only one input ranking (A3) significantly improves performance. For mcm, combSUM and combMNZ, performance improvement is more important (MAP increases by 20.27% on average) than for the markov run (MAP increases by 3.86%). • preserving the initial positions of documents (A4) or recomputing them (A1) does not have a noticeable influence on performance for both positional and majoritarian methods. • considering all the queries of the Web track of TREC2004 (A5) as well as the TD queries of the Web track of TREC-2002 (A6) does not alter the relative performance of the different data fusion methods. • considering the TD queries of the Web track of TREC2002, performances of all the data fusion methods are lower than that of the best performing input ranking for which the MAP value equals 18.58%. This is because most of the fused input rankings have very low performances compared to the best one, which brings more noise to the consensus ranking. • performances of the data fusion methods mcm and markov are significantly better than that of the best input ranking uogWebCAU150. This remains true for runs combSUM and combMNZ only under assumptions H1 all or H2 all. This shows that majoritarian methods are less sensitive to assumptions than positional methods. • outranking approach always performs significantly better than positional methods combSUM and combMNZ. It has also better performances than the Markov chain method, especially under assumption H2 all where difference of performances becomes significant. 6. CONCLUSIONS In this paper, we address the rank aggregation problem where different, but not disjoint, lists of documents are to be fused. We noticed that the input rankings can hide ties, so they should not be considered as complete orders. Only robust information should be used from each input ranking. Current rank aggregation methods, and especially positional methods (e.g. combSUM [15]), are not initially designed to work with such rankings. They should be adapted by considering specific working assumptions. We propose a new outranking method for rank aggregation which is well adapted to the IR context. Indeed, it ranks two documents w.r.t. the intensity of their positions difference in each input ranking and also considering the number of the input rankings that are concordant and discordant in favor of a specific document. There is also no need to make specific assumptions on the positions of the missing documents. This is an important feature since the absence of a document from a ranking should not be necessarily interpreted negatively. Experimental results show that the outranking method significantly out-performs popular classical positional data fusion methods like combSUM and combMNZ strategies. It also out-performs a good performing majoritarian methods which is the Markov chain method. These results are tested against different test collections and queries. From the experiments, we can also conclude that in order to improve the performances, we should fuse result lists of well performing IR models, and that majoritarian data fusion methods perform better than positional methods. The proposed method can have a real impact on Web metasearch performances since only ranks are available from most primary search engines, whereas most of the current approaches need scores to merge result lists into one single list. Further work involves investigating whether the outranking approach performs well in various other contexts, e.g. using the document scores or some combination of document ranks and scores. Acknowledgments The authors would like to thank Jacques Savoy for his valuable comments on a preliminary version of this paper. 7. REFERENCES [1] A. Aronson, D. Demner-Fushman, S. Humphrey, J. Lin, H. Liu, P. Ruch, M. Ruiz, L. Smith, L. Tanabe, and W. Wilbur. Fusion of knowledge-intensive and statistical approaches for retrieving and annotating textual genomics documents. In Proceedings TREC"2005. NIST Publication, 2005. [2] R. A. Baeza-Yates and B. A. Ribeiro-Neto. Modern Information Retrieval. ACM Press , 1999. [3] B. T. Bartell, G. W. Cottrell, and R. K. Belew. Automatic combination of multiple ranked retrieval systems. In Proceedings ACM-SIGIR"94, pages 173-181. Springer-Verlag, 1994. [4] N. J. Belkin, P. Kantor, E. A. Fox, and J. A. Shaw. Combining evidence of multiple query representations for information retrieval. IPM, 31(3):431-448, 1995. [5] J. Borda. M´emoire sur les ´elections au scrutin. Histoire de l"Acad´emie des Sciences, 1781. [6] J. P. Callan, Z. Lu, and W. B. Croft. Searching distributed collections with inference networks. In Proceedings ACM-SIGIR"95, pages 21-28, 1995. [7] M. Condorcet. Essai sur l"application de l"analyse `a la probabilit´e des d´ecisions rendues `a la pluralit´e des voix. Imprimerie Royale, Paris, 1785. [8] W. D. Cook and M. Kress. Ordinal ranking with intensity of preference. Management Science, 31(1):26-32, 1985. [9] N. Craswell and D. Hawking. Overview of the TREC-2002 Web Track. In Proceedings TREC"2002. NIST Publication, 2002. [10] N. Craswell and D. Hawking. Overview of the TREC-2004 Web Track. In Proceedings of TREC"2004. NIST Publication, 2004. [11] C. Dwork, S. R. Kumar, M. Naor, and D. Sivakumar. Rank aggregation methods for the Web. In Proceedings WWW"2001, pages 613-622, 2001. [12] R. Fagin. Combining fuzzy information from multiple systems. JCSS, 58(1):83-99, 1999. [13] R. Fagin, R. Kumar, M. Mahdian, D. Sivakumar, and E. Vee. Comparing and aggregating rankings with ties. In PODS, pages 47-58, 2004. [14] R. Fagin, R. Kumar, and D. Sivakumar. Comparing top k lists. SIAM J. on Discrete Mathematics, 17(1):134-160, 2003. [15] E. A. Fox and J. A. Shaw. Combination of multiple searches. In Proceedings of TREC"3. NIST Publication, 1994. [16] J. Katzer, M. McGill, J. Tessier, W. Frakes, and P. DasGupta. A study of the overlap among document representations. Information Technology: Research and Development, 1(4):261-274, 1982. [17] L. S. Larkey, M. E. Connell, and J. Callan. Collection selection and results merging with topically organized U.S. patents and TREC data. In Proceedings ACM-CIKM"2000, pages 282-289. ACM Press, 2000. [18] A. Le Calv´e and J. Savoy. Database merging strategy based on logistic regression. IPM, 36(3):341-359, 2000. [19] J. H. Lee. Analyses of multiple evidence combination. In Proceedings ACM-SIGIR"97, pages 267-276, 1997. [20] D. Lillis, F. Toolan, R. Collier, and J. Dunnion. Probfuse: a probabilistic approach to data fusion. In Proceedings ACM-SIGIR"2006, pages 139-146. ACM Press, 2006. [21] J. I. Marden. Analyzing and Modeling Rank Data. Number 64 in Monographs on Statistics and Applied Probability. Chapman & Hall, 1995. [22] M. Montague and J. A. Aslam. Metasearch consistency. In Proceedings ACM-SIGIR"2001, pages 386-387. ACM Press, 2001. [23] D. M. Pennock and E. Horvitz. Analysis of the axiomatic foundations of collaborative filtering. In Workshop on AI for Electronic Commerce at the 16th National Conference on Artificial Intelligence, 1999. [24] M. E. Renda and U. Straccia. Web metasearch: rank vs. score based rank aggregation methods. In Proceedings ACM-SAC"2003, pages 841-846. ACM Press, 2003. [25] W. H. Riker. Liberalism against populism. Waveland Press, 1982. [26] B. Roy. The outranking approach and the foundations of ELECTRE methods. Theory and Decision, 31:49-73, 1991. [27] B. Roy and J. Hugonnard. Ranking of suburban line extension projects on the Paris metro system by a multicriteria method. Transportation Research, 16A(4):301-312, 1982. [28] L. Si and J. Callan. Using sampled data and regression to merge search engine results. In Proceedings ACM-SIGIR"2002, pages 19-26. ACM Press, 2002. [29] M. Truchon. An extension of the Condorcet criterion and Kemeny orders. Cahier 9813, Centre de Recherche en Economie et Finance Appliqu´ees, Oct. 1998. [30] H. Turtle and W. B. Croft. Inference networks for document retrieval. In Proceedings of ACM-SIGIR"90, pages 1-24. ACM Press, 1990. [31] C. C. Vogt and G. W. Cottrell. Fusion via a linear combination of scores. Information Retrieval, 1(3):151-173, 1999.

rank aggregation;multiple criterion framework;metasearch engine;combsum and combmnz strategy;ir model;datum fusion;decision rule;multiple criterium approach;outrank method;information retrieval;datum fusion problem;outranking approach;majoritarian method

train_H-52

Vocabulary Independent Spoken Term Detection

We are interested in retrieving information from speech data like broadcast news, telephone conversations and roundtable meetings. Today, most systems use large vocabulary continuous speech recognition tools to produce word transcripts; the transcripts are indexed and query terms are retrieved from the index. However, query terms that are not part of the recognizer"s vocabulary cannot be retrieved, and the recall of the search is affected. In addition to the output word transcript, advanced systems provide also phonetic transcripts, against which query terms can be matched phonetically. Such phonetic transcripts suffer from lower accuracy and cannot be an alternative to word transcripts. We present a vocabulary independent system that can handle arbitrary queries, exploiting the information provided by having both word transcripts and phonetic transcripts. A speech recognizer generates word confusion networks and phonetic lattices. The transcripts are indexed for query processing and ranking purpose. The value of the proposed method is demonstrated by the relative high performance of our system, which received the highest overall ranking for US English speech data in the recent NIST Spoken Term Detection evaluation [1].

1. INTRODUCTION The rapidly increasing amount of spoken data calls for solutions to index and search this data. The classical approach consists of converting the speech to word transcripts using a large vocabulary continuous speech recognition (LVCSR) tool. In the past decade, most of the research efforts on spoken data retrieval have focused on extending classical IR techniques to word transcripts. Some of these works have been done in the framework of the NIST TREC Spoken Document Retrieval tracks and are described by Garofolo et al. [12]. These tracks focused on retrieval from a corpus of broadcast news stories spoken by professionals. One of the conclusions of those tracks was that the effectiveness of retrieval mostly depends on the accuracy of the transcripts. While the accuracy of automatic speech recognition (ASR) systems depends on the scenario and environment, state-of-the-art systems achieved better than 90% accuracy in transcription of such data. In 2000, Garofolo et al. concluded that Spoken document retrieval is a solved problem [12]. However, a significant drawback of such approaches is that search on queries containing out-of-vocabulary (OOV) terms will not return any results. OOV terms are missing words from the ASR system vocabulary and are replaced in the output transcript by alternatives that are probable, given the recognition acoustic model and the language model. It has been experimentally observed that over 10% of user queries can contain OOV terms [16], as queries often relate to named entities that typically have a poor coverage in the ASR vocabulary. The effects of OOV query terms in spoken data retrieval are discussed by Woodland et al. [28]. In many applications the OOV rate may get worse over time unless the recognizer"s vocabulary is periodically updated. Another approach consists of converting the speech to phonetic transcripts and representing the query as a sequence of phones. The retrieval is based on searching the sequence of phones representing the query in the phonetic transcripts. The main drawback of this approach is the inherent high error rate of the transcripts. Therefore, such approach cannot be an alternative to word transcripts, especially for in-vocabulary (IV) query terms that are part of the vocabulary of the ASR system. A solution would be to combine the two different approaches presented above: we index both word transcripts and phonetic transcripts; during query processing, the information is retrieved from the word index for IV terms and from the phonetic index for OOV terms. We would like to be able to process also hybrid queries, i.e, queries that include both IV and OOV terms. Consequently, we need to merge pieces of information retrieved from word index and phonetic index. Proximity information on the occurrences of the query terms is required for phrase search and for proximity-based ranking. In classical IR, the index stores for each occurrence of a term, its offset. Therefore, we cannot merge posting lists retrieved by phonetic index with those retrieved by word index since the offset of the occurrences retrieved from the two different indices are not comparable. The only element of comparison between phonetic and word transcripts are the timestamps. No previous work combining word and phonetic approach has been done on phrase search. We present a novel scheme for information retrieval that consists of storing, during the indexing process, for each unit of indexing (phone or word) its timestamp. We search queries by merging the information retrieved from the two different indices, word index and phonetic index, according to the timestamps of the query terms. We analyze the retrieval effectiveness of this approach on the NIST Spoken Term Detection 2006 evaluation data [1]. The paper is organized as follows. We describe the audio processing in Section 2. The indexing and retrieval methods are presented in section 3. Experimental setup and results are given in Section 4. In Section 5, we give an overview of related work. Finally, we conclude in Section 6. 2. AUTOMATIC SPEECH RECOGNITION SYSTEM We use an ASR system for transcribing speech data. It works in speaker-independent mode. For best recognition results, a speaker-independent acoustic model and a language model are trained in advance on data with similar characteristics. Typically, ASR generates lattices that can be considered as directed acyclic graphs. Each vertex in a lattice is associated with a timestamp and each edge (u, v) is labeled with a word or phone hypothesis and its prior probability, which is the probability of the signal delimited by the timestamps of the vertices u and v, given the hypothesis. The 1-best path transcript is obtained from the lattice using dynamic programming techniques. Mangu et al. [18] and Hakkani-Tur et al. [13] propose a compact representation of a word lattice called word confusion network (WCN). Each edge (u, v) is labeled with a word hypothesis and its posterior probability, i.e., the probability of the word given the signal. One of the main advantages of WCN is that it also provides an alignment for all of the words in the lattice. As explained in [13], the three main steps for building a WCN from a word lattice are as follows: 1. Compute the posterior probabilities for all edges in the word lattice. 2. Extract a path from the word lattice (which can be the 1-best, the longest or any random path), and call it the pivot path of the alignment. 3. Traverse the word lattice, and align all the transitions with the pivot, merging the transitions that correspond to the same word (or label) and occur in the same time interval by summing their posterior probabilities. The 1-best path of a WCN is obtained from the path containing the best hypotheses. As stated in [18], although WCNs are more compact than word lattices, in general the 1-best path obtained from WCN has a better word accuracy than the 1-best path obtained from the corresponding word lattice. Typical structures of a lattice and a WCN are given in Figure 1. Figure 1: Typical structures of a lattice and a WCN. 3. RETRIEVAL MODEL The main problem with retrieving information from spoken data is the low accuracy of the transcription particularly on terms of interest such as named entities and content words. Generally, the accuracy of a word transcript is characterized by its word error rate (WER). There are three kinds of errors that can occur in a transcript: substitution of a term that is part of the speech by another term, deletion of a spoken term that is part of the speech and insertion of a term that is not part of the speech. Substitutions and deletions reflect the fact that an occurrence of a term in the speech signal is not recognized. These misses reduce the recall of the search. Substitutions and insertions reflect the fact that a term which is not part of the speech signal appears in the transcript. These misses reduce the precision of the search. Search recall can be enhanced by expanding the transcript with extra words. These words can be taken from the other alternatives provided by the WCN; these alternatives may have been spoken but were not the top choice of the ASR. Such an expansion tends to correct the substitutions and the deletions and consequently, might improve recall but will probably reduce precision. Using an appropriate ranking model, we can avoid the decrease in precision. Mamou et al. have presented in [17] the enhancement in the recall and the MAP by searching on WCN instead of considering only the 1-best path word transcript in the context of spoken document retrieval. We have adapted this model of IV search to term detection. In word transcripts, OOV terms are deleted or substituted. Therefore, the usage of phonetic transcripts is more desirable. However, due to their low accuracy, we have preferred to use only the 1-best path extracted from the phonetic lattices. We will show that the usage of phonetic transcripts tends to improve the recall without affecting the precision too much, using an appropriate ranking. 3.1 Spoken document detection task As stated in the STD 2006 evaluation plan [2], the task consists in finding all the exact matches of a specific query in a given corpus of speech data. A query is a phrase containing several words. The queries are text and not speech. Note that this task is different from the more classical task of spoken document retrieval. Manual transcripts of the speech are not provided but are used by the evaluators to find true occurrences. By definition, true occurrences of a query are found automatically by searching the manual transcripts using the following rule: the gap between adjacent words in a query must be less than 0.5 seconds in the corresponding speech. For evaluating the results, each system output occurrence is judged as correct or not according to whether it is close in time to a true occurrence of the query retrieved from manual transcripts; it is judged as correct if the midpoint of the system output occurrence is less than or equal to 0.5 seconds from the time span of a true occurrence of the query. 3.2 Indexing We have used the same indexing process for WCN and phonetic transcripts. Each occurrence of a unit of indexing (word or phone) u in a transcript D is indexed with the following information: • the begin time t of the occurrence of u, • the duration d of the occurrence of u. In addition, for WCN indexing, we store • the confidence level of the occurrence of u at the time t that is evaluated by its posterior probability Pr(u|t, D), • the rank of the occurrence of u among the other hypotheses beginning at the same time t, rank(u|t, D). Note that since the task is to find exact matches of the phrase queries, we have not filtered stopwords and the corpus is not stemmed before indexing. 3.3 Search In the following, we present our approach for accomplishing the STD task using the indices described above. The terms are extracted from the query. The vocabulary of the ASR system building word transcripts is given. Terms that are part of this vocabulary are IV terms; the other terms are OOV. For an IV query term, the posting list is extracted from the word index. For an OOV query term, the term is converted to a sequence of phones using a joint maximum entropy N-gram model [10]. For example, the term prosody is converted to the sequence of phones (p, r, aa, z, ih, d, iy). The posting list of each phone is extracted from the phonetic index. The next step consists of merging the different posting lists according to the timestamp of the occurrences in order to create results matching the query. First, we check that the words and phones appear in the right order according to their begin times. Second, we check that the gap in time between adjacent words and phones is reasonable. Conforming to the requirements of the STD evaluation, the distance in time between two adjacent query terms must be less than 0.5 seconds. For OOV search, we check that the distance in time between two adjacent phones of a query term is less that 0.2 seconds; this value has been determined empirically. In such a way, we can reduce the effect of insertion errors since we allow insertions between the adjacent words and phones. Our query processing does not allow substitutions and deletions. Example: Let us consider the phrase query prosody research. The term prosody is OOV and the term research is IV. The term prosody is converted to the sequence of phones (p, r, aa, z, ih, d, iy). The posting list of each phone is extracted from the phonetic index. We merge the posting lists of the phones such that the sequence of phones appears in the right order and the gap in time between the pairs of phones (p, r), (r, aa), (aa, z), (z, ih), (ih, d), (d, iy) is less than 0.2 seconds. We obtain occurrences of the term prosody. The posting list of research is extracted from the word index and we merge it with the occurrences found for prosody such that they appear in the right order and the distance in time between prosody and research is less than 0.5 seconds. Note that our indexing model allows to search for different types of queries: 1. queries containing only IV terms using the word index. 2. queries containing only OOV terms using the phonetic index. 3. keyword queries containing both IV and OOV terms using the word index for IV terms and the phonetic index for OOV terms; for query processing, the different sets of matches are unified if the query terms have OR semantics and intersected if the query terms have AND semantics. 4. phrase queries containing both IV and OOV terms; for query processing, the posting lists of the IV terms retrieved from the word index are merged with the posting lists of the OOV terms retrieved from the phonetic index. The merging is possible since we have stored the timestamps for each unit of indexing (word and phone) in both indices. The STD evaluation has focused on the fourth query type. It is the hardest task since we need to combine posting lists retrieved from phonetic and word indices. 3.4 Ranking Since IV terms and OOV terms are retrieved from two different indices, we propose two different functions for scoring an occurrence of a term; afterward, an aggregate score is assigned to the query based on the scores of the query terms. Because the task is term detection, we do not use a document frequency criterion for ranking the occurrences. Let us consider a query Q = (k0, ..., kn), associated with a boosting vector B = (B1, ..., Bj). This vector associates a boosting factor to each rank of the different hypotheses; the boosting factors are normalized between 0 and 1. If the rank r is larger than j, we assume Br = 0. 3.4.1 In vocabulary term ranking For IV term ranking, we extend the work of Mamou et al. [17] on spoken document retrieval to term detection. We use the information provided by the word index. We define the score score(k, t, D) of a keyword k occurring at a time t in the transcript D, by the following formula: score(k, t, D) = Brank(k|t,D) × Pr(k|t, D) Note that 0 ≤ score(k, t, D) ≤ 1. 3.4.2 Out of vocabulary term ranking For OOV term ranking, we use the information provided by the phonetic index. We give a higher rank to occurrences of OOV terms that contain phones close (in time) to each other. We define a scoring function that is related to the average gap in time between the different phones. Let us consider a keyword k converted to the sequence of phones (pk 0 , ..., pk l ). We define the normalized score score(k, tk 0 , D) of a keyword k = (pk 0 , ..., pk l ), where each pk i occurs at time tk i with a duration of dk i in the transcript D, by the following formula: score(k, tk 0 , D) = 1 − l i=1 5 × (tk i − (tk i−1 + dk i−1)) l Note that according to what we have ex-plained in Section 3.3, we have ∀1 ≤ i ≤ l, 0 < tk i − (tk i−1 + dk i−1) < 0.2 sec, 0 < 5 × (tk i − (tk i−1 + dk i−1)) < 1, and consequently, 0 < score(k, tk 0 , D) ≤ 1. The duration of the keyword occurrence is tk l − tk 0 + dk l . Example: let us consider the sequence (p, r, aa, z, ih, d, iy) and two different occurrences of the sequence. For each phone, we give the begin time and the duration in second. Occurrence 1: (p, 0.25, 0.01), (r, 0.36, 0.01), (aa, 0.37, 0.01), (z, 0.38, 0.01), (ih, 0.39, 0.01), (d, 0.4, 0.01), (iy, 0.52, 0.01). Occurrence 2: (p, 0.45, 0.01), (r, 0.46, 0.01), (aa, 0.47, 0.01), (z, 0.48, 0.01), (ih, 0.49, 0.01), (d, 0.5, 0.01), (iy, 0.51, 0.01). According to our formula, the score of the first occurrence is 0.83 and the score of the second occurrence is 1. In the first occurrence, there are probably some insertion or silence between the phone p and r, and between the phone d and iy. The silence can be due to the fact that the phones belongs to two different words ans therefore, it is not an occurrence of the term prosody. 3.4.3 Combination The score of an occurrence of a query Q at time t0 in the document D is determined by the multiplication of the score of each keyword ki, where each ki occurs at time ti with a duration di in the transcript D: score(Q, t0, D) = n i=0 score(ki, ti, D)γn Note that according to what we have ex-plained in Section 3.3, we have ∀1 ≤ i ≤ n, 0 < ti −(ti−1 +di−1) < 0.5 sec. Our goal is to estimate for each found occurrence how likely the query appears. It is different from classical IR that aims to rank the results and not to score them. Since the probability to have a false alarm is inversely proportional to the length of the phrase query, we have boosted the score of queries by a γn exponent, that is related to the number of keywords in the phrase. We have determined empirically the value of γn = 1/n. The begin time of the query occurrence is determined by the begin time t0 of the first query term and the duration of the query occurrence by tn − t0 + dn. 4. EXPERIMENTS 4.1 Experimental setup Our corpus consists of the evaluation set provided by NIST for the STD 2006 evaluation [1]. It includes three different source types in US English: three hours of broadcast news (BNEWS), three hours of conversational telephony speech (CTS) and two hours of conference room meetings (CONFMTG). As shown in Section 4.2, these different collections have different accuracies. CTS and CONFMTG are spontaneous speech. For the experiments, we have processed the query set provided by NIST that includes 1100 queries. Each query is a phrase containing between one to five terms, common and rare terms, terms that are in the manual transcripts and those that are not. Testing and determination of empirical values have been achieved on another set of speech data and queries, the development set, also provided by NIST. We have used the IBM research prototype ASR system, described in [26], for transcribing speech data. We have produced WCNs for the three different source types. 1-best phonetic transcripts were generated only for BNEWS and CTS, since CONFMTG phonetic transcripts have too low accuracy. We have adapted Juru [7], a full-text search library written in Java, to index the transcripts and to store the timestamps of the words and phones; search results have been retrieved as described in Section 3. For each found occurrence of the given query, our system outputs: the location of the term in the audio recording (begin time and duration), the score indicating how likely is the occurrence of query, (as defined in Section 3.4) and a hard (binary) decision as to whether the detection is correct. We measure precision and recall by comparing the results obtained over the automatic transcripts (only the results having true hard decision) to the results obtained over the reference manual transcripts. Our aim is to evaluate the ability of the suggested retrieval approach to handle transcribed speech data. Thus, the closer the automatic results to the manual results is, the better the search effectiveness over the automatic transcripts will be. The results returned from the manual transcription for a given query are considered relevant and are expected to be retrieved with highest scores. This approach for measuring search effectiveness using manual data as a reference is very common in speech retrieval research [25, 22, 8, 9, 17]. Beside the recall and the precision, we use the evaluation measures defined by NIST for the 2006 STD evaluation [2]: the Actual Term-Weighted Value (ATWV) and the Maximum Term-Weighted Value (MTWV). The term-weighted value (TWV) is computed by first computing the miss and false alarm probabilities for each query separately, then using these and an (arbitrarily chosen) prior probability to compute query-specific values, and finally averaging these query-specific values over all queries q to produce an overall system value: TWV (θ) = 1 − averageq{Pmiss(q, θ) + β × PF A(q, θ)} where β = C V (Pr−1 q − 1). θ is the detection threshold. For the evaluation, the cost/value ratio, C/V , has been determined to 0.1 and the prior probability of a query Prq to 10−4 . Therefore, β = 999.9. Miss and false alarm probabilities for a given query q are functions of θ: Pmiss(q, θ) = 1 − Ncorrect(q, θ) Ntrue(q) PF A(q, θ) = Nspurious(q, θ) NNT (q) corpus WER(%) SUBR(%) DELR(%) INSR(%) BNEWS WCN 12.7 49 42 9 CTS WCN 19.6 51 38 11 CONFMTG WCN 47.4 47 49 3 Table 1: WER and distribution of the error types over word 1-best path extracted from WCNs for the different source types. where: • Ncorrect(q, θ) is the number of correct detections (retrieved by the system) of the query q with a score greater than or equal to θ. • Nspurious(q, θ) is the number of spurious detections of the query q with a score greater than or equal to θ. • Ntrue(q) is the number of true occurrences of the query q in the corpus. • NNT (q) is the number of opportunities for incorrect detection of the query q in the corpus; it is the NonTarget query trials. It has been defined by the following formula: NNT (q) = Tspeech − Ntrue(q). Tspeech is the total amount of speech in the collection (in seconds). ATWV is the actual term-weighted value; it is the detection value attained by the system as a result of the system output and the binary decision output for each putative occurrence. It ranges from −∞ to +1. MTWV is the maximum term-weighted value over the range of all possible values of θ. It ranges from 0 to +1. We have also provided the detection error tradeoff (DET) curve [19] of miss probability (Pmiss) vs. false alarm probability (PF A). We have used the STDEval tool to extract the relevant results from the manual transcripts and to compute ATWV, MTWV and the DET curve. We have determined empirically the following values for the boosting vector defined in Section 3.4: Bi = 1 i . 4.2 WER analysis We use the word error rate (WER) in order to characterize the accuracy of the transcripts. WER is defined as follows: S + D + I N × 100 where N is the total number of words in the corpus, and S, I, and D are the total number of substitution, insertion, and deletion errors, respectively. The substitution error rate (SUBR) is defined by S S + D + I × 100. Deletion error rate (DELR) and insertion error rate (INSR) are defined in a similar manner. Table 1 gives the WER and the distribution of the error types over 1-best path transcripts extracted from WCNs. The WER of the 1-best path phonetic transcripts is approximately two times worse than the WER of word transcripts. That is the reason why we have not retrieved from phonetic transcripts on CONFMTG speech data. 4.3 Theta threshold We have determined empirically a detection threshold θ per source type and the hard decision of the occurrences having a score less than θ is set to false; false occurrences returned by the system are not considered as retrieved and therefore, are not used for computing ATWV, precision and recall. The value of the threshold θ per source type is reported in Table 2. It is correlated to the accuracy of the transcripts. Basically, setting a threshold aims to eliminate from the retrieved occurrences, false alarms without adding misses. The higher the WER is, the higher the θ threshold should be. BNEWS CTS CONFMTG 0.4 0.61 0.91 Table 2: Values of the θ threshold per source type. 4.4 Processing resource profile We report in Table 3 the processing resource profile. Concerning the index size, note that our index is compressed using IR index compression techniques. The indexing time includes both audio processing (generation of word and phonetic transcripts) and building of the searchable indices. Index size 0.3267 MB/HS Indexing time 7.5627 HP/HS Index Memory Usage 1653.4297 MB Search speed 0.0041 sec.P/HS Search Memory Usage 269.1250 MB Table 3: Processing resource profile. (HS: Hours of Speech. HP: Processing Hours. sec.P: Processing seconds) 4.5 Retrieval measures We compare our approach (WCN phonetic) presented in Section 4.1 with another approach (1-best-WCN phonetic). The only difference between these two approaches is that, in 1-best-WCN phonetic, we index only the 1-best path extracted from the WCN instead of indexing all the WCN. WCN phonetic was our primary system for the evaluation and 1-best-WCN phonetic was one of our contrastive systems. Average precision and recall, MTWV and ATWV on the 1100 queries are given in Table 4. We provide also the DET curve for WCN phonetic approach in Figure 2. The point that maximizes the TWV, the MTWV, is specified on each curve. Note that retrieval performance has been evaluated separately for each source type since the accuracy of the speech differs per source type as shown in Section 4.2. As expected, we can see that MTWV and ATWV decrease in higher WER. The retrieval performance is improved when measure BNEWS CTS CONFMTG WCN phonetic ATWV 0.8485 0.7392 0.2365 MTWV 0.8532 0.7408 0.2508 precision 0.94 0.90 0.65 recall 0.89 0.81 0.37 1-best-WCN phonetic ATWV 0.8279 0.7102 0.2381 MTWV 0.8319 0.7117 0.2512 precision 0.95 0.91 0.66 recall 0.84 0.75 0.37 Table 4: ATWV, MTWV, precision and recall per source type. Figure 2: DET curve for WCN phonetic approach. using WCNs relatively to 1-best path. It is due to the fact that miss probability is improved by indexing all the hypotheses provided by the WCNs. This observation confirms the results shown by Mamou et al. [17] in the context of spoken document retrieval. The ATWV that we have obtained is close to the MTWV; we have combined our ranking model with appropriate threshold θ to eliminate results with lower score. Therefore, the effect of false alarms added by WCNs is reduced. WCN phonetic approach was used in the recent NIST STD evaluation and received the highest overall ranking among eleven participants. For comparison, the system that ranked at the third place, obtained an ATWV of 0.8238 for BNEWS, 0.6652 for CTS and 0.1103 for CONFMTG. 4.6 Influence of the duration of the query on the retrieval performance We have analysed the retrieval performance according to the average duration of the occurrences in the manual transcripts. The query set was divided into three different quantiles according to the duration; we have reported in Table 5 ATWV and MTWV according to the duration. We can see that we performed better on longer queries. One of the reasons is the fact that the ASR system is more accurate on long words. Hence, it was justified to boost the score of the results with the exponent γn, as explained in Section 3.4.3, according to the length of the query. quantile 0-33 33-66 66-100 BNEWS ATWV 0.7655 0.8794 0.9088 MTWV 0.7819 0.8914 0.9124 CTS ATWV 0.6545 0.8308 0.8378 MTWV 0.6551 0.8727 0.8479 CONFMTG ATWV 0.1677 0.3493 0.3651 MTWV 0.1955 0.4109 0.3880 Table 5: ATWV, MTWV according to the duration of the query occurrences per source type. 4.7 OOV vs. IV query processing We have randomly chosen three sets of queries from the query sets provided by NIST: 50 queries containing only IV terms; 50 queries containing only OOV terms; and 50 hybrid queries containing both IV and OOV terms. The following experiment has been achieved on the BNEWS collection and IV and OOV terms has been determined according to the vocabulary of BNEWS ASR system. We would like to compare three different approaches of retrieval: using only word index; using only phonetic index; combining word and phonetic indices. Table 6 summarizes the retrieval performance according to each approach and to each type of queries. Using a word-based approach for dealing with OOV and hybrid queries affects drastically the performance of the retrieval; precision and recall are null. Using a phone-based approach for dealing with IV queries affects also the performance of the retrieval relatively to the word-based approach. As expected, the approach combining word and phonetic indices presented in Section 3 leads to the same retrieval performance as the word approach for IV queries and to the same retrieval performance as the phonetic approach for OOV queries. This approach always outperforms the others and it justifies the fact that we need to combine word and phonetic search. 5. RELATED WORK In the past decade, the research efforts on spoken data retrieval have focused on extending classical IR techniques to spoken documents. Some of these works have been done in the context of the TREC Spoken Document Retrieval evaluations and are described by Garofolo et al. [12]. An LVCSR system is used to transcribe the speech into 1-best path word transcripts. The transcripts are indexed as clean text: for each occurrence, its document, its word offset and additional information are stored in the index. A generic IR system over the text is used for word spotting and search as described by Brown et al. [6] and James [14]. This stratindex word phonetic word and phonetic precision recall precision recall precision recall IV queries 0.8 0.96 0.11 0.77 0.8 0.96 OOV queries 0 0 0.13 0.79 0.13 0.79 hybrid queries 0 0 0.15 0.71 0.89 0.83 Table 6: Comparison of word and phonetic approach on IV and OOV queries egy works well for transcripts like broadcast news collections that have a low WER (in the range of 15%-30%) and are redundant by nature (the same piece of information is spoken several times in different manners). Moreover, the algorithms have been mostly tested over long queries stated in plain English and retrieval for such queries is more robust against speech recognition errors. An alternative approach consists of using word lattices in order to improve the effectiveness of SDR. Singhal et al. [24, 25] propose to add some terms to the transcript in order to alleviate the retrieval failures due to ASR errors. From an IR perspective, a classical way to bring new terms is document expansion using a similar corpus. Their approach consists in using word lattices in order to determine which words returned by a document expansion algorithm should be added to the original transcript. The necessity to use a document expansion algorithm was justified by the fact that the word lattices they worked with, lack information about word probabilities. Chelba and Acero in [8, 9] propose a more compact word lattice, the position specific posterior lattice (PSPL). This data structure is similar to WCN and leads to a more compact index. The offset of the terms in the speech documents is also stored in the index. However, the evaluation framework is carried out on lectures that are relatively planned, in contrast to conversational speech. Their ranking model is based on the term confidence level but does not take into consideration the rank of the term among the other hypotheses. Mamou et al. [17] propose a model for spoken document retrieval using WCNs in order to improve the recall and the MAP of the search. However, in the above works, the problem of queries containing OOV terms is not addressed. Popular approaches to deal with OOV queries are based on sub-words transcripts, where the sub-words are typically phones, syllables or word fragments (sequences of phones) [11, 20, 23]. The classical approach consists of using phonetic transcripts. The transcripts are indexed in the same manner as words in using classical text retrieval techniques; during query processing, the query is represented as a sequence of phones. The retrieval is based on searching the string of phones representing the query in the phonetic transcript. To account for the high recognition error rates, some other systems use richer transcripts like phonetic lattices. They are attractive as they accommodate high error rate conditions as well as allow for OOV queries to be used [15, 3, 20, 23, 21, 27]. However, phonetic lattices contain many edges that overlap in time with the same phonetic label, and are difficult to index. Moreover, beside the improvement in the recall of the search, the precision is affected since phonetic lattices are often inaccurate. Consequently, phonetic approaches should be used only for OOV search; for searching queries containing also IV terms, this technique affects the performance of the retrieval in comparison to the word based approach. Saraclar and Sproat in [22] show improvement in word spotting accuracy for both IV and OOV queries, using phonetic and word lattices, where a confidence measure of a word or a phone can be derived. They propose three different retrieval strategies: search both the word and the phonetic indices and unify the two different sets of results; search the word index for IV queries, search the phonetic index for OOV queries; search the word index and if no result is returned, search the phonetic index. However, no strategy is proposed to deal with phrase queries containing both IV and OOV terms. Amir et al. in [5, 4] propose to merge a word approach with a phonetic approach in the context of video retrieval. However, the phonetic transcript is obtained from a text to phonetic conversion of the 1-best path of the word transcript and is not based on a phonetic decoding of the speech data. An important issue to be considered when looking at the state-of-the-art in retrieval of spoken data, is the lack of a common test set and appropriate query terms. This paper uses such a task and the STD evaluation is a good summary of the performance of different approaches on the same test conditions. 6. CONCLUSIONS This work studies how vocabulary independent spoken term detection can be performed efficiently over different data sources. Previously, phonetic-based and word-based approaches have been used for IR on speech data. The former suffers from low accuracy and the latter from limited vocabulary of the recognition system. In this paper, we have presented a vocabulary independent model of indexing and search that combines both the approaches. The system can deal with all kinds of queries although the phrases that need to combine for the retrieval, information extracted from two different indices, a word index and a phonetic index. The scoring of OOV terms is based on the proximity (in time) between the different phones. The scoring of IV terms is based on information provided by the WCNs. We have shown an improvement in the retrieval performance when using all the WCN and not only the 1-best path and when using phonetic index for search of OOV query terms. This approach always outperforms the other approaches using only word index or phonetic index. As a future work, we will compare our model for OOV search on phonetic transcripts with a retrieval model based on the edit distance. 7. ACKNOWLEDGEMENTS Jonathan Mamou is grateful to David Carmel and Ron Hoory for helpful and interesting discussions. 8. REFERENCES [1] NIST Spoken Term Detection 2006 Evaluation Website, http://www.nist.gov/speech/tests/std/. [2] NIST Spoken Term Detection (STD) 2006 Evaluation Plan, http://www.nist.gov/speech/tests/std/docs/std06-evalplan-v10.pdf. [3] C. Allauzen, M. Mohri, and M. Saraclar. General indexation of weighted automata - application to spoken utterance retrieval. In Proceedings of the HLT-NAACL 2004 Workshop on Interdiciplinary Approaches to Speech Indexing and Retrieval, Boston, MA, USA, 2004. [4] A. Amir, M. Berg, and H. Permuter. Mutual relevance feedback for multimodal query formulation in video retrieval. In MIR "05: Proceedings of the 7th ACM SIGMM international workshop on Multimedia information retrieval, pages 17-24, New York, NY, USA, 2005. ACM Press. [5] A. Amir, A. Efrat, and S. Srinivasan. Advances in phonetic word spotting. In CIKM "01: Proceedings of the tenth international conference on Information and knowledge management, pages 580-582, New York, NY, USA, 2001. ACM Press. [6] M. Brown, J. Foote, G. Jones, K. Jones, and S. Young. Open-vocabulary speech indexing for voice and video mail retrieval. In Proceedings ACM Multimedia 96, pages 307-316, Hong-Kong, November 1996. [7] D. Carmel, E. Amitay, M. Herscovici, Y. S. Maarek, Y. Petruschka, and A. Soffer. Juru at TREC 10Experiments with Index Pruning. In Proceedings of the Tenth Text Retrieval Conference (TREC-10). National Institute of Standards and Technology. NIST, 2001. [8] C. Chelba and A. Acero. Indexing uncertainty for spoken document search. In Interspeech 2005, pages 61-64, Lisbon, Portugal, 2005. [9] C. Chelba and A. Acero. Position specific posterior lattices for indexing speech. In Proceedings of the 43rd Annual Conference of the Association for Computational Linguistics (ACL), Ann Arbor, MI, 2005. [10] S. Chen. Conditional and joint models for grapheme-to-phoneme conversion. In Eurospeech 2003, Geneva, Switzerland, 2003. [11] M. Clements, S. Robertson, and M. Miller. Phonetic searching applied to on-line distance learning modules. In Digital Signal Processing Workshop, 2002 and the 2nd Signal Processing Education Workshop. Proceedings of 2002 IEEE 10th, pages 186-191, 2002. [12] J. Garofolo, G. Auzanne, and E. Voorhees. The TREC spoken document retrieval track: A success story. In Proceedings of the Ninth Text Retrieval Conference (TREC-9). National Institute of Standards and Technology. NIST, 2000. [13] D. Hakkani-Tur and G. Riccardi. A general algorithm for word graph matrix decomposition. In Proceedings of the IEEE Internation Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 596-599, Hong-Kong, 2003. [14] D. James. The application of classical information retrieval techniques to spoken documents. PhD thesis, University of Cambridge, Downing College, 1995. [15] D. A. James. A system for unrestricted topic retrieval from radio news broadcasts. In Proc. ICASSP "96, pages 279-282, Atlanta, GA, 1996. [16] B. Logan, P. Moreno, J. V. Thong, and E. Whittaker. An experimental study of an audio indexing system for the web. In Proceedings of ICSLP, 1996. [17] J. Mamou, D. Carmel, and R. Hoory. Spoken document retrieval from call-center conversations. In SIGIR "06: Proceedings of the 29th annual international ACM SIGIR conference on Research and development in information retrieval, pages 51-58, New York, NY, USA, 2006. ACM Press. [18] L. Mangu, E. Brill, and A. Stolcke. Finding consensus in speech recognition: word error minimization and other applications of confusion networks. Computer Speech and Language, 14(4):373-400, 2000. [19] A. Martin, G. Doddington, T. Kamm, M. Ordowski, and M. Przybocki. The DET curve in assessment of detection task performance. In Proc. Eurospeech "97, pages 1895-1898, Rhodes, Greece, 1997. [20] K. Ng and V. W. Zue. Subword-based approaches for spoken document retrieval. Speech Commun., 32(3):157-186, 2000. [21] Y. Peng and F. Seide. Fast two-stage vocabulary-independent search in spontaneous speech. In Acoustics, Speech, and Signal Processing. Proceedings. (ICASSP). IEEE International Conference, volume 1, pages 481-484, 2005. [22] M. Saraclar and R. Sproat. Lattice-based search for spoken utterance retrieval. In HLT-NAACL 2004: Main Proceedings, pages 129-136, Boston, Massachusetts, USA, 2004. [23] F. Seide, P. Yu, C. Ma, and E. Chang. Vocabulary-independent search in spontaneous speech. In ICASSP-2004, IEEE International Conference on Acoustics, Speech, and Signal Processing, 2004. [24] A. Singhal, J. Choi, D. Hindle, D. Lewis, and F. Pereira. AT&T at TREC-7. In Proceedings of the Seventh Text Retrieval Conference (TREC-7). National Institute of Standards and Technology. NIST, 1999. [25] A. Singhal and F. Pereira. Document expansion for speech retrieval. In SIGIR "99: Proceedings of the 22nd annual international ACM SIGIR conference on research and development in information retrieval, pages 34-41, New York, NY, USA, 1999. ACM Press. [26] H. Soltau, B. Kingsbury, L. Mangu, D. Povey, G. Saon, and G. Zweig. The IBM 2004 conversational telephony system for rich transcription. In Proceedings of the IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), March 2005. [27] K. Thambiratnam and S. Sridharan. Dynamic match phone-lattice searches for very fast and accurate unrestricted vocabulary keyword spotting. In Acoustics, Speech, and Signal Processing. Proceedings. (ICASSP). IEEE International Conference, 2005. [28] P. C. Woodland, S. E. Johnson, P. Jourlin, and K. S. Jones. Effects of out of vocabulary words in spoken document retrieval (poster session). In SIGIR "00: Proceedings of the 23rd annual international ACM SIGIR conference on Research and development in information retrieval, pages 372-374, New York, NY, USA, 2000. ACM Press.

vocabulary independent system;out-of-vocabulary;vocabulary;indexing timestamp;phonetic index;oov search;speech retrieval;speak term detection;speech datum retrieval;index merging;automatic speech recognition;speech recognizer;word index;phonetic transcript;spoken term detection

train_H-53

Context Sensitive Stemming for Web Search

Traditionally, stemming has been applied to Information Retrieval tasks by transforming words in documents to the their root form before indexing, and applying a similar transformation to query terms. Although it increases recall, this naive strategy does not work well for Web Search since it lowers precision and requires a significant amount of additional computation. In this paper, we propose a context sensitive stemming method that addresses these two issues. Two unique properties make our approach feasible for Web Search. First, based on statistical language modeling, we perform context sensitive analysis on the query side. We accurately predict which of its morphological variants is useful to expand a query term with before submitting the query to the search engine. This dramatically reduces the number of bad expansions, which in turn reduces the cost of additional computation and improves the precision at the same time. Second, our approach performs a context sensitive document matching for those expanded variants. This conservative strategy serves as a safeguard against spurious stemming, and it turns out to be very important for improving precision. Using word pluralization handling as an example of our stemming approach, our experiments on a major Web search engine show that stemming only 29% of the query traffic, we can improve relevance as measured by average Discounted Cumulative Gain (DCG5) by 6.1% on these queries and 1.8% over all query traffic.

1. INTRODUCTION Web search has now become a major tool in our daily lives for information seeking. One of the important issues in Web search is that user queries are often not best formulated to get optimal results. For example, running shoe is a query that occurs frequently in query logs. However, the query running shoes is much more likely to give better search results than the original query because documents matching the intent of this query usually contain the words running shoes. Correctly formulating a query requires the user to accurately predict which word form is used in the documents that best satisfy his or her information needs. This is difficult even for experienced users, and especially difficult for non-native speakers. One traditional solution is to use stemming [16, 18], the process of transforming inflected or derived words to their root form so that a search term will match and retrieve documents containing all forms of the term. Thus, the word run will match running, ran, runs, and shoe will match shoes and shoeing. Stemming can be done either on the terms in a document during indexing (and applying the same transformation to the query terms during query processing) or by expanding the query with the variants during query processing. Stemming during indexing allows very little flexibility during query processing, while stemming by query expansion allows handling each query differently, and hence is preferred. Although traditional stemming increases recall by matching word variants [13], it can reduce precision by retrieving too many documents that have been incorrectly matched. When examining the results of applying stemming to a large number of queries, one usually finds that nearly equal numbers of queries are helped and hurt by the technique [6]. In addition, it reduces system performance because the search engine has to match all the word variants. As we will show in the experiments, this is true even if we simplify stemming to pluralization handling, which is the process of converting a word from its plural to singular form, or vice versa. Thus, one needs to be very cautious when using stemming in Web search engines. One problem of traditional stemming is its blind transformation of all query terms, that is, it always performs the same transformation for the same query word without considering the context of the word. For example, the word book has four forms book, books, booking, booked, and store has four forms store, stores, storing, stored. For the query book store, expanding both words to all of their variants significantly increases computation cost and hurts precision, since not all of the variants are useful for this query. Transforming book store to match book stores is fine, but matching book storing or booking store is not. A weighting method that gives variant words smaller weights alleviates the problems to a certain extent if the weights accurately reflect the importance of the variant in this particular query. However uniform weighting is not going to work and a query dependent weighting is still a challenging unsolved problem [20]. A second problem of traditional stemming is its blind matching of all occurrences in documents. For the query book store, a transformation that allows the variant stores to be matched will cause every occurrence of stores in the document to be treated equivalent to the query term store. Thus, a document containing the fragment reading a book in coffee stores will be matched, causing many wrong documents to be selected. Although we hope the ranking function can correctly handle these, with many more candidates to rank, the risk of making mistakes increases. To alleviate these two problems, we propose a context sensitive stemming approach for Web search. Our solution consists of two context sensitive analysis, one on the query side and the other on the document side. On the query side, we propose a statistical language modeling based approach to predict which word variants are better forms than the original word for search purpose and expanding the query with only those forms. On the document side, we propose a conservative context sensitive matching for the transformed word variants, only matching document occurrences in the context of other terms in the query. Our model is simple yet effective and efficient, making it feasible to be used in real commercial Web search engines. We use pluralization handling as a running example for our stemming approach. The motivation for using pluralization handling as an example is to show that even such simple stemming, if handled correctly, can give significant benefits to search relevance. As far as we know, no previous research has systematically investigated the usage of pluralization in Web search. As we have to point out, the method we propose is not limited to pluralization handling, it is a general stemming technique, and can also be applied to general query expansion. Experiments on general stemming yield additional significant improvements over pluralization handling for long queries, although details will not be reported in this paper. In the rest of the paper, we first present the related work and distinguish our method from previous work in Section 2. We describe the details of the context sensitive stemming approach in Section 3. We then perform extensive experiments on a major Web search engine to support our claims in Section 4, followed by discussions in Section 5. Finally, we conclude the paper in Section 6. 2. RELATED WORK Stemming is a long studied technology. Many stemmers have been developed, such as the Lovins stemmer [16] and the Porter stemmer [18]. The Porter stemmer is widely used due to its simplicity and effectiveness in many applications. However, the Porter stemming makes many mistakes because its simple rules cannot fully describe English morphology. Corpus analysis is used to improve Porter stemmer [26] by creating equivalence classes for words that are morphologically similar and occur in similar context as measured by expected mutual information [23]. We use a similar corpus based approach for stemming by computing the similarity between two words based on their distributional context features which can be more than just adjacent words [15], and then only keep the morphologically similar words as candidates. Using stemming in information retrieval is also a well known technique [8, 10]. However, the effectiveness of stemming for English query systems was previously reported to be rather limited. Lennon et al. [17] compared the Lovins and Porter algorithms and found little improvement in retrieval performance. Later, Harman [9] compares three general stemming techniques in text retrieval experiments including pluralization handing (called S stemmer in the paper). They also proposed selective stemming based on query length and term importance, but no positive results were reported. On the other hand, Krovetz [14] performed comparisons over small numbers of documents (from 400 to 12k) and showed dramatic precision improvement (up to 45%). However, due to the limited number of tested queries (less than 100) and the small size of the collection, the results are hard to generalize to Web search. These mixed results, mostly failures, led early IR researchers to deem stemming irrelevant in general for English [4], although recent research has shown stemming has greater benefits for retrieval in other languages [2]. We suspect the previous failures were mainly due to the two problems we mentioned in the introduction. Blind stemming, or a simple query length based selective stemming as used in [9] is not enough. Stemming has to be decided on case by case basis, not only at the query level but also at the document level. As we will show, if handled correctly, significant improvement can be achieved. A more general problem related to stemming is query reformulation [3, 12] and query expansion which expands words not only with word variants [7, 22, 24, 25]. To decide which expanded words to use, people often use pseudorelevance feedback techniquesthat send the original query to a search engine and retrieve the top documents, extract relevant words from these top documents as additional query words, and resubmit the expanded query again [21]. This normally requires sending a query multiple times to search engine and it is not cost effective for processing the huge amount of queries involved in Web search. In addition, query expansion, including query reformulation [3, 12], has a high risk of changing the user intent (called query drift). Since the expanded words may have different meanings, adding them to the query could potentially change the intent of the original query. Thus query expansion based on pseudorelevance and query reformulation can provide suggestions to users for interactive refinement but can hardly be directly used for Web search. On the other hand, stemming is much more conservative since most of the time, stemming preserves the original search intent. While most work on query expansion focuses on recall enhancement, our work focuses on increasing both recall and precision. The increase on recall is obvious. With quality stemming, good documents which were not selected before stemming will be pushed up and those low quality documents will be degraded. On selective query expansion, Cronen-Townsend et al. [6] proposed a method for selective query expansion based on comparing the Kullback-Leibler divergence of the results from the unexpanded query and the results from the expanded query. This is similar to the relevance feedback in the sense that it requires multiple passes retrieval. If a word can be expanded into several words, it requires running this process multiple times to decide which expanded word is useful. It is expensive to deploy this in production Web search engines. Our method predicts the quality of expansion based on offline information without sending the query to a search engine. In summary, we propose a novel approach to attack an old, yet still important and challenging problem for Web search - stemming. Our approach is unique in that it performs predictive stemming on a per query basis without relevance feedback from the Web, using the context of the variants in documents to preserve precision. It"s simple, yet very efficient and effective, making real time stemming feasible for Web search. Our results will affirm researchers that stemming is indeed very important to large scale information retrieval. 3. CONTEXT SENSITIVE STEMMING 3.1 Overview Our system has four components as illustrated in Figure 1: candidate generation, query segmentation and head word detection, context sensitive query stemming and context sensitive document matching. Candidate generation (component 1) is performed offline and generated candidates are stored in a dictionary. For an input query, we first segment query into concepts and detect the head word for each concept (component 2). We then use statistical language modeling to decide whether a particular variant is useful (component 3), and finally for the expanded variants, we perform context sensitive document matching (component 4). Below we discuss each of the components in more detail. Component 4: context sensitive document matching Input Query: and head word detection Component 2: segment Component 1: candidate generation comparisons −> comparison Component 3: selective word expansion decision: comparisons −> comparison example: hotel price comparisons output: "hotel" "comparisons" hotel −> hotels Figure 1: System Components 3.2 Expansion candidate generation One of the ways to generate candidates is using the Porter stemmer [18]. The Porter stemmer simply uses morphological rules to convert a word to its base form. It has no knowledge of the semantic meaning of the words and sometimes makes serious mistakes, such as executive to execution, news to new, and paste to past. A more conservative way is based on using corpus analysis to improve the Porter stemmer results [26]. The corpus analysis we do is based on word distributional similarity [15]. The rationale of using distributional word similarity is that true variants tend to be used in similar contexts. In the distributional word similarity calculation, each word is represented with a vector of features derived from the context of the word. We use the bigrams to the left and right of the word as its context features, by mining a huge Web corpus. The similarity between two words is the cosine similarity between the two corresponding feature vectors. The top 20 similar words to develop is shown in the following table. rank candidate score rank candidate score 0 develop 1 10 berts 0.119 1 developing 0.339 11 wads 0.116 2 developed 0.176 12 developer 0.107 3 incubator 0.160 13 promoting 0.100 4 develops 0.150 14 developmental 0.091 5 development 0.148 15 reengineering 0.090 6 tutoring 0.138 16 build 0.083 7 analyzing 0.128 17 construct 0.081 8 developement 0.128 18 educational 0.081 9 automation 0.126 19 institute 0.077 Table 1: Top 20 most similar candidates to word develop. Column score is the similarity score. To determine the stemming candidates, we apply a few Porter stemmer [18] morphological rules to the similarity list. After applying these rules, for the word develop, the stemming candidates are developing, developed, develops, development, developement, developer, developmental. For the pluralization handling purpose, only the candidate develops is retained. One thing we note from observing the distributionally similar words is that they are closely related semantically. These words might serve as candidates for general query expansion, a topic we will investigate in the future. 3.3 Segmentation and headword identification For long queries, it is quite important to detect the concepts in the query and the most important words for those concepts. We first break a query into segments, each segment representing a concept which normally is a noun phrase. For each of the noun phrases, we then detect the most important word which we call the head word. Segmentation is also used in document sensitive matching (section 3.5) to enforce proximity. To break a query into segments, we have to define a criteria to measure the strength of the relation between words. One effective method is to use mutual information as an indicator on whether or not to split two words [19]. We use a log of 25M queries and collect the bigram and unigram frequencies from it. For every incoming query, we compute the mutual information of two adjacent words; if it passes a predefined threshold, we do not split the query between those two words and move on to next word. We continue this process until the mutual information between two words is below the threshold, then create a concept boundary here. Table 2 shows some examples of query segmentation. The ideal way of finding the head word of a concept is to do syntactic parsing to determine the dependency structure of the query. Query parsing is more difficult than sentence [running shoe] [best] [new york] [medical schools] [pictures] [of] [white house] [cookies] [in] [san francisco] [hotel] [price comparison] Table 2: Query segmentation: a segment is bracketed. parsing since many queries are not grammatical and are very short. Applying a parser trained on sentences from documents to queries will have poor performance. In our solution, we just use simple heuristics rules, and it works very well in practice for English. For an English noun phrase, the head word is typically the last nonstop word, unless the phrase is of a particular pattern, like XYZ of/in/at/from UVW. In such cases, the head word is typically the last nonstop word of XYZ. 3.4 Context sensitive word expansion After detecting which words are the most important words to expand, we have to decide whether the expansions will be useful. Our statistics show that about half of the queries can be transformed by pluralization via naive stemming. Among this half, about 25% of the queries improve relevance when transformed, the majority (about 50%) do not change their top 5 results, and the remaining 25% perform worse. Thus, it is extremely important to identify which queries should not be stemmed for the purpose of maximizing relevance improvement and minimizing stemming cost. In addition, for a query with multiple words that can be transformed, or a word with multiple variants, not all of the expansions are useful. Taking query hotel price comparison as an example, we decide that hotel and price comparison are two concepts. Head words hotel and comparison can be expanded to hotels and comparisons. Are both transformations useful? To test whether an expansion is useful, we have to know whether the expanded query is likely to get more relevant documents from the Web, which can be quantified by the probability of the query occurring as a string on the Web. The more likely a query to occur on the Web, the more relevant documents this query is able to return. Now the whole problem becomes how to calculate the probability of query to occur on the Web. Calculating the probability of string occurring in a corpus is a well known language modeling problem. The goal of language modeling is to predict the probability of naturally occurring word sequences, s = w1w2...wN ; or more simply, to put high probability on word sequences that actually occur (and low probability on word sequences that never occur). The simplest and most successful approach to language modeling is still based on the n-gram model. By the chain rule of probability one can write the probability of any word sequence as Pr(w1w2...wN ) = NY i=1 Pr(wi|w1...wi−1) (1) An n-gram model approximates this probability by assuming that the only words relevant to predicting Pr(wi|w1...wi−1) are the previous n − 1 words; i.e. Pr(wi|w1...wi−1) = Pr(wi|wi−n+1...wi−1) A straightforward maximum likelihood estimate of n-gram probabilities from a corpus is given by the observed frequency of each of the patterns Pr(wi|wi−n+1...wi−1) = #(wi−n+1...wi) #(wi−n+1...wi−1) (2) where #(.) denotes the number of occurrences of a specified gram in the training corpus. Although one could attempt to use simple n-gram models to capture long range dependencies in language, attempting to do so directly immediately creates sparse data problems: Using grams of length up to n entails estimating the probability of Wn events, where W is the size of the word vocabulary. This quickly overwhelms modern computational and data resources for even modest choices of n (beyond 3 to 6). Also, because of the heavy tailed nature of language (i.e. Zipf"s law) one is likely to encounter novel n-grams that were never witnessed during training in any test corpus, and therefore some mechanism for assigning non-zero probability to novel n-grams is a central and unavoidable issue in statistical language modeling. One standard approach to smoothing probability estimates to cope with sparse data problems (and to cope with potentially missing n-grams) is to use some sort of back-off estimator. Pr(wi|wi−n+1...wi−1) = 8 >>< >>: ˆPr(wi|wi−n+1...wi−1), if #(wi−n+1...wi) > 0 β(wi−n+1...wi−1) × Pr(wi|wi−n+2...wi−1), otherwise (3) where ˆPr(wi|wi−n+1...wi−1) = discount #(wi−n+1...wi) #(wi−n+1...wi−1) (4) is the discounted probability and β(wi−n+1...wi−1) is a normalization constant β(wi−n+1...wi−1) = 1 − X x∈(wi−n+1...wi−1x) ˆPr(x|wi−n+1...wi−1) 1 − X x∈(wi−n+1...wi−1x) ˆPr(x|wi−n+2...wi−1) (5) The discounted probability (4) can be computed with different smoothing techniques, including absolute smoothing, Good-Turing smoothing, linear smoothing, and Witten-Bell smoothing [5]. We used absolute smoothing in our experiments. Since the likelihood of a string, Pr(w1w2...wN ), is a very small number and hard to interpret, we use entropy as defined below to score the string. Entropy = − 1 N log2 Pr(w1w2...wN ) (6) Now getting back to the example of the query hotel price comparisons, there are four variants of this query, and the entropy of these four candidates are shown in Table 3. We can see that all alternatives are less likely than the input query. It is therefore not useful to make an expansion for this query. On the other hand, if the input query is hotel price comparisons which is the second alternative in the table, then there is a better alternative than the input query, and it should therefore be expanded. To tolerate the variations in probability estimation, we relax the selection criterion to those query alternatives if their scores are within a certain distance (10% in our experiments) to the best score. Query variations Entropy hotel price comparison 6.177 hotel price comparisons 6.597 hotels price comparison 6.937 hotels price comparisons 7.360 Table 3: Variations of query hotel price comparison ranked by entropy score, with the original query in bold face. 3.5 Context sensitive document matching Even after we know which word variants are likely to be useful, we have to be conservative in document matching for the expanded variants. For the query hotel price comparisons, we decided that word comparisons is expanded to include comparison. However, not every occurrence of comparison in the document is of interest. A page which is about comparing customer service can contain all of the words hotel price comparisons comparison. This page is not a good page for the query. If we accept matches of every occurrence of comparison, it will hurt retrieval precision and this is one of the main reasons why most stemming approaches do not work well for information retrieval. To address this problem, we have a proximity constraint that considers the context around the expanded variant in the document. A variant match is considered valid only if the variant occurs in the same context as the original word does. The context is the left or the right non-stop segments 1 of the original word. Taking the same query as an example, the context of comparisons is price. The expanded word comparison is only valid if it is in the same context of comparisons, which is after the word price. Thus, we should only match those occurrences of comparison in the document if they occur after the word price. Considering the fact that queries and documents may not represent the intent in exactly the same way, we relax this proximity constraint to allow variant occurrences within a window of some fixed size. If the expanded word comparison occurs within the context of price within a window, it is considered valid. The smaller the window size is, the more restrictive the matching. We use a window size of 4, which typically captures contexts that include the containing and adjacent noun phrases. 4. EXPERIMENTAL EVALUATION 4.1 Evaluation metrics We will measure both relevance improvement and the stemming cost required to achieve the relevance. 1 a context segment can not be a single stop word. 4.1.1 Relevance measurement We use a variant of the average Discounted Cumulative Gain (DCG), a recently popularized scheme to measure search engine relevance [1, 11]. Given a query and a ranked list of K documents (K is set to 5 in our experiments), the DCG(K) score for this query is calculated as follows: DCG(K) = KX k=1 gk log2(1 + k) . (7) where gk is the weight for the document at rank k. Higher degree of relevance corresponds to a higher weight. A page is graded into one of the five scales: Perfect, Excellent, Good, Fair, Bad, with corresponding weights. We use dcg to represent the average DCG(5) over a set of test queries. 4.1.2 Stemming cost Another metric is to measure the additional cost incurred by stemming. Given the same level of relevance improvement, we prefer a stemming method that has less additional cost. We measure this by the percentage of queries that are actually stemmed, over all the queries that could possibly be stemmed. 4.2 Data preparation We randomly sample 870 queries from a three month query log, with 290 from each month. Among all these 870 queries, we remove all misspelled queries since misspelled queries are not of interest to stemming. We also remove all one word queries since stemming one word queries without context has a high risk of changing query intent, especially for short words. In the end, we have 529 correctly spelled queries with at least 2 words. 4.3 Naive stemming for Web search Before explaining the experiments and results in detail, we"d like to describe the traditional way of using stemming for Web search, referred as the naive model. This is to treat every word variant equivalent for all possible words in the query. The query book store will be transformed into (book OR books)(store OR stores) when limiting stemming to pluralization handling only, where OR is an operator that denotes the equivalence of the left and right arguments. 4.4 Experimental setup The baseline model is the model without stemming. We first run the naive model to see how well it performs over the baseline. Then we improve the naive stemming model by document sensitive matching, referred as document sensitive matching model. This model makes the same stemming as the naive model on the query side, but performs conservative matching on the document side using the strategy described in section 3.5. The naive model and document sensitive matching model stem the most queries. Out of the 529 queries, there are 408 queries that they stem, corresponding to 46.7% query traffic (out of a total of 870). We then further improve the document sensitive matching model from the query side with selective word stemming based on statistical language modeling (section 3.4), referred as selective stemming model. Based on language modeling prediction, this model stems only a subset of the 408 queries stemmed by the document sensitive matching model. We experiment with unigram language model and bigram language model. Since we only care how much we can improve the naive model, we will only use these 408 queries (all the queries that are affected by the naive stemming model) in the experiments. To get a sense of how these models perform, we also have an oracle model that gives the upper-bound performance a stemmer can achieve on this data. The oracle model only expands a word if the stemming will give better results. To analyze the pluralization handling influence on different query categories, we divide queries into short queries and long queries. Among the 408 queries stemmed by the naive model, there are 272 short queries with 2 or 3 words, and 136 long queries with at least 4 words. 4.5 Results We summarize the overall results in Table 4, and present the results on short queries and long queries separately in Table 5. Each row in Table 4 is a stemming strategy described in section 4.4. The first column is the name of the strategy. The second column is the number of queries affected by this strategy; this column measures the stemming cost, and the numbers should be low for the same level of dcg. The third column is the average dcg score over all tested queries in this category (including the ones that were not stemmed by the strategy). The fourth column is the relative improvement over the baseline, and the last column is the p-value of Wilcoxon significance test. There are several observations about the results. We can see the naively stemming only obtains a statistically insignificant improvement of 1.5%. Looking at Table 5, it gives an improvement of 2.7% on short queries. However, it also hurts long queries by -2.4%. Overall, the improvement is canceled out. The reason that it improves short queries is that most short queries only have one word that can be stemmed. Thus, blindly pluralizing short queries is relatively safe. However for long queries, most queries can have multiple words that can be pluralized. Expanding all of them without selection will significantly hurt precision. Document context sensitive stemming gives a significant lift to the performance, from 2.7% to 4.2% for short queries and from -2.4% to -1.6% for long queries, with an overall lift from 1.5% to 2.8%. The improvement comes from the conservative context sensitive document matching. An expanded word is valid only if it occurs within the context of original query in the document. This reduces many spurious matches. However, we still notice that for long queries, context sensitive stemming is not able to improve performance because it still selects too many documents and gives the ranking function a hard problem. While the chosen window size of 4 works the best amongst all the choices, it still allows spurious matches. It is possible that the window size needs to be chosen on a per query basis to ensure tighter proximity constraints for different types of noun phrases. Selective word pluralization further helps resolving the problem faced by document context sensitive stemming. It does not stem every word that places all the burden on the ranking algorithm, but tries to eliminate unnecessary stemming in the first place. By predicting which word variants are going to be useful, we can dramatically reduce the number of stemmed words, thus improving both the recall and the precision. With the unigram language model, we can reduce the stemming cost by 26.7% (from 408/408 to 300/408) and lift the overall dcg improvement from 2.8% to 3.4%. In particular, it gives significant improvements on long queries. The dcg gain is turned from negative to positive, from −1.6% to 1.1%. This confirms our hypothesis that reducing unnecessary word expansion leads to precision improvement. For short queries too, we observe both dcg improvement and stemming cost reduction with the unigram language model. The advantages of predictive word expansion with a language model is further boosted with a better bigram language model. The overall dcg gain is lifted from 3.4% to 3.9%, and stemming cost is dramatically reduced from 408/408 to 250/408, corresponding to only 29% of query traffic (250 out of 870) and an overall 1.8% dcg improvement overall all query traffic. For short queries, bigram language model improves the dcg gain from 4.4% to 4.7%, and reduces stemming cost from 272/272 to 150/272. For long queries, bigram language model improves dcg gain from 1.1% to 2.5%, and reduces stemming cost from 136/136 to 100/136. We observe that the bigram language model gives a larger lift for long queries. This is because the uncertainty in long queries is larger and a more powerful language model is needed. We hypothesize that a trigram language model would give a further lift for long queries and leave this for future investigation. Considering the tight upper-bound 2 on the improvement to be gained from pluralization handling (via the oracle model), the current performance on short queries is very satisfying. For short queries, the dcg gain upper-bound is 6.3% for perfect pluralization handling, our current gain is 4.7% with a bigram language model. For long queries, the dcg gain upper-bound is 4.6% for perfect pluralization handling, our current gain is 2.5% with a bigram language model. We may gain additional benefit with a more powerful language model for long queries. However, the difficulties of long queries come from many other aspects including the proximity and the segmentation problem. These problems have to be addressed separately. Looking at the the upper-bound of overhead reduction for oracle stemming, 75% (308/408) of the naive stemmings are wasteful. We currently capture about half of them. Further reduction of the overhead requires sacrificing the dcg gain. Now we can compare the stemming strategies from a different aspect. Instead of looking at the influence over all queries as we described above, Table 6 summarizes the dcg improvements over the affected queries only. We can see that the number of affected queries decreases as the stemming strategy becomes more accurate (dcg improvement). For the bigram language model, over the 250/408 stemmed queries, the dcg improvement is 6.1%. An interesting observation is the average dcg decreases with a better model, which indicates a better stemming strategy stems more difficult queries (low dcg queries). 5. DISCUSSIONS 5.1 Language models from query vs. from Web As we mentioned in Section 1, we are trying to predict the probability of a string occurring on the Web. The language model should describe the occurrence of the string on the Web. However, the query log is also a good resource. 2 Note that this upperbound is for pluralization handling only, not for general stemming. General stemming gives a 8% upperbound, which is quite substantial in terms of our metrics. Affected Queries dcg dcg Improvement p-value baseline 0/408 7.102 N/A N/A naive model 408/408 7.206 1.5% 0.22 document context sensitive model 408/408 7.302 2.8% 0.014 selective model: unigram LM 300/408 7.321 3.4% 0.001 selective model: bigram LM 250/408 7.381 3.9% 0.001 oracle model 100/408 7.519 5.9% 0.001 Table 4: Results comparison of different stemming strategies over all queries affected by naive stemming Short Query Results Affected Queries dcg Improvement p-value baseline 0/272 N/A N/A naive model 272/272 2.7% 0.48 document context sensitive model 272/272 4.2% 0.002 selective model: unigram LM 185/272 4.4% 0.001 selective model: bigram LM 150/272 4.7% 0.001 oracle model 71/272 6.3% 0.001 Long Query Results Affected Queries dcg Improvement p-value baseline 0/136 N/A N/A naive model 136/136 -2.4% 0.25 document context sensitive model 136/136 -1.6% 0.27 selective model: unigram LM 115/136 1.1% 0.001 selective model: bigram LM 100/136 2.5% 0.001 oracle model 29/136 4.6% 0.001 Table 5: Results comparison of different stemming strategies overall short queries and long queries Users reformulate a query using many different variants to get good results. To test the hypothesis that we can learn reliable transformation probabilities from the query log, we trained a language model from the same query top 25M queries as used to learn segmentation, and use that for prediction. We observed a slight performance decrease compared to the model trained on Web frequencies. In particular, the performance for unigram LM was not affected, but the dcg gain for bigram LM changed from 4.7% to 4.5% for short queries. Thus, the query log can serve as a good approximation of the Web frequencies. 5.2 How linguistics helps Some linguistic knowledge is useful in stemming. For the pluralization handling case, pluralization and de-pluralization is not symmetric. A plural word used in a query indicates a special intent. For example, the query new york hotels is looking for a list of hotels in new york, not the specific new york hotel which might be a hotel located in California. A simple equivalence of hotel to hotels might boost a particular page about new york hotel to top rank. To capture this intent, we have to make sure the document is a general page about hotels in new york. We do this by requiring that the plural word hotels appears in the document. On the other hand, converting a singular word to plural is safer since a general purpose page normally contains specific information. We observed a slight overall dcg decrease, although not statistically significant, for document context sensitive stemming if we do not consider this asymmetric property. 5.3 Error analysis One type of mistakes we noticed, though rare but seriously hurting relevance, is the search intent change after stemming. Generally speaking, pluralization or depluralization keeps the original intent. However, the intent could change in a few cases. For one example of such a query, job at apple, we pluralize job to jobs. This stemming makes the original query ambiguous. The query job OR jobs at apple has two intents. One is the employment opportunities at apple, and another is a person working at Apple, Steve Jobs, who is the CEO and co-founder of the company. Thus, the results after query stemming returns Steve Jobs as one of the results in top 5. One solution is performing results set based analysis to check if the intent is changed. This is similar to relevance feedback and requires second phase ranking. A second type of mistakes is the entity/concept recognition problem, These include two kinds. One is that the stemmed word variant now matches part of an entity or concept. For example, query cookies in san francisco is pluralized to cookies OR cookie in san francisco. The results will match cookie jar in san francisco. Although cookie still means the same thing as cookies, cookie jar is a different concept. Another kind is the unstemmed word matches an entity or concept because of the stemming of the other words. For example, quote ICE is pluralized to quote OR quotes ICE. The original intent for this query is searching for stock quote for ticker ICE. However, we noticed that among the top results, one of the results is Food quotes: Ice cream. This is matched because of Affected Queries old dcg new dcg dcg Improvement naive model 408/408 7.102 7.206 1.5% document context sensitive model 408/408 7.102 7.302 2.8% selective model: unigram LM 300/408 5.904 6.187 4.8% selective model: bigram LM 250/408 5.551 5.891 6.1% Table 6: Results comparison over the stemmed queries only: column old/new dcg is the dcg score over the affected queries before/after applying stemming the pluralized word quotes. The unchanged word ICE matches part of the noun phrase ice cream here. To solve this kind of problem, we have to analyze the documents and recognize cookie jar and ice cream as concepts instead of two independent words. A third type of mistakes occurs in long queries. For the query bar code reader software, two words are pluralized. code to codes and reader to readers. In fact, bar code reader in the original query is a strong concept and the internal words should not be changed. This is the segmentation and entity and noun phrase detection problem in queries, which we actively are attacking. For long queries, we should correctly identify the concepts in the query, and boost the proximity for the words within a concept. 6. CONCLUSIONS AND FUTURE WORK We have presented a simple yet elegant way of stemming for Web search. It improves naive stemming in two aspects: selective word expansion on the query side and conservative word occurrence matching on the document side. Using pluralization handling as an example, experiments on a major Web search engine data show it significantly improves the Web relevance and reduces the stemming cost. It also significantly improves Web click through rate (details not reported in the paper). For the future work, we are investigating the problems we identified in the error analysis section. These include: entity and noun phrase matching mistakes, and improved segmentation. 7. REFERENCES [1] E. Agichtein, E. Brill, and S. T. Dumais. Improving Web Search Ranking by Incorporating User Behavior Information. In SIGIR, 2006. [2] E. Airio. Word Normalization and Decompounding in Mono- and Bilingual IR. Information Retrieval, 9:249-271, 2006. [3] P. Anick. Using Terminological Feedback for Web Search Refinement: a Log-based Study. In SIGIR, 2003. [4] R. Baeza-Yates and B. Ribeiro-Neto. Modern Information Retrieval. ACM Press/Addison Wesley, 1999. [5] S. Chen and J. Goodman. An Empirical Study of Smoothing Techniques for Language Modeling. Technical Report TR-10-98, Harvard University, 1998. [6] S. Cronen-Townsend, Y. Zhou, and B. Croft. A Framework for Selective Query Expansion. In CIKM, 2004. [7] H. Fang and C. Zhai. Semantic Term Matching in Axiomatic Approaches to Information Retrieval. In SIGIR, 2006. [8] W. B. Frakes. Term Conflation for Information Retrieval. In C. J. Rijsbergen, editor, Research and Development in Information Retrieval, pages 383-389. Cambridge University Press, 1984. [9] D. Harman. How Effective is Suffixing? JASIS, 42(1):7-15, 1991. [10] D. Hull. Stemming Algorithms - A Case Study for Detailed Evaluation. JASIS, 47(1):70-84, 1996. [11] K. Jarvelin and J. Kekalainen. Cumulated Gain-Based Evaluation Evaluation of IR Techniques. ACM TOIS, 20:422-446, 2002. [12] R. Jones, B. Rey, O. Madani, and W. Greiner. Generating Query Substitutions. In WWW, 2006. [13] W. Kraaij and R. Pohlmann. Viewing Stemming as Recall Enhancement. In SIGIR, 1996. [14] R. Krovetz. Viewing Morphology as an Inference Process. In SIGIR, 1993. [15] D. Lin. Automatic Retrieval and Clustering of Similar Words. In COLING-ACL, 1998. [16] J. B. Lovins. Development of a Stemming Algorithm. Mechanical Translation and Computational Linguistics, II:22-31, 1968. [17] M. Lennon and D. Peirce and B. Tarry and P. Willett. An Evaluation of Some Conflation Algorithms for Information Retrieval. Journal of Information Science, 3:177-188, 1981. [18] M. Porter. An Algorithm for Suffix Stripping. Program, 14(3):130-137, 1980. [19] K. M. Risvik, T. Mikolajewski, and P. Boros. Query Segmentation for Web Search. In WWW, 2003. [20] S. E. Robertson. On Term Selection for Query Expansion. Journal of Documentation, 46(4):359-364, 1990. [21] G. Salton and C. Buckley. Improving Retrieval Performance by Relevance Feedback. JASIS, 41(4):288 - 297, 1999. [22] R. Sun, C.-H. Ong, and T.-S. Chua. Mining Dependency Relations for Query Expansion in Passage Retrieval. In SIGIR, 2006. [23] C. Van Rijsbergen. Information Retrieval. Butterworths, second version, 1979. [24] B. V´elez, R. Weiss, M. A. Sheldon, and D. K. Gifford. Fast and Effective Query Refinement. In SIGIR, 1997. [25] J. Xu and B. Croft. Query Expansion using Local and Global Document Analysis. In SIGIR, 1996. [26] J. Xu and B. Croft. Corpus-based Stemming using Cooccurrence of Word Variants. ACM TOIS, 16 (1):61-81, 1998.

stem;language model;bigram language model;head word detection;context sensitive document matching;lovin stemmer;porter stemmer;web search;candidate generation;query segmentation;unigram language model;context sensitive query stemming;stemming

train_H-54

Knowledge-intensive Conceptual Retrieval and Passage Extraction of Biomedical Literature

This paper presents a study of incorporating domain-specific knowledge (i.e., information about concepts and relationships between concepts in a certain domain) in an information retrieval (IR) system to improve its effectiveness in retrieving biomedical literature. The effects of different types of domain-specific knowledge in performance contribution are examined. Based on the TREC platform, we show that appropriate use of domainspecific knowledge in a proposed conceptual retrieval model yields about 23% improvement over the best reported result in passage retrieval in the Genomics Track of TREC 2006.

1. INTRODUCTION Biologists search for literature on a daily basis. For most biologists, PubMed, an online service of U.S. National Library of Medicine (NLM), is the most commonly used tool for searching the biomedical literature. PubMed allows for keyword search by using Boolean operators. For example, if one desires documents on the use of the drug propanolol in the disease hypertension, a typical PubMed query might be propanolol AND hypertension, which will return all the documents having the two keywords. Keyword search in PubMed is effective if the query is well-crafted by the users using their expertise. However, information needs of biologists, in some cases, are expressed as complex questions [8][9], which PubMed is not designed to handle. While NLM does maintain an experimental tool for free-text queries [6], it is still based on PubMed keyword search. The Genomics track of the 2006 Text REtrieval Conference (TREC) provides a common platform to assess the methods and techniques proposed by various groups for biomedical information retrieval. The queries were collected from real biologists and they are expressed as complex questions, such as How do mutations in the Huntingtin gene affect Huntington"s disease?. The document collection contains 162,259 Highwire full-text documents in HTML format. Systems from participating groups are expected to find relevant passages within the full-text documents. A passage is defined as any span of text that does not include the HTML paragraph tag (i.e., <P> or </P>). We approached the problem by utilizing domain-specific knowledge in a conceptual retrieval model. Domain-specific knowledge, in this paper, refers to information about concepts and relationships between concepts in a certain domain. We assume that appropriate use of domain-specific knowledge might improve the effectiveness of retrieval. For example, given a query What is the role of gene PRNP in the Mad Cow Disease?, expanding the gene symbol PRNP with its synonyms Prp, PrPSc, and prion protein, more relevant documents might be retrieved. PubMed and many other biomedical systems [8][9][10][13] also make use of domain-specific knowledge to improve retrieval effectiveness. Intuitively, retrieval on the level of concepts should outperform bag-of-words approaches, since the semantic relationships among words in a concept are utilized. In some recent studies [13][15], positive results have been reported for this hypothesis. In this paper, concepts are entry terms of the ontology Medical Subject Headings (MeSH), a controlled vocabulary maintained by NLM for indexing biomedical literature, or gene symbols in the Entrez gene database also from NLM. A concept could be a word, such as the gene symbol PRNP, or a phrase, such as Mad cow diseases. In the conceptual retrieval model presented in this paper, the similarity between a query and a document is measured on both concept and word levels. This paper makes two contributions: 1. We propose a conceptual approach to utilize domain-specific knowledge in an IR system to improve its effectiveness in retrieving biomedical literature. Based on this approach, our system achieved significant improvement (23%) over the best reported result in passage retrieval in the Genomics track of TREC 2006. 2. We examine the effects of utilizing concepts and of different types of domain-specific knowledge in performance contribution. This paper is organized as follows: problem statement is given in the next section. The techniques are introduced in section 3. In section 4, we present the experimental results. Related works are given in section 5 and finally, we conclude the paper in section 6. 2. PROBLEM STATEMENT We describe the queries, document collection and the system output in this section. The query set used in the Genomics track of TREC 2006 consists of 28 questions collected from real biologists. As described in [8], these questions all have the following general format: Biological object (1..m) Relationship ←⎯⎯⎯⎯→ Biological process (1..n) (1) where a biological object might be a gene, protein, or gene mutation and a biological process can be a physiological process or disease. A question might involve multiple biological objects (m) and multiple biological processes (n). These questions were derived from four templates (Table 2). Table 2 Query templates and examples in the Genomics track of TREC 2006 Template Example What is the role of gene in disease? What is the role of DRD4 in alcoholism? What effect does gene have on biological process? What effect does the insulin receptor gene have on tumorigenesis? How do genes interact in organ function? How do HMG and HMGB1 interact in hepatitis? How does a mutation in gene influence biological process? How does a mutation in Ret influence thyroid function? Features of the queries: 1) They are different from the typical Web queries and the PubMed queries, both of which usually consist of 1 to 3 keywords; 2) They are generated from structural templates which can be used by a system to identify the query components, the biological object or process. The document collection contains 162,259 Highwire full-text documents in HTML format. The output of the system is a list of passages ranked according to their similarities with the query. A passage is defined as any span of text that does not include the HTML paragraph tag (i.e., <P> or </P>). A passage could be a part of a sentence, a sentence, a set of consecutive sentences or a paragraph (i.e., the whole span of text that are inside of <P> and </P> HTML tags). This is a passage-level information retrieval problem with the attempt to put biologists in contexts where relevant information is provided. 3. TECHNIQUES AND METHODS We approached the problem by first retrieving the top-k most relevant paragraphs, then extracting passages from these paragraphs, and finally ranking the passages. In this process, we employed several techniques and methods, which will be introduced in this section. First, we give two definitions: Definition 3.1 A concept is 1) a entry term in the MeSH ontology, or 2) a gene symbol in the Entrez gene database. This definition of concept can be generalized to include other biomedical dictionary terms. Definition 3.2 A semantic type is a category defined in the Semantic Network of the Unified Medical Language System (UMLS) [14]. The current release of the UMLS Semantic Network contains 135 semantic types such as Disease or Syndrome. Each entry term in the MeSH ontology is assigned one or more semantic types. Each gene symbol in the Entrez gene database maps to the semantic type Gene or Genome. In addition, these semantic types are linked by 54 relationships. For example, Antibiotic prevents Disease or Syndrome. These relationships among semantic types represent general biomedical knowledge. We utilized these semantic types and their relationships to identify related concepts. The rest of this section is organized as follows: in section 3.1, we explain how the concepts are identified within a query. In section 3.2, we specify five different types of domain-specific knowledge and introduce how they are compiled. In section 3.3, we present our conceptual IR model. Finally, our strategy for passage extraction is described in section 3.4. 3.1 Identifying concepts within a query A concept, defined in Definition 3.1, is a gene symbol or a MeSH term. We make use of the query templates to identify gene symbols. For example, the query How do HMG and HMGB1 interact in hepatitis? is derived from the template How do genes interact in organ function?. In this case, HMG and HMGB1 will be identified as gene symbols. In cases where the query templates are not provided, programs for recognition of gene symbols within texts are needed. We use the query translation functionality of PubMed to extract MeSH terms in a query. This is done by submitting the whole query to PubMed, which will then return a file in which the MeSH terms in the query are labeled. In Table 3.1, three MeSH terms within the query What is the role of gene PRNP in the Mad cow disease? are found in the PubMed translation: "encephalopathy, bovine spongiform" for Mad cow disease, genes for gene, and role for role. Table 3.1 The PubMed translation of the query "What is the role of gene PRNP in the Mad cow disease?". Term PubMed translation Mad cow disease "bovine spongiform encephalopathy"[Text Word] OR "encephalopathy, bovine spongiform"[MeSH Terms] OR Mad cow disease[Text Word] gene ("genes"[TIAB] NOT Medline[SB]) OR "genes"[MeSH Terms] OR gene[Text Word] role "role"[MeSH Terms] OR role[Text Word] 3.2 Compiling domain-specific knowledge In this paper, domain-specific knowledge refers to information about concepts and their relationships in a certain domain. We used five types of domain-specific knowledge in the domain of genomics: Type 1. Synonyms (terms listed in the thesauruses that refer to the same meaning) Type 2. Hypernyms (more generic terms, one level only) Type 3. Hyponyms (more specific terms, one level only) Type 4. Lexical variants (different forms of the same concept, such as abbreviations. They are commonly used in the literature, but might not be listed in the thesauruses) Type 5. Implicitly related concepts (terms that are semantically related and also co-occur more frequently than being independent in the biomedical texts) Knowledge of type 1-3 is retrieved from the following two thesauruses: 1) MeSH, a controlled vocabulary maintained by NLM for indexing biomedical literature. The 2007 version of MeSH contains information about 190,000 concepts. These concepts are organized in a tree hierarchy; 2) Entrez Gene, one of the most widely used searchable databases of genes. The current version of Entrez Gene contains information about 1.7 million genes. It does not have a hierarchy. Only synonyms are retrieved from Entrez Gene. The compiling of type 4-5 knowledge is introduced in section 3.2.1 and 3.2.2, respectively. 3.2.1 Lexical variants Lexical variants of gene symbols New gene symbols and their lexical variants are regularly introduced into the biomedical literature [7]. However, many reference databases, such as UMLS and Entrez Gene, may not be able to keep track of all this kind of variants. For example, for the gene symbol "NF-kappa B", at least 5 different lexical variants can be found in the biomedical literature: "NF-kappaB", "NFkappaB", "NFkappa B", "NF-kB", and "NFkB", three of which are not in the current UMLS and two not in the Entrez Gene. [3][21] have shown that expanding gene symbols with their lexical variants improved the retrieval effectiveness of their biomedical IR systems. In our system, we employed the following two strategies to retrieve lexical variants of gene symbols. Strategy I: This strategy is to automatically generate lexical variants according to a set of manually crafted heuristics [3][21]. For example, given a gene symbol PLA2, a variant PLAII is generated according to the heuristic that Roman numerals and Arabic numerals are convertible when naming gene symbols. Another variant, PLA 2, is also generated since a hyphen or a space could be inserted at the transition between alphabetic and numerical characters in a gene symbol. Strategy II: This strategy is to retrieve lexical variants from an abbreviation database. ADAM [22] is an abbreviation database which covers frequently used abbreviations and their definitions (or long-forms) within MEDLINE, the authoritative repository of citations from the biomedical literature maintained by the NLM. Given a query How does nucleoside diphosphate kinase (NM23) contribute to tumor progression?, we first identify the abbreviation NM23 and its long-form nucleoside diphosphate kinase using the abbreviation identification program from [4]. Searching the long-form nucleoside diphosphate kinase in ADAM, other abbreviations, such as NDPK or NDK, are retrieved. These abbreviations are considered as the lexical variants of NM23. Lexical variants of MeSH concepts ADAM is used to obtain the lexical variants of MeSH concepts as well. All the abbreviations of a MeSH concept in ADAM are considered as lexical variants to each other. In addition, those long-forms that share the same abbreviation with the MeSH concept and are different by an edit distance of 1 or 2 are also considered as its lexical variants. As an example, "human papilloma viruses" and "human papillomaviruses" have the same abbreviation HPV in ADAM and their edit distance is 1. Thus they are considered as lexical variants to each other. The edit distance between two strings is measured by the minimum number of insertions, deletions, and substitutions of a single character required to transform one string into the other [12]. 3.2.2 Implicitly related concepts Motivation: In some cases, there are few documents in the literature that directly answer a given query. In this situation, those documents that implicitly answer their questions or provide supporting information would be very helpful. For example, there are few documents in PubMed that directly answer the query "What is the role of the genes HNF4 and COUP-tf I in the suppression in the function of the liver?". However, there exist some documents about the role of "HNF4" and "COUP-tf I" in regulating "hepatitis B virus" transcription. It is very likely that the biologists would be interested in these documents because "hepatitis B virus" is known as a virus that could cause serious damage to the function of liver. In the given example, "hepatitis B virus" is not a synonym, hypernym, hyponym, nor a lexical variant of any of the query concepts, but it is semantically related to the query concepts according to the UMLS Semantic Network. We call this type of concepts implicitly related concepts of the query. This notion is similar to the B-term used in [19] for relating two disjoint literatures for biomedical hypothesis generation. The difference is that we utilize the semantic relationships among query concepts to exclusively focus on concepts of certain semantic types. A query q in format (1) of section 2 can be represented by q = (A, C) where A is the set of biological objects and C is the set of biological processes. Those concepts that are semantically related to both A and C according to the UMLS Semantic Network are considered as the implicitly related concepts of the query. In the above example, A = {HNF4, COUP-tf I}, C = {function of liver}, and "hepatitis B virus" is one of the implicitly related concepts. We make use of the MEDLINE database to extract the implicitly related concepts. The 2006 version of MEDLINE database contains citations (i.e., abstracts, titles, and etc.) of over 15 million biomedical articles. Each document in MEDLINE is manually indexed by a list of MeSH terms to describe the topics covered by that document. Implicitly related concepts are extracted and ranked in the following steps: Step 1. Let list_A be the set of MeSH terms that are 1) used for indexing those MEDLINE citations having A, and 2) semantically related to A according to the UMLS Semantic Network. Similarly, list_C is created for C. Concepts in B = list_A ∩ list_C are considered as implicitly related concepts of the query. Step 2. For each concept b∈B, compute the association between b and A using the mutual information measure [5]: P( , ) ( , ) log P( )P( ) b A I b A b A = where P(x) = n/N, n is the number of MEDLINE citations having x and N is the size of MEDLINE. A large value for I(b, A) means that b and A co-occur much more often than being independent. I(b, C) is computed similarly. Step 3. Let r(b) = (I(b, A), I(b, C)), for b∈ B. Given b1, b2 ∈ B, we say r(b1) ≤ r(b2) if I(b1, A) ≤ I(b2, A) and I(b1, C) ≤ I(b2, C). Then the association between b and the query q is measured by: { : and ( ) ( )} ( , ) { : and ( ) ( )} x x B r x r b score b q x x B r b r x ∈ ≤ = ∈ ≤ (2) The numerator in Formula 2 is the number of the concepts in B that are associated with both A and C equally with or less than b. The denominator is the number of the concepts in B that are associated with both A and C equally with or more than b. Figure 3.2.2 shows the top 4 implicitly related concepts for the sample query. Figure 3.2.2 Top 4 implicitly related concepts for the query "How do interactions between HNF4 and COUP-TF1 suppress liver function?". In Figure 3.2.2, the top 4 implicitly related concepts are all highly associated with liver: Hepatocytes are liver cells; Hepatoblastoma is a malignant liver neoplasm occurring in young children; the vast majority of Gluconeogenesis takes place in the liver; and Hepatitis B virus is a virus that could cause serious damage to the function of liver. The top-k ranked concepts in B are used for query expansion: if I(b, A) ≥ I(b, C), then b is considered as an implicit related concept of A. A document having b but not A will receive a partial weight of A. The expansion is similar for C when I(b, A) < I(b, C). 3.3 Conceptual IR model We now discuss our conceptual IR model. We first give the basic conceptual IR model in section 3.3.1. Then we explain how the domain-specific knowledge is incorporated in the model using query expansion in section 3.3.2. A pseudo-feedback strategy is introduced in section 3.3.3. In section 3.3.4, we give a strategy to improve the ranking by avoiding incorrect match of abbreviations. 3.3.1 Basic model Given a query q and a document d, our model measures two similarities, concept similarity and word similarity: ( , ) ( , )( , ) ( , ) concept word sim q d sim q d sim q d= Concept similarity Two vectors are derived from a query q, 1 2 1 11 12 1 2 21 22 2 ( , ) ( , ,..., ) ( , ,..., ) m n q v v v c c c v c c c = = = where v1 is a vector of concepts describing the biological object(s) and v2 is a vector of concepts describing the biological process(es). Given a vector of concepts v, let s(v) be the set of concepts in v. The weight of vi is then measured by: ( ) max{log : ( ) ( ) and 0}i i v v N w v s v s v n n = ⊆ > where v is a vector that contains a subset of concepts in vi and nv is the number of documents having all the concepts in v. The concept similarity between q and d is then computed by 2 1 ( )( , ) i i concept i w vsim q d α = = ×∑ where αi is a parameter to indicate the completeness of vi that document d has covered. αi is measured by: and i i i c c d c v c c v idf idf α ∈ ∈ ∈ = ∑ ∑ (3) where idfc is the inverse document frequency of concept c. An example: suppose we have a query How does Nurr-77 delete T cells before they migrate to the spleen or lymph nodes and how does this impact autoimmunity?. After identifying the concepts in the query, we have: 1 2 ('Nurr-77') ('T cells', 'spleen', 'autoimmunity', 'lymph nodes') v v = = Suppose that some document frequencies of different combinations of concepts are as follows: 25 df('Nurr-77') 0 df('T cells', 'spleen', 'autoimmunity', 'lymph nodes') 326 df('T cells', 'spleen', 'autoimmunity') 82 df('spleen', 'autoimmunity', 'lymph nodes') 147 df('T cells', 'autoimmunity', 'lymph nodes') 2332 df('T cells', 'spleen', 'lymph nodes') The weight of vi is then computed by (note that there does not exist a document having all the concepts in v2): 1 2 ( ) log( / 25) ( ) log( /82) w v N w v N = = . Now suppose a document d contains concepts ‘Nurr-77", 'T cells', 'spleen', and 'lymph nodes', but not ‘autoimmunity", then the value of parameter αi is computed as follows: 1 2 1 ('T cells')+ ('spleen')+ ('lymph nodes') ('T cells')+ ('spleen')+ ('lymph nodes')+ ('autoimmunity') idf idf idf idf idf idf idf α α = = Word similarity The similarity between q and d on the word level is computed using Okapi [17]: 10.5 ( 1) log( )( , ) 0.5word w q N n k tf sim q d n K tf∈ − + + = + + ∑ (4) where N is the size of the document collection; n is the number of documents containing w; K=k1 × ((1-b)+b × dl/avdl) and k1=1.2, C Function of Liver Implicitly related concepts (B) Hepatocytes Hepatoblastoma Gluconeogenesis Hepatitis B virus HNF4 and COUP-tf I A b=0.75 are constants. dl is the document length of d and avdl is the average document length; tf is the term frequency of w within d. The model Given two documents d1 and d2, we say 1 2( , ) ( , )sim q d sim q d> or d1 will be ranked higher than d2, with respect to the same query q, if either 1) 1 2( , ) ( , ) concept concept sim q d sim q d> or 2) 1 2 1 2and( , ) ( , ) ( , ) ( , ) concept concept word word sim q d sim q d sim q d sim q d= > This conceptual IR model emphasizes the similarity on the concept level. A similar model but applied to non-biomedical domain has been given in [15]. 3.3.2 Incorporating domain-specific knowledge Given a concept c, a vector u is derived by incorporating its domain-specific knowledge: 1 2 3( , , , )u c u u u= where u1 is a vector of its synonyms, hyponyms, and lexical variants; u2 is a vector of its hypernyms; and u3 is a vector of its implicitly related concepts. An occurrence of any term in u1 will be counted as an occurrence of c. idfc in Formula 3 is updated as: 1, logc c u N D idf = 1,c uD is the set of documents having c or any term in 1u . The weight that a document d receives from u is given by: max{ : and }tw t u t d∈ ∈ where wt = β .cidf× The weighting factor β is an empirical tuning parameter determined as: 1. β = 1 if t is the original concept, its synonym, its hyponym, or its lexical variant; 2. β = 0.95 if t is a hypernym; 3. β = 0.90× (k-i+1)/k if t is an implicitly related concept. k is the number of selected top ranked implicitly related concepts (see section 3.2.2); i is the position of t in the ranking of implicitly related concepts. 3.3.3 Pseudo-feedback Pseudo-feedback is a technique commonly used to improve retrieval performance by adding new terms into the original query. We used a modified pseudo-feedback strategy described in [2]. Step 1. Let C be the set of concepts in the top 15 ranked documents. For each concept c in C, compute the similarity between c and the query q, the computation of sim(q,c) can be found in [2]. Step 2. The top-k ranked concepts by sim(q,c) are selected. Step 3. Associate each selected concept c' with the concept cq in q that 1) has the same semantic type as c', and 2) is most related to c' among all the concepts in q. The association between c' and cq is computed by: P( ', ) ( ', ) log P( ')P( ) q q q c c I c c c c = where P(x) = n/N, n is the number of documents having x and N is the size of the document collection. A document having c' but not cq receives a weight given by: (0.5× (k-i+1)/k) ,qcidf× where i is the position of c' in the ranking of step 2. 3.3.4 Avoid incorrect match of abbreviations Some gene symbols are very short and thus ambiguous. For example, the gene symbol APC could be the abbreviation for many non-gene long-forms, such as air pollution control, aerobic plate count, or argon plasma coagulation. This step is to avoid incorrect match of abbreviations in the top ranked documents. Given an abbreviation X with the long-form L in the query, we scan the top-k ranked (k=1000) documents and when a document is found with X, we compare L with all the long-forms of X in that document. If none of these long-forms is equal or close to L (i.e., the edit distance between L and the long-form of X in that document is 1 or 2), then the concept similarity of X is subtracted. 3.4 Passage extraction The goal of passage extraction is to highlight the most relevant fragments of text in paragraphs. A passage is defined as any span of text that does not include the HTML paragraph tag (i.e., <P> or </P>). A passage could be a part of a sentence, a sentence, a set of consecutive sentences or a paragraph (i.e., the whole span of text that are inside of <P> and </P> HTML tags). It is also possible to have more than one relevant passage in a single paragraph. Our strategy for passage extraction assumes that the optimal passage(s) in a paragraph should have all the query concepts that the whole paragraph has. Also they should have higher density of query concepts than other fragments of text in the paragraph. Suppose we have a query q and a paragraph p represented by a sequence of sentences 1 2... .np s s s= Let C be the set of concepts in q that occur in p and S = Φ. Step 1. For each sequence of consecutive sentences 1... ,i i js s s+ 1 ≤ i ≤ j ≤ n, let S = S 1{ ... }i i js s s+∪ if 1...i i js s s+ satisfies that: 1) Every query concept in C occurs in 1...i i js s s+ and 2) There does not exist k, such that i < k < j and every query concept in C occurs in 1...i i ks s s+ or 1 2... .k k js s s+ + Condition 1 requires 1...i i js s s+ having all the query concepts in p and condition 2 requires 1...i i js s s+ be the minimal. Step 2. Let 1min{ 1: ... }i i jL j i s s s S+= − + ∈ . For every 1...i i js s s+ in S, let 1{ ... }i i jS S s s s+= − if (j - i + 1) > L. This step is to remove those sequences of sentences in S that have lower density of query concepts. Step 3. For every two sequences of consecutive sentences 1 1 1 2 2 21 1... , and ...i i j i i js s s S s s s S+ +∈ ∈ , if 1 2 1 2 2 1 , and 1 i i j j i j ≤ ≤ ≤ + (5) then do Repeat this step until for every two sequences of consecutive sentences in S, condition (5) does not apply. This step is to merge those sequences of sentences in S that are adjacent or overlapped. Finally the remaining sequences of sentences in S are returned as the optimal passages in the paragraph p with respect to the query. 1 1 2 1 1 2 2 2 1 1 1 1 { ... } { ... } { ... } i i j i i j i i j S S s s s S S s s s S S s s s + + + = ∪ = − = − 4. EXPERIMENTAL RESULTS The evaluation of our techniques and the experimental results are given in this section. We first describe the datasets and evaluation metrics used in our experiments and then present the results. 4.1 Data sets and evaluation metrics Our experiments were performed on the platform of the Genomics track of TREC 2006. The document collection contains 162,259 full-text documents from 49 Highwire biomedical journals. The set of queries consists of 28 queries collected from real biologists. The performance is measured on three different levels (passage, aspect, and document) to provide better insight on how the question is answered from different perspectives. Passage MAP: As described in [8], this is a character-based precision calculated as follows: At each relevant retrieved passage, precision will be computed as the fraction of characters overlapping with the gold standard passages divided by the total number of characters included in all nominated passages from this system for the topic up until that point. Similar to regular MAP, relevant passages that were not retrieved will be added into the calculation as well, with precision set to 0 for relevant passages not retrieved. Then the mean of these average precisions over all topics will be calculated to compute the mean average passage precision. Aspect MAP: A question could be addressed from different aspects. For example, the question what is the role of gene PRNP in the Mad cow disease? could be answered from aspects like Diagnosis, Neurologic manifestations, or Prions/Genetics. This measure indicates how comprehensive the question is answered. Document MAP: This is the standard IR measure. The precision is measured at every point where a relevant document is obtained and then averaged over all relevant documents to obtain the average precision for a given query. For a set of queries, the mean of the average precision for all queries is the MAP of that IR system. The output of the system is a list of passages ranked according to their similarities with the query. The performances on the three levels are then calculated based on the ranking of the passages. 4.2 Results The Wilcoxon signed-rank test was employed to determine the statistical significance of the results. In the tables of the following sections, statistically significant improvements (at the 5% level) are marked with an asterisk. 4.2.1 Conceptual IR model vs. term-based model The initial baseline was established using word similarity only computed by the Okapi (Formula 4). Another run based on our basic conceptual IR model was performed without using query expansion, pseudo-feedback, or abbreviation correction. The experimental result is shown in Table 4.2.1. Our basic conceptual IR model significantly outperforms the Okapi on all three levels, which suggests that, although it requires additional efforts to identify concepts, retrieval on the concept level can achieve substantial improvements over purely term-based retrieval model. 4.2.2 Contribution of different types of knowledge A series of experiments were performed to examine how each type of domain-specific knowledge contributes to the retrieval performance. A new baseline was established using the basic conceptual IR model without incorporating any type of domainspecific knowledge. Then five runs were conducted by adding each individual type of domain-specific knowledge. We also conducted a run by adding all types of domain-specific knowledge. Results of these experiments are shown in Table 4.2.2. We found that any available type of domain-specific knowledge improved the performance in passage retrieval. The biggest improvement comes from the lexical variants, which is consistent with the result reported in [3]. This result also indicates that biologists are likely to use different variants of the same concept according to their own writing preferences and these variants might not be collected in the existing biomedical thesauruses. It also suggests that the biomedical IR systems can benefit from the domain-specific knowledge extracted from the literature by text mining systems. Synonyms provided the second biggest improvement. Hypernyms, hyponyms, and implicitly related concepts provided similar degrees of improvement. The overall performance is an accumulative result of adding different types of domain-specific knowledge and it is better than any individual addition. It is clearly shown that the performance is significantly improved (107% on passage level, 63.1% on aspect level, and 49.6% on document level) when the domain-specific knowledge is appropriately incorporated. Although it is not explicitly shown in Table 4.2.3, different types of domain-specific knowledge affect different subsets of queries. More specifically, each of these types (with the exception of the lexical variants which affects a large number of queries) affects only a few queries. But for those affected queries, their improvement is significant. As a consequence, the accumulative improvement is very significant. 4.2.3 Pseudo-feedback and abbreviation correction Using the Baseline+All in Table 4.2.2 as a new baseline, the contribution of abbreviation correction and pseudo-feedback is given in Table 4.2.3. There is little improvement by avoiding incorrect matching of abbreviations. The pseudo-feedback contributed about 4.6% improvement in passage retrieval. 4.2.4 Performance compared with best-reported results We compared our result with the results reported in the Genomics track of TREC 2006 [8] on the conditions that 1) systems are automatic systems and 2) passages are extracted from paragraphs. The performance of our system relative to the best reported results is shown in Table 4.2.4 (in TREC 2006, some systems returned the whole paragraphs as passages. As a consequence, excellent retrieval results were obtained on document and aspect levels at the expense of performance on the passage level. We do not include the results of such systems here). Table 4.2.4 Performance compared with best-reported results. Passage MAP Aspect MAP Document MAP Best reported results 0.1486 0.3492 0.5320 Our results 0.1823 0.3811 0.5391 Improvement 22.68% 9.14% 1.33% The best reported results in the first row of Table 4.2.4 on three levels (passage, aspect, and document) are from different systems. Our result is from a single run on passage retrieval in which it is better than the best reported result by 22.68% in passage retrieval and at the same time, 9.14% better in aspect retrieval, and 1.33% better in document retrieval (Since the average precision of each individual query was not reported, we can not apply the Wilcoxon signed-rank test to calculate the significance of difference between our performance and the best reported result.). Table 4.2.1 Basic conceptual IR model vs. term-based model Run Passage Aspect Document MAP Imprvd qs # (%) MAP Imprvd qs # (%) MAP Imprvd qs # (%) Okapi 0.064 N/A 0.175 N/A 0.285 N/A Basic conceptual IR model 0.084* (+31.3%) 17 (65.4%) 0.233* (+33.1%) 12 (46.2%) 0.359* (+26.0%) 15 (57.7%) Table 4.2.2 Contribution of different types of domain-specific knowledge Run Passage Aspect Document MAP Imprvd qs # (%) MAP Imprvd qs # (%) MAP Imprvd qs # (%) Baseline = Basic conceptual IR model 0.084 N/A 0.233 N/A 0.359 N/A Baseline+Synonyms 0.105 (+25%) 11 (42.3%) 0.246 (+5.6%) 9 (34.6%) 0.420 (+17%) 13 (50%) Baseline+Hypernyms 0.088 (+4.8%) 11 (42.3%) 0.225 (-3.4%) 9 (34.6%) 0.390 (+8.6%) 16 (61.5%) Baseline+Hyponyms 0.087 (+3.6%) 10 (38.5%) 0.217 (-6.9%) 7 (26.9%) 0.389 (+8.4%) 10 (38.5%) Baseline+Variants 0.150* (+78.6%) 16 (61.5%) 0.348* (+49.4%) 13 (50%) 0.495* (+37.9%) 10 (38.5%) Baseline+Related 0.086 (+2.4%) 9 (34.6%) 0.220 (-5.6%) 9 (34.6%) 0.387 (+7.8%) 13 (50%) Baseline+All 0.174* (107%) 25 (96.2%) 0.380* (+63.1%) 19 (73.1%) 0.537* (+49.6%) 14 (53.8%) Table 4.2.3 Contribution of abbreviation correction and pseudo-feedback Run Passage Aspect Document MAP Imprvd qs # (%) MAP Imprvd qs # (%) MAP Imprvd qs # (%) Baseline+All 0.174 N/A 0.380 N/A 0.537 N/A Baseline+All+Abbr 0.175 (+0.6%) 5 (19.2%) 0.375 (-1.3%) 4 (15.4%) 0.535 (-0.4%) 4 (15.4%) Baseline+All+Abbr+PF 0.182 (+4.6%) 10 (38.5%) 0.381 (+0.3%) 6 (23.1%) 0.539 (+0.4%) 9 (34.6%) A separate experiment has been done using a second testbed, the ad-hoc Task of TREC Genomics 2005, to evaluate our knowledge-intensive conceptual IR model for document retrieval of biomedical literature. The overall performance in terms of MAP is 35.50%, which is about 22.92% above the best reported result [9]. Notice that the performance was only measured on the document level for the ad-hoc Task of TREC Genomics 2005. 5. RELATED WORKS Many studies used manually-crafted thesauruses or knowledge databases created by text mining systems to improve retrieval effectiveness based on either word-statistical retrieval systems or conceptual retrieval systems. [11][1] assessed query expansion using the UMLS Metathesaurus. Based on a word-statistical retrieval system, [11] used definitions and different types of thesaurus relationships for query expansion and a deteriorated performance was reported. [1] expanded queries with phrases and UMLS concepts determined by the MetaMap, a program which maps biomedical text to UMLS concepts, and no significant improvement was shown. We used MeSH, Entrez gene, and other non-thesaurus knowledge resources such as an abbreviation database for query expansion. A critical difference between our work and those in [11][1] is that our retrieval model is based on concepts, not on individual words. The Genomics track in TREC provides a common platform to evaluate methods and techniques proposed by various groups for biomedical information retrieval. As summarized in [8][9][10], many groups utilized domain-specific knowledge to improve retrieval effectiveness. Among these groups, [3] assessed both thesaurus-based knowledge, such as gene information, and non thesaurus-based knowledge, such as lexical variants of gene symbols, for query expansion. They have shown that query expansion with acronyms and lexical variants of gene symbols produced the biggest improvement, whereas, the query expansion with gene information from gene databases deteriorated the performance. [21] used a similar approach for generating lexical variants of gene symbols and reported significant improvements. Our system utilized more types of domain-specific knowledge, including hyponyms, hypernyms and implicitly related concepts. In addition, under the conceptual retrieval framework, we examined more comprehensively the effects of different types of domain-specific knowledge in performance contribution. [20][15] utilized WordNet, a database of English words and their lexical relationships developed by Princeton University, for query expansion in the non-biomedical domain. In their studies, queries were expanded using the lexical semantic relations such as synonyms, hypernyms, or hyponyms. Little benefit has been shown in [20]. This has been due to ambiguity of the query terms which have different meanings in different contexts. When these synonyms having multiple meanings are added to the query, substantial irrelevant documents are retrieved. In the biomedical domain, this kind of ambiguity of query terms is relatively less frequent, because, although the abbreviations are highly ambiguous, general biomedical concepts usually have only one meaning in the thesaurus, such as UMLS, whereas a term in WordNet usually have multiple meanings (represented as synsets in WordNet). Besides, we have implemented a post-ranking step to reduce the number of incorrect matches of abbreviations, which will hopefully decrease the negative impact caused by the abbreviation ambiguity. Besides, we have implemented a postranking step to reduce the number of incorrect matches of abbreviations, which will hopefully decrease the negative impact caused by the abbreviation ambiguity. The retrieval model in [15] emphasized the similarity between a query and a document on the phrase level assuming that phrases are more important than individual words when retrieving documents. Although the assumption is similar, our conceptual model is based on the biomedical concepts, not phrases. [13] presented a good study of the role of knowledge in the document retrieval of clinical medicine. They have shown that appropriate use of semantic knowledge in a conceptual retrieval framework can yield substantial improvements. Although the retrieval model is similar, we made a study in the domain of genomics, in which the problem structure and task knowledge is not as well-defined as in the domain of clinical medicine [18]. Also, our similarity function is very different from that in [13]. In summary, our approach differs from previous works in four important ways: First, we present a case study of conceptual retrieval in the domain of genomics, where many knowledge resources can be used to improve the performance of biomedical IR systems. Second, we have studied more types of domainspecific knowledge than previous researchers and carried out more comprehensive experiments to look into the effects of different types of domain-specific knowledge in performance contribution. Third, although some of the techniques seem similar to previously published ones, they are actually quite different in details. For example, in our pseudo-feedback process, we require that the unit of feedback is a concept and the concept has to be of the same semantic type as a query concept. This is to ensure that our conceptual model of retrieval can be applied. As another example, the way in which implicitly related concepts are extracted in this paper is significantly different from that given in [19]. Finally, our conceptual IR model is actually based on complex concepts because some biomedical meanings, such as biological processes, are represented by multiple simple concepts. 6. CONCLUSION This paper proposed a conceptual approach to utilize domainspecific knowledge in an IR system to improve its effectiveness in retrieving biomedical literature. We specified five different types of domain-specific knowledge (i.e., synonyms, hyponyms, hypernyms, lexical variants, and implicitly related concepts) and examined their effects in performance contribution. We also evaluated other two techniques, pseudo-feedback and abbreviation correction. Experimental results have shown that appropriate use of domain-specific knowledge in a conceptual IR model yields significant improvements (23%) in passage retrieval over the best known results. In our future work, we will explore the use of other existing knowledge resources, such as UMLS and the Wikipedia, and evaluate techniques such as disambiguation of gene symbols for improving retrieval effectiveness. The application of our conceptual IR model in other domains such as clinical medicine will be investigated. 7. ACKNOWLEDGMENTS insightful discussion. 8. REFERENCES [1] Aronson A.R., Rindflesch T.C. Query expansion using the UMLS Metathesaurus. Proc AMIA Annu Fall Symp. 1997. 485-9. [2] Baeza-Yates R., Ribeiro-Neto B. Modern Information Retrieval. Addison-Wesley, 1999, 129-131. [3] Buttcher S., Clarke C.L.A., Cormack G.V. Domain-specific synonym expansion and validation for biomedical information retrieval (MultiText experiments for TREC 2004). TREC"04. [4] Chang J.T., Schutze H., Altman R.B. Creating an online dictionary of abbreviations from MEDLINE. Journal of the American Medical Informatics Association. 2002 9(6). [5] Church K.W., Hanks P. Word association norms, mutual information and lexicography. Computational Linguistics. 1990;16:22, C29. [6] Fontelo P., Liu F., Ackerman M. askMEDLINE: a free-text, natural language query tool for MEDLINE/PubMed. BMC Med Inform Decis Mak. 2005 Mar 10;5(1):5. [7] Fukuda K., Tamura A., Tsunoda T., Takagi T. Toward information extraction: identifying protein names from biological papers. Pac Symp Biocomput. 1998;:707-18. [8] Hersh W.R., and etc. TREC 2006 Genomics Track Overview. TREC"06. [9] Hersh W.R., and etc. TREC 2005 Genomics Track Overview. In TREC"05. [10] Hersh W.R., and etc. TREC 2004 Genomics Track Overview. In TREC"04. [11] Hersh W.R., Price S., Donohoe L. Assessing thesaurus-based query expansion using the UMLS Metathesaurus. Proc AMIA Symp. 344-8. 2000. [12] Levenshtein, V. Binary codes capable of correcting deletions, insertions, and reversals. Soviet Physics - Doklady 10, 10 (1996), 707-710. [13] Lin J., Demner-Fushman D. The Role of Knowledge in Conceptual Retrieval: A Study in the Domain of Clinical Medicine. SIGIR"06. 99-06. [14] Lindberg D., Humphreys B., and McCray A. The Unified Medical Language System. Methods of Information in Medicine. 32(4):281-291, 1993. [15] Liu S., Liu F., Yu C., and Meng W.Y. An Effective Approach to Document Retrieval via Utilizing WordNet and Recognizing Phrases. SIGIR"04. 266-272 [16] Proux D., Rechenmann F., Julliard L., Pillet V.V., Jacq B. Detecting Gene Symbols and Names in Biological Texts: A First Step toward Pertinent Information Extraction. Genome Inform Ser Workshop Genome Inform. 1998;9:72-80. [17] Robertson S.E., Walker S. Okapi/Keenbow at TREC-8. NIST Special Publication 500-246: TREC 8. [18] Sackett D.L., and etc. Evidence-Based Medicine: How to Practice and Teach EBM. Churchill Livingstone. Second edition, 2000. [19] Swanson,D.R., Smalheiser,N.R. An interactive system for finding complemen-tary literatures: a stimulus to scientific discovery. Artificial Intelligence, 1997; 91,183-203. [20] Voorhees E. Query expansion using lexical-semantic relations. SIGIR 1994. 61-9 [21] Zhong M., Huang X.J. Concept-based biomedical text retrieval. SIGIR"06. 723-4 [22] Zhou W., Torvik V.I., Smalheiser N.R. ADAM: Another Database of Abbreviations in MEDLINE. Bioinformatics. 2006; 22(22): 2813-2818.

keyword search;document collection;document map;domain-specific knowledge;retrieval model;biomedical document;document retrieval;query concept;conceptual ir model;passage-level information retrieval;passage map;passage extraction;aspect map

train_H-60

A Frequency-based and a Poisson-based Definition of the Probability of Being Informative

This paper reports on theoretical investigations about the assumptions underlying the inverse document frequency (idf ). We show that an intuitive idf -based probability function for the probability of a term being informative assumes disjoint document events. By assuming documents to be independent rather than disjoint, we arrive at a Poisson-based probability of being informative. The framework is useful for understanding and deciding the parameter estimation and combination in probabilistic retrieval models.

1. INTRODUCTION AND BACKGROUND The inverse document frequency (idf ) is one of the most successful parameters for a relevance-based ranking of retrieved objects. With N being the total number of documents, and n(t) being the number of documents in which term t occurs, the idf is defined as follows: idf(t) := − log n(t) N , 0 <= idf(t) < ∞ Ranking based on the sum of the idf -values of the query terms that occur in the retrieved documents works well, this has been shown in numerous applications. Also, it is well known that the combination of a document-specific term weight and idf works better than idf alone. This approach is known as tf-idf , where tf(t, d) (0 <= tf(t, d) <= 1) is the so-called term frequency of term t in document d. The idf reflects the discriminating power (informativeness) of a term, whereas the tf reflects the occurrence of a term. The idf alone works better than the tf alone does. An explanation might be the problem of tf with terms that occur in many documents; let us refer to those terms as noisy terms. We use the notion of noisy terms rather than frequent terms since frequent terms leaves open whether we refer to the document frequency of a term in a collection or to the so-called term frequency (also referred to as withindocument frequency) of a term in a document. We associate noise with the document frequency of a term in a collection, and we associate occurrence with the withindocument frequency of a term. The tf of a noisy term might be high in a document, but noisy terms are not good candidates for representing a document. Therefore, the removal of noisy terms (known as stopword removal) is essential when applying tf . In a tf-idf approach, the removal of stopwords is conceptually obsolete, if stopwords are just words with a low idf . From a probabilistic point of view, tf is a value with a frequency-based probabilistic interpretation whereas idf has an informative rather than a probabilistic interpretation. The missing probabilistic interpretation of idf is a problem in probabilistic retrieval models where we combine uncertain knowledge of different dimensions (e.g.: informativeness of terms, structure of documents, quality of documents, age of documents, etc.) such that a good estimate of the probability of relevance is achieved. An intuitive solution is a normalisation of idf such that we obtain values in the interval [0; 1]. For example, consider a normalisation based on the maximal idf -value. Let T be the set of terms occurring in a collection. Pfreq (t is informative) := idf(t) maxidf maxidf := max({idf(t)|t ∈ T}), maxidf <= − log(1/N) minidf := min({idf(t)|t ∈ T}), minidf >= 0 minidf maxidf ≤ Pfreq (t is informative) ≤ 1.0 This frequency-based probability function covers the interval [0; 1] if the minimal idf is equal to zero, which is the case if we have at least one term that occurs in all documents. Can we interpret Pfreq , the normalised idf , as the probability that the term is informative? When investigating the probabilistic interpretation of the 227 normalised idf , we made several observations related to disjointness and independence of document events. These observations are reported in section 3. We show in section 3.1 that the frequency-based noise probability n(t) N used in the classic idf -definition can be explained by three assumptions: binary term occurrence, constant document containment and disjointness of document containment events. In section 3.2 we show that by assuming independence of documents, we obtain 1 − e−1 ≈ 1 − 0.37 as the upper bound of the noise probability of a term. The value e−1 is related to the logarithm and we investigate in section 3.3 the link to information theory. In section 4, we link the results of the previous sections to probability theory. We show the steps from possible worlds to binomial distribution and Poisson distribution. In section 5, we emphasise that the theoretical framework of this paper is applicable for both idf and tf . Finally, in section 6, we base the definition of the probability of being informative on the results of the previous sections and compare frequency-based and Poisson-based definitions. 2. BACKGROUND The relationship between frequencies, probabilities and information theory (entropy) has been the focus of many researchers. In this background section, we focus on work that investigates the application of the Poisson distribution in IR since a main part of the work presented in this paper addresses the underlying assumptions of Poisson. [4] proposes a 2-Poisson model that takes into account the different nature of relevant and non-relevant documents, rare terms (content words) and frequent terms (noisy terms, function words, stopwords). [9] shows experimentally that most of the terms (words) in a collection are distributed according to a low dimension n-Poisson model. [10] uses a 2-Poisson model for including term frequency-based probabilities in the probabilistic retrieval model. The non-linear scaling of the Poisson function showed significant improvement compared to a linear frequency-based probability. The Poisson model was here applied to the term frequency of a term in a document. We will generalise the discussion by pointing out that document frequency and term frequency are dual parameters in the collection space and the document space, respectively. Our discussion of the Poisson distribution focuses on the document frequency in a collection rather than on the term frequency in a document. [7] and [6] address the deviation of idf and Poisson, and apply Poisson mixtures to achieve better Poisson-based estimates. The results proved again experimentally that a onedimensional Poisson does not work for rare terms, therefore Poisson mixtures and additional parameters are proposed. [3], section 3.3, illustrates and summarises comprehensively the relationships between frequencies, probabilities and Poisson. Different definitions of idf are put into context and a notion of noise is defined, where noise is viewed as the complement of idf . We use in our paper a different notion of noise: we consider a frequency-based noise that corresponds to the document frequency, and we consider a term noise that is based on the independence of document events. [11], [12], [8] and [1] link frequencies and probability estimation to information theory. [12] establishes a framework in which information retrieval models are formalised based on probabilistic inference. A key component is the use of a space of disjoint events, where the framework mainly uses terms as disjoint events. The probability of being informative defined in our paper can be viewed as the probability of the disjoint terms in the term space of [12]. [8] address entropy and bibliometric distributions. Entropy is maximal if all events are equiprobable and the frequency-based Lotka law (N/iλ is the number of scientists that have written i publications, where N and λ are distribution parameters), Zipf and the Pareto distribution are related. The Pareto distribution is the continuous case of the Lotka and Lotka and Zipf show equivalences. The Pareto distribution is used by [2] for term frequency normalisation. The Pareto distribution compares to the Poisson distribution in the sense that Pareto is fat-tailed, i. e. Pareto assigns larger probabilities to large numbers of events than Poisson distributions do. This makes Pareto interesting since Poisson is felt to be too radical on frequent events. We restrict in this paper to the discussion of Poisson, however, our results show that indeed a smoother distribution than Poisson promises to be a good candidate for improving the estimation of probabilities in information retrieval. [1] establishes a theoretical link between tf-idf and information theory and the theoretical research on the meaning of tf-idf clarifies the statistical model on which the different measures are commonly based. This motivation matches the motivation of our paper: We investigate theoretically the assumptions of classical idf and Poisson for a better understanding of parameter estimation and combination. 3. FROM DISJOINT TO INDEPENDENT We define and discuss in this section three probabilities: The frequency-based noise probability (definition 1), the total noise probability for disjoint documents (definition 2). and the noise probability for independent documents (definition 3). 3.1 Binary occurrence, constant containment and disjointness of documents We show in this section, that the frequency-based noise probability n(t) N in the idf definition can be explained as a total probability with binary term occurrence, constant document containment and disjointness of document containments. We refer to a probability function as binary if for all events the probability is either 1.0 or 0.0. The occurrence probability P(t|d) is binary, if P(t|d) is equal to 1.0 if t ∈ d, and P(t|d) is equal to 0.0, otherwise. P(t|d) is binary : ⇐⇒ P(t|d) = 1.0 ∨ P(t|d) = 0.0 We refer to a probability function as constant if for all events the probability is equal. The document containment probability reflect the chance that a document occurs in a collection. This containment probability is constant if we have no information about the document containment or we ignore that documents differ in containment. Containment could be derived, for example, from the size, quality, age, links, etc. of a document. For a constant containment in a collection with N documents, 1 N is often assumed as the containment probability. We generalise this definition and introduce the constant λ where 0 ≤ λ ≤ N. The containment of a document d depends on the collection c, this is reflected by the notation P(d|c) used for the containment 228 of a document. P(d|c) is constant : ⇐⇒ ∀d : P(d|c) = λ N For disjoint documents that cover the whole event space, we set λ = 1 and obtain Èd P(d|c) = 1.0. Next, we define the frequency-based noise probability and the total noise probability for disjoint documents. We introduce the event notation t is noisy and t occurs for making the difference between the noise probability P(t is noisy|c) in a collection and the occurrence probability P(t occurs|d) in a document more explicit, thereby keeping in mind that the noise probability corresponds to the occurrence probability of a term in a collection. Definition 1. The frequency-based term noise probability: Pfreq (t is noisy|c) := n(t) N Definition 2. The total term noise probability for disjoint documents: Pdis (t is noisy|c) := d P(t occurs|d) · P(d|c) Now, we can formulate a theorem that makes assumptions explicit that explain the classical idf . Theorem 1. IDF assumptions: If the occurrence probability P(t|d) of term t over documents d is binary, and the containment probability P(d|c) of documents d is constant, and document containments are disjoint events, then the noise probability for disjoint documents is equal to the frequency-based noise probability. Pdis (t is noisy|c) = Pfreq (t is noisy|c) Proof. The assumptions are: ∀d : (P(t occurs|d) = 1 ∨ P(t occurs|d) = 0) ∧ P(d|c) = λ N ∧ d P(d|c) = 1.0 We obtain: Pdis (t is noisy|c) = d|t∈d 1 N = n(t) N = Pfreq (t is noisy|c) The above result is not a surprise but it is a mathematical formulation of assumptions that can be used to explain the classical idf . The assumptions make explicit that the different types of term occurrence in documents (frequency of a term, importance of a term, position of a term, document part where the term occurs, etc.) and the different types of document containment (size, quality, age, etc.) are ignored, and document containments are considered as disjoint events. From the assumptions, we can conclude that idf (frequencybased noise, respectively) is a relatively simple but strict estimate. Still, idf works well. This could be explained by a leverage effect that justifies the binary occurrence and constant containment: The term occurrence for small documents tends to be larger than for large documents, whereas the containment for small documents tends to be smaller than for large documents. From that point of view, idf means that P(t ∧ d|c) is constant for all d in which t occurs, and P(t ∧ d|c) is zero otherwise. The occurrence and containment can be term specific. For example, set P(t∧d|c) = 1/ND(c) if t occurs in d, where ND(c) is the number of documents in collection c (we used before just N). We choose a document-dependent occurrence P(t|d) := 1/NT (d), i. e. the occurrence probability is equal to the inverse of NT (d), which is the total number of terms in document d. Next, we choose the containment P(d|c) := NT (d)/NT (c)·NT (c)/ND(c) where NT (d)/NT (c) is a document length normalisation (number of terms in document d divided by the number of terms in collection c), and NT (c)/ND(c) is a constant factor of the collection (number of terms in collection c divided by the number of documents in collection c). We obtain P(t∧d|c) = 1/ND(c). In a tf-idf -retrieval function, the tf -component reflects the occurrence probability of a term in a document. This is a further explanation why we can estimate the idf with a simple P(t|d), since the combined tf-idf contains the occurrence probability. The containment probability corresponds to a document normalisation (document length normalisation, pivoted document length) and is normally attached to the tf -component or the tf-idf -product. The disjointness assumption is typical for frequency-based probabilities. From a probability theory point of view, we can consider documents as disjoint events, in order to achieve a sound theoretical model for explaining the classical idf . But does disjointness reflect the real world where the containment of a document appears to be independent of the containment of another document? In the next section, we replace the disjointness assumption by the independence assumption. 3.2 The upper bound of the noise probability for independent documents For independent documents, we compute the probability of a disjunction as usual, namely as the complement of the probability of the conjunction of the negated events: P(d1 ∨ . . . ∨ dN ) = 1 − P(¬d1 ∧ . . . ∧ ¬dN ) = 1 − d (1 − P(d)) The noise probability can be considered as the conjunction of the term occurrence and the document containment. P(t is noisy|c) := P(t occurs ∧ (d1 ∨ . . . ∨ dN )|c) For disjoint documents, this view of the noise probability led to definition 2. For independent documents, we use now the conjunction of negated events. Definition 3. The term noise probability for independent documents: Pin (t is noisy|c) := d (1 − P(t occurs|d) · P(d|c)) With binary occurrence and a constant containment P(d|c) := λ/N, we obtain the term noise of a term t that occurs in n(t) documents: Pin (t is noisy|c) = 1 − 1 − λ N n(t) 229 For binary occurrence and disjoint documents, the containment probability was 1/N. Now, with independent documents, we can use λ as a collection parameter that controls the average containment probability. We show through the next theorem that the upper bound of the noise probability depends on λ. Theorem 2. The upper bound of being noisy: If the occurrence P(t|d) is binary, and the containment P(d|c) is constant, and document containments are independent events, then 1 − e−λ is the upper bound of the noise probability. ∀t : Pin (t is noisy|c) < 1 − e−λ Proof. The upper bound of the independent noise probability follows from the limit limN→∞(1 + x N )N = ex (see any comprehensive math book, for example, [5], for the convergence equation of the Euler function). With x = −λ, we obtain: lim N→∞ 1 − λ N N = e−λ For the term noise, we have: Pin (t is noisy|c) = 1 − 1 − λ N n(t) Pin (t is noisy|c) is strictly monotonous: The noise of a term tn is less than the noise of a term tn+1, where tn occurs in n documents and tn+1 occurs in n + 1 documents. Therefore, a term with n = N has the largest noise probability. For a collection with infinite many documents, the upper bound of the noise probability for terms tN that occur in all documents becomes: lim N→∞ Pin (tN is noisy) = lim N→∞ 1 − 1 − λ N N = 1 − e−λ By applying an independence rather a disjointness assumption, we obtain the probability e−1 that a term is not noisy even if the term does occur in all documents. In the disjoint case, the noise probability is one for a term that occurs in all documents. If we view P(d|c) := λ/N as the average containment, then λ is large for a term that occurs mostly in large documents, and λ is small for a term that occurs mostly in small documents. Thus, the noise of a term t is large if t occurs in n(t) large documents and the noise is smaller if t occurs in small documents. Alternatively, we can assume a constant containment and a term-dependent occurrence. If we assume P(d|c) := 1, then P(t|d) := λ/N can be interpreted as the average probability that t represents a document. The common assumption is that the average containment or occurrence probability is proportional to n(t). However, here is additional potential: The statistical laws (see [3] on Luhn and Zipf) indicate that the average probability could follow a normal distribution, i. e. small probabilities for small n(t) and large n(t), and larger probabilities for medium n(t). For the monotonous case we investigate here, the noise of a term with n(t) = 1 is equal to 1 − (1 − λ/N) = λ/N and the noise of a term with n(t) = N is close to 1− e−λ . In the next section, we relate the value e−λ to information theory. 3.3 The probability of a maximal informative signal The probability e−1 is special in the sense that a signal with that probability is a signal with maximal information as derived from the entropy definition. Consider the definition of the entropy contribution H(t) of a signal t. H(t) := P(t) · − ln P(t) We form the first derivation for computing the optimum. ∂H(t) ∂P(t) = − ln P(t) + −1 P(t) · P(t) = −(1 + ln P(t)) For obtaining optima, we use: 0 = −(1 + ln P(t)) The entropy contribution H(t) is maximal for P(t) = e−1 . This result does not depend on the base of the logarithm as we see next: ∂H(t) ∂P(t) = − logb P(t) + −1 P(t) · ln b · P(t) = − 1 ln b + logb P(t) = − 1 + ln P(t) ln b We summarise this result in the following theorem: Theorem 3. The probability of a maximal informative signal: The probability Pmax = e−1 ≈ 0.37 is the probability of a maximal informative signal. The entropy of a maximal informative signal is Hmax = e−1 . Proof. The probability and entropy follow from the derivation above. The complement of the maximal noise probability is e−λ and we are looking now for a generalisation of the entropy definition such that e−λ is the probability of a maximal informative signal. We can generalise the entropy definition by computing the integral of λ+ ln P(t), i. e. this derivation is zero for e−λ . We obtain a generalised entropy: −(λ + ln P(t)) d(P(t)) = P(t) · (1 − λ − ln P(t)) The generalised entropy corresponds for λ = 1 to the classical entropy. By moving from disjoint to independent documents, we have established a link between the complement of the noise probability of a term that occurs in all documents and information theory. Next, we link independent documents to probability theory. 4. THE LINK TO PROBABILITY THEORY We review for independent documents three concepts of probability theory: possible worlds, binomial distribution and Poisson distribution. 4.1 Possible Worlds Each conjunction of document events (for each document, we consider two document events: the document can be true or false) is associated with a so-called possible world. For example, consider the eight possible worlds for three documents (N = 3). 230 world w conjunction w7 d1 ∧ d2 ∧ d3 w6 d1 ∧ d2 ∧ ¬d3 w5 d1 ∧ ¬d2 ∧ d3 w4 d1 ∧ ¬d2 ∧ ¬d3 w3 ¬d1 ∧ d2 ∧ d3 w2 ¬d1 ∧ d2 ∧ ¬d3 w1 ¬d1 ∧ ¬d2 ∧ d3 w0 ¬d1 ∧ ¬d2 ∧ ¬d3 With each world w, we associate a probability µ(w), which is equal to the product of the single probabilities of the document events. world w probability µ(w) w7 λ N ¡3 · 1 − λ N ¡0 w6 λ N ¡2 · 1 − λ N ¡1 w5 λ N ¡2 · 1 − λ N ¡1 w4 λ N ¡1 · 1 − λ N ¡2 w3 λ N ¡2 · 1 − λ N ¡1 w2 λ N ¡1 · 1 − λ N ¡2 w1 λ N ¡1 · 1 − λ N ¡2 w0 λ N ¡0 · 1 − λ N ¡3 The sum over the possible worlds in which k documents are true and N −k documents are false is equal to the probability function of the binomial distribution, since the binomial coefficient yields the number of possible worlds in which k documents are true. 4.2 Binomial distribution The binomial probability function yields the probability that k of N events are true where each event is true with the single event probability p. P(k) := binom(N, k, p) := N k pk (1 − p)N −k The single event probability is usually defined as p := λ/N, i. e. p is inversely proportional to N, the total number of events. With this definition of p, we obtain for an infinite number of documents the following limit for the product of the binomial coefficient and pk : lim N→∞ N k pk = = lim N→∞ N · (N −1) · . . . · (N −k +1) k! λ N k = λk k! The limit is close to the actual value for k << N. For large k, the actual value is smaller than the limit. The limit of (1−p)N −k follows from the limit limN→∞(1+ x N )N = ex . lim N→∞ (1 − p)N−k = lim N→∞ 1 − λ N N −k = lim N→∞ e−λ · 1 − λ N −k = e−λ Again, the limit is close to the actual value for k << N. For large k, the actual value is larger than the limit. 4.3 Poisson distribution For an infinite number of events, the Poisson probability function is the limit of the binomial probability function. lim N→∞ binom(N, k, p) = λk k! · e−λ P(k) = poisson(k, λ) := λk k! · e−λ The probability poisson(0, 1) is equal to e−1 , which is the probability of a maximal informative signal. This shows the relationship of the Poisson distribution and information theory. After seeing the convergence of the binomial distribution, we can choose the Poisson distribution as an approximation of the independent term noise probability. First, we define the Poisson noise probability: Definition 4. The Poisson term noise probability: Ppoi (t is noisy|c) := e−λ · n(t) k=1 λk k! For independent documents, the Poisson distribution approximates the probability of the disjunction for large n(t), since the independent term noise probability is equal to the sum over the binomial probabilities where at least one of n(t) document containment events is true. Pin (t is noisy|c) = n(t) k=1 n(t) k pk (1 − p)N −k Pin (t is noisy|c) ≈ Ppoi (t is noisy|c) We have defined a frequency-based and a Poisson-based probability of being noisy, where the latter is the limit of the independence-based probability of being noisy. Before we present in the final section the usage of the noise probability for defining the probability of being informative, we emphasise in the next section that the results apply to the collection space as well as to the the document space. 5. THE COLLECTION SPACE AND THE DOCUMENT SPACE Consider the dual definitions of retrieval parameters in table 1. We associate a collection space D × T with a collection c where D is the set of documents and T is the set of terms in the collection. Let ND := |D| and NT := |T| be the number of documents and terms, respectively. We consider a document as a subset of T and a term as a subset of D. Let nT (d) := |{t|d ∈ t}| be the number of terms that occur in the document d, and let nD(t) := |{d|t ∈ d}| be the number of documents that contain the term t. In a dual way, we associate a document space L × T with a document d where L is the set of locations (also referred to as positions, however, we use the letters L and l and not P and p for avoiding confusion with probabilities) and T is the set of terms in the document. The document dimension in a collection space corresponds to the location (position) dimension in a document space. The definition makes explicit that the classical notion of term frequency of a term in a document (also referred to as the within-document term frequency) actually corresponds to the location frequency of a term in a document. For the 231 space collection document dimensions documents and terms locations and terms document/location frequency nD(t, c): Number of documents in which term t occurs in collection c nL(t, d): Number of locations (positions) at which term t occurs in document d ND(c): Number of documents in collection c NL(d): Number of locations (positions) in document d term frequency nT (d, c): Number of terms that document d contains in collection c nT (l, d): Number of terms that location l contains in document d NT (c): Number of terms in collection c NT (d): Number of terms in document d noise/occurrence P(t|c) (term noise) P(t|d) (term occurrence) containment P(d|c) (document) P(l|d) (location) informativeness − ln P(t|c) − ln P(t|d) conciseness − ln P(d|c) − ln P(l|d) P(informative) ln(P(t|c))/ ln(P(tmin, c)) ln(P(t|d))/ ln(P(tmin, d)) P(concise) ln(P(d|c))/ ln(P(dmin|c)) ln(P(l|d))/ ln(P(lmin|d)) Table 1: Retrieval parameters actual term frequency value, it is common to use the maximal occurrence (number of locations; let lf be the location frequency). tf(t, d):=lf(t, d):= Pfreq (t occurs|d) Pfreq (tmax occurs|d) = nL(t, d) nL(tmax , d) A further duality is between informativeness and conciseness (shortness of documents or locations): informativeness is based on occurrence (noise), conciseness is based on containment. We have highlighted in this section the duality between the collection space and the document space. We concentrate in this paper on the probability of a term to be noisy and informative. Those probabilities are defined in the collection space. However, the results regarding the term noise and informativeness apply to their dual counterparts: term occurrence and informativeness in a document. Also, the results can be applied to containment of documents and locations. 6. THE PROBABILITY OF BEING INFORMATIVE We showed in the previous sections that the disjointness assumption leads to frequency-based probabilities and that the independence assumption leads to Poisson probabilities. In this section, we formulate a frequency-based definition and a Poisson-based definition of the probability of being informative and then we compare the two definitions. Definition 5. The frequency-based probability of being informative: Pfreq (t is informative|c) := − ln n(t) N − ln 1 N = − logN n(t) N = 1 − logN n(t) = 1 − ln n(t) ln N We define the Poisson-based probability of being informative analogously to the frequency-based probability of being informative (see definition 5). Definition 6. The Poisson-based probability of being informative: Ppoi (t is informative|c) := − ln e−λ · Èn(t) k=1 λk k! − ln(e−λ · λ) = λ − ln Èn(t) k=1 λk k! λ − ln λ For the sum expression, the following limit holds: lim n(t)→∞ n(t) k=1 λk k! = eλ − 1 For λ >> 1, we can alter the noise and informativeness Poisson by starting the sum from 0, since eλ >> 1. Then, the minimal Poisson informativeness is poisson(0, λ) = e−λ . We obtain a simplified Poisson probability of being informative: Ppoi (t is informative|c) ≈ λ − ln Èn(t) k=0 λk k! λ = 1 − ln Èn(t) k=0 λk k! λ The computation of the Poisson sum requires an optimisation for large n(t). The implementation for this paper exploits the nature of the Poisson density: The Poisson density yields only values significantly greater than zero in an interval around λ. Consider the illustration of the noise and informativeness definitions in figure 1. The probability functions displayed are summarised in figure 2 where the simplified Poisson is used in the noise and informativeness graphs. The frequency-based noise corresponds to the linear solid curve in the noise figure. With an independence assumption, we obtain the curve in the lower triangle of the noise figure. By changing the parameter p := λ/N of the independence probability, we can lift or lower the independence curve. The noise figure shows the lifting for the value λ := ln N ≈ 9.2. The setting λ = ln N is special in the sense that the frequency-based and the Poisson-based informativeness have the same denominator, namely ln N, and the Poisson sum converges to λ. Whether we can draw more conclusions from this setting is an open question. We can conclude, that the lifting is desirable if we know for a collection that terms that occur in relatively few doc232 0 0.2 0.4 0.6 0.8 1 0 2000 4000 6000 8000 10000 Probabilityofbeingnoisy n(t): Number of documents with term t frequency independence: 1/N independence: ln(N)/N poisson: 1000 poisson: 2000 poisson: 1000,2000 0 0.2 0.4 0.6 0.8 1 0 2000 4000 6000 8000 10000 Probabilityofbeinginformative n(t): Number of documents with term t frequency independence: 1/N independence: ln(N)/N poisson: 1000 poisson: 2000 poisson: 1000,2000 Figure 1: Noise and Informativeness Probability function Noise Informativeness Frequency Pfreq Def n(t)/N ln(n(t)/N)/ ln(1/N) Interval 1/N ≤ Pfreq ≤ 1.0 0.0 ≤ Pfreq ≤ 1.0 Independence Pin Def 1 − (1 − p)n(t) ln(1 − (1 − p)n(t) )/ ln(p) Interval p ≤ Pin < 1 − e−λ ln(p) ≤ Pin ≤ 1.0 Poisson Ppoi Def e−λ Èn(t) k=1 λk k! (λ − ln Èn(t) k=1 λk k! )/(λ − ln λ) Interval e−λ · λ ≤ Ppoi < 1 − e−λ (λ − ln(eλ − 1))/(λ − ln λ) ≤ Ppoi ≤ 1.0 Poisson Ppoi simplified Def e−λ Èn(t) k=0 λk k! (λ − ln Èn(t) k=0 λk k! )/λ Interval e−λ ≤ Ppoi < 1.0 0.0 < Ppoi ≤ 1.0 Figure 2: Probability functions uments are no guarantee for finding relevant documents, i. e. we assume that rare terms are still relatively noisy. On the opposite, we could lower the curve when assuming that frequent terms are not too noisy, i. e. they are considered as being still significantly discriminative. The Poisson probabilities approximate the independence probabilities for large n(t); the approximation is better for larger λ. For n(t) < λ, the noise is zero whereas for n(t) > λ the noise is one. This radical behaviour can be smoothened by using a multi-dimensional Poisson distribution. Figure 1 shows a Poisson noise based on a two-dimensional Poisson: poisson(k, λ1, λ2) := π · e−λ1 · λk 1 k! + (1 − π) · e−λ2 · λk 2 k! The two dimensional Poisson shows a plateau between λ1 = 1000 and λ2 = 2000, we used here π = 0.5. The idea behind this setting is that terms that occur in less than 1000 documents are considered to be not noisy (i.e. they are informative), that terms between 1000 and 2000 are half noisy, and that terms with more than 2000 are definitely noisy. For the informativeness, we observe that the radical behaviour of Poisson is preserved. The plateau here is approximately at 1/6, and it is important to realise that this plateau is not obtained with the multi-dimensional Poisson noise using π = 0.5. The logarithm of the noise is normalised by the logarithm of a very small number, namely 0.5 · e−1000 + 0.5 · e−2000 . That is why the informativeness will be only close to one for very little noise, whereas for a bit of noise, informativeness will drop to zero. This effect can be controlled by using small values for π such that the noise in the interval [λ1; λ2] is still very little. The setting π = e−2000/6 leads to noise values of approximately e−2000/6 in the interval [λ1; λ2], the logarithms lead then to 1/6 for the informativeness. The indepence-based and frequency-based informativeness functions do not differ as much as the noise functions do. However, for the indepence-based probability of being informative, we can control the average informativeness by the definition p := λ/N whereas the control on the frequencybased is limited as we address next. For the frequency-based idf , the gradient is monotonously decreasing and we obtain for different collections the same distances of idf -values, i. e. the parameter N does not affect the distance. For an illustration, consider the distance between the value idf(tn+1) of a term tn+1 that occurs in n+1 documents, and the value idf(tn) of a term tn that occurs in n documents. idf(tn+1) − idf(tn) = ln n n + 1 The first three values of the distance function are: idf(t2) − idf(t1) = ln(1/(1 + 1)) = 0.69 idf(t3) − idf(t2) = ln(1/(2 + 1)) = 0.41 idf(t4) − idf(t3) = ln(1/(3 + 1)) = 0.29 For the Poisson-based informativeness, the gradient decreases first slowly for small n(t), then rapidly near n(t) ≈ λ and then it grows again slowly for large n(t). In conclusion, we have seen that the Poisson-based definition provides more control and parameter possibilities than 233 the frequency-based definition does. Whereas more control and parameter promises to be positive for the personalisation of retrieval systems, it bears at the same time the danger of just too many parameters. The framework presented in this paper raises the awareness about the probabilistic and information-theoretic meanings of the parameters. The parallel definitions of the frequency-based probability and the Poisson-based probability of being informative made the underlying assumptions explicit. The frequency-based probability can be explained by binary occurrence, constant containment and disjointness of documents. Independence of documents leads to Poisson, where we have to be aware that Poisson approximates the probability of a disjunction for a large number of events, but not for a small number. This theoretical result explains why experimental investigations on Poisson (see [7]) show that a Poisson estimation does work better for frequent (bad, noisy) terms than for rare (good, informative) terms. In addition to the collection-wide parameter setting, the framework presented here allows for document-dependent settings, as explained for the independence probability. This is in particular interesting for heterogeneous and structured collections, since documents are different in nature (size, quality, root document, sub document), and therefore, binary occurrence and constant containment are less appropriate than in relatively homogeneous collections. 7. SUMMARY The definition of the probability of being informative transforms the informative interpretation of the idf into a probabilistic interpretation, and we can use the idf -based probability in probabilistic retrieval approaches. We showed that the classical definition of the noise (document frequency) in the inverse document frequency can be explained by three assumptions: the term within-document occurrence probability is binary, the document containment probability is constant, and the document containment events are disjoint. By explicitly and mathematically formulating the assumptions, we showed that the classical definition of idf does not take into account parameters such as the different nature (size, quality, structure, etc.) of documents in a collection, or the different nature of terms (coverage, importance, position, etc.) in a document. We discussed that the absence of those parameters is compensated by a leverage effect of the within-document term occurrence probability and the document containment probability. By applying an independence rather a disjointness assumption for the document containment, we could establish a link between the noise probability (term occurrence in a collection), information theory and Poisson. From the frequency-based and the Poisson-based probabilities of being noisy, we derived the frequency-based and Poisson-based probabilities of being informative. The frequency-based probability is relatively smooth whereas the Poisson probability is radical in distinguishing between noisy or not noisy, and informative or not informative, respectively. We showed how to smoothen the radical behaviour of Poisson with a multidimensional Poisson. The explicit and mathematical formulation of idf - and Poisson-assumptions is the main result of this paper. Also, the paper emphasises the duality of idf and tf , collection space and document space, respectively. Thus, the result applies to term occurrence and document containment in a collection, and it applies to term occurrence and position containment in a document. This theoretical framework is useful for understanding and deciding the parameter estimation and combination in probabilistic retrieval models. The links between indepence-based noise as document frequency, probabilistic interpretation of idf , information theory and Poisson described in this paper may lead to variable probabilistic idf and tf definitions and combinations as required in advanced and personalised information retrieval systems. Acknowledgment: I would like to thank Mounia Lalmas, Gabriella Kazai and Theodora Tsikrika for their comments on the as they said heavy pieces. My thanks also go to the meta-reviewer who advised me to improve the presentation to make it less formidable and more accessible for those without a theoretic bent. This work was funded by a research fellowship from Queen Mary University of London. 8. REFERENCES [1] A. Aizawa. An information-theoretic perspective of tf-idf measures. Information Processing and Management, 39:45-65, January 2003. [2] G. Amati and C. J. Rijsbergen. Term frequency normalization via Pareto distributions. In 24th BCS-IRSG European Colloquium on IR Research, Glasgow, Scotland, 2002. [3] R. K. Belew. Finding out about. Cambridge University Press, 2000. [4] A. Bookstein and D. Swanson. Probabilistic models for automatic indexing. Journal of the American Society for Information Science, 25:312-318, 1974. [5] I. N. Bronstein. Taschenbuch der Mathematik. Harri Deutsch, Thun, Frankfurt am Main, 1987. [6] K. Church and W. Gale. Poisson mixtures. Natural Language Engineering, 1(2):163-190, 1995. [7] K. W. Church and W. A. Gale. Inverse document frequency: A measure of deviations from poisson. In Third Workshop on Very Large Corpora, ACL Anthology, 1995. [8] T. Lafouge and C. Michel. Links between information construction and information gain: Entropy and bibliometric distribution. Journal of Information Science, 27(1):39-49, 2001. [9] E. Margulis. N-poisson document modelling. In Proceedings of the 15th Annual International ACM SIGIR Conference on Research and Development in Information Retrieval, pages 177-189, 1992. [10] S. E. Robertson and S. Walker. Some simple effective approximations to the 2-poisson model for probabilistic weighted retrieval. In Proceedings of the 17th Annual International ACM SIGIR Conference on Research and Development in Information Retrieval, pages 232-241, London, et al., 1994. Springer-Verlag. [11] S. Wong and Y. Yao. An information-theoric measure of term specificity. Journal of the American Society for Information Science, 43(1):54-61, 1992. [12] S. Wong and Y. Yao. On modeling information retrieval with probabilistic inference. ACM Transactions on Information Systems, 13(1):38-68, 1995. 234

idf;informativeness;document disjointness;poisson distribution;probability theory;probabilistic information retrieval;information theory;independence assumption;inverse document frequency;information retrieval;poisson-based probability;collection space;frequency-based probability;noise probability;probability function;disjointness of document

train_H-61

Impedance Coupling in Content-targeted Advertising

The current boom of the Web is associated with the revenues originated from on-line advertising. While search-based advertising is dominant, the association of ads with a Web page (during user navigation) is becoming increasingly important. In this work, we study the problem of associating ads with a Web page, referred to as content-targeted advertising, from a computer science perspective. We assume that we have access to the text of the Web page, the keywords declared by an advertiser, and a text associated with the advertiser"s business. Using no other information and operating in fully automatic fashion, we propose ten strategies for solving the problem and evaluate their effectiveness. Our methods indicate that a matching strategy that takes into account the semantics of the problem (referred to as AAK for ads and keywords) can yield gains in average precision figures of 60% compared to a trivial vector-based strategy. Further, a more sophisticated impedance coupling strategy, which expands the text of the Web page to reduce vocabulary impedance with regard to an advertisement, can yield extra gains in average precision of 50%. These are first results. They suggest that great accuracy in content-targeted advertising can be attained with appropriate algorithms.

1. INTRODUCTION The emergence of the Internet has opened up new marketing opportunities. In fact, a company has now the possibility of showing its advertisements (ads) to millions of people at a low cost. During the 90"s, many companies invested heavily on advertising in the Internet with apparently no concerns about their investment return [16]. This situation radically changed in the following decade when the failure of many Web companies led to a dropping in supply of cheap venture capital and a considerable reduction in on-line advertising investments [15,16]. It was clear then that more effective strategies for on-line advertising were required. For that, it was necessary to take into account short-term and long-term interests of the users related to their information needs [9,14]. As a consequence, many companies intensified the adoption of intrusive techniques for gathering information of users mostly without their consent [8]. This raised privacy issues which stimulated the research for less invasive measures [16]. More recently, Internet information gatekeepers as, for example, search engines, recommender systems, and comparison shopping services, have employed what is called paid placement strategies [3]. In such methods, an advertiser company is given prominent positioning in advertisement lists in return for a placement fee. Amongst these methods, the most popular one is a non-intrusive technique called keyword targeted marketing [16]. In this technique, keywords extracted from the user"s search query are matched against keywords associated with ads provided by advertisers. A ranking of the ads, which also takes into consideration the amount that each advertiser is willing to pay, is computed. The top ranked ads are displayed in the search result page together with the answers for the user query. The success of keyword targeted marketing has motivated information gatekeepers to offer their advertisement services in different contexts. For example, as shown in Figure 1, relevant ads could be shown to users directly in the pages of information portals. The motivation is to take advantage of 496 the users immediate information interests at browsing time. The problem of matching ads to a Web page that is browsed, which we also refer to as content-targeted advertising [1], is different from that of keyword marketing. In this case, instead of dealing with users" keywords, we have to use the contents of a Web page to decide which ads to display. Figure 1: Example of content-based advertising in the page of a newspaper. The middle slice of the page shows the beginning of an article about the launch of a DVD movie. At the bottom slice, we can see advertisements picked for this page by Google"s content-based advertising system, AdSense. It is important to notice that paid placement advertising strategies imply some risks to information gatekeepers. For instance, there is the possibility of a negative impact on their credibility which, at long term, can demise their market share [3]. This makes investments in the quality of ad recommendation systems even more important to minimize the possibility of exhibiting ads unrelated to the user"s interests. By investing in their ad systems, information gatekeepers are investing in the maintenance of their credibility and in the reinforcement of a positive user attitude towards the advertisers and their ads [14]. Further, that can translate into higher clickthrough rates that lead to an increase in revenues for information gatekeepers and advertisers, with gains to all parts [3]. In this work, we focus on the problem of content-targeted advertising. We propose new strategies for associating ads with a Web page. Five of these strategies are referred to as matching strategies. They are based on the idea of matching the text of the Web page directly to the text of the ads and its associated keywords. Five other strategies, which we here introduce, are referred to as impedance coupling strategies. They are based on the idea of expanding the Web page with new terms to facilitate the task of matching ads and Web pages. This is motivated by the observation that there is frequently a mismatch between the vocabulary of a Web page and the vocabulary of an advertisement. We say that there is a vocabulary impedance problem and that our technique provides a positive effect of impedance coupling by reducing the vocabulary impedance. Further, all our strategies rely on information that is already available to information gatekeepers that operate keyword targeted advertising systems. Thus, no other data from the advertiser is required. Using a sample of a real case database with over 93,000 ads and 100 Web pages selected for testing, we evaluate our ad recommendation strategies. First, we evaluate the five matching strategies. They match ads to a Web page using a standard vector model and provide what we may call trivial solutions. Our results indicate that a strategy that matches the ad plus its keywords to a Web page, requiring the keywords to appear in the Web page, provides improvements in average precision figures of roughly 60% relative to a strategy that simply matches the ads to the Web page. Such strategy, which we call AAK (for ads and keywords), is then taken as our baseline. Following we evaluate the five impedance coupling strategies. They are based on the idea of expanding the ad and the Web page with new terms to reduce the vocabulary impedance between their texts. Our results indicate that it is possible to generate extra improvements in average precision figures of roughly 50% relative to the AAK strategy. The paper is organized as follows. In section 2, we introduce five matching strategies to solve content-targeted advertising. In section 3, we present our impedance coupling strategies. In section 4, we describe our experimental methodology and datasets and discuss our results. In section 5 we discuss related work. In section 6 we present our conclusions. 2. MATCHING STRATEGIES Keyword advertising relies on matching search queries to ads and its associated keywords. Context-based advertising, which we address here, relies on matching ads and its associated keywords to the text of a Web page. Given a certain Web page p, which we call triggering page, our task is to select advertisements related to the contents of p. Without loss of generality, we consider that an advertisement ai is composed of a title, a textual description, and a hyperlink. To illustrate, for the first ad by Google shown in Figure 1, the title is Star Wars Trilogy Full, the description is Get this popular DVD free. Free w/ free shopping. Sign up now, and the hyperlink points to the site www.freegiftworld.com. Advertisements can be grouped by advertisers in groups called campaigns, such that a campaign can have one or more advertisements. Given our triggering page p and a set A of ads, a simple way of ranking ai ∈ A with regard to p is by matching the contents of p to the contents of ai. For this, we use the vector space model [2], as discussed in the immediately following. In the vector space model, queries and documents are represented as weighted vectors in an n-dimensional space. Let wiq be the weight associated with term ti in the query q and wij be the weight associated with term ti in the document dj. Then, q = (w1q, w2q, ..., wiq, ..., wnq) and dj = (w1j, w2j, ..., wij, ..., wnj) are the weighted vectors used to represent the query q and the document dj. These weights can be computed using classic tf-idf schemes. In such schemes, weights are taken as the product between factors that quantify the importance of a term in a document (given by the term frequency, or tf, factor) and its rarity in the whole collection (given by the inverse document factor, or idf, factor), see [2] for details. The ranking of the query q with regard to the document dj is computed by the cosine similarity 497 formula, that is, the cosine of the angle between the two corresponding vectors: sim(q, dj) = q • dj |q| × |dj| = Pn i=1 wiq · wij qPn i=1 w2 iq qPn i=1 w2 ij (1) By considering p as the query and ai as the document, we can rank the ads with regard to the Web page p. This is our first matching strategy. It is represented by the function AD given by: AD(p, ai) = sim(p, ai) where AD stands for direct match of the ad, composed by title and description and sim(p, ai) is computed according to Eq. (1). In our second method, we use other source of evidence provided by the advertisers: the keywords. With each advertisement ai an advertiser associates a keyword ki, which may be composed of one or more terms. We denote the association between an advertisement ai and a keyword ki as the pair (ai, ki) ∈ K, where K is the set of associations made by the advertisers. In the case of keyword targeted advertising, such keywords are used to match the ads to the user queries. In here, we use them to match ads to the Web page p. This provides our second method for ad matching given by: KW(p, ai) = sim(p, ki) where (ai, ki) ∈ K and KW stands for match the ad keywords. We notice that most of the keywords selected by advertisers are also present in the ads associated with those keywords. For instance, in our advertisement test collection, this is true for 90% of the ads. Thus, instead of using the keywords as matching devices, we can use them to emphasize the main concepts in an ad, in an attempt to improve our AD strategy. This leads to our third method of ad matching given by: AD KW(p, ai) = sim(p, ai ∪ ki) where (ai, ki) ∈ K and AD KW stands for match the ad and its keywords. Finally, it is important to notice that the keyword ki associated with ai could not appear at all in the triggering page p, even when ai is highly ranked. However, if we assume that ki summarizes the main topic of ai according to an advertiser viewpoint, it can be interesting to assure its presence in p. This reasoning suggests that requiring the occurrence of the keyword ki in the triggering page p as a condition to associate ai with p might lead to improved results. This leads to two extra matching strategies as follows: ANDKW(p, ai) =  sim(p, ai) if ki p 0 if otherwise AD ANDKW(p, ai) = AAK(p, ai) =  sim(p, ai ∪ ki) if ki p 0 if otherwise where (ai, ki) ∈ K, ANDKW stands for match the ad keywords and force their appearance, and AD ANDKW (or AAK for ads and keywords) stands for match the ad, its keywords, and force their appearance. As we will see in our results, the best among these simple methods is AAK. Thus, it will be used as baseline for our impedance coupling strategies which we now discuss. 3. IMPEDANCE COUPLING STRATEGIES Two key issues become clear as one plays with the contenttargeted advertising problem. First, the triggering page normally belongs to a broader contextual scope than that of the advertisements. Second, the association between a good advertisement and the triggering page might depend on a topic that is not mentioned explicitly in the triggering page. The first issue is due to the fact that Web pages can be about any subject and that advertisements are concise in nature. That is, ads tend to be more topic restricted than Web pages. The second issue is related to the fact that, as we later discuss, most advertisers place a small number of advertisements. As a result, we have few terms describing their interest areas. Consequently, these terms tend to be of a more general nature. For instance, a car shop probably would prefer to use car instead of super sport to describe its core business topic. As a consequence, many specific terms that appear in the triggering page find no match in the advertisements. To make matters worst, a page might refer to an entity or subject of the world through a label that is distinct from the label selected by an advertiser to refer to the same entity. A consequence of these two issues is that vocabularies of pages and ads have low intersection, even when an ad is related to a page. We cite this problem from now on as the vocabulary impedance problem. In our experiments, we realized that this problem limits the final quality of direct matching strategies. Therefore, we studied alternatives to reduce the referred vocabulary impedance. For this, we propose to expand the triggering pages with new terms. Figure 2 illustrates our intuition. We already know that the addition of keywords (selected by the advertiser) to the ads leads to improved results. We say that a keyword reduces the vocabulary impedance by providing an alternative matching path. Our idea is to add new terms (words) to the Web page p to also reduce the vocabulary impedance by providing a second alternative matching path. We refer to our expansion technique as impedance coupling. For this, we proceed as follows. expansion terms keyword vocabulary impedance triggering page p ad Figure 2: Addition of new terms to a Web page to reduce the vocabulary impedance. An advertiser trying to describe a certain topic in a concise way probably will choose general terms to characterize that topic. To facilitate the matching between this ad and our triggering page p, we need to associate new general terms with p. For this, we assume that Web documents similar to the triggering page p share common topics. Therefore, 498 by inspecting the vocabulary of these similar documents we might find good terms for better characterizing the main topics in the page p. We now describe this idea using a Bayesian network model [10,11,13] depicted in Figure 3. R D0 D1 Dj Dk T1 T2 T3 Ti Tm ... ... ... ... Figure 3: Bayesian network model for our impedance coupling technique. In our model, which is based on the belief network in [11], the nodes represent pieces of information in the domain. With each node is associated a binary random variable, which takes the value 1 to mean that the corresponding entity (a page or terms) is observed and, thus, relevant in our computations. In this case, we say that the information was observed. Node R represents the page r, a new representation for the triggering page p. Let N be the set of the k most similar documents to the triggering page, including the triggering page p itself, in a large enough Web collection C. Root nodes D0 through Dk represent the documents in N, that is, the triggering page D0 and its k nearest neighbors, D1 through Dk, among all pages in C. There is an edge from node Dj to node R if document dj is in N. Nodes T1 through Tm represent the terms in the vocabulary of C. There is an edge from node Dj to a node Ti if term ti occurs in document dj. In our model, the observation of the pages in N leads to the observation of a new representation of the triggering page p and to a set of terms describing the main topics associated with p and its neighbors. Given these definitions, we can now use the network to determine the probability that a term ti is a good term for representing a topic of the triggering page p. In other words, we are interested in the probability of observing the final evidence regarding a term ti, given that the new representation of the page p has been observed, P(Ti = 1|R = 1). This translates into the following equation1 : P(Ti|R) = 1 P(R) X d P(Ti|d)P(R|d)P(d) (2) where d represents the set of states of the document nodes. Since we are interested just in the states in which only a single document dj is observed and P(d) can be regarded as a constant, we can rewrite Eq. (2) as: P(Ti|R) = ν P(R) kX j=0 P(Ti|dj)P(R|dj) (3) where dj represents the state of the document nodes in which only document dj is observed and ν is a constant 1 To simplify our notation we represent the probabilities P(X = 1) as P(X) and P(X = 0) as P(X). associated with P(dj). Eq. (3) is the general equation to compute the probability that a term ti is related to the triggering page. We now define the probabilities P(Ti|dj) and P(R|dj) as follows: P(Ti|dj) = η wij (4) P(R|dj) =  (1 − α) j = 0 α sim(r, dj) 1 ≤ j ≤ k (5) where η is a normalizing constant, wij is the weight associated with term ti in the document dj, and sim(p, dj) is given by Eq. (1), i.e., is the cosine similarity between p and dj. The weight wij is computed using a classic tf-idf scheme and is zero if term ti does not occur in document dj. Notice that P(Ti|dj) = 1 − P(Ti|dj) and P(R|dj) = 1 − P(R|dj). By defining the constant α, it is possible to determine how important should be the influence of the triggering page p to its new representation r. By substituting Eq. (4) and Eq. (5) into Eq. (3), we obtain: P(Ti|R) = ρ ((1 − α) wi0 + α kX j=1 wij sim(r, dj)) (6) where ρ = η ν is a normalizing constant. We use Eq. (6) to determine the set of terms that will compose r, as illustrated in Figure 2. Let ttop be the top ranked term according to Eq. (6). The set r is composed of the terms ti such that P (Ti|R) P (Ttop|R) ≥ β, where β is a given threshold. In our experiments, we have used β = 0.05. Notice that the set r might contain terms that already occur in p. That is, while we will refer to the set r as expansion terms, it should be clear that p ∩ r = ∅. By using α = 0, we simply consider the terms originally in page p. By increasing α, we relax the context of the page p, adding terms from neighbor pages, turning page p into its new representation r. This is important because, sometimes, a topic apparently not important in the triggering page offers a good opportunity for advertising. For example, consider a triggering page that describes a congress in London about digital photography. Although London is probably not an important topic in this page, advertisements about hotels in London would be appropriate. Thus, adding hotels to page p is important. This suggests using α > 0, that is, preserving the contents of p and using the terms in r to expand p. In this paper, we examine both approaches. Thus, in our sixth method we match r, the set of new expansion terms, directly to the ads, as follows: AAK T(p, ai) = AAK(r, ai) where AAK T stands for match the ad and keywords to the set r of expansion terms. In our seventh method, we match an expanded page p to the ads as follows: AAK EXP(p, ai) = AAK(p ∪ r, ai) where AAK EXP stands for match the ad and keywords to the expanded triggering page. 499 To improve our ad placement methods, other external source that we can use is the content of the page h pointed to by the advertisement"s hyperlink, that is, its landing page. After all, this page comprises the real target of the ad and perhaps could present a more detailed description of the product or service being advertised. Given that the advertisement ai points to the landing page hi, we denote this association as the pair (ai, hi) ∈ H, where H is the set of associations between the ads and the pages they point to. Our eighth method consists of matching the triggering page p to the landing pages pointed to by the advertisements, as follows: H(p, ai) = sim(p, hi) where (ai, hi) ∈ H and H stands for match the hyperlink pointed to by the ad. We can also combine this information with the more promising methods previously described, AAK and AAK EXP as follows. Given that (ai, hi) ∈ H and (ai, ki) ∈ K, we have our last two methods: AAK H(p, ai) =  sim(p, ai ∪ hi ∪ ki) if ki p 0 if otherwise AAK EXP H(p, ai) =  sim(p ∪ r, ai ∪ hi ∪ ki) if ki (p ∪ r) 0 if otherwise where AAK H stands for match ads and keywords also considering the page pointed by the ad and AAH EXP H stands for match ads and keywords with expanded triggering page, also considering the page pointed by the ad. Notice that other combinations were not considered in this study due to space restrictions. These other combinations led to poor results in our experimentation and for this reason were discarded. 4. EXPERIMENTS 4.1 Methodology To evaluate our ad placement strategies, we performed a series of experiments using a sample of a real case ad collection with 93,972 advertisements, 1,744 advertisers, and 68,238 keywords2 . The advertisements are grouped in 2,029 campaigns with an average of 1.16 campaigns per advertiser. For the strategies AAK T and AAK EXP, we had to generate a set of expansion terms. For that, we used a database of Web pages crawled by the TodoBR search engine [12] (http://www.todobr.com.br/). This database is composed of 5,939,061 pages of the Brazilian Web, under the domain .br. For the strategies H, AAK H, and AAK EXP H, we also crawled the pages pointed to by the advertisers. No other filtering method was applied to these pages besides the removal of HTML tags. Since we are initially interested in the placement of advertisements in the pages of information portals, our test collection was composed of 100 pages extracted from a Brazilian newspaper. These are our triggering pages. They were crawled in such a way that only the contents of their articles was preserved. As we have no preferences for particular 2 Data in portuguese provided by an on-line advertisement company that operates in Brazil. topics, the crawled pages cover topics as diverse as politics, economy, sports, and culture. For each of our 100 triggering pages, we selected the top three ranked ads provided by each of our 10 ad placement strategies. Thus, for each triggering page we select no more than 30 ads. These top ads were then inserted in a pool for that triggering page. Each pool contained an average of 15.81 advertisements. All advertisements in each pool were submitted to a manual evaluation by a group of 15 users. The average number of relevant advertisements per page pool was 5.15. Notice that we adopted the same pooling method used to evaluate the TREC Web-based collection [6]. To quantify the precision of our results, we used 11-point average figures [2]. Since we are not able to evaluate the entire ad collection, recall values are relative to the set of evaluated advertisements. 4.2 Tuning Idf factors We start by analyzing the impact of different idf factors in our advertisement collection. Idf factors are important because they quantify how discriminative is a term in the collection. In our ad collection, idf factors can be computed by taking ads, advertisers or campaigns as documents. To exemplify, consider the computation of ad idf for a term ti that occurs 9 times in a collection of 100 ads. Then, the inverse document frequency of ti is given by: idfi = log 100 9 Hence, we can compute ad, advertiser or campaign idf factors. As we observe in Figure 4, for the AD strategy, the best ranking is obtained by the use of campaign idf, that is, by calculating our idf factor so that it discriminates campaigns. Similar results were obtained for all the other methods. 0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0 0.2 0.4 0.6 0.8 1 precision recall Campaign idf Advertiser idf Ad idf Figure 4: Precision-recall curves obtained for the AD strategy using ad, advertiser, and campaign idf factors. This reflects the fact that terms might be better discriminators for a business topic than for an specific ad. This effect can be accomplished by calculating the factor relative to idf advertisers or campaigns instead of ads. In fact, campaign idf factors yielded the best results. Thus, they will be used in all the experiments reported from now on. 500 4.3 Results Matching Strategies Figure 5 displays the results for the matching strategies presented in Section 2. As shown, directly matching the contents of the ad to the triggering page (AD strategy) is not so effective. The reason is that the ad contents are very noisy. It may contain messages that do not properly describe the ad topics such as requisitions for user actions (e.g, visit our site) and general sentences that could be applied to any product or service (e.g, we delivery for the whole country). On the other hand, an advertiser provided keyword summarizes well the topic of the ad. As a consequence, the KW strategy is superior to the AD and AD KW strategies. This situation changes when we require the keywords to appear in the target Web page. By filtering out ads whose keywords do not occur in the triggering page, much noise is discarded. This makes ANDKW a better alternative than KW. Further, in this new situation, the contents of the ad becomes useful to rank the most relevant ads making AD ANDKW (or AAK for ads and keywords) the best among all described methods. For this reason, we adopt AAK as our baseline in the next set of experiments. 0 0.1 0.2 0.3 0.4 0.5 0.6 0 0.2 0.4 0.6 0.8 1 precision recall AAK ANDKW KW AD_KW AD Figure 5: Comparison among our five matching strategies. AAK (ads and keywords) is superior. Table 1 illustrates average precision figures for Figure 5. We also present actual hits per advertisement slot. We call hit an assignment of an ad (to the triggering page) that was considered relevant by the evaluators. We notice that our AAK strategy provides a gain in average precision of 60% relative to the trivial AD strategy. This shows that careful consideration of the evidence related to the problem does pay off. Impedance Coupling Strategies Table 2 shows top ranked terms that occur in a page covering Argentinean wines produced using grapes derived from the Bordeaux region of France. The p column includes the top terms for this page ranked according to our tf-idf weighting scheme. The r column includes the top ranked expansion terms generated according to Eq. (6). Notice that the expansion terms not only emphasize important terms of the target page (by increasing their weights) such as wines and Methods Hits 11-pt average #1 #2 #3 total score gain(%) AD 41 32 13 86 0.104 AD KW 51 28 17 96 0.106 +1.9 KW 46 34 28 108 0.125 +20.2 ANDKW 49 37 35 121 0.153 +47.1 AD ANDKW (AAK) 51 48 39 138 0.168 +61.5 Table 1: Average precision figures, corresponding to Figure 5, for our five matching strategies. Columns labelled #1, #2, and #3 indicate total of hits in first, second, and third advertisement slots, respectively. The AAK strategy provides improvements of 60% relative to the AD strategy. Rank p r term score term score 1 argentina 0.090 wines 0.251 2 obtained* 0.047 wine* 0.140 3 class* 0.036 whites 0.091 4 whites 0.035 red* 0.057 5 french* 0.031 grape 0.051 6 origin* 0.029 bordeaux 0.045 7 france* 0.029 acideness* 0.038 8 grape 0.017 argentina 0.037 9 sweet* 0.016 aroma* 0.037 10 country* 0.013 blanc* 0.036 ... 35 wines 0.010 -... Table 2: Top ranked terms for the triggering page p according to our tf-idf weighting scheme and top ranked terms for r, the expansion terms for p, generated according to Eq. (6). Ranking scores were normalized in order to sum up to 1. Terms marked with ‘*" are not shared by the sets p and r. whites, but also reveal new terms related to the main topic of the page such as aroma and red. Further, they avoid some uninteresting terms such as obtained and country. Figure 6 illustrates our results when the set r of expansion terms is used. They show that matching the ads to the terms in the set r instead of to the triggering page p (AAK T strategy) leads to a considerable improvement over our baseline, AAK. The gain is even larger when we use the terms in r to expand the triggering page (AAK EXP method). This confirms our hypothesis that the triggering page could have some interesting terms that should not be completely discarded. Finally, we analyze the impact on the ranking of using the contents of pages pointed by the ads. Figure 7 displays our results. It is clear that using only the contents of the pages pointed by the ads (H strategy) yields very poor results. However, combining evidence from the pages pointed by the ads with our baseline yields improved results. Most important, combining our best strategy so far (AAK EXP) with pages pointed by ads (AAK EXP H strategy) leads to superior results. This happens because the two additional sources of evidence, expansion terms and pages pointed by the ads, are distinct and complementary, providing extra and valuable information for matching ads to a Web page. 501 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0 0.2 0.4 0.6 0.8 1 precision recall AAK_EXP AAK_T AAK Figure 6: Impact of using a new representation for the triggering page, one that includes expansion terms. 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0 0.2 0.4 0.6 0.8 1 precision recall AAK_EXP_H AAK_H AAK H Figure 7: Impact of using the contents of the page pointed by the ad (the hyperlink). Figure 8 and Table 3 summarize all results described in this section. In Figure 8 we show precision-recall curves and in Table 3 we show 11-point average figures. We also present actual hits per advertisement slot and gains in average precision relative to our baseline, AAK. We notice that the highest number of hits in the first slot was generated by the method AAK EXP. However, the method with best overall retrieval performance was AAK EXP H, yielding a gain in average precision figures of roughly 50% over the baseline (AAK). 4.4 Performance Issues In a keyword targeted advertising system, ads are assigned at query time, thus the performance of the system is a very important issue. In content-targeted advertising systems, we can associate ads with a page at publishing (or updating) time. Also, if a new ad comes in we might consider assigning this ad to already published pages in offline mode. That is, we might design the system such that its performance depends fundamentally on the rate that new pages 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0 0.2 0.4 0.6 0.8 1 precision recall AAK_EXP_H AAK_EXP AAK_T AAK_H AAK H Figure 8: Comparison among our ad placement strategies. Methods Hits 11-pt average #1 #2 #3 total score gain(%) H 28 5 6 39 0.026 -84.3 AAK 51 48 39 138 0.168 AAK H 52 50 46 148 0.191 +13.5 AAK T 65 49 43 157 0.226 +34.6 AAK EXP 70 52 53 175 0.242 +43.8 AAK EXP H 64 61 51 176 0.253 +50.3 Table 3: Results for our impedance coupling strategies. are published and the rate that ads are added or modified. Further, the data needed by our strategies (page crawling, page expansion, and ad link crawling) can be gathered and processed offline, not affecting the user experience. Thus, from this point of view, the performance is not critical and will not be addressed in this work. 5. RELATED WORK Several works have stressed the importance of relevance in advertising. For example, in [14] it was shown that advertisements that are presented to users when they are not interested on them are viewed just as annoyance. Thus, in order to be effective, the authors conclude that advertisements should be relevant to consumer concerns at the time of exposure. The results in [9] enforce this conclusion by pointing out that the more targeted the advertising, the more effective it is. Therefore it is not surprising that other works have addressed the relevance issue. For instance, in [8] it is proposed a system called ADWIZ that is able to adapt online advertisement to a user"s short-term interests in a non-intrusive way. Contrary to our work, ADWIZ does not directly use the content of the page viewed by the user. It relies on search keywords supplied by the user to search engines and on the URL of the page requested by the user. On the other hand, in [7] the authors presented an intrusive approach in which an agent sits between advertisers and the user"s browser allowing a banner to be placed into the currently viewed page. In spite of having the opportunity to use the page"s content, 502 the agent infers relevance based on category information and user"s private information collected along the time. In [5] the authors provide a comparison between the ranking strategies used by Google and Overture for their keyword advertising systems. Both systems select advertisements by matching them to the keywords provided by the user in a search query and rank the resulting advertisement list according to the advertisers" willingness to pay. In particular, Google approach also considers the clickthrough rate of each advertisement as an additional evidence for its relevance. The authors conclude that Google"s strategy is better than that used by Overture. As mentioned before, the ranking problem in keyword advertising is different from that of content-targeted advertising. Instead of dealing with keywords provided by users in search queries, we have to deal with the contents of a page which can be very diffuse. Finally, the work in [4] focuses on improving search engine results in a TREC collection by means of an automatic query expansion method based on kNN [17]. Such method resembles our expansion approach presented in section 3. Our method is different from that presented by [4]. They expand user queries applied to a document collection with terms extracted from the top k documents returned as answer to the query in the same collection. In our case, we use two collections: an advertisement and a Web collection. We expand triggering pages with terms extracted from the Web collection and then we match these expanded pages to the ads from the advertisement collection. By doing this, we emphasize the main topics of the triggering pages, increasing the possibility of associating relevant ads with them. 6. CONCLUSIONS In this work we investigated ten distinct strategies for associating ads with a Web page that is browsed (contenttargeted advertising). Five of our strategies attempt to match the ads directly to the Web page. Because of that, they are called matching strategies. The other five strategies recognize that there is a vocabulary impedance problem among ads and Web pages and attempt to solve the problem by expanding the Web pages and the ads with new terms. Because of that they are called impedance coupling strategies. Using a sample of a real case database with over 93 thousand ads, we evaluated our strategies. For the five matching strategies, our results indicated that planned consideration of additional evidence (such as the keywords provided by the advertisers) yielded gains in average precision figures (for our test collection) of 60%. This was obtained by a strategy called AAK (for ads and keywords), which is taken as the baseline for evaluating our more advanced impedance coupling strategies. For our five impedance coupling strategies, the results indicate that additional gains in average precision of 50% (now relative to the AAK strategy) are possible. These were generated by expanding the Web page with new terms (obtained using a sample Web collection containing over five million pages) and the ads with the contents of the page they point to (a hyperlink provided by the advertisers). These are first time results that indicate that high quality content-targeted advertising is feasible and practical. 7. ACKNOWLEDGEMENTS This work was supported in part by the GERINDO project, grant MCT/CNPq/CT-INFO 552.087/02-5, by CNPq grant 300.188/95-1 (Berthier Ribeiro-Neto), and by CNPq grant 303.576/04-9 (Edleno Silva de Moura). Marco Cristo is supported by Fucapi, Manaus, AM, Brazil. 8. REFERENCES [1] The Google adwords. Google content-targeted advertising. http://adwords.google.com/select/ct_faq.html, November 2004. [2] R. Baeza-Yates and B. Ribeiro-Neto. Modern Information Retrieval. Addison-Wesley-Longman, 1st edition, 1999. [3] H. K. Bhargava and J. Feng. Paid placement strategies for internet search engines. In Proceedings of the eleventh international conference on World Wide Web, pages 117-123. ACM Press, 2002. [4] E. P. Chan, S. Garcia, and S. Roukos. Trec-5 ad-hoc retrieval using k nearest-neighbors re-scoring. In The Fifth Text REtrieval Conference (TREC-5). National Institute of Standards and Technology (NIST), November 1996. [5] J. Feng, H. K. Bhargava, and D. Pennock. Comparison of allocation rules for paid placement advertising in search engines. In Proceedings of the 5th international conference on Electronic commerce, pages 294-299. ACM Press, 2003. [6] D. Hawking, N. Craswell, and P. B. Thistlewaite. Overview of TREC-7 very large collection track. In The Seventh Text REtrieval Conference (TREC-7), pages 91-104, Gaithersburg, Maryland, USA, November 1998. [7] Y. Kohda and S. Endo. Ubiquitous advertising on the www: merging advertisement on the browser. Comput. Netw. ISDN Syst., 28(7-11):1493-1499, 1996. [8] M. Langheinrich, A. Nakamura, N. Abe, T. Kamba, and Y. Koseki. Unintrusive customization techniques for web advertising. Comput. Networks, 31(11-16):1259-1272, 1999. [9] T. P. Novak and D. L. Hoffman. New metrics for new media: toward the development of web measurement standards. World Wide Web J., 2(1):213-246, 1997. [10] J. Pearl. Probabilistic Reasoning in Intelligent Systems: Networks of plausible inference. Morgan Kaufmann Publishers, 2nd edition, 1988. [11] B. Ribeiro-Neto and R. Muntz. A belief network model for IR. In Proceedings of the 19th Annual International ACM SIGIR Conference on Research and Development in Information Retrieval, pages 253-260, Zurich, Switzerland, August 1996. [12] A. Silva, E. Veloso, P. Golgher, B. Ribeiro-Neto, A. Laender, and N. Ziviani. CobWeb - a crawler for the brazilian web. In Proceedings of the String Processing and Information Retrieval Symposium (SPIRE"99), pages 184-191, Cancun, Mexico, September 1999. [13] H. Turtle and W. B. Croft. Evaluation of an inference network-based retrieval model. ACM Transactions on Information Systems, 9(3):187-222, July 1991. [14] C. Wang, P. Zhang, R. Choi, and M. Daeredita. Understanding consumers attitude toward advertising. In Eighth Americas Conference on Information Systems, pages 1143-1148, August 2002. [15] M. Weideman. Ethical issues on content distribution to digital consumers via paid placement as opposed to website visibility in search engine results. In The Seventh ETHICOMP International Conference on the Social and Ethical Impacts of Information and Communication Technologies, pages 904-915. Troubador Publishing Ltd, April 2004. [16] M. Weideman and T. Haig-Smith. An investigation into search engines as a form of targeted advert delivery. In Proceedings of the 2002 annual research conference of the South African institute of computer scientists and information technologists on Enablement through technology, pages 258-258. South African Institute for Computer Scientists and Information Technologists, 2002. [17] Y. Yang. Expert network: Effective and efficient learning from human decisions in text categorization and retrieval. In W. B. Croft and e. C. J. van Rijsbergen, editors, Proceedings of the 17rd annual international ACM SIGIR conference on Research and development in information retrieval, pages 13-22. Springer-Verlag, 1994. 503

paid placement strategy;web;bayesian network model;bayesian network;keyword targeted advertising;expansion term;ad and keyword;knn;on-line advertising;matching strategy;ad placement strategy;advertise;content-targeted advertising;impedance coupling strategy

train_H-62

Implicit User Modeling for Personalized Search

Information retrieval systems (e.g., web search engines) are critical for overcoming information overload. A major deficiency of existing retrieval systems is that they generally lack user modeling and are not adaptive to individual users, resulting in inherently non-optimal retrieval performance. For example, a tourist and a programmer may use the same word java to search for different information, but the current search systems would return the same results. In this paper, we study how to infer a user"s interest from the user"s search context and use the inferred implicit user model for personalized search . We present a decision theoretic framework and develop techniques for implicit user modeling in information retrieval. We develop an intelligent client-side web search agent (UCAIR) that can perform eager implicit feedback, e.g., query expansion based on previous queries and immediate result reranking based on clickthrough information. Experiments on web search show that our search agent can improve search accuracy over the popular Google search engine.

1. INTRODUCTION Although many information retrieval systems (e.g., web search engines and digital library systems) have been successfully deployed, the current retrieval systems are far from optimal. A major deficiency of existing retrieval systems is that they generally lack user modeling and are not adaptive to individual users [17]. This inherent non-optimality is seen clearly in the following two cases: (1) Different users may use exactly the same query (e.g., Java) to search for different information (e.g., the Java island in Indonesia or the Java programming language), but existing IR systems return the same results for these users. Without considering the actual user, it is impossible to know which sense Java refers to in a query. (2) A user"s information needs may change over time. The same user may use Java sometimes to mean the Java island in Indonesia and some other times to mean the programming language. Without recognizing the search context, it would be again impossible to recognize the correct sense. In order to optimize retrieval accuracy, we clearly need to model the user appropriately and personalize search according to each individual user. The major goal of user modeling for information retrieval is to accurately model a user"s information need, which is, unfortunately, a very difficult task. Indeed, it is even hard for a user to precisely describe what his/her information need is. What information is available for a system to infer a user"s information need? Obviously, the user"s query provides the most direct evidence. Indeed, most existing retrieval systems rely solely on the query to model a user"s information need. However, since a query is often extremely short, the user model constructed based on a keyword query is inevitably impoverished . An effective way to improve user modeling in information retrieval is to ask the user to explicitly specify which documents are relevant (i.e., useful for satisfying his/her information need), and then to improve user modeling based on such examples of relevant documents. This is called relevance feedback, which has been proved to be quite effective for improving retrieval accuracy [19, 20]. Unfortunately, in real world applications, users are usually reluctant to make the extra effort to provide relevant examples for feedback [11]. It is thus very interesting to study how to infer a user"s information need based on any implicit feedback information, which naturally exists through user interactions and thus does not require any extra user effort. Indeed, several previous studies have shown that implicit user modeling can improve retrieval accuracy. In [3], a web browser (Curious Browser) is developed to record a user"s explicit relevance ratings of web pages (relevance feedback) and browsing behavior when viewing a page, such as dwelling time, mouse click, mouse movement and scrolling (implicit feedback). It is shown that the dwelling time on a page, amount of scrolling on a page and the combination of time and scrolling have a strong correlation with explicit relevance ratings, which suggests that implicit feedback may be helpful for inferring user information need. In [10], user clickthrough data is collected as training data to learn a retrieval function, which is used to produce a customized ranking of search results that suits a group of users" preferences. In [25], the clickthrough data collected over a long time period is exploited through query expansion to improve retrieval accuracy. 824 While a user may have general long term interests and preferences for information, often he/she is searching for documents to satisfy an ad-hoc information need, which only lasts for a short period of time; once the information need is satisfied, the user would generally no longer be interested in such information. For example, a user may be looking for information about used cars in order to buy one, but once the user has bought a car, he/she is generally no longer interested in such information. In such cases, implicit feedback information collected over a long period of time is unlikely to be very useful, but the immediate search context and feedback information, such as which of the search results for the current information need are viewed, can be expected to be much more useful. Consider the query Java again. Any of the following immediate feedback information about the user could potentially help determine the intended meaning of Java in the query: (1) The previous query submitted by the user is hashtable (as opposed to, e.g., travel Indonesia). (2) In the search results, the user viewed a page where words such as programming, software, and applet occur many times. To the best of our knowledge, how to exploit such immediate and short-term search context to improve search has so far not been well addressed in the previous work. In this paper, we study how to construct and update a user model based on the immediate search context and implicit feedback information and use the model to improve the accuracy of ad-hoc retrieval. In order to maximally benefit the user of a retrieval system through implicit user modeling, we propose to perform eager implicit feedback. That is, as soon as we observe any new piece of evidence from the user, we would update the system"s belief about the user"s information need and respond with improved retrieval results based on the updated user model. We present a decision-theoretic framework for optimizing interactive information retrieval based on eager user model updating, in which the system responds to every action of the user by choosing a system action to optimize a utility function. In a traditional retrieval paradigm, the retrieval problem is to match a query with documents and rank documents according to their relevance values. As a result, the retrieval process is a simple independent cycle of query and result display. In the proposed new retrieval paradigm, the user"s search context plays an important role and the inferred implicit user model is exploited immediately to benefit the user. The new retrieval paradigm is thus fundamentally different from the traditional paradigm, and is inherently more general. We further propose specific techniques to capture and exploit two types of implicit feedback information: (1) identifying related immediately preceding query and using the query and the corresponding search results to select appropriate terms to expand the current query, and (2) exploiting the viewed document summaries to immediately rerank any documents that have not yet been seen by the user. Using these techniques, we develop a client-side web search agent UCAIR (User-Centered Adaptive Information Retrieval) on top of a popular search engine (Google). Experiments on web search show that our search agent can improve search accuracy over Google. Since the implicit information we exploit already naturally exists through user interactions, the user does not need to make any extra effort. Thus the developed search agent can improve existing web search performance without additional effort from the user. The remaining sections are organized as follows. In Section 2, we discuss the related work. In Section 3, we present a decisiontheoretic interactive retrieval framework for implicit user modeling. In Section 4, we present the design and implementation of an intelligent client-side web search agent (UCAIR) that performs eager implicit feedback. In Section 5, we report our experiment results using the search agent. Section 6 concludes our work. 2. RELATED WORK Implicit user modeling for personalized search has been studied in previous work, but our work differs from all previous work in several aspects: (1) We emphasize the exploitation of immediate search context such as the related immediately preceding query and the viewed documents in the same session, while most previous work relies on long-term collection of implicit feedback information [25]. (2) We perform eager feedback and bring the benefit of implicit user modeling as soon as any new implicit feedback information is available, while the previous work mostly exploits longterm implicit feedback [10]. (3) We propose a retrieval framework to integrate implicit user modeling with the interactive retrieval process, while the previous work either studies implicit user modeling separately from retrieval [3] or only studies specific retrieval models for exploiting implicit feedback to better match a query with documents [23, 27, 22]. (4) We develop and evaluate a personalized Web search agent with online user studies, while most existing work evaluates algorithms offline without real user interactions. Currently some search engines provide rudimentary personalization, such as Google Personalized web search [6], which allows users to explicitly describe their interests by selecting from predefined topics, so that those results that match their interests are brought to the top, and My Yahoo! search [16], which gives users the option to save web sites they like and block those they dislike. In contrast, UCAIR personalizes web search through implicit user modeling without any additional user efforts. Furthermore, the personalization of UCAIR is provided on the client side. There are two remarkable advantages on this. First, the user does not need to worry about the privacy infringement, which is a big concern for personalized search [26]. Second, both the computation of personalization and the storage of the user profile are done at the client side so that the server load is reduced dramatically [9]. There have been many works studying user query logs [1] or query dynamics [13]. UCAIR makes direct use of a user"s query history to benefit the same user immediately in the same search session. UCAIR first judges whether two neighboring queries belong to the same information session and if so, it selects terms from the previous query to perform query expansion. Our query expansion approach is similar to automatic query expansion [28, 15, 5], but instead of using pseudo feedback to expand the query, we use user"s implicit feedback information to expand the current query. These two techniques may be combined. 3. OPTIMIZATION IN INTERACTIVE IR In interactive IR, a user interacts with the retrieval system through an action dialogue, in which the system responds to each user action with some system action. For example, the user"s action may be submitting a query and the system"s response may be returning a list of 10 document summaries. In general, the space of user actions and system responses and their granularities would depend on the interface of a particular retrieval system. In principle, every action of the user can potentially provide new evidence to help the system better infer the user"s information need. Thus in order to respond optimally, the system should use all the evidence collected so far about the user when choosing a response. When viewed in this way, most existing search engines are clearly non-optimal. For example, if a user has viewed some documents on the first page of search results, when the user clicks on the Next link to fetch more results, an existing retrieval system would still return the next page of results retrieved based on the original query without considering the new evidence that a particular result has been viewed by the user. 825 We propose to optimize retrieval performance by adapting system responses based on every action that a user has taken, and cast the optimization problem as a decision task. Specifically, at any time, the system would attempt to do two tasks: (1) User model updating: Monitor any useful evidence from the user regarding his/her information need and update the user model as soon as such evidence is available; (2) Improving search results: Rerank immediately all the documents that the user has not yet seen, as soon as the user model is updated. We emphasize eager updating and reranking, which makes our work quite different from any existing work. Below we present a formal decision theoretic framework for optimizing retrieval performance through implicit user modeling in interactive information retrieval. 3.1 A decision-theoretic framework Let A be the set of all user actions and R(a) be the set of all possible system responses to a user action a ∈ A. At any time, let At = (a1, ..., at) be the observed sequence of user actions so far (up to time point t) and Rt−1 = (r1, ..., rt−1) be the responses that the system has made responding to the user actions. The system"s goal is to choose an optimal response rt ∈ R(at) for the current user action at. Let M be the space of all possible user models. We further define a loss function L(a, r, m) ∈ , where a ∈ A is a user action, r ∈ R(a) is a system response, and m ∈ M is a user model. L(a, r, m) encodes our decision preferences and assesses the optimality of responding with r when the current user model is m and the current user action is a. According to Bayesian decision theory, the optimal decision at time t is to choose a response that minimizes the Bayes risk, i.e., r∗ t = argmin r∈R(at) M L(at, r, mt)P(mt|U, D, At, Rt−1)dmt (1) where P(mt|U, D, At, Rt−1) is the posterior probability of the user model mt given all the observations about the user U we have made up to time t. To simplify the computation of Equation 1, let us assume that the posterior probability mass P(mt|U, D, At, Rt−1) is mostly concentrated on the mode m∗ t = argmaxmt P(mt|U, D, At, Rt−1). We can then approximate the integral with the value of the loss function at m∗ t . That is, r∗ t ≈ argminr∈R(at)L(at, r, m∗ t ) (2) where m∗ t = argmaxmt P(mt|U, D, At, Rt−1). Leaving aside how to define and estimate these probabilistic models and the loss function, we can see that such a decision-theoretic formulation suggests that, in order to choose the optimal response to at, the system should perform two tasks: (1) compute the current user model and obtain m∗ t based on all the useful information. (2) choose a response rt to minimize the loss function value L(at, rt, m∗ t ). When at does not affect our belief about m∗ t , the first step can be omitted and we may reuse m∗ t−1 for m∗ t . Note that our framework is quite general since we can potentially model any kind of user actions and system responses. In most cases, as we may expect, the system"s response is some ranking of documents, i.e., for most actions a, R(a) consists of all the possible rankings of the unseen documents, and the decision problem boils down to choosing the best ranking of unseen documents based on the most current user model. When a is the action of submitting a keyword query, such a response is exactly what a current retrieval system would do. However, we can easily imagine that a more intelligent web search engine would respond to a user"s clicking of the Next link (to fetch more unseen results) with a more optimized ranking of documents based on any viewed documents in the current page of results. In fact, according to our eager updating strategy, we may even allow a system to respond to a user"s clicking of browser"s Back button after viewing a document in the same way, so that the user can maximally benefit from implicit feedback. These are precisely what our UCAIR system does. 3.2 User models A user model m ∈ M represents what we know about the user U, so in principle, it can contain any information about the user that we wish to model. We now discuss two important components in a user model. The first component is a component model of the user"s information need. Presumably, the most important factor affecting the optimality of the system"s response is how well the response addresses the user"s information need. Indeed, at any time, we may assume that the system has some belief about what the user is interested in, which we model through a term vector x = (x1, ..., x|V |), where V = {w1, ..., w|V |} is the set of all terms (i.e., vocabulary) and xi is the weight of term wi. Such a term vector is commonly used in information retrieval to represent both queries and documents. For example, the vector-space model, assumes that both the query and the documents are represented as term vectors and the score of a document with respect to a query is computed based on the similarity between the query vector and the document vector [21]. In a language modeling approach, we may also regard the query unigram language model [12, 29] or the relevance model [14] as a term vector representation of the user"s information need. Intuitively, x would assign high weights to terms that characterize the topics which the user is interested in. The second component we may include in our user model is the documents that the user has already viewed. Obviously, even if a document is relevant, if the user has already seen the document, it would not be useful to present the same document again. We thus introduce another variable S ⊂ D (D is the whole set of documents in the collection) to denote the subset of documents in the search results that the user has already seen/viewed. In general, at time t, we may represent a user model as mt = (S, x, At, Rt−1), where S is the seen documents, x is the system"s understanding of the user"s information need, and (At, Rt−1) represents the user"s interaction history. Note that an even more general user model may also include other factors such as the user"s reading level and occupation. If we assume that the uncertainty of a user model mt is solely due to the uncertainty of x, the computation of our current estimate of user model m∗ t will mainly involve computing our best estimate of x. That is, the system would choose a response according to r∗ t = argminr∈R(at)L(at, r, S, x∗ , At, Rt−1) (3) where x∗ = argmaxx P(x|U, D, At, Rt−1). This is the decision mechanism implemented in the UCAIR system to be described later. In this system, we avoided specifying the probabilistic model P(x|U, D, At, Rt−1) by computing x∗ directly with some existing feedback method. 3.3 Loss functions The exact definition of loss function L depends on the responses, thus it is inevitably application-specific. We now briefly discuss some possibilities when the response is to rank all the unseen documents and present the top k of them. Let r = (d1, ..., dk) be the top k documents, S be the set of seen documents by the user, and x∗ be the system"s best guess of the user"s information need. We 826 may simply define the loss associated with r as the negative sum of the probability that each of the di is relevant, i.e., L(a, r, m) = − k i=1 P(relevant|di, m). Clearly, in order to minimize this loss function, the optimal response r would contain the k documents with the highest probability of relevance, which is intuitively reasonable. One deficiency of this top-k loss function is that it is not sensitive to the internal order of the selected top k documents, so switching the ranking order of a non-relevant document and a relevant one would not affect the loss, which is unreasonable. To model ranking, we can introduce a factor of the user model - the probability of each of the k documents being viewed by the user, P(view|di), and define the following ranking loss function: L(a, r, m) = − k i=1 P(view|di)P(relevant|di, m) Since in general, if di is ranked above dj (i.e., i < j), P(view|di) > P(view|dj), this loss function would favor a decision to rank relevant documents above non-relevant ones, as otherwise, we could always switch di with dj to reduce the loss value. Thus the system should simply perform a regular retrieval and rank documents according to the probability of relevance [18]. Depending on the user"s retrieval preferences, there can be many other possibilities. For example, if the user does not want to see redundant documents, the loss function should include some redundancy measure on r based on the already seen documents S. Of course, when the response is not to choose a ranked list of documents, we would need a different loss function. We discuss one such example that is relevant to the search agent that we implement. When a user enters a query qt (current action), our search agent relies on some existing search engine to actually carry out search. In such a case, even though the search agent does not have control of the retrieval algorithm, it can still attempt to optimize the search results through refining the query sent to the search engine and/or reranking the results obtained from the search engine. The loss functions for reranking are already discussed above; we now take a look at the loss functions for query refinement. Let f be the retrieval function of the search engine that our agent uses so that f(q) would give us the search results using query q. Given that the current action of the user is entering a query qt (i.e., at = qt), our response would be f(q) for some q. Since we have no choice of f, our decision is to choose a good q. Formally, r∗ t = argminrt L(a, rt, m) = argminf(q)L(a, f(q), m) = f(argminqL(qt, f(q), m)) which shows that our goal is to find q∗ = argminqL(qt, f(q), m), i.e., an optimal query that would give us the best f(q). A different choice of loss function L(qt, f(q), m) would lead to a different query refinement strategy. In UCAIR, we heuristically compute q∗ by expanding qt with terms extracted from rt−1 whenever qt−1 and qt have high similarity. Note that rt−1 and qt−1 are contained in m as part of the user"s interaction history. 3.4 Implicit user modeling Implicit user modeling is captured in our framework through the computation of x∗ = argmaxx P(x|U, D, At, Rt−1), i.e., the system"s current belief of what the user"s information need is. Here again there may be many possibilities, leading to different algorithms for implicit user modeling. We now discuss a few of them. First, when two consecutive queries are related, the previous query can be exploited to enrich the current query and provide more search context to help disambiguation. For this purpose, instead of performing query expansion as we did in the previous section, we could also compute an updated x∗ based on the previous query and retrieval results. The computed new user model can then be used to rank the documents with a standard information retrieval model. Second, we can also infer a user"s interest based on the summaries of the viewed documents. When a user is presented with a list of summaries of top ranked documents, if the user chooses to skip the first n documents and to view the (n+1)-th document, we may infer that the user is not interested in the displayed summaries for the first n documents, but is attracted by the displayed summary of the (n + 1)-th document. We can thus use these summaries as negative and positive examples to learn a more accurate user model x∗ . Here many standard relevance feedback techniques can be exploited [19, 20]. Note that we should use the displayed summaries, as opposed to the actual contents of those documents, since it is possible that the displayed summary of the viewed document is relevant, but the document content is actually not. Similarly, a displayed summary may mislead a user to skip a relevant document. Inferring user models based on such displayed information, rather than the actual content of a document is an important difference between UCAIR and some other similar systems. In UCAIR, both of these strategies for inferring an implicit user model are implemented. 4. UCAIR: A PERSONALIZED SEARCH AGENT 4.1 Design In this section, we present a client-side web search agent called UCAIR, in which we implement some of the methods discussed in the previous section for performing personalized search through implicit user modeling. UCAIR is a web browser plug-in 1 that acts as a proxy for web search engines. Currently, it is only implemented for Internet Explorer and Google, but it is a matter of engineering to make it run on other web browsers and interact with other search engines. The issue of privacy is a primary obstacle for deploying any real world applications involving serious user modeling, such as personalized search. For this reason, UCAIR is strictly running as a client-side search agent, as opposed to a server-side application. This way, the captured user information always resides on the computer that the user is using, thus the user does not need to release any information to the outside. Client-side personalization also allows the system to easily observe a lot of user information that may not be easily available to a server. Furthermore, performing personalized search on the client-side is more scalable than on the serverside, since the overhead of computation and storage is distributed among clients. As shown in Figure 1, the UCAIR toolbar has 3 major components: (1) The (implicit) user modeling module captures a user"s search context and history information, including the submitted queries and any clicked search results and infers search session boundaries. (2) The query modification module selectively improves the query formulation according to the current user model. (3) The result re-ranking module immediately re-ranks any unseen search results whenever the user model is updated. In UCAIR, we consider four basic user actions: (1) submitting a keyword query; (2) viewing a document; (3) clicking the Back button; (4) clicking the Next link on a result page. For each of these four actions, the system responds with, respectively, (1) 1 UCAIR is available at: http://sifaka.cs.uiuc.edu/ir/ucair/download.html 827 Search Engine (e.g., Google) Search History Log (e.g.,past queries, clicked results) Query Modification Result Re-Ranking User Modeling Result Buffer UCAIR Userquery results clickthrough… Figure 1: UCAIR architecture generating a ranked list of results by sending a possibly expanded query to a search engine; (2) updating the information need model x; (3) reranking the unseen results on the current result page based on the current model x; and (4) reranking the unseen pages and generating the next page of results based on the current model x. Behind these responses, there are three basic tasks: (1) Decide whether the previous query is related to the current query and if so expand the current query with useful terms from the previous query or the results of the previous query. (2) Update the information need model x based on a newly clicked document summary. (3) Rerank a set of unseen documents based on the current model x. Below we describe our algorithms for each of them. 4.2 Session boundary detection and query expansion To effectively exploit previous queries and their corresponding clickthrough information, UCAIR needs to judge whether two adjacent queries belong to the same search session (i.e., detect session boundaries). Existing work on session boundary detection is mostly in the context of web log analysis (e.g., [8]), and uses statistical information rather than textual features. Since our clientside agent does not have access to server query logs, we make session boundary decisions based on textual similarity between two queries. Because related queries do not necessarily share the same words (e.g., java island and travel Indonesia), it is insufficient to use only query text. Therefore we use the search results of the two queries to help decide whether they are topically related. For example, for the above queries java island and travel Indonesia", the words java, bali, island, indonesia and travel may occur frequently in both queries" search results, yielding a high similarity score. We only use the titles and summaries of the search results to calculate the similarity since they are available in the retrieved search result page and fetching the full text of every result page would significantly slow down the process. To compensate for the terseness of titles and summaries, we retrieve more results than a user would normally view for the purpose of detecting session boundaries (typically 50 results). The similarity between the previous query q and the current query q is computed as follows. Let {s1, s2, . . . , sn } and {s1, s2, . . . , sn} be the result sets for the two queries. We use the pivoted normalization TF-IDF weighting formula [24] to compute a term weight vector si for each result si. We define the average result savg to be the centroid of all the result vectors, i.e., (s1 + s2 + . . . + sn)/n. The cosine similarity between the two average results is calculated as s avg · savg/ s 2 avg · s2 avg If the similarity value exceeds a predefined threshold, the two queries will be considered to be in the same information session. If the previous query and the current query are found to belong to the same search session, UCAIR would attempt to expand the current query with terms from the previous query and its search results. Specifically, for each term in the previous query or the corresponding search results, if its frequency in the results of the current query is greater than a preset threshold (e.g. 5 results out of 50), the term would be added to the current query to form an expanded query. In this case, UCAIR would send this expanded query rather than the original one to the search engine and return the results corresponding to the expanded query. Currently, UCAIR only uses the immediate preceding query for query expansion; in principle, we could exploit all related past queries. 4.3 Information need model updating Suppose at time t, we have observed that the user has viewed k documents whose summaries are s1, ..., sk. We update our user model by computing a new information need vector with a standard feedback method in information retrieval (i.e., Rocchio [19]). According to the vector space retrieval model, each clicked summary si can be represented by a term weight vector si with each term weighted by a TF-IDF weighting formula [21]. Rocchio computes the centroid vector of all the summaries and interpolates it with the original query vector to obtain an updated term vector. That is, x = αq + (1 − α) 1 k k i=1 si where q is the query vector, k is the number of summaries the user clicks immediately following the current query and α is a parameter that controls the influence of the clicked summaries on the inferred information need model. In our experiments, α is set to 0.5. Note that we update the information need model whenever the user views a document. 4.4 Result reranking In general, we want to rerank all the unseen results as soon as the user model is updated. Currently, UCAIR implements reranking in two cases, corresponding to the user clicking the Back button and Next link in the Internet Explorer. In both cases, the current (updated) user model would be used to rerank the unseen results so that the user would see improved search results immediately. To rerank any unseen document summaries, UCAIR uses the standard vector space retrieval model and scores each summary based on the similarity of the result and the current user information need vector x [21]. Since implicit feedback is not completely reliable, we bring up only a small number (e.g. 5) of highest reranked results to be followed by any originally high ranked results. 828 Google result (user query = java map) UCAIR result (user query =java map) previous query = travel Indonesia previous query = hashtable expanded user query = java map Indonesia expanded user query = java map class 1 Java map projections of the world ... Lonely Planet - Indonesia Map Map (Java 2 Platform SE v1.4.2) www.btinternet.com/ se16/js/mapproj.htm www.lonelyplanet.com/mapshells/... java.sun.com/j2se/1.4.2/docs/... 2 Java map projections of the world ... INDONESIA TOURISM : CENTRAL JAVA - MAP Java 2 Platform SE v1.3.1: Interface Map www.btinternet.com/ se16/js/oldmapproj.htm www.indonesia-tourism.com/... java.sun.com/j2se/1.3/docs/api/java/... 3 Java Map INDONESIA TOURISM : WEST JAVA - MAP An Introduction to Java Map Collection Classes java.sun.com/developer/... www.indonesia-tourism.com/ ... www.oracle.com/technology/... 4 Java Technology Concept Map IndoStreets - Java Map An Introduction to Java Map Collection Classes java.sun.com/developer/onlineTraining/... www.indostreets.com/maps/java/ www.theserverside.com/news/... 5 Science@NASA Home Indonesia Regions and Islands Maps, Bali, Java, ... Koders - Mappings.java science.nasa.gov/Realtime/... www.maps2anywhere.com/Maps/... www.koders.com/java/ 6 An Introduction to Java Map Collection Classes Indonesia City Street Map,... Hibernate simplifies inheritance mapping www.oracle.com/technology/... www.maps2anywhere.com/Maps/... www.ibm.com/developerworks/java/... 7 Lonely Planet - Java Map Maps Of Indonesia tmap 30.map Class Hierarchy www.lonelyplanet.com/mapshells/ www.embassyworld.com/maps/... tmap.pmel.noaa.gov/... 8 ONJava.com: Java API Map Maps of Indonesia by Peter Loud Class Scope www.onjava.com/pub/a/onjava/api map/ users.powernet.co.uk/... jalbum.net/api/se/datadosen/util/Scope.html 9 GTA San Andreas : Sam Maps of Indonesia by Peter Loud Class PrintSafeHashMap www.gtasanandreas.net/sam/ users.powernet.co.uk/mkmarina/indonesia/ jalbum.net/api/se/datadosen/... 10 INDONESIA TOURISM : WEST JAVA - MAP indonesiaphoto.com Java Pro - Union and Vertical Mapping of Classes www.indonesia-tourism.com/... www.indonesiaphoto.com/... www.fawcette.com/javapro/... Table 1: Sample results of query expansion 5. EVALUATION OF UCAIR We now present some results on evaluating the two major UCAIR functions: selective query expansion and result reranking based on user clickthrough data. 5.1 Sample results The query expansion strategy implemented in UCAIR is intentionally conservative to avoid misinterpretation of implicit user models. In practice, whenever it chooses to expand the query, the expansion usually makes sense. In Table 1, we show how UCAIR can successfully distinguish two different search contexts for the query java map, corresponding to two different previous queries (i.e., travel Indonesia vs. hashtable). Due to implicit user modeling, UCAIR intelligently figures out to add Indonesia and class, respectively, to the user"s query java map, which would otherwise be ambiguous as shown in the original results from Google on March 21, 2005. UCAIR"s results are much more accurate than Google"s results and reflect personalization in search. The eager implicit feedback component is designed to immediately respond to a user"s activity such as viewing a document. In Figure 2, we show how UCAIR can successfully disambiguate an ambiguous query jaguar by exploiting a viewed document summary. In this case, the initial retrieval results using jaguar (shown on the left side) contain two results about the Jaguar cars followed by two results about the Jaguar software. However, after the user views the web page content of the second result (about Jaguar car) and returns to the search result page by clicking Back button, UCAIR automatically nominates two new search results about Jaguar cars (shown on the right side), while the original two results about Jaguar software are pushed down on the list (unseen from the picture). 5.2 Quantitative evaluation To further evaluate UCAIR quantitatively, we conduct a user study on the effectiveness of the eager implicit feedback component. It is a challenge to quantitatively evaluate the potential performance improvement of our proposed model and UCAIR over Google in an unbiased way [7]. Here, we design a user study, in which participants would do normal web search and judge a randomly and anonymously mixed set of results from Google and UCAIR at the end of the search session; participants do not know whether a result comes from Google or UCAIR. We recruited 6 graduate students for this user study, who have different backgrounds (3 computer science, 2 biology, and 1 chem<top> <num> Number: 716 <title> Spammer arrest sue <desc> Description: Have any spammers been arrested or sued for sending unsolicited e-mail? <narr> Narrative: Instances of arrests, prosecutions, convictions, and punishments of spammers, and lawsuits against them are relevant. Documents which describe laws to limit spam without giving details of lawsuits or criminal trials are not relevant. </top> Figure 3: An example of TREC query topic, expressed in a form which might be given to a human assistant or librarian istry). We use query topics from TREC 2 2004 Terabyte track [2] and TREC 2003 Web track [4] topic distillation task in the way to be described below. An example topic from TREC 2004 Terabyte track appears in Figure 3. The title is a short phrase and may be used as a query to the retrieval system. The description field provides a slightly longer statement of the topic requirement, usually expressed as a single complete sentence or question. Finally the narrative supplies additional information necessary to fully specify the requirement, expressed in the form of a short paragraph. Initially, each participant would browse 50 topics either from Terabyte track or Web track and pick 5 or 7 most interesting topics. For each picked topic, the participant would essentially do the normal web search using UCAIR to find many relevant web pages by using the title of the query topic as the initial keyword query. During this process, the participant may view the search results and possibly click on some interesting ones to view the web pages, just as in a normal web search. There is no requirement or restriction on how many queries the participant must submit or when the participant should stop the search for one topic. When the participant plans to change the search topic, he/she will simply press a button 2 Text REtrieval Conference: http://trec.nist.gov/ 829 Figure 2: Screen shots for result reranking to evaluate the search results before actually switching to the next topic. At the time of evaluation, 30 top ranked results from Google and UCAIR (some are overlapping) are randomly mixed together so that the participant would not know whether a result comes from Google or UCAIR. The participant would then judge the relevance of these results. We measure precision at top n (n = 5, 10, 20, 30) documents of Google and UCAIR. We also evaluate precisions at different recall levels. Altogether, 368 documents judged as relevant from Google search results and 429 documents judged as relevant from UCAIR by participants. Scatter plots of precision at top 10 and top 20 documents are shown in Figure 4 and Figure 5 respectively (The scatter plot of precision at top 30 documents is very similar to precision at top 20 documents). Each point of the scatter plots represents the precisions of Google and UCAIR on one query topic. Table 2 shows the average precision at top n documents among 32 topics. From Figure 4, Figure 5 and Table 2, we see that the search results from UCAIR are consistently better than those from Google by all the measures. Moreover, the performance improvement is more dramatic for precision at top 20 documents than that at precision at top 10 documents. One explanation for this is that the more interaction the user has with the system, the more clickthrough data UCAIR can be expected to collect. Thus the retrieval system can build more precise implicit user models, which lead to better retrieval accuracy. Ranking Method prec@5 prec@10 prec@20 prec@30 Google 0.538 0.472 0.377 0.308 UCAIR 0.581 0.556 0.453 0.375 Improvement 8.0% 17.8% 20.2% 21.8% Table 2: Table of average precision at top n documents for 32 query topics The plot in Figure 6 shows the precision-recall curves for UCAIR and Google, where it is clearly seen that the performance of UCAIR 0 0.2 0.4 0.6 0.8 1 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 UCAIR prec@10 Googleprec@10 Scatterplot of Precision at Top 10 Documents Figure 4: Precision at top 10 documents of UCAIR and Google is consistently and considerably better than that of Google at all levels of recall. 6. CONCLUSIONS In this paper, we studied how to exploit implicit user modeling to intelligently personalize information retrieval and improve search accuracy. Unlike most previous work, we emphasize the use of immediate search context and implicit feedback information as well as eager updating of search results to maximally benefit a user. We presented a decision-theoretic framework for optimizing interactive information retrieval based on eager user model updating, in which the system responds to every action of the user by choosing a system action to optimize a utility function. We further propose specific techniques to capture and exploit two types of implicit feedback information: (1) identifying related immediately preceding query and using the query and the corresponding search results to select appropriate terms to expand the current query, and (2) exploiting the viewed document summaries to immediately rerank any documents that have not yet been seen by the user. Using these techniques, we develop a client-side web search agent (UCAIR) on top of a popular search engine (Google). Experiments on web search show that our search agent can improve search accuracy over 830 0 0.2 0.4 0.6 0.8 1 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 UCAIR prec@20 Googleprec@20 Scatterplot of Precision at Top 20 documents Figure 5: Precision at top 20 documents of UCAIR and Google 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0.4 0.45 0.5 0.55 0.6 0.65 0.7 0.75 0.8 0.85 0.9 recall precision Precision−Recall curves Google Result UCAIR Result Figure 6: Precision at top 20 result of UCAIR and Google Google. Since the implicit information we exploit already naturally exists through user interactions, the user does not need to make any extra effort. The developed search agent thus can improve existing web search performance without any additional effort from the user. 7. ACKNOWLEDGEMENT We thank the six participants of our evaluation experiments. This work was supported in part by the National Science Foundation grants IIS-0347933 and IIS-0428472. 8. REFERENCES [1] S. M. Beitzel, E. C. Jensen, A. Chowdhury, D. Grossman, and O. Frieder. Hourly analysis of a very large topically categorized web query log. In Proceedings of SIGIR 2004, pages 321-328, 2004. [2] C. Clarke, N. Craswell, and I. Soboroff. Overview of the TREC 2004 terabyte track. In Proceedings of TREC 2004, 2004. [3] M. Claypool, P. Le, M. Waseda, and D. Brown. Implicit interest indicators. In Proceedings of Intelligent User Interfaces 2001, pages 33-40, 2001. [4] N. Craswell, D. Hawking, R. Wilkinson, and M. Wu. Overview of the TREC 2003 web track. In Proceedings of TREC 2003, 2003. [5] W. B. Croft, S. Cronen-Townsend, and V. Larvrenko. Relevance feedback and personalization: A language modeling perspective. In Proeedings of Second DELOS Workshop: Personalisation and Recommender Systems in Digital Libraries, 2001. [6] Google Personalized. http://labs.google.com/personalized. [7] D. Hawking, N. Craswell, P. B. Thistlewaite, and D. Harman. Results and challenges in web search evaluation. Computer Networks, 31(11-16):1321-1330, 1999. [8] X. Huang, F. Peng, A. An, and D. Schuurmans. Dynamic web log session identification with statistical language models. Journal of the American Society for Information Science and Technology, 55(14):1290-1303, 2004. [9] G. Jeh and J. Widom. Scaling personalized web search. In Proceedings of WWW 2003, pages 271-279, 2003. [10] T. Joachims. Optimizing search engines using clickthrough data. In Proceedings of SIGKDD 2002, pages 133-142, 2002. [11] D. Kelly and J. Teevan. Implicit feedback for inferring user preference: A bibliography. SIGIR Forum, 37(2):18-28, 2003. [12] J. Lafferty and C. Zhai. Document language models, query models, and risk minimization for information retrieval. In Proceedings of SIGIR"01, pages 111-119, 2001. [13] T. Lau and E. Horvitz. Patterns of search: Analyzing and modeling web query refinement. In Proceedings of the Seventh International Conference on User Modeling (UM), pages 145 -152, 1999. [14] V. Lavrenko and B. Croft. Relevance-based language models. In Proceedings of SIGIR"01, pages 120-127, 2001. [15] M. Mitra, A. Singhal, and C. Buckley. Improving automatic query expansion. In Proceedings of SIGIR 1998, pages 206-214, 1998. [16] My Yahoo! http://mysearch.yahoo.com. [17] G. Nunberg. As google goes, so goes the nation. New York Times, May 2003. [18] S. E. Robertson. The probability ranking principle in ı˚. Journal of Documentation, 33(4):294-304, 1977. [19] J. J. Rocchio. Relevance feedback in information retrieval. In The SMART Retrieval System: Experiments in Automatic Document Processing, pages 313-323. Prentice-Hall Inc., 1971. [20] G. Salton and C. Buckley. Improving retrieval performance by retrieval feedback. Journal of the American Society for Information Science, 41(4):288-297, 1990. [21] G. Salton and M. J. McGill. Introduction to Modern Information Retrieval. McGraw-Hill, 1983. [22] X. Shen, B. Tan, and C. Zhai. Context-sensitive information retrieval using implicit feedback. In Proceedings of SIGIR 2005, pages 43-50, 2005. [23] X. Shen and C. Zhai. Exploiting query history for document ranking in interactive information retrieval (Poster). In Proceedings of SIGIR 2003, pages 377-378, 2003. [24] A. Singhal. Modern information retrieval: A brief overview. Bulletin of the IEEE Computer Society Technical Committee on Data Engineering, 24(4):35-43, 2001. [25] K. Sugiyama, K. Hatano, and M. Yoshikawa. Adaptive web search based on user profile constructed without any effort from users. In Proceedings of WWW 2004, pages 675-684, 2004. [26] E. Volokh. Personalization and privacy. Communications of the ACM, 43(8):84-88, 2000. [27] R. W. White, J. M. Jose, C. J. van Rijsbergen, and I. Ruthven. A simulated study of implicit feedback models. In Proceedings of ECIR 2004, pages 311-326, 2004. [28] J. Xu and W. B. Croft. Query expansion using local and global document analysis. In Proceedings of SIGIR 1996, pages 4-11, 1996. [29] C. Zhai and J. Lafferty. Model-based feedback in KL divergence retrieval model. In Proceedings of the CIKM 2001, pages 403-410, 2001. 831

interactive ir;personalize search;user model;interactive retrieval;query expansion;personalized web search;user-centered adaptive information retrieval;personalize information retrieval;implicit feedback;search accuracy;implicit user modeling;information retrieval system;retrieval performance;query refinement

train_H-63

Location based Indexing Scheme for DAYS

Data dissemination through wireless channels for broadcasting information to consumers is becoming quite common. Many dissemination schemes have been proposed but most of them push data to wireless channels for general consumption. Push based broadcast [1] is essentially asymmetric, i.e., the volume of data being higher from the server to the users than from the users back to the server. Push based scheme requires some indexing which indicates when the data will be broadcast and its position in the broadcast. Access latency and tuning time are the two main parameters which may be used to evaluate an indexing scheme. Two of the important indexing schemes proposed earlier were tree based and the exponential indexing schemes. None of these schemes were able to address the requirements of location dependent data (LDD) which is highly desirable feature of data dissemination. In this paper, we discuss the broadcast of LDD in our project DAta in Your Space (DAYS), and propose a scheme for indexing LDD. We argue that this scheme, when applied to LDD, significantly improves performance in terms of tuning time over the above mentioned schemes. We prove our argument with the help of simulation results.

1. INTRODUCTION Wireless data dissemination is an economical and efficient way to make desired data available to a large number of mobile or static users. The mode of data transfer is essentially asymmetric, that is, the capacity of the transfer of data (downstream communication) from the server to the client (mobile user) is significantly larger than the client or mobile user to the server (upstream communication). The effectiveness of a data dissemination system is judged by its ability to provide user the required data at anywhere and at anytime. One of the best ways to accomplish this is through the dissemination of highly personalized Location Based Services (LBS) which allows users to access personalized location dependent data. An example would be someone using their mobile device to search for a vegetarian restaurant. The LBS application would interact with other location technology components or use the mobile user's input to determine the user's location and download the information about the restaurants in proximity to the user by tuning into the wireless channel which is disseminating LDD. We see a limited deployment of LBS by some service providers. But there are every indications that with time some of the complex technical problems such as uniform location framework, calculating and tracking locations in all types of places, positioning in various environments, innovative location applications, etc., will be resolved and LBS will become a common facility and will help to improve market productivity and customer comfort. In our project called DAYS, we use wireless data broadcast mechanism to push LDD to users and mobile users monitor and tune the channel to find and download the required data. A simple broadcast, however, is likely to cause significant performance degradation in the energy constrained mobile devices and a common solution to this problem is the use of efficient air indexing. The indexing approach stores control information which tells the user about the data location in the broadcast and how and when he could access it. A mobile user, thus, has some free time to go into the doze mode which conserves valuable power. It also allows the user to personalize his own mobile device by selectively tuning to the information of his choice. Access efficiency and energy conservation are the two issues which are significant for data broadcast systems. Access efficiency refers to the latency experienced when a request is initiated till the response is received. Energy conservation [7, 10] refers to the efficient use of the limited energy of the mobile device in accessing broadcast data. Two parameters that affect these are the tuning time and the access latency. Tuning time refers to the time during which the mobile unit (MU) remains in active state to tune the channel and download its required data. It can also be defined as the number of buckets tuned by the mobile device in active state to get its required data. Access latency may be defined as the time elapsed since a request has been issued till the response has been received. 1 This research was supported by a grant from NSF IIS-0209170. Several indexing schemes have been proposed in the past and the prominent among them are the tree based and the exponential indexing schemes [17]. The main disadvantages of the tree based schemes are that they are based on centralized tree structures. To start a search, the MU has to wait until it reaches the root of the next broadcast tree. This significantly affects the tuning time of the mobile unit. The exponential schemes facilitate index replication by sharing links in different search trees. For broadcasts with large number of pages, the exponential scheme has been shown to perform similarly as the tree based schemes in terms of access latency. Also, the average length of broadcast increases due to the index replication and this may cause significant increase in the access latency. None of the above indexing schemes is equally effective in broadcasting location dependent data. In addition to providing low latency, they lack properties which are used to address LDD issues. We propose an indexing scheme in DAYS which takes care of some these problems. We show with simulation results that our scheme outperforms some of the earlier indexing schemes for broadcasting LDD in terms of tuning time. The rest of the paper is presented as follows. In section 2, we discuss previous work related to indexing of broadcast data. Section 3 describes our DAYS architecture. Location dependent data, its generation and subsequent broadcast is presented in section 4. Section 5 discusses our indexing scheme in detail. Simulation of our scheme and its performance evaluation is presented in section 6. Section 7 concludes the paper and mentions future related work. 2. PREVIOUS WORK Several disk-based indexing techniques have been used for air indexing. Imielinski et al. [5, 6] applied the B+ index tree, where the leaf nodes store the arrival times of the data items. The distributed indexing method was proposed to efficiently replicate and distribute the index tree in a broadcast. Specifically, the index tree is divided into a replicated part and a non replicated part. Each broadcast consists of the replicated part and the nonreplicated part that indexes the data items immediately following it. As such, each node in the non-replicated part appears only once in a broadcast and, hence, reduces the replication cost and access latency while achieving a good tuning time. Chen et al. [2] and Shivakumar et al. [8] considered unbalanced tree structures to optimize energy consumption for non-uniform data access. These structures minimize the average index search cost by reducing the number of index searches for hot data at the expense of spending more on cold data. Tan and Yu discussed data and index organization under skewed broadcast Hashing and signature methods have also been suggested for wireless broadcast that supports equality queries [9]. A flexible indexing method was proposed in [5]. The flexible index first sorts the data items in ascending (or descending) order of the search key values and then divides them into p segments. The first bucket in each data segment contains a control index, which is a binary index mapping a given key value to the segment containing that key, and a local index, which is an m-entry index mapping a given key value to the buckets within the current segment. By tuning the parameters of p and m, mobile clients can achieve either a good tuning time or good access latency. Another indexing technique proposed is the exponential indexing scheme [17]. In this scheme, a parameterized index, called the exponential index is used to optimize the access latency or the tuning time. It facilitates index replication by linking different search trees. All of the above mentioned schemes have been applied to data which are non related to each other. These non related data may be clustered or non clustered. However, none of them has specifically addressed the requirements of LDD. Location dependent data are data which are associated with a location. Presently there are several applications that deal with LDD [13, 16]. Almost all of them depict LDD with the help of hierarchical structures [3, 4]. This is based on the containment property of location dependent data. The Containment property helps determining relative position of an object by defining or identifying locations that contains those objects. The subordinate locations are hierarchically related to each other. Thus, Containment property limits the range of availability or operation of a service. We use this containment property in our indexing scheme to index LDD. 3. DAYS ARCHITECTURE DAYS has been conceptualized to disseminate topical and nontopical data to users in a local broadcast space and to accept queries from individual users globally. Topical data, for example, weather information, traffic information, stock information, etc., constantly changes over time. Non topical data such as hotel, restaurant, real estate prices, etc., do not change so often. Thus, we envision the presence of two types of data distribution: In the first case, server pushes data to local users through wireless channels. The other case deals with the server sending results of user queries through downlink wireless channels. Technically, we see the presence of two types of queues in the pull based data access. One is a heavily loaded queue containing globally uploaded queries. The other is a comparatively lightly loaded queue consisting of locally uploaded queries. The DAYS architecture [12] as shown in figure 1 consists of a Data Server, Broadcast Scheduler, DAYS Coordinator, Network of LEO satellites for global data delivery and a Local broadcast space. Data is pushed into the local broadcast space so that users may tune into the wireless channels to access the data. The local broadcast space consists of a broadcast tower, mobile units and a network of data staging machines called the surrogates. Data staging in surrogates has been earlier investigated as a successful technique [12, 15] to cache users' related data. We believe that data staging can be used to drastically reduce the latency time for both the local broadcast data as well as global responses. Query request in the surrogates may subsequently be used to generate the popularity patterns which ultimately decide the broadcast schedule [12]. 18 Popularity Feedback from Surrogates for Broadcast Scheduler Local Broadcast Space Broadcast Tower SurrogateMU MU MU MU Data ServerBroadcast schedulerDAYS Coordinator Local downlink channel Global downlink channel Pull request queue Global request queue Local request queue Location based index Starbucks Plaza Kansas City Figure 1. DAYS Architecture Figure 2. Location Structure ofStarbucks, Plaza 4. LOCATION DEPENDENT DATA (LDD) We argue that incorporating location information in wireless data broadcast can significantly decrease the access latency. This property becomes highly useful for mobile unit which has limited storage and processing capability. There are a variety of applications to obtain information about traffic, restaurant and hotel booking, fast food, gas stations, post office, grocery stores, etc. If these applications are coupled with location information, then the search will be fast and highly cost effective. An important property of the locations is Containment which helps to determine the relative location of an object with respect to its parent that contains the object. Thus, Containment limits the range of availability of a data. We use this property in our indexing scheme. The database contains the broadcast contents which are converted into LDD [14] by associating them with respective locations so that it can be broadcasted in a clustered manner. The clustering of LDD helps the user to locate information efficiently and supports containment property. We present an example to justify our proposition. Example: Suppose a user issues query Starbucks Coffee in Plaza please. to access information about the Plaza branch of Starbucks Coffee in Kansas City. In the case of location independent set up the system will list all Starbucks coffee shops in Kansas City area. It is obvious that such responses will increase access latency and are not desirable. These can be managed efficiently if the server has location dependent data, i.e., a mapping between a Starbucks coffee shop data and its physical location. Also, for a query including range of locations of Starbucks, a single query requesting locations for the entire region of Kansas City, as shown in Figure 2, will suffice. This will save enormous amount of bandwidth by decreasing the number of messages and at the same time will be helpful in preventing the scalability bottleneck in highly populated area. 4.1 Mapping Function for LDD The example justifies the need for a mapping function to process location dependent queries. This will be especially important for pull based queries across the globe for which the reply could be composed for different parts of the world. The mapping function is necessary to construct the broadcast schedule. We define Global Property Set (GPS) [11], Information Content (IC) set, and Location Hierarchy (LH) set where IC ⊆ GPS and LH ⊆ GPS to develop a mapping function. LH = {l1, l2, l3…,lk} where li represent locations in the location tree and IC = {ic1, ic2, ic3,…,icn} where ici represent information type. For example, if we have traffic, weather, and stock information are in broadcast then IC = {ictraffic, icweather, and icstock}. The mapping scheme must be able to identify and select an IC member and a LH node for (a) correct association, (b) granularity match, (c) and termination condition. For example, weather ∈ IC could be associated with a country or a state or a city or a town of LH. The granularity match between the weather and a LH node is as per user requirement. Thus, with a coarse granularity weather information is associated with a country to get country"s weather and with town in a finer granularity. If a town is the finest granularity, then it defines the terminal condition for association between IC and LH for weather. This means that a user cannot get weather information about subdivision of a town. In reality weather of a subdivision does not make any sense. We develop a simple heuristic mapping approach scheme based on user requirement. Let IC = {m1, m2,m3 .,..., mk}, where mi represent its element and let LH = {n1, n2, n3, ..., nl}, where ni represents LH"s member. We define GPS for IC (GPSIC) ⊆ GPS and for LH (GPSLH) ⊆ GPS as GPSIC = {P1, P2,…, Pn}, where P1, P2, P3,…, Pn are properties of its members and GPSLH = {Q1, Q2,…, Qm} where Q1, Q2,…, Qm are properties of its members. The properties of a particular member of IC are a subset of GPSIC. It is generally true that (property set (mi∈ IC) ∪ property set (mj∈ IC)) ≠ ∅, however, there may be cases where the intersection is not null. For example, stock ∈ IC and movie ∈ IC rating do not have any property in common. We assume that any two or more members of IC have at least one common geographical property (i.e. location) because DAYS broadcasts information about those categories, which are closely tied with a location. For example, stock of a company is related to a country, weather is related to a city or state, etc. We define the property subset of mi∈ IC as PSm i ∀ mi ∈ IC and PSm i = {P1, P2, ..., Pr} where r ≤ n. ∀ Pr {Pr ∈ PSm i → Pr∈ GPSIC} which implies that ∀ i, PSm i ⊆ GPSIC. The geographical properties of this set are indicative of whether mi ∈ IC can be mapped to only a single granularity level (i.e. a single location) in LH or a multiple granularity levels (i.e. more than one nodes in 19 the hierarchy) in LH. How many and which granularity levels should a mi map to, depends upon the level at which the service provider wants to provide information about the mi in question. Similarly we define a property subset of LH members as PSn j ∀ nj ∈ LH which can be written as PSn j ={Q1, Q2, Q3, …, Qs} where s ≤ m. In addition, ∀ Qs {Qs∈ PSn j → Qs∈ GPSLH} which implies that ∀j, PSn j ⊆ GPSLH. The process of mapping from IC to LH is then identifying for some mx∈ IC one or more ny∈ LH such that PSmx ∩ PSnv ≠ φ. This means that when mx maps to ny and all children of ny if mx can map to multiple granularity levels or mx maps only to ny if mx can map to a single granularity level. We assume that new members can join and old member can leave IC or LH any time. The deletion of members from the IC space is simple but addition of members to the IC space is more restrictive. If we want to add a new member to the IC space, then we first define a property set for the new member: PSmnew_m ={P1, P2, P3, …, Pt} and add it to the IC only if the condition:∀ Pw {Pw∈ PSpnew_m → Pw∈ GPSIC} is satisfied. This scheme has an additional benefit of allowing the information service providers to have a control over what kind of information they wish to provide to the users. We present the following example to illustrate the mapping concept. IC = {Traffic, Stock, Restaurant, Weather, Important history dates, Road conditions} LH = {Country, State, City, Zip-code, Major-roads} GPSIC = {Surface-mobility, Roads, High, Low, Italian-food, StateName, Temp, CityName, Seat-availability, Zip, Traffic-jams, Stock-price, CountryName, MajorRoadName, Wars, Discoveries, World} GPSLH = {Country, CountrySize, StateName, CityName, Zip, MajorRoadName} Ps(ICStock) = {Stock-price, CountryName, High, Low} Ps(ICTraffic) = {Surface-mobility, Roads, High, Low, Traffic-jams, CityName} Ps(ICImportant dates in history) = {World, Wars, Discoveries} Ps(ICRoad conditions) = {Precipitation, StateName, CityName} Ps(ICRestaurant) = {Italian-food, Zip code} Ps(ICWeather) = {StateName, CityName, Precipitation, Temperature} PS(LHCountry) = {CountryName, CountrySize} PS(LHState = {StateName, State size}, PS(LHCity) ={CityName, City size} PS(LHZipcode) = {ZipCodeNum } PS(LHMajor roads) = {MajorRoadName} Now, only PS(ICStock) ∩ PSCountry ≠φ. In addition, PS(ICStock) indicated that Stock can map to only a single location Country. When we consider the member Traffic of IC space, only PS(ICTraffic) ∩ PScity ≠ φ. As PS(ICTraffic) indicates that Traffic can map to only a single location, it maps only to City and none of its children. Now unlike Stock, mapping of Traffic with Major roads, which is a child of City, is meaningful. However service providers have right to control the granularity levels at which they want to provide information about a member of IC space. PS(ICRoad conditions) ∩ PSState ≠φ and PS(ICRoad conditions) ∩ PSCity≠φ. So Road conditions maps to State as well as City. As PS(ICRoad conditions) indicates that Road conditions can map to multiple granularity levels, Road conditions will also map to Zip Code and Major roads, which are the children of State and City. Similarly, Restaurant maps only to Zip code, and Weather maps to State, City and their children, Major Roads and Zip Code. 5. LOCATION BASED INDEXING SCHEME This section discusses our location based indexing scheme (LBIS). The scheme is designed to conform to the LDD broadcast in our project DAYS. As discussed earlier, we use the containment property of LDD in the indexing scheme. This significantly limits the search of our required data to a particular portion of broadcast. Thus, we argue that the scheme provides bounded tuning time. We describe the architecture of our indexing scheme. Our scheme contains separate data buckets and index buckets. The index buckets are of two types. The first type is called the Major index. The Major index provides information about the types of data broadcasted. For example, if we intend to broadcast information like Entertainment, Weather, Traffic etc., then the major index points to either these major types of information and/or their main subtypes of information, the number of main subtypes varying from one information to another. This strictly limits number of accesses to a Major index. The Major index never points to the original data. It points to the sub indexes called the Minor index. The minor indexes are the indexes which actually points to the original data. We called these minor index pointers as Location Pointers as they points to the data which are associated with a location. Thus, our search for a data includes accessing of a major index and some minor indexes, the number of minor index varying depending on the type of information. Thus, our indexing scheme takes into account the hierarchical nature of the LDD, the Containment property, and requires our broadcast schedule to be clustered based on data type and location. The structure of the location hierarchy requires the use of different types of index at different levels. The structure and positions of index strictly depend on the location hierarchy as described in our mapping scheme earlier. We illustrate the implementation of our scheme with an example. The rules for framing the index are mentioned subsequently. 20 A1 Entertainment Resturant Movie A2 A3 A4 R1 R2 R3 R4 R5 R6 R7 R8 Weather KC SL JC SF Entertainment R1 R2 R3 R4 R5 R6 R7 R8 KC SL JC SF (A, R, NEXT = 8) 3, R5 4, R7 Type (S, L) ER W E EM (1, 4) (5, 4) (1, 4), (9, 4) (9, 4) Type (S, L) W E EM ER (1, 4) (5, 8) (5, 4) (9, 4) Type (S, L) E EM ER W (1, 8) (1, 4) (5, 4) (9, 4) A1 A2 A3 A4 Movie Resturant Weather 1 2 3 4 5 6 7 8 9 10 11 12 Major index Major index Major index Minor index Major index Minor index Figure 3. Location Mapped Information for Broadcast Figure 4. Data coupled with Location based Index Example: Let us suppose that our broadcast content contains ICentertainment and ICweather which is represented as shown in Fig. 3. Ai represents Areas of City and Ri represents roads in a certain area. The leaves of Weather structure represent four cities. The index structure is given in Fig. 4 which shows the position of major and minor index and data in the broadcast schedule. We propose the following rules for the creation of the air indexed broadcast schedule: • The major index and the minor index are created. • The major index contains the position and range of different types of data items (Weather and Entertainment, Figure 3) and their categories. The sub categories of Entertainment, Movie and Restaurant, are also in the index. Thus, the major index contains Entertainment (E), Entertainment-Movie (EM), Entertainment-Restaurant (ER), and Weather (W). The tuple (S, L) represents the starting position (S) of the data item and L represents the range of the item in terms of number of data buckets. • The minor index contains the variables A, R and a pointer Next. In our example (Figure 3), road R represents the first node of area A. The minor index is used to point to actual data buckets present at the lowest levels of the hierarchy. In contrast, the major index points to a broader range of locations and so it contains information about main and sub categories of data. • Index information is not incorporated in the data buckets. Index buckets are separate containing only the control information. • The number of major index buckets m=#(IC), IC = {ic1, ic2, ic3,…,icn} where ici represent information type and # represents the cardinality of the Information Content set IC. In this example, IC= {icMovie, icWeather, icRestaurant} and so #(IC) =3. Hence, the number of major index buckets is 3. • Mechanism to resolve the query is present in the java based coordinator in MU. For example, if a query Q is presented as Q (Entertainment, Movie, Road_1), then the resultant search will be for the EM information in the major index. We say, Q EM. Our proposed index works as follows: Let us suppose that an MU issues a query which is represented by Java Coordinator present in the MU as Restaurant information on Road 7. This is resolved by the coordinator as Q ER. This means one has to search for ER unit of index in the major index. Let us suppose that the MU logs into the channel at R2. The first index it receives is a minor index after R2. In this index, value of Next variable = 4, which means that the next major index is present after bucket 4. The MU may go into doze mode. It becomes active after bucket 4 and receives the major index. It searches for ER information which is the first entry in this index. It is now certain that the MU will get the position of the data bucket in the adjoining minor index. The second unit in the minor index depicts the position of the required data R7. It tells that the data bucket is the first bucket in Area 4. The MU goes into doze mode again and becomes active after bucket 6. It gets the required data in the next bucket. We present the algorithm for searching the location based Index. Algorithm 1 Location based Index Search in DAYS 1. Scan broadcast for the next index bucket, found=false 2. While (not found) do 3. if bucket is Major Index then 4. Find the Type & Tuple (S, L) 5. if S is greater than 1, go into doze mode for S seconds 6. end if 7. Wake up at the Sth bucket and observe the Minor Index 8. end if 9. if bucket is Minor Index then 10. if TypeRequested not equal to Typefound and (A,R)Request not equal to (A,R)found then 11. Go into doze mode till NEXT & repeat from step 3 12. end if 13. else find entry in Minor Index which points to data 14. Compute time of arrival T of data bucket 15. Go into doze mode till T 16. Wake up at T and access data, found = true 17. end else 18. end if 19. end While 21 6. PERFORMANCE EVALUATION Conservation of energy is the main concern when we try to access data from wireless broadcast. An efficient scheme should allow the mobile device to access its required data by staying active for a minimum amount of time. This would save considerable amount of energy. Since items are distributed based on types and are mapped to suitable locations, we argue that our broadcast deals with clustered data types. The mobile unit has to access a larger major index and a relatively much smaller minor index to get information about the time of arrival of data. This is in contrast to the exponential scheme where the indexes are of equal sizes. The example discussed and Algorithm 1 reveals that to access any data, we need to access the major index only once followed by one or more accesses to the minor index. The number of minor index access depends on the number of internal locations. As the number of internal locations vary for item to item (for example, Weather is generally associated with a City whereas traffic is granulated up to major and minor roads of a city), we argue that the structure of the location mapped information may be visualized as a forest which is a collection of general trees, the number of general trees depending on the types of information broadcasted and depth of a tree depending on the granularity of the location information associated with the information. For our experiments, we assume the forest as a collection of balanced M-ary trees. We further assume the M-ary trees to be full by assuming the presence of dummy nodes in different levels of a tree. Thus, if the number of data items is d and the height of the tree is m, then n= (m*d-1)/(m-1) where n is the number of vertices in the tree and i= (d-1)/(m-1) where i is the number of internal vertices. Tuning time for a data item involves 1 unit of time required to access the major index plus time required to access the data items present in the leaves of the tree. Thus, tuning time with d data items is t = logmd+1 We can say that tuning time is bounded by O(logmd). We compare our scheme with the distributed indexing and exponential scheme. We assume a flat broadcast and number of pages varying from 5000 to 25000. The various simulation parameters are shown in Table 1. Figure 5-8 shows the relative tuning times of three indexing algorithms, ie, the LBIS, exponential scheme and the distributed tree scheme. Figure 5 shows the result for number of internal location nodes = 3. We can see that LBIS significantly outperforms both the other schemes. The tuning time in LBIS ranges from approx 6.8 to 8. This large tuning time is due to the fact that after reaching the lowest minor index, the MU may have to access few buckets sequentially to get the required data bucket. We can see that the tuning time tends to become stable as the length of broadcast increases. In figure 6 we consider m= 4. Here we can see that the exponential and the distributed perform almost similarly, though the former seems to perform slightly better as the broadcast length increases. A very interesting pattern is visible in figure 7. For smaller broadcast size, the LBIS seems to have larger tuning time than the other two schemes. But as the length of broadcast increases, it is clearly visible the LBIS outperforms the other two schemes. The Distributed tree indexing shows similar behavior like the LBIS. The tuning time in LBIS remains low because the algorithm allows the MU to skip some intermediate Minor Indexes. This allows the MU to move into lower levels directly after coming into active mode, thus saving valuable energy. This action is not possible in the distributed tree indexing and hence we can observe that its tuning time is more than the LBIS scheme, although it performs better than the exponential scheme. Figure 8, in contrast, shows us that the tuning time in LBIS, though less than the other two schemes, tends to increase sharply as the broadcast length becomes greater than the 15000 pages. This may be attributed both due to increase in time required to scan the intermediate Minor Indexes and the length of the broadcast. But we can observe that the slope of the LBIS curve is significantly less than the other two curves. Table 1 Simulation Parameters P Definition Values N Number of data Items 5000 - 25000 m Number of internal location nodes 3, 4, 5, 6 B Capacity of bucket without index (for exponential index) 10,64,128,256 i Index base for exponential index 2,4,6,8 k Index size for distributed tree 8 bytes The simulation results establish some facts about our location based indexing scheme. The scheme performs better than the other two schemes in terms of tuning time in most of the cases. As the length of the broadcast increases, after a certain point, though the tuning time increases as a result of factors which we have described before, the scheme always performs better than the other two schemes. Due to the prescribed limit of the number of pages in the paper, we are unable to show more results. But these omitted results show similar trend as the results depicted in figure 5-8. 7. CONCLUSION AND FUTURE WORK In this paper we have presented a scheme for mapping of wireless broadcast data with their locations. We have presented an example to show how the hierarchical structure of the location tree maps with the data to create LDD. We have presented a scheme called LBIS to index this LDD. We have used the containment property of LDD in the scheme that limits the search to a narrow range of data in the broadcast, thus saving valuable energy in the device. The mapping of data with locations and the indexing scheme will be used in our DAYS project to create the push based architecture. The LBIS has been compared with two other prominent indexing schemes, i.e., the distributed tree indexing scheme and the exponential indexing scheme. We showed in our simulations that the LBIS scheme has the lowest tuning time for broadcasts having large number of pages, thus saving valuable battery power in the MU. 22 In the future work we try to incorporate pull based architecture in our DAYS project. Data from the server is available for access by the global users. This may be done by putting a request to the source server. The query in this case is a global query. It is transferred from the user"s source server to the destination server through the use of LEO satellites. We intend to use our LDD scheme and data staging architecture in the pull based architecture. We will show that the LDD scheme together with the data staging architecture significantly improves the latency for global as well as local query. 8. REFERENCES [1] Acharya, S., Alonso, R. Franklin, M and Zdonik S. Broadcast disk: Data management for asymmetric communications environments. In Proceedings of ACM SIGMOD Conference on Management of Data, pages 199-210, San Jose, CA, May 1995. [2] Chen, M.S.,Wu, K.L. and Yu, P. S. Optimizing index allocation for sequential data broadcasting in wireless mobile computing. IEEE Transactions on Knowledge and Data Engineering (TKDE), 15(1):161-173, January/February 2003. Figure 5. Broadcast Size (# buckets) Dist tree Expo LBIS Figure 6. Broadcast Size (# buckets) Dist tree Expo LBIS Figure 7. Broadcast Size (# buckets) Dist tree Expo LBIS Figure 8. Broadcast Size (# buckets) Dist tree Expo LBIS Averagetuningtime Averagetuningtime Averagetuningtime Averagetuningtime 23 [3] Hu, Q. L., Lee, D. L. and Lee, W.C. Performance evaluation of a wireless hierarchical data dissemination system. In Proceedings of the 5th Annual ACM International Conference on Mobile Computing and Networking (MobiCom"99), pages 163-173, Seattle, WA, August 1999. [4] Hu, Q. L. Lee, W.C. and Lee, D. L. Power conservative multi-attribute queries on data broadcast. In Proceedings of the 16th International Conference on Data Engineering (ICDE"00), pages 157-166, San Diego, CA, February 2000. [5] Imielinski, T., Viswanathan, S. and Badrinath. B. R. Power efficient filtering of data on air. In Proceedings of the 4th International Conference on Extending Database Technology (EDBT"94), pages 245-258, Cambridge, UK, March 1994. [6] Imielinski, T., Viswanathan, S. and Badrinath. B. R. Data on air - Organization and access. IEEE Transactions on Knowledge and Data Engineering (TKDE), 9(3):353-372, May/June 1997. [7] Shih, E., Bahl, P. and Sinclair, M. J. Wake on wireless: An event driven energy saving strategy for battery operated devices. In Proceedings of the 8th Annual ACM International Conference on Mobile Computing and Networking (MobiCom"02), pages 160-171, Atlanta, GA, September 2002. [8] Shivakumar N. and Venkatasubramanian, S. Energy-efficient indexing for information dissemination in wireless systems. ACM/Baltzer Journal of Mobile Networks and Applications (MONET), 1(4):433-446, December 1996. [9] Tan K. L. and Yu, J. X. Energy efficient filtering of non uniform broadcast. In Proceedings of the 16th International Conference on Distributed Computing Systems (ICDCS"96), pages 520-527, Hong Kong, May 1996. [10] Viredaz, M. A., Brakmo, L. S. and Hamburgen, W. R. Energy management on handheld devices. ACM Queue, 1(7):44-52, October 2003. [11] Garg, N. Kumar, V., & Dunham, M.H. Information Mapping and Indexing in DAYS, 6th International Workshop on Mobility in Databases and Distributed Systems, in conjunction with the 14th International Conference on Database and Expert Systems Applications September 1-5, Prague, Czech Republic, 2003. [12] Acharya D., Kumar, V., & Dunham, M.H. InfoSpace: Hybrid and Adaptive Public Data Dissemination System for Ubiquitous Computing. Accepted for publication in the special issue of Pervasive Computing. Wiley Journal for Wireless Communications and Mobile Computing, 2004. [13] Acharya D., Kumar, V., & Prabhu, N. Discovering and using Web Services in M-Commerce, Proceedings for 5th VLDB Workshop on Technologies for E-Services, Toronto, Canada,2004. [14] Acharya D., Kumar, V. Indexing Location Dependent Data in broadcast environment. Accepted for publication, JDIM special issue on Distributed Data Management, 2005. [15] Flinn, J., Sinnamohideen, S., & Satyanarayan, M. Data Staging on Untrusted Surrogates, Intel Research, Pittsburg, Unpublished Report, 2003. [16] Seydim, A.Y., Dunham, M.H. & Kumar, V. Location dependent query processing, Proceedings of the 2nd ACM international workshop on Data engineering for wireless and mobile access, p.47-53, Santa Barbara, California, USA, 2001. [17] Xu, J., Lee, W.C., Tang., X. Exponential Index: A Parameterized Distributed Indexing Scheme for Data on Air. In Proceedings of the 2nd ACM/USENIX International Conference on Mobile Systems, Applications, and Services (MobiSys'04), Boston, MA, June 2004. 24

wireless broadcast datum mapping;location base service;datum stage;indexing scheme;pull based datum access;wireless datum dissemination;tree structure;day;wireless datum broadcast;datum broadcast system;location dependent datum;ldd;location based service;index;wireless channel;mapping of wireless broadcast datum;mobile user

train_H-64

Machine Learning for Information Architecture in a Large Governmental Website

This paper describes ongoing research into the application of machine learning techniques for improving access to governmental information in complex digital libraries. Under the auspices of the GovStat Project, our goal is to identify a small number of semantically valid concepts that adequately spans the intellectual domain of a collection. The goal of this discovery is twofold. First we desire a practical aid for information architects. Second, automatically derived documentconcept relationships are a necessary precondition for realworld deployment of many dynamic interfaces. The current study compares concept learning strategies based on three document representations: keywords, titles, and full-text. In statistical and user-based studies, human-created keywords provide significant improvements in concept learning over both title-only and full-text representations.

1. INTRODUCTION The GovStat Project is a joint effort of the University of North Carolina Interaction Design Lab and the University of Maryland Human-Computer Interaction Lab1 . Citing end-user difficulty in finding governmental information (especially statistical data) online, the project seeks to create an integrated model of user access to US government statistical information that is rooted in realistic data models and innovative user interfaces. To enable such models and interfaces, we propose a data-driven approach, based on data mining and machine learning techniques. In particular, our work analyzes a particular digital library-the website of the Bureau of Labor Statistics2 (BLS)-in efforts to discover a small number of linguistically meaningful concepts, or bins, that collectively summarize the semantic domain of the site. The project goal is to classify the site"s web content according to these inferred concepts as an initial step towards data filtering via active user interfaces (cf. [13]). Many digital libraries already make use of content classification, both explicitly and implicitly; they divide their resources manually by topical relation; they organize content into hierarchically oriented file systems. The goal of the present 1 http://www.ils.unc.edu/govstat 2 http://www.bls.gov 151 research is to develop another means of browsing the content of these collections. By analyzing the distribution of terms across documents, our goal is to supplement the agency"s pre-existing information structures. Statistical learning technologies are appealing in this context insofar as they stand to define a data-driven-as opposed to an agency-drivennavigational structure for a site. Our approach combines supervised and unsupervised learning techniques. A pure document clustering [12] approach to such a large, diverse collection as BLS led to poor results in early tests [6]. But strictly supervised techniques [5] are inappropriate, too. Although BLS designers have defined high-level subject headings for their collections, as we discuss in Section 2, this scheme is less than optimal. Thus we hope to learn an additional set of concepts by letting the data speak for themselves. The remainder of this paper describes the details of our concept discovery efforts and subsequent evaluation. In Section 2 we describe the previously existing, human-created conceptual structure of the BLS website. This section also describes evidence that this structure leaves room for improvement. Next (Sections 3-5), we turn to a description of the concepts derived via content clustering under three document representations: keyword, title only, and full-text. Section 6 describes a two-part evaluation of the derived conceptual structures. Finally, we conclude in Section 7 by outlining upcoming work on the project. 2. STRUCTURING ACCESS TO THE BLS WEBSITE The Bureau of Labor Statistics is a federal government agency charged with compiling and publishing statistics pertaining to labor and production in the US and abroad. Given this broad mandate, the BLS publishes a wide array of information, intended for diverse audiences. The agency"s website acts as a clearinghouse for this process. With over 15,000 text/html documents (and many more documents if spreadsheets and typeset reports are included), providing access to the collection provides a steep challenge to information architects. 2.1 The Relation Browser The starting point of this work is the notion that access to information in the BLS website could be improved by the addition of a dynamic interface such as the relation browser described by Marchionini and Brunk [13]. The relation browser allows users to traverse complex data sets by iteratively slicing the data along several topics. In Figure 1 we see a prototype instantiation of the relation browser, applied to the FedStats website3 . The relation browser supports information seeking by allowing users to form queries in a stepwise fashion, slicing and re-slicing the data as their interests dictate. Its motivation is in keeping with Shneiderman"s suggestion that queries and their results should be tightly coupled [2]. Thus in Fig3 http://www.fedstats.gov Figure 1: Relation Browser Prototype ure 1, users might limit their search set to those documents about energy. Within this subset of the collection, they might further eliminate documents published more than a year ago. Finally, they might request to see only documents published in PDF format. As Marchionini and Brunk discuss, capturing the publication date and format of documents is trivial. But successful implementations of the relation browser also rely on topical classification. This presents two stumbling blocks for system designers: • Information architects must define the appropriate set of topics for their collection • Site maintainers must classify each document into its appropriate categories These tasks parallel common problems in the metadata community: defining appropriate elements and marking up documents to support metadata-aware information access. Given a collection of over 15,000 documents, these hurdles are especially daunting, and automatic methods of approaching them are highly desirable. 2.2 A Pre-Existing Structure Prior to our involvement with the project, designers at BLS created a shallow classificatory structure for the most important documents in their website. As seen in Figure 2, the BLS home page organizes 65 top-level documents into 15 categories. These include topics such as Employment and Unemployment, Productivity, and Inflation and Spending. 152 Figure 2: The BLS Home Page We hoped initially that these pre-defined categories could be used to train a 15-way document classifier, thus automating the process of populating the relation browser altogether. However, this approach proved unsatisfactory. In personal meetings, BLS officials voiced dissatisfaction with the existing topics. Their form, it was argued, owed as much to the institutional structure of BLS as it did to the inherent topology of the website"s information space. In other words, the topics reflected official divisions rather than semantic clusters. The BLS agents suggested that re-designing this classification structure would be desirable. The agents" misgivings were borne out in subsequent analysis. The BLS topics comprise a shallow classificatory structure; each of the 15 top-level categories is linked to a small number of related pages. Thus there are 7 pages associated with Inflation. Altogether, the link structure of this classificatory system contains 65 documents; that is, excluding navigational links, there are 65 documents linked from the BLS home page, where each hyperlink connects a document to a topic (pages can be linked to multiple topics). Based on this hyperlink structure, we defined M, a symmetric 65×65 matrix, where mij counts the number of topics in which documents i and j are both classified on the BLS home page. To analyze the redundancy inherent in the pre-existing structure, we derived the principal components of M (cf. [11]). Figure 3 shows the resultant scree plot4 . Because all 65 documents belong to at least one BLS topic, 4 A scree plot shows the magnitude of the kth eigenvalue versus its rank. During principal component analysis scree plots visualize the amount of variance captured by each component. m00M0M 0 1010M10M 10 2020M20M 20 3030M30M 30 4040M40M 40 5050M50M 50 6060M60M 60 m00M0M 0 22M2M 2 44M4M 4 66M6M 6 88M8M 8 1010M10M 10 1212M12M 12 1414M14M 14 Eigenvalue RankMEigenvalue RankM Eigenvalue Rank Eigenvlue MagnitudeMEigenvlue MagnitudeM EigenvlueMagnitude Figure 3: Scree Plot of BLS Categories the rank of M is guaranteed to be less than or equal to 15 (hence, eigenvalues 16 . . . 65 = 0). What is surprising about Figure 3, however, is the precipitous decline in magnitude among the first four eigenvalues. The four largest eigenvlaues account for 62.2% of the total variance in the data. This fact suggests a high degree of redundancy among the topics. Topical redundancy is not in itself problematic. However, the documents in this very shallow classificatory structure are almost all gateways to more specific information. Thus the listing of the Producer Price Index under three categories could be confusing to the site"s users. In light of this potential for confusion and the agency"s own request for redesign, we undertook the task of topic discovery described in the following sections. 3. A HYBRID APPROACH TO TOPIC DISCOVERY To aid in the discovery of a new set of high-level topics for the BLS website, we turned to unsupervised machine learning methods. In efforts to let the data speak for themselves, we desired a means of concept discovery that would be based not on the structure of the agency, but on the content of the material. To begin this process, we crawled the BLS website, downloading all documents of MIME type text/html. This led to a corpus of 15,165 documents. Based on this corpus, we hoped to derive k ≈ 10 topical categories, such that each document di is assigned to one or more classes. 153 Document clustering (cf. [16]) provided an obvious, but only partial solution to the problem of automating this type of high-level information architecture discovery. The problems with standard clustering are threefold. 1. Mutually exclusive clusters are inappropriate for identifying the topical content of documents, since documents may be about many subjects. 2. Due to the heterogeneity of the data housed in the BLS collection (tables, lists, surveys, etc.), many documents" terms provide noisy topical information. 3. For application to the relation browser, we require a small number (k ≈ 10) of topics. Without significant data reduction, term-based clustering tends to deliver clusters at too fine a level of granularity. In light of these problems, we take a hybrid approach to topic discovery. First, we limit the clustering process to a sample of the entire collection, described in Section 4. Working on a focused subset of the data helps to overcome problems two and three, listed above. To address the problem of mutual exclusivity, we combine unsupervised with supervised learning methods, as described in Section 5. 4. FOCUSING ON CONTENT-RICH DOCUMENTS To derive empirically evidenced topics we initially turned to cluster analysis. Let A be the n×p data matrix with n observations in p variables. Thus aij shows the measurement for the ith observation on the jth variable. As described in [12], the goal of cluster analysis is to assign each of the n observations to one of a small number k groups, each of which is characterized by high intra-cluster correlation and low inter-cluster correlation. Though the algorithms for accomplishing such an arrangement are legion, our analysis focuses on k-means clustering5 , during which, each observation oi is assigned to the cluster Ck whose centroid is closest to it, in terms of Euclidean distance. Readers interested in the details of the algorithm are referred to [12] for a thorough treatment of the subject. Clustering by k-means is well-studied in the statistical literature, and has shown good results for text analysis (cf. [8, 16]). However, k-means clustering requires that the researcher specify k, the number of clusters to define. When applying k-means to our 15,000 document collection, indicators such as the gap statistic [17] and an analysis of the mean-squared distance across values of k suggested that k ≈ 80 was optimal. This paramterization led to semantically intelligible clusters. However, 80 clusters are far too many for application to an interface such as the relation 5 We have focused on k-means as opposed to other clustering algorithms for several reasons. Chief among these is the computational efficiency enjoyed by the k-means approach. Because we need only a flat clustering there is little to be gained by the more expensive hierarchical algorithms. In future work we will turn to model-based clustering [7] as a more principled method of selecting the number of clusters and of representing clusters. browser. Moreover, the granularity of these clusters was unsuitably fine. For instance, the 80-cluster solution derived a cluster whose most highly associated words (in terms of log-odds ratio [1]) were drug, pharmacy, and chemist. These words are certainly related, but they are related at a level of specificity far below what we sought. To remedy the high dimensionality of the data, we resolved to limit the algorithm to a subset of the collection. In consultation with employees of the BLS, we continued our analysis on documents that form a series titled From the Editor"s Desk6 . These are brief articles, written by BLS employees. BLS agents suggested that we focus on the Editor"s Desk because it is intended to span the intellectual domain of the agency. The column is published daily, and each entry describes an important current issue in the BLS domain. The Editor"s Desk column has been written daily (five times per week) since 1998. As such, we operated on a set of N = 1279 documents. Limiting attention to these 1279 documents not only reduced the dimensionality of the problem. It also allowed the clustering process to learn on a relatively clean data set. While the entire BLS collection contains a great deal of nonprose text (i.e. tables, lists, etc.), the Editor"s Desk documents are all written in clear, journalistic prose. Each document is highly topical, further aiding the discovery of termtopic relations. Finally, the Editor"s Desk column provided an ideal learning environment because it is well-supplied with topical metadata. Each of the 1279 documents contains a list of one or more keywords. Additionally, a subset of the documents (1112) contained a subject heading. This metadata informed our learning and evaluation, as described in Section 6.1. 5. COMBINING SUPERVISED AND UNSUPERVISED LEARNING FORTOPIC DISCOVERY To derive suitably general topics for the application of a dynamic interface to the BLS collection, we combined document clustering with text classification techniques. Specifically, using k-means, we clustered each of the 1279 documents into one of k clusters, with the number of clusters chosen by analyzing the within-cluster mean squared distance at different values of k (see Section 6.1). Constructing mutually exclusive clusters violates our assumption that documents may belong to multiple classes. However, these clusters mark only the first step in a two-phase process of topic identification. At the end of the process, documentcluster affinity is measured by a real-valued number. Once the Editor"s Desk documents were assigned to clusters, we constructed a k-way classifier that estimates the strength of evidence that a new document di is a member of class Ck. We tested three statistical classification techniques: probabilistic Rocchio (prind), naive Bayes, and support vector machines (SVMs). All were implemented using McCallum"s BOW text classification library [14]. Prind is a probabilistic version of the Rocchio classification algorithm [9]. Interested readers are referred to Joachims" article for 6 http://www.bls.gov/opub/ted 154 further details of the classification method. Like prind, naive Bayes attempts to classify documents into the most probable class. It is described in detail in [15]. Finally, support vector machines were thoroughly explicated by Vapnik [18], and applied specifically to text in [10]. They define a decision boundary by finding the maximally separating hyperplane in a high-dimensional vector space in which document classes become linearly separable. Having clustered the documents and trained a suitable classifier, the remaining 14,000 documents in the collection are labeled by means of automatic classification. That is, for each document di we derive a k-dimensional vector, quantifying the association between di and each class C1 . . . Ck. Deriving topic scores via naive Bayes for the entire 15,000document collection required less than two hours of CPU time. The output of this process is a score for every document in the collection on each of the automatically discovered topics. These scores may then be used to populate a relation browser interface, or they may be added to a traditional information retrieval system. To use these weights in the relation browser we currently assign to each document the two topics on which it scored highest. In future work we will adopt a more rigorous method of deriving documenttopic weight thresholds. Also, evaluation of the utility of the learned topics for users will be undertaken. 6. EVALUATION OF CONCEPT DISCOVERY Prior to implementing a relation browser interface and undertaking the attendant user studies, it is of course important to evaluate the quality of the inferred concepts, and the ability of the automatic classifier to assign documents to the appropriate subjects. To evaluate the success of the two-stage approach described in Section 5, we undertook two experiments. During the first experiment we compared three methods of document representation for the clustering task. The goal here was to compare the quality of document clusters derived by analysis of full-text documents, documents represented only by their titles, and documents represented by human-created keyword metadata. During the second experiment, we analyzed the ability of the statistical classifiers to discern the subject matter of documents from portions of the database in addition to the Editor"s Desk. 6.1 Comparing Document Representations Documents from The Editor"s Desk column came supplied with human-generated keyword metadata. Additionally, The titles of the Editor"s Desk documents tend to be germane to the topic of their respective articles. With such an array of distilled evidence of each document"s subject matter, we undertook a comparison of document representations for topic discovery by clustering. We hypothesized that keyword-based clustering would provide a useful model. But we hoped to see whether comparable performance could be attained by methods that did not require extensive human indexing, such as the title-only or full-text representations. To test this hypothesis, we defined three modes of document representation-full-text, title-only, and keyword only-we generated three sets of topics, Tfull, Ttitle, and Tkw, respectively. Topics based on full-text documents were derived by application of k-means clustering to the 1279 Editor"s Desk documents, where each document was represented by a 1908dimensional vector. These 1908 dimensions captured the TF.IDF weights [3] of each term ti in document dj, for all terms that occurred at least three times in the data. To arrive at the appropriate number of clusters for these data, we inspected the within-cluster mean-squared distance for each value of k = 1 . . . 20. As k approached 10 the reduction in error with the addition of more clusters declined notably, suggesting that k ≈ 10 would yield good divisions. To select a single integer value, we calculated which value of k led to the least variation in cluster size. This metric stemmed from a desire to suppress the common result where one large cluster emerges from the k-means algorithm, accompanied by several accordingly small clusters. Without reason to believe that any single topic should have dramatically high prior odds of document membership, this heuristic led to kfull = 10. Clusters based on document titles were constructed similarly. However, in this case, each document was represented in the vector space spanned by the 397 terms that occur at least twice in document titles. Using the same method of minimizing the variance in cluster membership ktitle-the number of clusters in the title-based representation-was also set to 10. The dimensionality of the keyword-based clustering was very similar to that of the title-based approach. There were 299 keywords in the data, all of which were retained. The median number of keywords per document was 7, where a keyword is understood to be either a single word, or a multiword term such as consumer price index. It is worth noting that the keywords were not drawn from any controlled vocabulary; they were assigned to documents by publishers at the BLS. Using the keywords, the documents were clustered into 10 classes. To evaluate the clusters derived by each method of document representation, we used the subject headings that were included with 1112 of the Editor"s Desk documents. Each of these 1112 documents was assigned one or more subject headings, which were withheld from all of the cluster applications. Like the keywords, subject headings were assigned to documents by BLS publishers. Unlike the keywords, however, subject headings were drawn from a controlled vocabulary. Our analysis began with the assumption that documents with the same subject headings should cluster together. To facilitate this analysis, we took a conservative approach; we considered multi-subject classifications to be unique. Thus if document di was assigned to a single subject prices, while document dj was assigned to two subjects, international comparisons, prices, documents di and dj are not considered to come from the same class. Table 1 shows all Editor"s Desk subject headings that were assigned to at least 10 documents. As noted in the table, 155 Table 1: Top Editor"s Desk Subject Headings Subject Count prices 92 unemployment 55 occupational safety & health 53 international comparisons, prices 48 manufacturing, prices 45 employment 44 productivity 40 consumer expenditures 36 earnings & wages 27 employment & unemployment 27 compensation costs 25 earnings & wages, metro. areas 18 benefits, compensation costs 18 earnings & wages, occupations 17 employment, occupations 14 benefits 14 earnings & wage, regions 13 work stoppages 12 earnings & wages, industries 11 Total 609 Table 2: Contingecy Table for Three Document Representations Representation Right Wrong Accuracy Full-text 392 217 0.64 Title 441 168 0.72 Keyword 601 8 0.98 there were 19 such subject headings, which altogether covered 609 (54%) of the documents with subjects assigned. These document-subject pairings formed the basis of our analysis. Limiting analysis to subjects with N > 10 kept the resultant χ2 tests suitably robust. The clustering derived by each document representation was tested by its ability to collocate documents with the same subjects. Thus for each of the 19 subject headings in Table 1, Si, we calculated the proportion of documents assigned to Si that each clustering co-classified. Further, we assumed that whichever cluster captured the majority of documents for a given class constituted the right answer for that class. For instance, There were 92 documents whose subject heading was prices. Taking the BLS editors" classifications as ground truth, all 92 of these documents should have ended up in the same cluster. Under the full-text representation 52 of these documents were clustered into category 5, while 35 were in category 3, and 5 documents were in category 6. Taking the majority cluster as the putative right home for these documents, we consider the accuracy of this clustering on this subject to be 52/92 = 0.56. Repeating this process for each topic across all three representations led to the contingency table shown in Table 2. The obvious superiority of the keyword-based clustering evidenced by Table 2 was borne out by a χ2 test on the accuracy proportions. Comparing the proportion right and Table 3: Keyword-Based Clusters benefits costs international jobs plans compensation import employment benefits costs prices jobs employees benefits petroleum youth occupations prices productivity safety workers prices productivity safety earnings index output health operators inflation nonfarm occupational spending unemployment expenditures unemployment consumer mass spending jobless wrong achieved by keyword and title-based clustering led to p 0.001. Due to this result, in the remainder of this paper, we focus our attention on the clusters derived by analysis of the Editor"s Desk keywords. The ten keyword-based clusters are shown in Table 3, represented by the three terms most highly associated with each cluster, in terms of the log-odds ratio. Additionally, each cluster has been given a label by the researchers. Evaluating the results of clustering is notoriously difficult. In order to lend our analysis suitable rigor and utility, we made several simplifying assumptions. Most problematic is the fact that we have assumed that each document belongs in only a single category. This assumption is certainly false. However, by taking an extremely rigid view of what constitutes a subject-that is, by taking a fully qualified and often multipart subject heading as our unit of analysis-we mitigate this problem. Analogically, this is akin to considering the location of books on a library shelf. Although a given book may cover many subjects, a classification system should be able to collocate books that are extremely similar, say books about occupational safety and health. The most serious liability with this evaluation, then, is the fact that we have compressed multiple subject headings, say prices : international into single subjects. This flattening obscures the multivalence of documents. We turn to a more realistic assessment of document-class relations in Section 6.2. 6.2 Accuracy of the Document Classifiers Although the keyword-based clusters appear to classify the Editor"s Desk documents very well, their discovery only solved half of the problem required for the successful implementation of a dynamic user interface such as the relation browser. The matter of roughly fourteen thousand unclassified documents remained to be addressed. To solve this problem, we trained the statistical classifiers described above in Section 5. For each document in the collection di, these classifiers give pi, a k-vector of probabilities or distances (depending on the classification method used), where pik quantifies the strength of association between the ith document and the kth class. All classifiers were trained on the full text of each document, regardless of the representation used to discover the initial clusters. The different training sets were thus constructed simply by changing the 156 Table 4: Cross Validation Results for 3 Classifiers Method Av. Percent Accuracy SE Prind 59.07 1.07 Naive Bayes 75.57 0.4 SVM 75.08 0.68 class variable for each instance (document) to reflect its assigned cluster under a given model. To test the ability of each classifier to locate documents correctly, we first performed a 10-fold cross validation on the Editor"s Desk documents. During cross-validation the data are split randomly into n subsets (in this case n = 10). The process proceeds by iteratively holding out each of the n subsets as a test collection for a model trained on the remaining n − 1 subsets. Cross validation is described in [15]. Using this methodology, we compared the performance of the three classification models described above. Table 4 gives the results from cross validation. Although naive Bayes is not significantly more accurate for these data than the SVM classifier, we limit the remainder of our attention to analysis of its performance. Our selection of naive Bayes is due to the fact that it appears to work comparably to the SVM approach for these data, while being much simpler, both in theory and implementation. Because we have only 1279 documents and 10 classes, the number of training documents per class is relatively small. In addition to models fitted to the Editor"s Desk data, then, we constructed a fourth model, supplementing the training sets of each class by querying the Google search engine7 and applying naive Bayes to the augmented test set. For each class, we created a query by submitting the three terms with the highest log-odds ratio with that class. Further, each query was limited to the domain www.bls.gov. For each class we retrieved up to 400 documents from Google (the actual number varied depending on the size of the result set returned by Google). This led to a training set of 4113 documents in the augmented model, as we call it below8 . Cross validation suggested that the augmented model decreased classification accuracy (accuracy= 58.16%, with standard error= 0.32). As we discuss below, however, augmenting the training set appeared to help generalization during our second experiment. The results of our cross validation experiment are encouraging. However, the success of our classifiers on the Editor"s Desk documents that informed the cross validation study may not be good predictors of the models" performance on the remainder to the BLS website. To test the generality of the naive Bayes classifier, we solicited input from 11 human judges who were familiar with the BLS website. The sample was chosen by convenience, and consisted of faculty and graduate students who work on the GovStat project. However, none of the reviewers had prior knowledge of the outcome of the classification before their participation. For the experiment, a random sample of 100 documents was drawn from the entire BLS collection. On average each re7 http://www.google.com 8 A more formal treatment of the combination of labeled and unlabeled data is available in [4]. Table 5: Human-Model Agreement on 100 Sample Docs. Human Judge 1st Choice Model Model 1st Choice Model 2nd Choice N. Bayes (aug.) 14 24 N. Bayes 24 1 Human Judge 2nd Choice Model Model 1st Choice Model 2nd Choice N. Bayes (aug.) 14 21 N. Bayes 21 4 viewer classified 83 documents, placing each document into as many of the categories shown in Table 3 as he or she saw fit. Results from this experiment suggest that room for improvement remains with respect to generalizing to the whole collection from the class models fitted to the Editor"s Desk documents. In Table 5, we see, for each classifier, the number of documents for which it"s first or second most probable class was voted best or second best by the 11 human judges. In the context of this experiment, we consider a first- or second-place classification by the machine to be accurate because the relation browser interface operates on a multiway classification, where each document is classified into multiple categories. Thus a document with the correct class as its second choice would still be easily available to a user. Likewise, a correct classification on either the most popular or second most popular category among the human judges is considered correct in cases where a given document was classified into multiple classes. There were 72 multiclass documents in our sample, as seen in Figure 4. The remaining 28 documents were assigned to 1 or 0 classes. Under this rationale, The augmented naive Bayes classifier correctly grouped 73 documents, while the smaller model (not augmented by a Google search) correctly classified 50. The resultant χ2 test gave p = 0.001, suggesting that increasing the training set improved the ability of the naive Bayes model to generalize from the Editor"s Desk documents to the collection as a whole. However, the improvement afforded by the augmented model comes at some cost. In particular, the augmented model is significantly inferior to the model trained solely on Editor"s Desk documents if we concern ourselves only with documents selected by the majority of human reviewers-i.e. only first-choice classes. Limiting the right answers to the left column of Table 5 gives p = 0.02 in favor of the non-augmented model. For the purposes of applying the relation browser to complex digital library content (where documents will be classified along multiple categories), the augmented model is preferable. But this is not necessarily the case in general. It must also be said that 73% accuracy under a fairly liberal test condition leaves room for improvement in our assignment of topics to categories. We may begin to understand the shortcomings of the described techniques by consulting Figure 5, which shows the distribution of categories across documents given by humans and by the augmented naive Bayes model. The majority of reviewers put 157 Number of Human-Assigned ClassesMNumber of Human-Assigned ClassesM Number of Human-Assigned Classes FrequencyMFrequencyM Frequency m00M0M 0 11M1M 1 22M2M 2 33M3M 3 44M4M 4 55M5M 5 66M6M 6 77M7M 7 m00M0M 055M5M 51010M10M 101515M15M 152020M20M 202525M25M 253030M30M 303535M35M 35 Figure 4: Number of Classes Assigned to Documents by Judges documents into only three categories, jobs, benefits, and occupations. On the other hand, the naive Bayes classifier distributed classes more evenly across the topics. This behavior suggests areas for future improvement. Most importantly, we observed a strong correlation among the three most frequent classes among the human judges (for instance, there was 68% correlation between benefits and occupations). This suggests that improving the clustering to produce topics that were more nearly orthogonal might improve performance. 7. CONCLUSIONS AND FUTURE WORK Many developers and maintainers of digital libraries share the basic problem pursued here. Given increasingly large, complex bodies of data, how may we improve access to collections without incurring extraordinary cost, and while also keeping systems receptive to changes in content over time? Data mining and machine learning methods hold a great deal of promise with respect to this problem. Empirical methods of knowledge discovery can aid in the organization and retrieval of information. As we have argued in this paper, these methods may also be brought to bear on the design and implementation of advanced user interfaces. This study explored a hybrid technique for aiding information architects as they implement dynamic interfaces such as the relation browser. Our approach combines unsupervised learning techniques, applied to a focused subset of the BLS website. The goal of this initial stage is to discover the most basic and far-reaching topics in the collection. Based mjobsjobsMjobsM jobs benefitsunemploymentpricespricesMpricesM prices safetyinternationalspendingspendingMspendingM spending occupationscostscostsMcostsM costs productivityHuman ClassificationsMHuman ClassificationsM Human Classifications m0.000.00M0.00M 0.00 0.050.100.150.15M0.15M 0.15 0.200.25mjobsjobsMjobsM jobs benefitsunemploymentpricespricesMpricesM prices safetyinternationalspendingspendingMspendingM spending occupationscostscostsMcostsM costs productivityMachine ClassificationsMMachine ClassificationsM Machine Classifications m0.000.00M0.00M 0.00 0.050.100.10M0.10M 0.10 0.15 Figure 5: Distribution of Classes Across Documents on a statistical model of these topics, the second phase of our approach uses supervised learning (in particular, a naive Bayes classifier, trained on individual words), to assign topical relations to the remaining documents in the collection. In the study reported here, this approach has demonstrated promise. In its favor, our approach is highly scalable. It also appears to give fairly good results. Comparing three modes of document representation-full-text, title only, and keyword-we found 98% accuracy as measured by collocation of documents with identical subject headings. While it is not surprising that editor-generated keywords should give strong evidence for such learning, their superiority over fulltext and titles was dramatic, suggesting that even a small amount of metadata can be very useful for data mining. However, we also found evidence that learning topics from a subset of the collection may lead to overfitted models. After clustering 1279 Editor"s Desk documents into 10 categories, we fitted a 10-way naive Bayes classifier to categorize the remaining 14,000 documents in the collection. While we saw fairly good results (classification accuracy of 75% with respect to a small sample of human judges), this experiment forced us to reconsider the quality of the topics learned by clustering. The high correlation among human judgments in our sample suggests that the topics discovered by analysis of the Editor"s Desk were not independent. While we do not desire mutually exclusive categories in our setting, we do desire independence among the topics we model. Overall, then, the techniques described here provide an encouraging start to our work on acquiring subject metadata for dynamic interfaces automatically. It also suggests that a more sophisticated modeling approach might yield 158 better results in the future. In upcoming work we will experiment with streamlining the two-phase technique described here. Instead of clustering documents to find topics and then fitting a model to the learned clusters, our goal is to expand the unsupervised portion of our analysis beyond a narrow subset of the collection, such as The Editor"s Desk. In current work we have defined algorithms to identify documents likely to help the topic discovery task. Supplied with a more comprehensive training set, we hope to experiment with model-based clustering, which combines the clustering and classification processes into a single modeling procedure. Topic discovery and document classification have long been recognized as fundamental problems in information retrieval and other forms of text mining. What is increasingly clear, however, as digital libraries grow in scope and complexity, is the applicability of these techniques to problems at the front-end of systems such as information architecture and interface design. Finally, then, in future work we will build on the user studies undertaken by Marchionini and Brunk in efforts to evaluate the utility of automatically populated dynamic interfaces for the users of digital libraries. 8. REFERENCES [1] A. Agresti. An Introduction to Categorical Data Analysis. Wiley, New York, 1996. [2] C. Ahlberg, C. Williamson, and B. Shneiderman. Dynamic queries for information exploration: an implementation and evaluation. In Proceedings of the SIGCHI conference on Human factors in computing systems, pages 619-626, 1992. [3] R. Baeza-Yates and B. Ribeiro-Neto. Modern Information Retrieval. ACM Press, 1999. [4] A. Blum and T. Mitchell. Combining labeled and unlabeled data with co-training. In Proceedings of the eleventh annual conference on Computational learning theory, pages 92-100. ACM Press, 1998. [5] H. Chen and S. Dumais. Hierarchical classification of web content. In Proceedings of the 23rd annual international ACM SIGIR conference on Research and development in information retrieval, pages 256-263, 2000. [6] M. Efron, G. Marchionini, and J. Zhang. Implications of the recursive representation problem for automatic concept identification in on-line governmental information. In Proceedings of the ASIST Special Interest Group on Classification Research (ASIST SIG-CR), 2003. [7] C. Fraley and A. E. Raftery. How many clusters? which clustering method? answers via model-based cluster analysis. The Computer Journal, 41(8):578-588, 1998. [8] A. K. Jain, M. N. Murty, and P. J. Flynn. Data clustering: a review. ACM Computing Surveys, 31(3):264-323, September 1999. [9] T. Joachims. A probabilistic analysis of the Rocchio algorithm with TFIDF for text categorization. In D. H. Fisher, editor, Proceedings of ICML-97, 14th International Conference on Machine Learning, pages 143-151, Nashville, US, 1997. Morgan Kaufmann Publishers, San Francisco, US. [10] T. Joachims. Text categorization with support vector machines: learning with many relevant features. In C. N´edellec and C. Rouveirol, editors, Proceedings of ECML-98, 10th European Conference on Machine Learning, pages 137-142, Chemnitz, DE, 1998. Springer Verlag, Heidelberg, DE. [11] I. T. Jolliffe. Principal Component Analysis. Springer, 2nd edition, 2002. [12] L. Kaufman and P. J. Rosseeuw. Finding Groups in Data: an Introduction to Cluster Analysis. Wiley, 1990. [13] G. Marchionini and B. Brunk. Toward a general relation browser: a GUI for information architects. Journal of Digital Information, 4(1), 2003. http://jodi.ecs.soton.ac.uk/Articles/v04/i01/Marchionini/. [14] A. K. McCallum. Bow: A toolkit for statistical language modeling, text retrieval, classification and clustering. http://www.cs.cmu.edu/˜mccallum/bow, 1996. [15] T. Mitchell. Machine Learning. McGraw Hill, 1997. [16] E. Rasmussen. Clustering algorithms. In W. B. Frakes and R. Baeza-Yates, editors, Information Retrieval: Data Structures and Algorithms, pages 419-442. Prentice Hall, 1992. [17] R. Tibshirani, G. Walther, and T. Hastie. Estimating the number of clusters in a dataset via the gap statistic, 2000. http://citeseer.nj.nec.com/tibshirani00estimating.html. [18] V. N. Vapnik. The Nature of Statistical Learning Theory. Springer, 2000. 159

machine learn;information architecture;interface design;multiway classification;access;bureau of labor statistics;bls collection;data-driven approach;digital library;k-means clustering;machine learning technique;eigenvalue;complex digital library;supervised and unsupervised learning technique

train_H-69

Ranking Web Objects from Multiple Communities

Vertical search is a promising direction as it leverages domainspecific knowledge and can provide more precise information for users. In this paper, we study the Web object-ranking problem, one of the key issues in building a vertical search engine. More specifically, we focus on this problem in cases when objects lack relationships between different Web communities, and take high-quality photo search as the test bed for this investigation. We proposed two score fusion methods that can automatically integrate as many Web communities (Web forums) with rating information as possible. The proposed fusion methods leverage the hidden links discovered by a duplicate photo detection algorithm, and aims at minimizing score differences of duplicate photos in different forums. Both intermediate results and user studies show the proposed fusion methods are practical and efficient solutions to Web object ranking in cases we have described. Though the experiments were conducted on high-quality photo ranking, the proposed algorithms are also applicable to other ranking problems, such as movie ranking and music ranking.

1. INTRODUCTION Despite numerous refinements and optimizations, general purpose search engines still fail to find relevant results for many queries. As a new trend, vertical search has shown promise because it can leverage domain-specific knowledge and is more effective in connecting users with the information they want. There are many vertical search engines, including some for paper search (e.g. Libra [21], Citeseer [7] and Google Scholar [4]), product search (e.g. Froogle [5]), movie search [6], image search [1, 8], video search [6], local search [2], as well as news search [3]. We believe the vertical search engine trend will continue to grow. Essentially, building vertical search engines includes data crawling, information extraction, object identification and integration, and object-level Web information retrieval (or Web object ranking) [20], among which ranking is one of the most important factors. This is because it deals with the core problem of how to combine and rank objects coming from multiple communities. Although object-level ranking has been well studied in building vertical search engines, there are still some kinds of vertical domains in which objects cannot be effectively ranked. For example, algorithms that evolved from PageRank [22], PopRank [21] and LinkFusion [27] were proposed to rank objects coming from multiple communities, but can only work on well-defined graphs of heterogeneous data. Well-defined means that like objects (e.g. authors in paper search) can be identified in multiple communities (e.g. conferences). This allows heterogeneous objects to be well linked to form a graph through leveraging all the relationships (e.g. cited-by, authored-by and published-by) among the multiple communities. However, this assumption does not always stand for some domains. High-quality photo search, movie search and news search are exceptions. For example, a photograph forum 377 website usually includes three kinds of objects: photos, authors and reviewers. Yet different photo forums seem to lack any relationships, as there are no cited-by relationships. This makes it difficult to judge whether two authors cited are the same author, or two photos are indeed identical photos. Consequently, although each photo has a rating score in a forum, it is non-trivial to rank photos coming from different photo forums. Similar problems also exist in movie search and news search. Although two movie titles can be identified as the same one by title and director in different movie discussion groups, it is non-trivial to combine rating scores from different discussion groups and rank movies effectively. We call such non-trivial object relationship in which identification is difficult, incomplete relationships. Other related work includes rank aggregation for the Web [13, 14], and learning algorithm for rank, such as RankBoost [15], RankSVM [17, 19], and RankNet [12]. We will contrast differences of these methods with the proposed methods after we have described the problem and our methods. We will specifically focus on Web object-ranking problem in cases that lack object relationships or have with incomplete object relationships, and take high-quality photo search as the test bed for this investigation. In the following, we will introduce rationale for building high-quality photo search. 1.1 High-Quality Photo Search In the past ten years, the Internet has grown to become an incredible resource, allowing users to easily access a huge number of images. However, compared to the more than 1 billion images indexed by commercial search engines, actual queries submitted to image search engines are relatively minor, and occupy only 8-10 percent of total image and text queries submitted to commercial search engines [24]. This is partially because user requirements for image search are far less than those for general text search. On the other hand, current commercial search engines still cannot well meet various user requirements, because there is no effective and practical solution to understand image content. To better understand user needs in image search, we conducted a query log analysis based on a commercial search engine. The result shows that more than 20% of image search queries are related to nature and places and daily life categories. Users apparently are interested in enjoying high-quality photos or searching for beautiful images of locations or other kinds. However, such user needs are not well supported by current image search engines because of the difficulty of the quality assessment problem. Ideally, the most critical part of a search engine - the ranking function - can be simplified as consisting of two key factors: relevance and quality. For the relevance factor, search in current commercial image search engines provide most returned images that are quite relevant to queries, except for some ambiguity. However, as to quality factor, there is still no way to give an optimal rank to an image. Though content-based image quality assessment has been investigated over many years [23, 25, 26], it is still far from ready to provide a realistic quality measure in the immediate future. Seemingly, it really looks pessimistic to build an image search engine that can fulfill the potentially large requirement of enjoying high-quality photos. Various proliferating Web communities, however, notices us that people today have created and shared a lot of high-quality photos on the Web on virtually any topics, which provide a rich source for building a better image search engine. In general, photos from various photo forums are of higher quality than personal photos, and are also much more appealing to public users than personal photos. In addition, photos uploaded to photo forums generally require rich metadata about title, camera setting, category and description to be provide by photographers. These metadata are actually the most precise descriptions for photos and undoubtedly can be indexed to help search engines find relevant results. More important, there are volunteer users in Web communities actively providing valuable ratings for these photos. The rating information is generally of great value in solving the photo quality ranking problem. Motivated by such observations, we have been attempting to build a vertical photo search engine by extracting rich metadata and integrating information form various photo Web forums. In this paper, we specifically focus on how to rank photos from multiple Web forums. Intuitively, the rating scores from different photo forums can be empirically normalized based on the number of photos and the number of users in each forum. However, such a straightforward approach usually requires large manual effort in both tedious parameter tuning and subjective results evaluation, which makes it impractical when there are tens or hundreds of photo forums to combine. To address this problem, we seek to build relationships/links between different photo forums. That is, we first adopt an efficient algorithm to find duplicate photos which can be considered as hidden links connecting multiple forums. We then formulate the ranking challenge as an optimization problem, which eventually results in an optimal ranking function. 1.2 Main Contributions and Organization. The main contributions of this paper are: 1. We have proposed and built a vertical image search engine by leveraging rich metadata from various photo forum Web sites to meet user requirements of searching for and enjoying high-quality photos, which is impossible in traditional image search engines. 2. We have proposed two kinds of Web object-ranking algorithms for photos with incomplete relationships, which can automatically and efficiently integrate as many as possible Web communities with rating information and achieves an equal qualitative result compared with the manually tuned fusion scheme. The rest of this paper is organized as follows. In Section 2, we present in detail the proposed solutions to the ranking problem, including how to find hidden links between different forums, normalize rating scores, obtain the optimal ranking function, and contrast our methods with some other related research. In Section 3, we describe the experimental setting and experiments and user studies conducted to evaluate our algorithm. Our conclusion and a discussion of future work is in Section 4. It is worth noting that although we treat vertical photo search as the test bed in this paper, the proposed ranking algorithm can also be applied to rank other content that includes video clips, poems, short stories, drawings, sculptures, music, and so on. 378 2. ALGORITHM 2.1 Overview The difficulty of integrating multiple Web forums is in their different rating systems, where there are generally two kinds of freedom. The first kind of freedom is the rating interval or rating scale including the minimal and maximal ratings for each Web object. For example, some forums use a 5-point rating scale whereas other forums use 3-point or 10-point rating scales. It seems easy to fix this freedom, but detailed analysis of the data and experiments show that it is a non-trivial problem. The second kind of freedom is the varying rating criteria found in different Web forums. That is, the same score does not mean the same quality in different forums. Intuitively, if we can detect same photographers or same photographs, we can build relationships between any two photo forums and therefore can standardize the rating criterion by score normalization and transformation. Fortunately, we find that quite a number of duplicate photographs exist in various Web photo forums. This fact is reasonable when considering that photographers sometimes submit a photo to more than one forum to obtain critiques or in hopes of widespread publicity. In this work, we adopt an efficient duplicate photo detection algorithm [10] to find these photos. The proposed methods below are based on the following considerations. Faced with the need to overcome a ranking problem, a standardized rating criterion rather than a reasonable rating criterion is needed. Therefore, we can take a large scale forum as the reference forum, and align other forums by taking into account duplicate Web objects (duplicate photos in this work). Ideally, the scores of duplicate photos should be equal even though they are in different forums. Yet we can deem that scores in different forumsexcept for the reference forum - can vary in a parametric space. This can be determined by minimizing the objective function defined by the sum of squares of the score differences. By formulating the ranking problem as an optimization problem that attempts to make the scores of duplicate photos in non-reference forums as close as possible to those in the reference forum, we can effectively solve the ranking problem. For convenience, the following notations are employed. Ski and ¯Ski denote the total score and mean score of ith Web object (photo) in the kth Web site, respectively. The total score refers to the sum of the various rating scores (e.g., novelty rating and aesthetic rating), and the mean score refers to the mean of the various rating scores. Suppose there are a total of K Web sites. We further use {Skl i |i = 1, ..., Ikl; k, l = 1, ..., K; k = l} to denote the set of scores for Web objects (photos) in kth Web forums that are duplicate with the lth Web forums, where Ikl is the total number of duplicate Web objects between these two Web sites. In general, score fusion can be seen as the procedure of finding K transforms ψk(Ski) = eSki, k = 1, ..., K such that eSki can be used to rank Web objects from different Web sites. The objective function described in the above Figure 1: Web community integration. Each Web community forms a subgraph, and all communities are linked together by some hidden links (dashed lines). paragraph can then be formulated as min {ψk|k=2,...,K} KX k=2 Ik1X i=1 ¯wk i S1k i − ψk(Sk1 i ) 2 (1) where we use k = 1 as the reference forum and thus ψ1(S1i) = S1i. ¯wk i (≥ 0) is the weight coefficient that can be set heuristically according to the numbers of voters (reviewers or commenters) in both the reference forum and the non-reference forum. The more reviewers, the more popular the photo is and the larger the corresponding weight ¯wk i should be. In this work, we do not inspect the problem of how to choose ¯wk i and simply set them to one. But we believe the proper use of ¯wk i , which leverages more information, can significantly improve the results. Figure 1 illustrates the aforementioned idea. The Web Community 1 is the reference community. The dashed lines are links indicating that the two linked Web objects are actually the same. The proposed algorithm will try to find the best ψk(k = 2, ..., K), which has certain parametric forms according to certain models. So as to minimize the cost function defined in Eq. 1, the summation is taken on all the red dashed lines. We will first discuss the score normalization methods in Section 2.2, which serves as the basis for the following work. Before we describe the proposed ranking algorithms, we first introduce a manually tuned method in Section 2.3, which is laborious and even impractical when the number of communities become large. In Section 2.4, we will briefly explain how to precisely find duplicate photos between Web forums. Then we will describe the two proposed methods: Linear fusion and Non-linear fusion, and a performance measure for result evaluation in Section 2.5. Finally, in Section 2.6 we will discuss the relationship of the proposed methods with some other related work. 2.2 Score Normalization Since different Web (photo) forums on the Web usually have different rating criteria, it is necessary to normalize them before applying different kinds of fusion methods. In addition, as there are many kinds of ratings, such as ratings for novelty, ratings for aesthetics etc, it is reasonable to choose a common one - total score or average scorethat can always be extracted in any Web forum or calculated by corresponding ratings. This allows the normaliza379 tion method on the total score or average score to be viewed as an impartial rating method between different Web forums. It is straightforward to normalize average scores by linearly transforming them to a fixed interval. We call this kind of score as Scaled Mean Score. The difficulty, however, of using this normalization method is that, if there are only a few users rating an object, say a photo in a photo forum, the average score for the object is likely to be spammed or skewed. Total score can avoid such drawbacks that contain more information such as a Web object"s quality and popularity. The problem is thus how to normalize total scores in different Web forums. The simplest way may be normalization by the maximal and minimal scores. The drawback of this normalization method is it is non robust, or in other words, it is sensitive to outliers. To make the normalization insensitive to unusual data, we propose the Mode-90% Percentile normalization method. Here, the mode score represents the total score that has been assigned to more photos than any other total score. And The high percentile score (e.g.,90%) represents the total score for which the high percentile of images have a lower total score. This normalization method utilizes the mode and 90% percentile as two reference points to align two rating systems, which makes the distributions of total scores in different forums more consistent. The underlying assumption, for example in different photo forums, is that even the qualities of top photos in different forums may vary greatly and be less dependent on the forum quality, the distribution of photos of middle-level quality (from mode to 90% percentile) should be almost of the same quality up to the freedom which reflects the rating criterion (strictness) of Web forums. Photos of this middle-level in a Web forum usually occupy more than 70 % of total photos in that forum. We will give more detailed analysis of the scores in Section 3.2. 2.3 Manual Fusion The Web movie forum, IMDB [16], proposed to use a Bayesian-ranking function to normalize rating scores within one community. Motivated by this ranking function, we propose this manual fusion method: For the kth Web site, we use the following formula eSki = αk · „ nk · ¯Ski nk + n∗ k + n∗ k · S∗ k nk + n∗ k « (2) to rank photos, where nk is the number of votes and n∗ k, S∗ k and αk are three parameters. This ranking function first takes a balance between the original mean score ¯Ski and a reference score S∗ k to get a weighted mean score which may be more reliable than ¯Ski. Then the weighted mean score is scaled by αk to get the final score fSki. For n Web communities, there are then about 3n parameters in {(αk, n∗ k, S∗ k)|k = 1, ..., n} to tune. Though this method can achieves pretty good results after careful and thorough manual tuning on these parameters, when n becomes increasingly large, say there are tens or hundreds of Web communities crawled and indexed, this method will become more and more laborious and will eventually become impractical. It is therefore desirable to find an effective fusion method whose parameters can be automatically determined. 2.4 Duplicate Photo Detection We use Dedup [10], an efficient and effective duplicate image detection algorithm, to find duplicate photos between any two photo forums. This algorithm uses hash function to map a high dimensional feature to a 32 bits hash code (see below for how to construct the hash code). Its computational complexity to find all the duplicate images among n images is about O(n log n). The low-level visual feature for each photo is extracted on k × k regular grids. Based on all features extracted from the image database, a PCA model is built. The visual features are then transformed to a relatively low-dimensional and zero mean PCA space, or 29 dimensions in our system. Then the hash code for each photo is built as follows: each dimension is transformed to one, if the value in this dimension is greater than 0, and 0 otherwise. Photos in the same bucket are deemed potential duplicates and are further filtered by a threshold in terms of Euclidean similarity in the visual feature space. Figure 2 illustrates the hashing procedure, where visual features - mean gray values - are extracted on both 6 × 6 and 7×7 grids. The 85-dimensional features are transformed to a 32-dimensional vector, and the hash code is generated according to the signs. Figure 2: Hashing procedure for duplicate photo dectection 2.5 Score Fusion In this section, we will present two solutions on score fusion based on different parametric form assumptions of ψk in Eq. 1. 2.5.1 Linear Fusion by Duplicate Photos Intuitively, the most straightforward way to factor out the uncertainties caused by the different criterion is to scale, rel380 ative to a given center, the total scores of each unreferenced Web photo forum with respect to the reference forum. More strictly, we assume ψk has the following form ψk(Ski) = αkSki + tk, k = 2, ..., K (3) ψ1(S1i) = S1i (4) which means that the scores of k(= 1)th forum should be scaled by αk relative to the center tk 1−αk as shown in Figure 3. Then, if we substitute above ψk to Eq. 1, we get the following objective function, min {αk,tk|k=2,...,K} KX k=2 Ik1X i=1 ¯wk i h S1k i − αkSk1 i − tk i2 . (5) By solving the following set of functions, ( ∂f ∂αk = = 0 ∂f ∂tk = 0 , k = 1, ..., K where f is the objective function defined in Eq. 5, we get the closed form solution as: „ αk tk « = A−1 k Lk (6) where Ak = „ P i ¯wi(Sk1 i )2 P i ¯wiSk1 iP i ¯wiSk1 i P i ¯wi « (7) Lk = „ P i ¯wiS1k i Sk1 iP i ¯wiS1k i « (8) and k = 2, ..., K. This is a linear fusion method. It enjoys simplicity and excellent performance in the following experiments. Figure 3: Linear Fusion method 2.5.2 Nonlinear Fusion by Duplicate Photos Sometimes we want a method which can adjust scores on intervals with two endpoints unchanged. As illustrated in Figure 4, the method can tune scores between [C0, C1] while leaving scores C0 and C1 unchanged. This kind of fusion method is then much finer than the linear ones and contains many more parameters to tune and expect to further improve the results. Here, we propose a nonlinear fusion solution to satisfy such constraints. First, we introduce a transform: ηc0,c1,α(x) = ( x−c0 c1−c0 α (c1 − c0) + c0, if x ∈ (c0, c1] x otherwise where α > 0. This transform satisfies that for x ∈ [c0, c1], ηc0,c1,α(x) ∈ [c0, c1] with ηc0,c1,α(c0) = c0 and ηc0,c1,α(c1) = c1. Then we can utilize this nonlinear transform to adjust the scores in certain interval, say (M, T], ψk(Ski) = ηM,T,α(Ski) . (9) Figure 4: Nonlinear Fusion method. We intent to finely adjust the shape of the curves in each segment. Even there is no closed-form solution for the following optimization problem, min {αk|k∈[2,K]} KX k=2 Ik1X i=1 ¯wk i h S1k i − ηM,T,α(Ski) i2 it is not hard to get the numeric one. Under the same assumptions made in Section 2.2, we can use this method to adjust scores of the middle-level (from the mode point to the 90 % percentile). This more complicated non-linear fusion method is expected to achieve better results than the linear one. However, difficulties in evaluating the rank results block us from tuning these parameters extensively. The current experiments in Section 3.5 do not reveal any advantages over the simple linear model. 2.5.3 Performance Measure of the Fusion Results Since our objective function is to make the scores of the same Web objects (e.g. duplicate photos) between a nonreference forum and the reference forum as close as possible, it is natural to investigate how close they become to each other and how the scores of the same Web objects change between the two non-reference forums before and after score fusion. Taken Figure 1 as an example, the proposed algorithms minimize the score differences of the same Web objects in two Web forums: the reference forum (the Web Community 1) and a non-reference forum, which corresponds to minimizing the objective function on the red dashed (hidden) links. After the optimization, we must ask what happens to the score differences of the same Web objects in two nonreference forums? Or, in other words, whether the scores of two objects linked by the green dashed (hidden) links become more consistent? We therefore define the following performance measureδ measure - to quantify the changes for scores of the same Web objects in different Web forums as δkl = Sim(Slk , Skl ) − Sim(Slk ∗ , Skl ∗ ) (10) 381 where Skl = (Skl 1 , ..., Skl Ikl )T , Skl ∗ = (eSkl 1 , ..., eSkl Ikl )T and Sim(a, b) = a · b ||a||||b|| . δkl > 0 means after score fusion, scores on the same Web objects between kth and lth Web forum become more consistent, which is what we expect. On the contrary, if δkl < 0, those scores become more inconsistent. Although we cannot rely on this measure to evaluate our final fusion results as ranking photos by their popularity and qualities is such a subjective process that every person can have its own results, it can help us understand the intermediate ranking results and provide insights into the final performances of different ranking methods. 2.6 Contrasts with Other Related Work We have already mentioned the differences of the proposed methods with the traditional methods, such as PageRank [22], PopRank [21], and LinkFusion [27] algorithms in Section 1. Here, we discuss some other related works. The current problem can also be viewed as a rank aggregation one [13, 14] as we deal with the problem of how to combine several rank lists. However, there are fundamental differences between them. First of all, unlike the Web pages, which can be easily and accurately detected as the same pages, detecting the same photos in different Web forums is a non-trivial work, and can only be implemented by some delicate algorithms while with certain precision and recall. Second, the numbers of the duplicate photos from different Web forums are small relative to the whole photo sets (see Table 1). In another words, the top K rank lists of different Web forums are almost disjointed for a given query. Under this condition, both the algorithms proposed in [13] and their measurements - Kendall tau distance or Spearman footrule distance - will degenerate to some trivial cases. Another category of rank fusion (aggregation) methods is based on machine learning algorithms, such as RankSVM [17, 19], RankBoost [15], and RankNet [12]. All of these methods entail some labelled datasets to train a model. In current settings, it is difficult or even impossible to get these datasets labelled as to their level of professionalism or popularity, since the photos are too vague and subjective to rank. Instead, the problem here is how to combine several ordered sub lists to form a total order list. 3. EXPERIMENTS In this section, we carry out our research on high-quality photo search. We first briefly introduce the newly proposed vertical image search engine - EnjoyPhoto in section 3.1. Then we focus on how to rank photos from different Web forums. In order to do so, we first normalize the scores (ratings) for photos from different multiple Web forums in section 3.2. Then we try to find duplicate photos in section 3.3. Some intermediate results are discussed using δ measure in section 3.4. Finally a set of user studies is carried out carefully to justify our proposed method in section 3.5. 3.1 EnjoyPhoto: high-quality Photo Search Engine In order to meet user requirement of enjoying high-quality photos, we propose and build a high-quality photo search engine - EnjoyPhoto, which accounts for the following three key issues: 1. how to crawl and index photos, 2. how to determine the qualities of each photo and 3. how to display the search results in order to make the search process enjoyable. For a given text based query, this system ranks the photos based on certain combination of relevance of the photo to this query (Issue 1) and the quality of the photo (Issue 2), and finally displays them in an enjoyable manner (Issue 3). As for Issue 3, we devise the interface of the system deliberately in order to smooth the users" process of enjoying high-quality photos. Techniques, such as Fisheye and slides show, are utilized in current system. Figure 5 shows the interface. We will not talk more about this issue as it is not an emphasis of this paper. Figure 5: EnjoyPhoto: an enjoyable high-quality photo search engine, where 26,477 records are returned for the query fall in about 0.421 seconds As for Issue 1, we extracted from a commercial search engine a subset of photos coming from various photo forums all over the world, and explicitly parsed the Web pages containing these photos. The number of photos in the data collection is about 2.5 million. After the parsing, each photo was associated with its title, category, description, camera setting, EXIF data 1 (when available for digital images), location (when available in some photo forums), and many kinds of ratings. All these metadata are generally precise descriptions or annotations for the image content, which are then indexed by general text-based search technologies [9, 18, 11]. In current system, the ranking function was specifically tuned to emphasize title, categorization, and rating information. Issue 2 is essentially dealt with in the following sections which derive the quality of photos by analyzing ratings provided by various Web photo forums. Here we chose six photo forums to study the ranking problem and denote them as Web-A, Web-B, Web-C, Web-D, Web-E and Web-F. 3.2 Photo Score Normalization Detailed analysis of different score normalization methods are analyzed in this section. In this analysis, the zero 1 Digital cameras save JPEG (.jpg) files with EXIF (Exchangeable Image File) data. Camera settings and scene information are recorded by the camera into the image file. www.digicamhelp.com/what-is-exif/ 382 0 2 4 6 8 10 0 1000 2000 3000 4000 Normalized Score TotalNumber (a) Web-A 0 2 4 6 8 10 0 0.5 1 1.5 2 2.5 3 x 10 4 Normalized Score TotalNumber (b) Web-B 0 2 4 6 8 10 0 0.5 1 1.5 2 x 10 5 Normalized Score TotalNumber (c) Web-C 0 2 4 6 8 10 0 2 4 6 8 10 x 10 4 Normalized Score TotalNumber (d) Web-D 0 2 4 6 8 10 0 2000 4000 6000 8000 10000 12000 14000 Normalized Score TotalNumber (e) Web-E 0 2 4 6 8 10 0 1 2 3 4 5 6 x 10 4 Normalized Score TotalNumber (f) Web-F Figure 6: Distributions of mean scores normalized to [0, 10] scores that usually occupy about than 30% of the total number of photos for some Web forums are not currently taken into account. How to utilize these photos is left for future explorations. In Figure 6, we list the distributions of the mean score, which is transformed to a fixed interval [0, 10]. The distributions of the average scores of these Web forums look quite different. Distributions in Figure 6(a), 6(b), and 6(e) looks like Gaussian distributions, while those in Figure 6(d) and 6(f) are dominated by the top score. The reason of these eccentric distributions for Web-D and Web-F lies in their coarse rating systems. In fact, Web-D and Web-F use 2 or 3 point rating scales whereas other Web forums use 7 or 14 point rating scales. Therefore, it will be problematic if we directly use these averaged scores. Furthermore the average score is very likely to be spammed, if there are only a few users rating a photo. Figure 7 shows the total score normalization method by maximal and minimal scores, which is one of our base line system. All the total scores of a given Web forum are normalized to [0, 100] according to the maximal score and minimal score of corresponding Web forum. We notice that total score distribution of Web-A in Figure 7(a) has two larger tails than all the others. To show the shape of the distributions more clearly, we only show the distributions on [0, 25] in Figure 7(b),7(c),7(d),7(e), and 7(f). Figure 8 shows the Mode-90% Percentile normalization method, where the modes of the six distributions are normalized to 5 and the 90% percentile to 8. We can see that this normalization method makes the distributions of total scores in different forums more consistent. The two proposed algorithms are all based on these normalization methods. 3.3 Duplicate photo detection Targeting at computational efficiency, the Dedup algorithm may lose some recall rate, but can achieve a high precision rate. We also focus on finding precise hidden links rather than all hidden links. Figure 9 shows some duplicate detection examples. The results are shown in Table 1 and verify that large numbers of duplicate photos exist in any two Web forums even with the strict condition for Dedup where we chose first 29 bits as the hash code. Since there are only a few parameters to estimate in the proposed fusion methods, the numbers of duplicate photos shown Table 1 are 0 20 40 60 80 100 0 100 200 300 400 500 600 Normalized Score TotalNumber (a) Web-A 0 5 10 15 20 25 0 1 2 3 4 5 x 10 4 Normalized Score TotalNumber (b) Web-B 0 5 10 15 20 25 0 1 2 3 4 5 x 10 5 Normalized Score TotalNumber (c) Web-C 0 5 10 15 20 25 0 0.5 1 1.5 2 2.5 x 10 4 Normalized Score TotalNumber (d) Web-D 0 5 10 15 20 25 0 2000 4000 6000 8000 10000 Normalized Score TotalNumber (e) Web-E 0 5 10 15 20 25 0 0.5 1 1.5 2 2.5 3 x 10 4 Normalized Score TotalNumber (f) Web-F Figure 7: Maxmin Normalization 0 5 10 15 0 200 400 600 800 1000 1200 1400 Normalized Score TotalNumber (a) Web-A 0 5 10 15 0 1 2 3 4 5 x 10 4 Normalized Score TotalNumber (b) Web-B 0 5 10 15 0 2 4 6 8 10 12 14 x 10 4 Normalized Score TotalNumber (c) Web-C 0 5 10 15 0 0.5 1 1.5 2 2.5 x 10 4 Normalized Score TotalNumber (d) Web-D 0 5 10 15 0 2000 4000 6000 8000 10000 12000 Normalized Score TotalNumber (e) Web-E 0 5 10 15 0 2000 4000 6000 8000 10000 Normalized Score TotalNumber (f) Web-F Figure 8: Mode-90% Percentile Normalization sufficient to determine these parameters. The last table column lists the total number of photos in the corresponding Web forums. 3.4 δ Measure The parameters of the proposed linear and nonlinear algorithms are calculated using the duplicate data shown in Table 1, where the Web-C is chosen as the reference Web forum since it shares the most duplicate photos with other forums. Table 2 and 3 show the δ measure on the linear model and nonlinear model. As δkl is symmetric and δkk = 0, we only show the upper triangular part. The NaN values in both tables lie in that no duplicate photos have been detected by the Dedup algorithm as reported in Table 1. The linear model guarantees that the δ measures related Table 1: Number of duplicate photos between each pair of Web forums A B C D E F Scale A 0 316 1,386 178 302 0 130k B 316 0 14,708 909 8,023 348 675k C 1,386 14,708 0 1,508 19,271 1,083 1,003k D 178 909 1,508 0 1,084 21 155k E 302 8,023 19,271 1,084 0 98 448k F 0 348 1,083 21 98 0 122k 383 Figure 9: Some results of duplicate photo detection Table 2: δ measure on the linear model. Web-B Web-C Web-D Web-E Web-F Web-A 0.0659 0.0911 0.0956 0.0928 NaN Web-B - 0.0672 0.0578 0.0791 0.4618 Web-C - - 0.0105 0.0070 0.2220 Web-D - - - 0.0566 0.0232 Web-E - - - - 0.6525 to the reference community should be no less than 0 theoretically. It is indeed the case (see the underlined numbers in Table 2). But this model can not guarantee that the δ measures on the non-reference communities can also be no less than 0, as the normalization steps are based on duplicate photos between the reference community and a nonreference community. Results shows that all the numbers in the δ measure are greater than 0 (see all the non-underlined numbers in Table 2), which indicates that it is probable that this model will give optimal results. On the contrary, the nonlinear model does not guarantee that δ measures related to the reference community should be no less than 0, as not all duplicate photos between the two Web forums can be used when optimizing this model. In fact, the duplicate photos that lie in different intervals will not be used in this model. It is these specific duplicate photos that make the δ measure negative. As a result, there are both negative and positive items in Table 3, but overall the number of positive ones are greater than negative ones (9:5), that indicates the model may be better than the normalization only method (see next subsection) which has an all-zero δ measure, and worse than the linear model. 3.5 User Study Because it is hard to find an objective criterion to evaluate Table 3: δ measure on the nonlinear model. Web-B Web-C Web-D Web-E Web-F Web-A 0.0559 0.0054 -0.0185 -0.0054 NaN Web-B - -0.0162 -0.0345 -0.0301 0.0466 Web-C - - 0.0136 0.0071 0.1264 Web-D - - - 0.0032 0.0143 Web-E - - - - 0.214 which ranking function is better, we chose to employ user studies for subjective evaluations. Ten subjects were invited to participate in the user study. They were recruited from nearby universities. As search engines of both text search and image search are familiar to university students, there was no prerequisite criterion for choosing students. We conducted user studies using Internet Explorer 6.0 on Windows XP with 17-inch LCD monitors set at 1,280 pixels by 1,024 pixels in 32-bit color. Data was recorded with server logs and paper-based surveys after each task. Figure 10: User study interface We specifically device an interface for user study as shown in Figure 10. For each pair of fusion methods, participants were encouraged to try any query they wished. For those without specific ideas, two combo boxes (category list and query list) were listed on the bottom panel, where the top 1,000 image search queries from a commercial search engine were provided. After a participant submitted a query, the system randomly selected the left or right frame to display each of the two ranking results. The participant were then required to judge which ranking result was better of the two ranking results, or whether the two ranking results were of equal quality, and submit the judgment by choosing the corresponding radio button and clicking the Submit button. For example, in Figure 10, query sunset is submitted to the system. Then, 79,092 photos were returned and ranked by the Minmax fusion method in the left frame and linear fusion method in the right frame. A participant then compares the two ranking results (without knowing the ranking methods) and submits his/her feedback by choosing answers in the Your option. Table 4: Results of user study Norm.Only Manually Linear Linear 29:13:10 14:22:15Nonlinear 29:15:9 12:27:12 6:4:45 Table 4 shows the experimental results, where Linear denotes the linear fusion method, Nonlinear denotes the non linear fusion method, Norm. Only means Maxmin normalization method, Manually means the manually tuned method. The three numbers in each item, say 29:13:10, mean that 29 judgments prefer the linear fusion results, 10 384 judgments prefer the normalization only method, and 13 judgments consider these two methods as equivalent. We conduct the ANOVA analysis, and obtain the following conclusions: 1. Both the linear and nonlinear methods are significantly better than the Norm. Only method with respective P-values 0.00165(< 0.05) and 0.00073(<< 0.05). This result is consistent with the δ-measure evaluation result. The Norm. Only method assumes that the top 10% photos in different forums are of the same quality. However, this assumption does not stand in general. For example, a top 10% photo in a top tier photo forum is generally of higher quality than a top 10% photo in a second-tier photo forum. This is similar to that, those top 10% students in a top-tier university and those in a second-tier university are generally of different quality. Both linear and nonlinear fusion methods acknowledge the existence of such differences and aim at quantizing the differences. Therefore, they perform better than the Norm. Only method. 2. The linear fusion method is significantly better than the nonlinear one with P-value 1.195 × 10−10 . This result is rather surprising as this more complicated ranking method is expected to tune the ranking more finely than the linear one. The main reason for this result may be that it is difficult to find the best intervals where the nonlinear tuning should be carried out and yet simply the middle part of the Mode-90% Percentile Normalization method was chosen. The timeconsuming and subjective evaluation methods - user studies - blocked us extensively tuning these parameters. 3. The proposed linear and nonlinear methods perform almost the same with or slightly better than the manually tuned method. Given that the linear/nonlinear fusion methods are fully automatic approaches, they are considered practical and efficient solutions when more communities (e.g. dozens of communities) need to be integrated. 4. CONCLUSIONS AND FUTURE WORK In this paper, we studied the Web object-ranking problem in the cases of lacking object relationships where traditional ranking algorithms are no longer valid, and took high-quality photo search as the test bed for this investigation. We have built a vertical high-quality photo search engine, and proposed score fusion methods which can automatically integrate as many data sources (Web forums) as possible. The proposed fusion methods leverage the hidden links discovered by duplicate photo detection algorithm, and minimize score differences of duplicate photos in different forums. Both the intermediate results and the user studies show that the proposed fusion methods are a practical and efficient solution to Web object ranking in the aforesaid relationships. Though the experiments were conducted on high-quality photo ranking, the proposed algorithms are also applicable to other kinds of Web objects including video clips, poems, short stories, music, drawings, sculptures, and so on. Current system is far from being perfect. In order to make this system more effective, more delicate analysis for the vertical domain (e.g., Web photo forums) are needed. The following points, for example, may improve the searching results and will be our future work: 1. more subtle analysis and then utilization of different kinds of ratings (e.g., novelty ratings, aesthetic ratings); 2. differentiating various communities who may have different interests and preferences or even distinct culture understandings; 3. incorporating more useful information, including photographers" and reviewers" information, to model the photos in a heterogeneous data space instead of the current homogeneous one. We will further utilize collaborative filtering to recommend relevant high-quality photos to browsers. One open problem is whether we can find an objective and efficient criterion for evaluating the ranking results, instead of employing subjective and inefficient user studies, which blocked us from trying more ranking algorithms and tuning parameters in one algorithm. 5. ACKNOWLEDGMENTS We thank Bin Wang and Zhi Wei Li for providing Dedup codes to detect duplicate photos; Zhen Li for helping us design the interface of EnjoyPhoto; Ming Jing Li, Longbin Chen, Changhu Wang, Yuanhao Chen, and Li Zhuang etc. for useful discussions. Special thanks go to Dwight Daniels for helping us revise the language of this paper. 6. REFERENCES [1] Google image search. http://images.google.com. [2] Google local search. http://local.google.com/. [3] Google news search. http://news.google.com. [4] Google paper search. http://Scholar.google.com. [5] Google product search. http://froogle.google.com. [6] Google video search. http://video.google.com. [7] Scientific literature digital library. http://citeseer.ist.psu.edu. [8] Yahoo image search. http://images.yahoo.com. [9] R. Baeza-Yates and B. Ribeiro-Neto. Modern Information Retrieval. New York: ACM Press; Harlow, England: Addison-Wesley, 1999. [10] W. Bin, L. Zhiwei, L. Ming Jing, and M. Wei-Ying. Large-scale duplicate detection for web image search. In Proceedings of the International Conference on Multimedia and Expo, page 353, 2006. [11] S. Brin and L. Page. The anatomy of a large-scale hypertextual web search engine. In Computer Networks, volume 30, pages 107-117, 1998. [12] C. Burges, T. Shaked, E. Renshaw, A. Lazier, M. Deeds, N. Hamilton, and G. Hullender. Learning to rank using gradient descent. In Proceedings of the 22nd international conference on Machine learning, pages 89 - 96, 2005. [13] C. Dwork, R. Kumar, M. Naor, and D. Sivakumar. Rank aggregation methods for the web. In Proceedings 10th International Conference on World Wide Web, pages 613 - 622, Hong-Kong, 2001. [14] R. Fagin, R. Kumar, and D. Sivakumar. Comparing top k lists. SIAM Journal on Discrete Mathematics, 17(1):134 - 160, 2003. [15] Y. Freund, R. Iyer, R. E. Schapire, and Y. Singer. An efficient boosting algorithm for combining preferences. 385 Journal of Machine Learning Research, 4(1):933-969(37), 2004. [16] IMDB. Formula for calculating the top rated 250 titles in imdb. http://www.imdb.com/chart/top. [17] T. Joachims. Optimizing search engines using clickthrough data. In Proceedings of the eighth ACM SIGKDD international conference on Knowledge discovery and data mining, pages 133 - 142, 2002. [18] J. M. Kleinberg. Authoritative sources in a hyperlinked environment. Journal of the ACM, 46(5):604-632, 1999. [19] R. Nallapati. Discriminative models for information retrieval. In Proceedings of the 25th annual international ACM SIGIR conference on Research and development in information retrieval, pages 64 - 71, 2004. [20] Z. Nie, Y. Ma, J.-R. Wen, and W.-Y. Ma. Object-level web information retrieval. In Technical Report of Microsoft Research, volume MSR-TR-2005-11, 2005. [21] Z. Nie, Y. Zhang, J.-R. Wen, and W.-Y. Ma. Object-level ranking: Bringing order to web objects. In Proceedings of the 14th international conference on World Wide Web, pages 567 - 574, Chiba, Japan, 2005. [22] L. Page, S. Brin, R. Motwani, and T. Winograd. The pagerank citation ranking: Bringing order to the web. In Technical report, Stanford Digital Libraries, 1998. [23] A. Savakis, S. Etz, and A. Loui. Evaluation of image appeal in consumer photography. In SPIE Human Vision and Electronic Imaging, pages 111-120, 2000. [24] D. Sullivan. Hitwise search engine ratings. Search Engine Watch Articles, http://searchenginewatch. com/reports/article.php/3099931, August 23, 2005. [25] S. Susstrunk and S. Winkler. Color image quality on the internet. In IS&T/SPIE Electronic Imaging 2004: Internet Imaging V, volume 5304, pages 118-131, 2004. [26] H. Tong, M. Li, Z. H.J., J. He, and Z. C.S. Classification of digital photos taken by photographers or home users. In Pacific-Rim Conference on Multimedia (PCM), pages 198-205, 2004. [27] W. Xi, B. Zhang, Z. Chen, Y. Lu, S. Yan, W.-Y. Ma, and E. A. Fox. Link fusion: a unified link analysis framework for multi-type interrelated data objects. In Proceedings of the 13th international conference on World Wide Web, pages 319 - 327, 2004. 386

duplicate photo detection algorithm;algorithm;image search query;score fusion method;nonlinear fusion method;multiple web forum;rank photo;domain specific knowledge;web object;rank;high-quality photo search;ranking function;image search;web object-ranking problem;web object-ranking;vertical search

train_H-73

Unified Utility Maximization Framework for Resource Selection

This paper presents a unified utility framework for resource selection of distributed text information retrieval. This new framework shows an efficient and effective way to infer the probabilities of relevance of all the documents across the text databases. With the estimated relevance information, resource selection can be made by explicitly optimizing the goals of different applications. Specifically, when used for database recommendation, the selection is optimized for the goal of highrecall (include as many relevant documents as possible in the selected databases); when used for distributed document retrieval, the selection targets the high-precision goal (high precision in the final merged list of documents). This new model provides a more solid framework for distributed information retrieval. Empirical studies show that it is at least as effective as other state-of-the-art algorithms.

1. INTRODUCTION Conventional search engines such as Google or AltaVista use ad-hoc information retrieval solution by assuming all the searchable documents can be copied into a single centralized database for the purpose of indexing. Distributed information retrieval, also known as federated search [1,4,7,11,14,22] is different from ad-hoc information retrieval as it addresses the cases when documents cannot be acquired and stored in a single database. For example, Hidden Web contents (also called invisible or deep Web contents) are information on the Web that cannot be accessed by the conventional search engines. Hidden web contents have been estimated to be 2-50 [19] times larger than the contents that can be searched by conventional search engines. Therefore, it is very important to search this type of valuable information. The architecture of distributed search solution is highly influenced by different environmental characteristics. In a small local area network such as small company environments, the information providers may cooperate to provide corpus statistics or use the same type of search engines. Early distributed information retrieval research focused on this type of cooperative environments [1,8]. On the other side, in a wide area network such as very large corporate environments or on the Web there are many types of search engines and it is difficult to assume that all the information providers can cooperate as they are required. Even if they are willing to cooperate in these environments, it may be hard to enforce a single solution for all the information providers or to detect whether information sources provide the correct information as they are required. Many applications fall into the latter type of uncooperative environments such as the Mind project [16] which integrates non-cooperating digital libraries or the QProber system [9] which supports browsing and searching of uncooperative hidden Web databases. In this paper, we focus mainly on uncooperative environments that contain multiple types of independent search engines. There are three important sub-problems in distributed information retrieval. First, information about the contents of each individual database must be acquired (resource representation) [1,8,21]. Second, given a query, a set of resources must be selected to do the search (resource selection) [5,7,21]. Third, the results retrieved from all the selected resources have to be merged into a single final list before it can be presented to the end user (retrieval and results merging) [1,5,20,22]. Many types of solutions exist for distributed information retrieval. Invisible-web.net1 provides guided browsing of hidden Web databases by collecting the resource descriptions of these databases and building hierarchies of classes that group them by similar topics. A database recommendation system goes a step further than a browsing system like Invisible-web.net by recommending most relevant information sources to users" queries. It is composed of the resource description and the resource selection components. This solution is useful when the users want to browse the selected databases by themselves instead of asking the system to retrieve relevant documents automatically. Distributed document retrieval is a more sophisticated task. It selects relevant information sources for users" queries as the database recommendation system does. Furthermore, users" queries are forwarded to the corresponding selected databases and the returned individual ranked lists are merged into a single list to present to the users. The goal of a database recommendation system is to select a small set of resources that contain as many relevant documents as possible, which we call a high-recall goal. On the other side, the effectiveness of distributed document retrieval is often measured by the Precision of the final merged document result list, which we call a high-precision goal. Prior research indicated that these two goals are related but not identical [4,21]. However, most previous solutions simply use effective resource selection algorithm of database recommendation system for distributed document retrieval system or solve the inconsistency with heuristic methods [1,4,21]. This paper presents a unified utility maximization framework to integrate the resource selection problem of both database recommendation and distributed document retrieval together by treating them as different optimization goals. First, a centralized sample database is built by randomly sampling a small amount of documents from each database with query-based sampling [1]; database size statistics are also estimated [21]. A logistic transformation model is learned off line with a small amount of training queries to map the centralized document scores in the centralized sample database to the corresponding probabilities of relevance. Second, after a new query is submitted, the query can be used to search the centralized sample database which produces a score for each sampled document. The probability of relevance for each document in the centralized sample database can be estimated by applying the logistic model to each document"s score. Then, the probabilities of relevance of all the (mostly unseen) documents among the available databases can be estimated using the probabilities of relevance of the documents in the centralized sample database and the database size estimates. For the task of resource selection for a database recommendation system, the databases can be ranked by the expected number of relevant documents to meet the high-recall goal. For resource selection for a distributed document retrieval system, databases containing a small number of documents with large probabilities of relevance are favored over databases containing many documents with small probabilities of relevance. This selection criterion meets the high-precision goal of distributed document retrieval application. Furthermore, the Semi-supervised learning (SSL) [20,22] algorithm is applied to merge the returned documents into a final ranked list. The unified utility framework makes very few assumptions and works in uncooperative environments. Two key features make it a more solid model for distributed information retrieval: i) It formalizes the resource selection problems of different applications as various utility functions, and optimizes the utility functions to achieve the optimal results accordingly; and ii) It shows an effective and efficient way to estimate the probabilities of relevance of all documents across databases. Specifically, the framework builds logistic models on the centralized sample database to transform centralized retrieval scores to the corresponding probabilities of relevance and uses the centralized sample database as the bridge between individual databases and the logistic model. The human effort (relevance judgment) required to train the single centralized logistic model does not scale with the number of databases. This is a large advantage over previous research, which required the amount of human effort to be linear with the number of databases [7,15]. The unified utility framework is not only more theoretically solid but also very effective. Empirical studies show the new model to be at least as accurate as the state-of-the-art algorithms in a variety of configurations. The next section discusses related work. Section 3 describes the new unified utility maximization model. Section 4 explains our experimental methodology. Sections 5 and 6 present our experimental results for resource selection and document retrieval. Section 7 concludes. 2. PRIOR RESEARCH There has been considerable research on all the sub-problems of distributed information retrieval. We survey the most related work in this section. The first problem of distributed information retrieval is resource representation. The STARTS protocol is one solution for acquiring resource descriptions in cooperative environments [8]. However, in uncooperative environments, even the databases are willing to share their information, it is not easy to judge whether the information they provide is accurate or not. Furthermore, it is not easy to coordinate the databases to provide resource representations that are compatible with each other. Thus, in uncooperative environments, one common choice is query-based sampling, which randomly generates and sends queries to individual search engines and retrieves some documents to build the descriptions. As the sampled documents are selected by random queries, query-based sampling is not easily fooled by any adversarial spammer that is interested to attract more traffic. Experiments have shown that rather accurate resource descriptions can be built by sending about 80 queries and downloading about 300 documents [1]. Many resource selection algorithms such as gGlOSS/vGlOSS [8] and CORI [1] have been proposed in the last decade. The CORI algorithm represents each database by its terms, the document frequencies and a small number of corpus statistics (details in [1]). As prior research on different datasets has shown the CORI algorithm to be the most stable and effective of the three algorithms [1,17,18], we use it as a baseline algorithm in this work. The relevant document distribution estimation (ReDDE [21]) resource selection algorithm is a recent algorithm that tries to estimate the distribution of relevant documents across the available databases and ranks the databases accordingly. Although the ReDDE algorithm has been shown to be effective, it relies on heuristic constants that are set empirically [21]. The last step of the document retrieval sub-problem is results merging, which is the process of transforming database-specific 33 document scores into comparable database-independent document scores. The semi supervised learning (SSL) [20,22] result merging algorithm uses the documents acquired by querybased sampling as training data and linear regression to learn the database-specific, query-specific merging models. These linear models are used to convert the database-specific document scores into the approximated centralized document scores. The SSL algorithm has been shown to be effective [22]. It serves as an important component of our unified utility maximization framework (Section 3). In order to achieve accurate document retrieval results, many previous methods simply use resource selection algorithms that are effective of database recommendation system. But as pointed out above, a good resource selection algorithm optimized for high-recall may not work well for document retrieval, which targets the high-precision goal. This type of inconsistency has been observed in previous research [4,21]. The research in [21] tried to solve the problem with a heuristic method. The research most similar to what we propose here is the decision-theoretic framework (DTF) [7,15]. This framework computes a selection that minimizes the overall costs (e.g., retrieval quality, time) of document retrieval system and several methods [15] have been proposed to estimate the retrieval quality. However, two points distinguish our research from the DTF model. First, the DTF is a framework designed specifically for document retrieval, but our new model integrates two distinct applications with different requirements (database recommendation and distributed document retrieval) into the same unified framework. Second, the DTF builds a model for each database to calculate the probabilities of relevance. This requires human relevance judgments for the results retrieved from each database. In contrast, our approach only builds one logistic model for the centralized sample database. The centralized sample database can serve as a bridge to connect the individual databases with the centralized logistic model, thus the probabilities of relevance of documents in different databases can be estimated. This strategy can save large amount of human judgment effort and is a big advantage of the unified utility maximization framework over the DTF especially when there are a large number of databases. 3. UNIFIED UTILITY MAXIMIZATION FRAMEWORK The Unified Utility Maximization (UUM) framework is based on estimating the probabilities of relevance of the (mostly unseen) documents available in the distributed search environment. In this section we describe how the probabilities of relevance are estimated and how they are used by the Unified Utility Maximization model. We also describe how the model can be optimized for the high-recall goal of a database recommendation system and the high-precision goal of a distributed document retrieval system. 3.1 Estimating Probabilities of Relevance As pointed out above, the purpose of resource selection is highrecall and the purpose of document retrieval is high-precision. In order to meet these diverse goals, the key issue is to estimate the probabilities of relevance of the documents in various databases. This is a difficult problem because we can only observe a sample of the contents of each database using query-based sampling. Our strategy is to make full use of all the available information to calculate the probability estimates. 3.1.1 Learning Probabilities of Relevance In the resource description step, the centralized sample database is built by query-based sampling and the database sizes are estimated using the sample-resample method [21]. At the same time, an effective retrieval algorithm (Inquery [2]) is applied on the centralized sample database with a small number (e.g., 50) of training queries. For each training query, the CORI resource selection algorithm [1] is applied to select some number (e.g., 10) of databases and retrieve 50 document ids from each database. The SSL results merging algorithm [20,22] is used to merge the results. Then, we can download the top 50 documents in the final merged list and calculate their corresponding centralized scores using Inquery and the corpus statistics of the centralized sample database. The centralized scores are further normalized (divided by the maximum centralized score for each query), as this method has been suggested to improve estimation accuracy in previous research [15]. Human judgment is acquired for those documents and a logistic model is built to transform the normalized centralized document scores to probabilities of relevance as follows: ( ) ))(exp(1 ))(exp( |)( _ _ dSba dSba drelPdR ccc ccc ++ + == (1) where )( _ dSc is the normalized centralized document score and ac and bc are the two parameters of the logistic model. These two parameters are estimated by maximizing the probabilities of relevance of the training queries. The logistic model provides us the tool to calculate the probabilities of relevance from centralized document scores. 3.1.2 Estimating Centralized Document Scores When the user submits a new query, the centralized document scores of the documents in the centralized sample database are calculated. However, in order to calculate the probabilities of relevance, we need to estimate centralized document scores for all documents across the databases instead of only the sampled documents. This goal is accomplished using: the centralized scores of the documents in the centralized sample database, and the database size statistics. We define the database scale factor for the ith database as the ratio of the estimated database size and the number of documents sampled from this database as follows: ^ _ i i i db db db samp N SF N = (2) where ^ idbN is the estimated database size and _idb sampN is the number of documents from the ith database in the centralized sample database. The intuition behind the database scale factor is that, for a database whose scale factor is 50, if one document from this database in the centralized sample database has a centralized document score of 0.5, we may guess that there are about 50 documents in that database which have scores of about 0.5. Actually, we can apply a finer non-parametric linear interpolation method to estimate the centralized document score curve for each database. Formally, we rank all the sampled documents from the ith database by their centralized document 34 scores to get the sampled centralized document score list {Sc(dsi1), Sc(dsi2), Sc(dsi3),…..} for the ith database; we assume that if we could calculate the centralized document scores for all the documents in this database and get the complete centralized document score list, the top document in the sampled list would have rank SFdbi/2, the second document in the sampled list would rank SFdbi3/2, and so on. Therefore, the data points of sampled documents in the complete list are: {(SFdbi/2, Sc(dsi1)), (SFdbi3/2, Sc(dsi2)), (SFdbi5/2, Sc(dsi3)),…..}. Piecewise linear interpolation is applied to estimate the centralized document score curve, as illustrated in Figure 1. The complete centralized document score list can be estimated by calculating the values of different ranks on the centralized document curve as: ],1[,)(S ^^ c idbij Njd ∈ . It can be seen from Figure 1 that more sample data points produce more accurate estimates of the centralized document score curves. However, for databases with large database scale ratios, this kind of linear interpolation may be rather inaccurate, especially for the top ranked (e.g., [1, SFdbi/2]) documents. Therefore, an alternative solution is proposed to estimate the centralized document scores of the top ranked documents for databases with large scale ratios (e.g., larger than 100). Specifically, a logistic model is built for each of these databases. The logistic model is used to estimate the centralized document score of the top 1 document in the corresponding database by using the two sampled documents from that database with highest centralized scores. ))()(exp(1 ))()(exp( )( 22110 22110 ^ 1 iciicii iciicii ic dsSdsS dsSdsS dS ααα ααα +++ ++ = (3) 0iα , 1iα and 2iα are the parameters of the logistic model. For each training query, the top retrieved document of each database is downloaded and the corresponding centralized document score is calculated. Together with the scores of the top two sampled documents, these parameters can be estimated. After the centralized score of the top document is estimated, an exponential function is fitted for the top part ([1, SFdbi/2]) of the centralized document score curve as: ]2/,1[)*exp()( 10 ^ idbiiijc SFjjdS ∈+= ββ (4) ^ 0 1 1log( ( ))i c i iS dβ β= − (5) )12/( ))(log()((log( ^ 11 1 − − = idb icic i SF dSdsS β (6) The two parameters 0iβ and 1iβ are fitted to make sure the exponential function passes through the two points (1, ^ 1)( ic dS ) and (SFdbi/2, Sc(dsi1)). The exponential function is only used to adjust the top part of the centralized document score curve and the lower part of the curve is still fitted with the linear interpolation method described above. The adjustment by fitting exponential function of the top ranked documents has been shown empirically to produce more accurate results. From the centralized document score curves, we can estimate the complete centralized document score lists accordingly for all the available databases. After the estimated centralized document scores are normalized, the complete lists of probabilities of relevance can be constructed out of the complete centralized document score lists by Equation 1. Formally for the ith database, the complete list of probabilities of relevance is: ],1[,)(R ^^ idbij Njd ∈ . 3.2 The Unified Utility Maximization Model In this section, we formally define the new unified utility maximization model, which optimizes the resource selection problems for two goals of high-recall (database recommendation) and high-precision (distributed document retrieval) in the same framework. In the task of database recommendation, the system needs to decide how to rank databases. In the task of document retrieval, the system not only needs to select the databases but also needs to decide how many documents to retrieve from each selected database. We generalize the database recommendation selection process, which implicitly recommends all documents in every selected database, as a special case of the selection decision for the document retrieval task. Formally, we denote di as the number of documents we would like to retrieve from the ith database and ,.....},{ 21 ddd = as a selection action for all the databases. The database selection decision is made based on the complete lists of probabilities of relevance for all the databases. The complete lists of probabilities of relevance are inferred from all the available information specifically sR , which stands for the resource descriptions acquired by query-based sampling and the database size estimates acquired by sample-resample; cS stands for the centralized document scores of the documents in the centralized sample database. If the method of estimating centralized document scores and probabilities of relevance in Section 3.1 is acceptable, then the most probable complete lists of probabilities of relevance can be derived and we denote them as 1 ^ ^ * 1{(R( ), [1, ]),dbjd j Nθ = ∈ 2 ^ ^ 2(R( ), [1, ]),.......}dbjd j N∈ . Random vector denotes an arbitrary set of complete lists of probabilities of relevance and ),|( cs SRP θ as the probability of generating this set of lists. Finally, to each selection action d and a set of complete lists of Figure 1. Linear interpolation construction of the complete centralized document score list (database scale factor is 50). 35 probabilities of relevance θ , we associate a utility function ),( dU θ which indicates the benefit from making the d selection when the true complete lists of probabilities of relevance are θ . Therefore, the selection decision defined by the Bayesian framework is: θθθ θ dSRPdUd cs d ).|(),(maxarg * = (7) One common approach to simplify the computation in the Bayesian framework is to only calculate the utility function at the most probable parameter values instead of calculating the whole expectation. In other words, we only need to calculate ),( * dU θ and Equation 7 is simplified as follows: ),(maxarg * * θdUd d = (8) This equation serves as the basic model for both the database recommendation system and the document retrieval system. 3.3 Resource Selection for High-Recall High-recall is the goal of the resource selection algorithm in federated search tasks such as database recommendation. The goal is to select a small set of resources (e.g., less than Nsdb databases) that contain as many relevant documents as possible, which can be formally defined as: = = i N j iji idb ddIdU ^ 1 ^ * )(R)(),( θ (9) I(di) is the indicator function, which is 1 when the ith database is selected and 0 otherwise. Plug this equation into the basic model in Equation 8 and associate the selected database number constraint to obtain the following: sdb i i i N j iji d NdItoSubject ddId idb = = = )(: )(R)(maxarg ^ 1 ^* (10) The solution of this optimization problem is very simple. We can calculate the expected number of relevant documents for each database as follows: = = idb i N j ijRd dN ^ 1 ^^ )(R (11) The Nsdb databases with the largest expected number of relevant documents can be selected to meet the high-recall goal. We call this the UUM/HR algorithm (Unified Utility Maximization for High-Recall). 3.4 Resource Selection for High-Precision High-Precision is the goal of resource selection algorithm in federated search tasks such as distributed document retrieval. It is measured by the Precision at the top part of the final merged document list. This high-precision criterion is realized by the following utility function, which measures the Precision of retrieved documents from the selected databases. = = i d j iji i ddIdU 1 ^ * )(R)(),( θ (12) Note that the key difference between Equation 12 and Equation 9 is that Equation 9 sums up the probabilities of relevance of all the documents in a database, while Equation 12 only considers a much smaller part of the ranking. Specifically, we can calculate the optimal selection decision by: = = i d j iji d i ddId 1 ^* )(R)(maxarg (13) Different kinds of constraints caused by different characteristics of the document retrieval tasks can be associated with the above optimization problem. The most common one is to select a fixed number (Nsdb) of databases and retrieve a fixed number (Nrdoc) of documents from each selected database, formally defined as: 0, )(: )(R)(maxarg 1 ^* ≠= = = = irdoci sdb i i i d j iji d difNd NdItoSubject ddId i (14) This optimization problem can be solved easily by calculating the number of expected relevant documents in the top part of the each database"s complete list of probabilities of relevance: = = rdoc i N j ijRdTop dN 1 ^^ _ )(R (15) Then the databases can be ranked by these values and selected. We call this the UUM/HP-FL algorithm (Unified Utility Maximization for High-Precision with Fixed Length document rankings from each selected database). A more complex situation is to vary the number of retrieved documents from each selected database. More specifically, we allow different selected databases to return different numbers of documents. For simplification, the result list lengths are required to be multiples of a baseline number 10. (This value can also be varied, but for simplification it is set to 10 in this paper.) This restriction is set to simulate the behavior of commercial search engines on the Web. (Search engines such as Google and AltaVista return only 10 or 20 document ids for every result page.) This procedure saves the computation time of calculating optimal database selection by allowing the step of dynamic programming to be 10 instead of 1 (more detail is discussed latterly). For further simplification, we restrict to select at most 100 documents from each database (di<=100) Then, the selection optimization problem is formalized as follows: ]10..,,2,1,0[,*10 )(: )(R)(maxarg _ 1 ^* ∈= = = = = kkd Nd NdItoSubject ddId i rdocTotal i i sdb i i i d j iji d i (16) NTotal_rdoc is the total number of documents to be retrieved. Unfortunately, there is no simple solution for this optimization problem as there are for Equations 10 and 14. However, a 36 dynamic programming algorithm can be applied to calculate the optimal solution. The basic steps of this dynamic programming method are described in Figure 2. As this algorithm allows retrieving result lists of varying lengths from each selected database, it is called UUM/HP-VL algorithm. After the selection decisions are made, the selected databases are searched and the corresponding document ids are retrieved from each database. The final step of document retrieval is to merge the returned results into a single ranked list with the semisupervised learning algorithm. It was pointed out before that the SSL algorithm maps the database-specific scores into the centralized document scores and builds the final ranked list accordingly, which is consistent with all our selection procedures where documents with higher probabilities of relevance (thus higher centralized document scores) are selected. 4. EXPERIMENTAL METHODOLOGY 4.1 Testbeds It is desirable to evaluate distributed information retrieval algorithms with testbeds that closely simulate the real world applications. The TREC Web collections WT2g or WT10g [4,13] provide a way to partition documents by different Web servers. In this way, a large number (O(1000)) of databases with rather diverse contents could be created, which may make this testbed a good candidate to simulate the operational environments such as open domain hidden Web. However, two weakness of this testbed are: i) Each database contains only a small amount of document (259 documents by average for WT2g) [4]; and ii) The contents of WT2g or WT10g are arbitrarily crawled from the Web. It is not likely for a hidden Web database to provide personal homepages or web pages indicating that the pages are under construction and there is no useful information at all. These types of web pages are contained in the WT2g/WT10g datasets. Therefore, the noisy Web data is not similar with that of high-quality hidden Web database contents, which are usually organized by domain experts. Another choice is the TREC news/government data [1,15,17, 18,21]. TREC news/government data is concentrated on relatively narrow topics. Compared with TREC Web data: i) The news/government documents are much more similar to the contents provided by a topic-oriented database than an arbitrary web page, ii) A database in this testbed is larger than that of TREC Web data. By average a database contains thousands of documents, which is more realistic than a database of TREC Web data with about 250 documents. As the contents and sizes of the databases in the TREC news/government testbed are more similar with that of a topic-oriented database, it is a good candidate to simulate the distributed information retrieval environments of large organizations (companies) or domainspecific hidden Web sites, such as West that provides access to legal, financial and news text databases [3]. As most current distributed information retrieval systems are developed for the environments of large organizations (companies) or domainspecific hidden Web other than open domain hidden Web, TREC news/government testbed was chosen in this work. Trec123-100col-bysource testbed is one of the most used TREC news/government testbed [1,15,17,21]. It was chosen in this work. Three testbeds in [21] with skewed database size distributions and different types of relevant document distributions were also used to give more thorough simulation for real environments. Trec123-100col-bysource: 100 databases were created from TREC CDs 1, 2 and 3. They were organized by source and publication date [1]. The sizes of the databases are not skewed. Details are in Table 1. Three testbeds built in [21] were based on the trec123-100colbysource testbed. Each testbed contains many small databases and two large databases created by merging about 10-20 small databases together. Input: Complete lists of probabilities of relevance for all the |DB| databases. Output: Optimal selection solution for Equation 16. i) Create the three-dimensional array: Sel (1..|DB|, 1..NTotal_rdoc/10, 1..Nsdb) Each Sel (x, y, z) is associated with a selection decision xyzd , which represents the best selection decision in the condition: only databases from number 1 to number x are considered for selection; totally y*10 documents will be retrieved; only z databases are selected out of the x database candidates. And Sel (x, y, z) is the corresponding utility value by choosing the best selection. ii) Initialize Sel (1, 1..NTotal_rdoc/10, 1..Nsdb) with only the estimated relevance information of the 1st database. iii) Iterate the current database candidate i from 2 to |DB| For each entry Sel (i, y, z): Find k such that: )10,min(1: ))()1,,1((maxarg *10 ^ * yktosubject dRzkyiSelk kj ij k ≤≤ +−−−= ≤ ),,1())()1,,1(( * *10 ^ * zyiSeldRzkyiSelIf kj ij −>+−−− ≤ This means that we should retrieve * 10 k∗ documents from the ith database, otherwise we should not select this database and the previous best solution Sel (i-1, y, z) should be kept. Then set the value of iyzd and Sel (i, y, z) accordingly. iv) The best selection solution is given by _ /10| | Toral rdoc sdbDB N Nd and the corresponding utility value is Sel (|DB|, NTotal_rdoc/10, Nsdb). Figure 2. The dynamic programming optimization procedure for Equation 16. Table1: Testbed statistics. Number of documents Size (MB) Testbed Size (GB) Min Avg Max Min Avg Max Trec123 3.2 752 10782 39713 28 32 42 Table2: Query set statistics. Name TREC Topic Set TREC Topic Field Average Length (Words) Trec123 51-150 Title 3.1 37 Trec123-2ldb-60col (representative): The databases in the trec123-100col-bysource were sorted with alphabetical order. Two large databases were created by merging 20 small databases with the round-robin method. Thus, the two large databases have more relevant documents due to their large sizes, even though the densities of relevant documents are roughly the same as the small databases. Trec123-AP-WSJ-60col (relevant): The 24 Associated Press collections and the 16 Wall Street Journal collections in the trec123-100col-bysource testbed were collapsed into two large databases APall and WSJall. The other 60 collections were left unchanged. The APall and WSJall databases have higher densities of documents relevant to TREC queries than the small databases. Thus, the two large databases have many more relevant documents than the small databases. Trec123-FR-DOE-81col (nonrelevant): The 13 Federal Register collections and the 6 Department of Energy collections in the trec123-100col-bysource testbed were collapsed into two large databases FRall and DOEall. The other 80 collections were left unchanged. The FRall and DOEall databases have lower densities of documents relevant to TREC queries than the small databases, even though they are much larger. 100 queries were created from the title fields of TREC topics 51-150. The queries 101-150 were used as training queries and the queries 51-100 were used as test queries (details in Table 2). 4.2 Search Engines In the uncooperative distributed information retrieval environments of large organizations (companies) or domainspecific hidden Web, different databases may use different types of search engine. To simulate the multiple type-engine environment, three different types of search engines were used in the experiments: INQUERY [2], a unigram statistical language model with linear smoothing [12,20] and a TFIDF retrieval algorithm with ltc weight [12,20]. All these algorithms were implemented with the Lemur toolkit [12]. These three kinds of search engines were assigned to the databases among the four testbeds in a round-robin manner. 5. RESULTS: RESOURCE SELECTION OF DATABASE RECOMMENDATION All four testbeds described in Section 4 were used in the experiments to evaluate the resource selection effectiveness of the database recommendation system. The resource descriptions were created using query-based sampling. About 80 queries were sent to each database to download 300 unique documents. The database size statistics were estimated by the sample-resample method [21]. Fifty queries (101-150) were used as training queries to build the relevant logistic model and to fit the exponential functions of the centralized document score curves for large ratio databases (details in Section 3.1). Another 50 queries (51-100) were used as test data. Resource selection algorithms of database recommendation systems are typically compared using the recall metric nR [1,17,18,21]. Let B denote a baseline ranking, which is often the RBR (relevance based ranking), and E as a ranking provided by a resource selection algorithm. And let Bi and Ei denote the number of relevant documents in the ith ranked database of B or E. Then Rn is defined as follows: = = = k i i k i i k B E R 1 1 (17) Usually the goal is to search only a few databases, so our figures only show results for selecting up to 20 databases. The experiments summarized in Figure 3 compared the effectiveness of the three resource selection algorithms, namely the CORI, ReDDE and UUM/HR. The UUM/HR algorithm is described in Section 3.3. It can be seen from Figure 3 that the ReDDE and UUM/HR algorithms are more effective (on the representative, relevant and nonrelevant testbeds) or as good as (on the Trec123-100Col testbed) the CORI resource selection algorithm. The UUM/HR algorithm is more effective than the ReDDE algorithm on the representative and relevant testbeds and is about the same as the ReDDE algorithm on the Trec123100Col and the nonrelevant testbeds. This suggests that the UUM/HR algorithm is more robust than the ReDDE algorithm. It can be noted that when selecting only a few databases on the Trec123-100Col or the nonrelevant testbeds, the ReDEE algorithm has a small advantage over the UUM/HR algorithm. We attribute this to two causes: i) The ReDDE algorithm was tuned on the Trec123-100Col testbed; and ii) Although the difference is small, this may suggest that our logistic model of estimating probabilities of relevance is not accurate enough. More training data or a more sophisticated model may help to solve this minor puzzle. Collections Selected. Collections Selected. Trec123-100Col Testbed. Representative Testbed. Collection Selected. Collection Selected. Relevant Testbed. Nonrelevant Testbed. Figure 3. Resource selection experiments on the four testbeds. 38 6. RESULTS: DOCUMENT RETRIEVAL EFFECTIVENESS For document retrieval, the selected databases are searched and the returned results are merged into a single final list. In all of the experiments discussed in this section the results retrieved from individual databases were combined by the semisupervised learning results merging algorithm. This version of the SSL algorithm [22] is allowed to download a small number of returned document texts on the fly to create additional training data in the process of learning the linear models which map database-specific document scores into estimated centralized document scores. It has been shown to be very effective in environments where only short result-lists are retrieved from each selected database [22]. This is a common scenario in operational environments and was the case for our experiments. Document retrieval effectiveness was measured by Precision at the top part of the final document list. The experiments in this section were conducted to study the document retrieval effectiveness of five selection algorithms, namely the CORI, ReDDE, UUM/HR, UUM/HP-FL and UUM/HP-VL algorithms. The last three algorithms were proposed in Section 3. All the first four algorithms selected 3 or 5 databases, and 50 documents were retrieved from each selected database. The UUM/HP-FL algorithm also selected 3 or 5 databases, but it was allowed to adjust the number of documents to retrieve from each selected database; the number retrieved was constrained to be from 10 to 100, and a multiple of 10. The Trec123-100Col and representative testbeds were selected for document retrieval as they represent two extreme cases of resource selection effectiveness; in one case the CORI algorithm is as good as the other algorithms and in the other case it is quite Table 5. Precision on the representative testbed when 3 databases were selected. (The first baseline is CORI; the second baseline for UUM/HP methods is UUM/HR.) Precision at Doc Rank CORI ReDDE UUM/HR UUM/HP-FL UUM/HP-VL 5 docs 0.3720 0.4080 (+9.7%) 0.4640 (+24.7%) 0.4600 (+23.7%)(-0.9%) 0.5000 (+34.4%)(+7.8%) 10 docs 0.3400 0.4060 (+19.4%) 0.4600 (+35.3%) 0.4540 (+33.5%)(-1.3%) 0.4640 (+36.5%)(+0.9%) 15 docs 0.3120 0.3880 (+24.4%) 0.4320 (+38.5%) 0.4240 (+35.9%)(-1.9%) 0.4413 (+41.4%)(+2.2) 20 docs 0.3000 0.3750 (+25.0%) 0.4080 (+36.0%) 0.4040 (+34.7%)(-1.0%) 0.4240 (+41.3%)(+4.0%) 30 docs 0.2533 0.3440 (+35.8%) 0.3847 (+51.9%) 0.3747 (+47.9%)(-2.6%) 0.3887 (+53.5%)(+1.0%) Table 6. Precision on the representative testbed when 5 databases were selected. (The first baseline is CORI; the second baseline for UUM/HP methods is UUM/HR.) Precision at Doc Rank CORI ReDDE UUM/HR UUM/HP-FL UUM/HP-VL 5 docs 0.3960 0.4080 (+3.0%) 0.4560 (+15.2%) 0.4280 (+8.1%)(-6.1%) 0.4520 (+14.1%)(-0.9%) 10 docs 0.3880 0.4060 (+4.6%) 0.4280 (+10.3%) 0.4460 (+15.0%)(+4.2%) 0.4560 (+17.5%)(+6.5%) 15 docs 0.3533 0.3987 (+12.9%) 0.4227 (+19.6%) 0.4440 (+25.7%)(+5.0%) 0.4453 (+26.0%)(+5.4%) 20 docs 0.3330 0.3960 (+18.9%) 0.4140 (+24.3%) 0.4290 (+28.8%)(+3.6%) 0.4350 (+30.6%)(+5.1%) 30 docs 0.2967 0.3740 (+26.1%) 0.4013 (+35.3%) 0.3987 (+34.4%)(-0.7%) 0.4060 (+36.8%)(+1.2%) Table 3. Precision on the trec123-100col-bysource testbed when 3 databases were selected. (The first baseline is CORI; the second baseline for UUM/HP methods is UUM/HR.) Precision at Doc Rank CORI ReDDE UUM/HR UUM/HP-FL UUM/HP-VL 5 docs 0.3640 0.3480 (-4.4%) 0.3960 (+8.8%) 0.4680 (+28.6%)(+18.1%) 0.4640 (+27.5%)(+17.2%) 10 docs 0.3360 0.3200 (-4.8%) 0.3520 (+4.8%) 0.4240 (+26.2%)(+20.5%) 0.4220 (+25.6%)(+19.9%) 15 docs 0.3253 0.3187 (-2.0%) 0.3347 (+2.9%) 0.3973 (+22.2%)(+15.7%) 0.3920 (+20.5%)(+17.1%) 20 docs 0.3140 0.2980 (-5.1%) 0.3270 (+4.1%) 0.3720 (+18.5%)(+13.8%) 0.3700 (+17.8%)(+13.2%) 30 docs 0.2780 0.2660 (-4.3%) 0.2973 (+6.9%) 0.3413 (+22.8%)(+14.8%) 0.3400 (+22.3%)(+14.4%) Table 4. Precision on the trec123-100col-bysource testbed when 5 databases were selected. (The first baseline is CORI; the second baseline for UUM/HP methods is UUM/HR.) Precision at Doc Rank CORI ReDDE UUM/HR UUM/HP-FL UUM/HP-VL 5 docs 0.4000 0.3920 (-2.0%) 0.4280 (+7.0%) 0.4680 (+17.0%)(+9.4%) 0.4600 (+15.0%)(+7.5%) 10 docs 0.3800 0.3760 (-1.1%) 0.3800 (+0.0%) 0.4180 (+10.0%)(+10.0%) 0.4320 (+13.7%)(+13.7%) 15 docs 0.3560 0.3560 (+0.0%) 0.3720 (+4.5%) 0.3920 (+10.1%)(+5.4%) 0.4080 (+14.6%)(+9.7%) 20 docs 0.3430 0.3390 (-1.2%) 0.3550 (+3.5%) 0.3710 (+8.2%)(+4.5%) 0.3830 (+11.7%)(+7.9%) 30 docs 0.3240 0.3140 (-3.1%) 0.3313 (+2.3%) 0.3500 (+8.0%)(+5.6%) 0.3487 (+7.6%)(+5.3%) 39 a lot worse than the other algorithms. Tables 3 and 4 show the results on the Trec123-100Col testbed, and Tables 5 and 6 show the results on the representative testbed. On the Trec123-100Col testbed, the document retrieval effectiveness of the CORI selection algorithm is roughly the same or a little bit better than the ReDDE algorithm but both of them are worse than the other three algorithms (Tables 3 and 4). The UUM/HR algorithm has a small advantage over the CORI and ReDDE algorithms. One main difference between the UUM/HR algorithm and the ReDDE algorithm was pointed out before: The UUM/HR uses training data and linear interpolation to estimate the centralized document score curves, while the ReDDE algorithm [21] uses a heuristic method, assumes the centralized document score curves are step functions and makes no distinction among the top part of the curves. This difference makes UUM/HR better than the ReDDE algorithm at distinguishing documents with high probabilities of relevance from low probabilities of relevance. Therefore, the UUM/HR reflects the high-precision retrieval goal better than the ReDDE algorithm and thus is more effective for document retrieval. The UUM/HR algorithm does not explicitly optimize the selection decision with respect to the high-precision goal as the UUM/HP-FL and UUM/HP-VL algorithms are designed to do. It can be seen that on this testbed, the UUM/HP-FL and UUM/HP-VL algorithms are much more effective than all the other algorithms. This indicates that their power comes from explicitly optimizing the high-precision goal of document retrieval in Equations 14 and 16. On the representative testbed, CORI is much less effective than other algorithms for distributed document retrieval (Tables 5 and 6). The document retrieval results of the ReDDE algorithm are better than that of the CORI algorithm but still worse than the results of the UUM/HR algorithm. On this testbed the three UUM algorithms are about equally effective. Detailed analysis shows that the overlap of the selected databases between the UUM/HR, UUM/HP-FL and UUM/HP-VL algorithms is much larger than the experiments on the Trec123-100Col testbed, since all of them tend to select the two large databases. This explains why they are about equally effective for document retrieval. In real operational environments, databases may return no document scores and report only ranked lists of results. As the unified utility maximization model only utilizes retrieval scores of sampled documents with a centralized retrieval algorithm to calculate the probabilities of relevance, it makes database selection decisions without referring to the document scores from individual databases and can be easily generalized to this case of rank lists without document scores. The only adjustment is that the SSL algorithm merges ranked lists without document scores by assigning the documents with pseudo-document scores normalized for their ranks (In a ranked list of 50 documents, the first one has a score of 1, the second has a score of 0.98 etc) ,which has been studied in [22]. The experiment results on trec123-100Col-bysource testbed with 3 selected databases are shown in Table 7. The experiment setting was the same as before except that the document scores were eliminated intentionally and the selected databases only return ranked lists of document ids. It can be seen from the results that the UUM/HP-FL and UUM/HP-VL work well with databases returning no document scores and are still more effective than other alternatives. Other experiments with databases that return no document scores are not reported but they show similar results to prove the effectiveness of UUM/HP-FL and UUM/HPVL algorithms. The above experiments suggest that it is very important to optimize the high-precision goal explicitly in document retrieval. The new algorithms based on this principle achieve better or at least as good results as the prior state-of-the-art algorithms in several environments. 7. CONCLUSION Distributed information retrieval solves the problem of finding information that is scattered among many text databases on local area networks and Internets. Most previous research use effective resource selection algorithm of database recommendation system for distributed document retrieval application. We argue that the high-recall resource selection goal of database recommendation and high-precision goal of document retrieval are related but not identical. This kind of inconsistency has also been observed in previous work, but the prior solutions either used heuristic methods or assumed cooperation by individual databases (e.g., all the databases used the same kind of search engines), which is frequently not true in the uncooperative environment. In this work we propose a unified utility maximization model to integrate the resource selection of database recommendation and document retrieval tasks into a single unified framework. In this framework, the selection decisions are obtained by optimizing different objective functions. As far as we know, this is the first work that tries to view and theoretically model the distributed information retrieval task in an integrated manner. The new framework continues a recent research trend studying the use of query-based sampling and a centralized sample database. A single logistic model was trained on the centralized Table 7. Precision on the trec123-100col-bysource testbed when 3 databases were selected (The first baseline is CORI; the second baseline for UUM/HP methods is UUM/HR.) (Search engines do not return document scores) Precision at Doc Rank CORI ReDDE UUM/HR UUM/HP-FL UUM/HP-VL 5 docs 0.3520 0.3240 (-8.0%) 0.3680 (+4.6%) 0.4520 (+28.4%)(+22.8%) 0.4520 (+28.4%)(+22.8) 10 docs 0.3320 0.3140 (-5.4%) 0.3340 (+0.6%) 0.4120 (+24.1%)(+23.4%) 0.4020 (+21.1%)(+20.4%) 15 docs 0.3227 0.2987 (-7.4%) 0.3280 (+1.6%) 0.3920 (+21.5%)(+19.5%) 0.3733 (+15.7%)(+13.8%) 20 docs 0.3030 0.2860 (-5.6%) 0.3130 (+3.3%) 0.3670 (+21.2%)(+17.3%) 0.3590 (+18.5%)(+14.7%) 30 docs 0.2727 0.2640 (-3.2%) 0.2900 (+6.3%) 0.3273 (+20.0%)(+12.9%) 0.3273 (+20.0%)(+12.9%) 40 sample database to estimate the probabilities of relevance of documents by their centralized retrieval scores, while the centralized sample database serves as a bridge to connect the individual databases with the centralized logistic model. Therefore, the probabilities of relevance for all the documents across the databases can be estimated with very small amount of human relevance judgment, which is much more efficient than previous methods that build a separate model for each database. This framework is not only more theoretically solid but also very effective. One algorithm for resource selection (UUM/HR) and two algorithms for document retrieval (UUM/HP-FL and UUM/HP-VL) are derived from this framework. Empirical studies have been conducted on testbeds to simulate the distributed search solutions of large organizations (companies) or domain-specific hidden Web. Furthermore, the UUM/HP-FL and UUM/HP-VL resource selection algorithms are extended with a variant of SSL results merging algorithm to address the distributed document retrieval task when selected databases do not return document scores. Experiments have shown that these algorithms achieve results that are at least as good as the prior state-of-the-art, and sometimes considerably better. Detailed analysis indicates that the advantage of these algorithms comes from explicitly optimizing the goals of the specific tasks. The unified utility maximization framework is open for different extensions. When cost is associated with searching the online databases, the utility framework can be adjusted to automatically estimate the best number of databases to search so that a large amount of relevant documents can be retrieved with relatively small costs. Another extension of the framework is to consider the retrieval effectiveness of the online databases, which is an important issue in the operational environments. All of these are the directions of future research. ACKNOWLEDGEMENT This research was supported by NSF grants EIA-9983253 and IIS-0118767. Any opinions, findings, conclusions, or recommendations expressed in this paper are the authors", and do not necessarily reflect those of the sponsor. REFERENCES [1] J. Callan. (2000). Distributed information retrieval. In W.B. Croft, editor, Advances in Information Retrieval. Kluwer Academic Publishers. (pp. 127-150). [2] J. Callan, W.B. Croft, and J. Broglio. (1995). TREC and TIPSTER experiments with INQUERY. Information Processing and Management, 31(3). (pp. 327-343). [3] J. G. Conrad, X. S. Guo, P. Jackson and M. Meziou. (2002). Database selection using actual physical and acquired logical collection resources in a massive domainspecific operational environment. Distributed search over the hidden web: Hierarchical database sampling and selection. In Proceedings of the 28th International Conference on Very Large Databases (VLDB). [4] N. Craswell. (2000). Methods for distributed information retrieval. Ph. D. thesis, The Australian Nation University. [5] N. Craswell, D. Hawking, and P. Thistlewaite. (1999). Merging results from isolated search engines. In Proceedings of 10th Australasian Database Conference. [6] D. D'Souza, J. Thom, and J. Zobel. (2000). A comparison of techniques for selecting text collections. In Proceedings of the 11th Australasian Database Conference. [7] N. Fuhr. (1999). A Decision-Theoretic approach to database selection in networked IR. ACM Transactions on Information Systems, 17(3). (pp. 229-249). [8] L. Gravano, C. Chang, H. Garcia-Molina, and A. Paepcke. (1997). STARTS: Stanford proposal for internet metasearching. In Proceedings of the 20th ACM-SIGMOD International Conference on Management of Data. [9] L. Gravano, P. Ipeirotis and M. Sahami. (2003). QProber: A System for Automatic Classification of Hidden-Web Databases. ACM Transactions on Information Systems, 21(1). [10] P. Ipeirotis and L. Gravano. (2002). Distributed search over the hidden web: Hierarchical database sampling and selection. In Proceedings of the 28th International Conference on Very Large Databases (VLDB). [11] InvisibleWeb.com. http://www.invisibleweb.com [12] The lemur toolkit. http://www.cs.cmu.edu/~lemur [13] J. Lu and J. Callan. (2003). Content-based information retrieval in peer-to-peer networks. In Proceedings of the 12th International Conference on Information and Knowledge Management. [14] W. Meng, C.T. Yu and K.L. Liu. (2002) Building efficient and effective metasearch engines. ACM Comput. Surv. 34(1). [15] H. Nottelmann and N. Fuhr. (2003). Evaluating different method of estimating retrieval quality for resource selection. In Proceedings of the 25th Annual International ACM SIGIR Conference on Research and Development in Information Retrieval. [16] H., Nottelmann and N., Fuhr. (2003). The MIND architecture for heterogeneous multimedia federated digital libraries. ACM SIGIR 2003 Workshop on Distributed Information Retrieval. [17] A.L. Powell, J.C. French, J. Callan, M. Connell, and C.L. Viles. (2000). The impact of database selection on distributed searching. In Proceedings of the 23rd Annual International ACM SIGIR Conference on Research and Development in Information Retrieval. [18] A.L. Powell and J.C. French. (2003). Comparing the performance of database selection algorithms. ACM Transactions on Information Systems, 21(4). (pp. 412-456). [19] C. Sherman (2001). Search for the invisible web. Guardian Unlimited. [20] L. Si and J. Callan. (2002). Using sampled data and regression to merge search engine results. In Proceedings of the 25th Annual International ACM SIGIR Conference on Research and Development in Information Retrieval. [21] L. Si and J. Callan. (2003). Relevant document distribution estimation method for resource selection. In Proceedings of the 26th Annual International ACM SIGIR Conference on Research and Development in Information Retrieval. [22] L. Si and J. Callan. (2003). A Semi-Supervised learning method to merge search engine results. ACM Transactions on Information Systems, 21(4). (pp. 457-491). 41

distributed document retrieval;resource selection;retrieval and result merging;unified utility maximization model;distributed text information retrieval resource selection;semi-supervised learning;resource representation;logistic transformation model;distribute information retrieval;hidden web content;resource selection of distributed text information retrieval;federated search;database recommendation

train_H-77

Automatic Extraction of Titles from General Documents using Machine Learning

In this paper, we propose a machine learning approach to title extraction from general documents. By general documents, we mean documents that can belong to any one of a number of specific genres, including presentations, book chapters, technical papers, brochures, reports, and letters. Previously, methods have been proposed mainly for title extraction from research papers. It has not been clear whether it could be possible to conduct automatic title extraction from general documents. As a case study, we consider extraction from Office including Word and PowerPoint. In our approach, we annotate titles in sample documents (for Word and PowerPoint respectively) and take them as training data, train machine learning models, and perform title extraction using the trained models. Our method is unique in that we mainly utilize formatting information such as font size as features in the models. It turns out that the use of formatting information can lead to quite accurate extraction from general documents. Precision and recall for title extraction from Word is 0.810 and 0.837 respectively, and precision and recall for title extraction from PowerPoint is 0.875 and 0.895 respectively in an experiment on intranet data. Other important new findings in this work include that we can train models in one domain and apply them to another domain, and more surprisingly we can even train models in one language and apply them to another language. Moreover, we can significantly improve search ranking results in document retrieval by using the extracted titles.

1. INTRODUCTION Metadata of documents is useful for many kinds of document processing such as search, browsing, and filtering. Ideally, metadata is defined by the authors of documents and is then used by various systems. However, people seldom define document metadata by themselves, even when they have convenient metadata definition tools [26]. Thus, how to automatically extract metadata from the bodies of documents turns out to be an important research issue. Methods for performing the task have been proposed. However, the focus was mainly on extraction from research papers. For instance, Han et al. [10] proposed a machine learning based method to conduct extraction from research papers. They formalized the problem as that of classification and employed Support Vector Machines as the classifier. They mainly used linguistic features in the model.1 In this paper, we consider metadata extraction from general documents. By general documents, we mean documents that may belong to any one of a number of specific genres. General documents are more widely available in digital libraries, intranets and the internet, and thus investigation on extraction from them is sorely needed. Research papers usually have well-formed styles and noticeable characteristics. In contrast, the styles of general documents can vary greatly. It has not been clarified whether a machine learning based approach can work well for this task. There are many types of metadata: title, author, date of creation, etc. As a case study, we consider title extraction in this paper. General documents can be in many different file formats: Microsoft Office, PDF (PS), etc. As a case study, we consider extraction from Office including Word and PowerPoint. We take a machine learning approach. We annotate titles in sample documents (for Word and PowerPoint respectively) and take them as training data to train several types of models, and perform title extraction using any one type of the trained models. In the models, we mainly utilize formatting information such as font size as features. We employ the following models: Maximum Entropy Model, Perceptron with Uneven Margins, Maximum Entropy Markov Model, and Voted Perceptron. In this paper, we also investigate the following three problems, which did not seem to have been examined previously. (1) Comparison between models: among the models above, which model performs best for title extraction; (2) Generality of model: whether it is possible to train a model on one domain and apply it to another domain, and whether it is possible to train a model in one language and apply it to another language; (3) Usefulness of extracted titles: whether extracted titles can improve document processing such as search. Experimental results indicate that our approach works well for title extraction from general documents. Our method can significantly outperform the baselines: one that always uses the first lines as titles and the other that always uses the lines in the largest font sizes as titles. Precision and recall for title extraction from Word are 0.810 and 0.837 respectively, and precision and recall for title extraction from PowerPoint are 0.875 and 0.895 respectively. It turns out that the use of format features is the key to successful title extraction. (1) We have observed that Perceptron based models perform better in terms of extraction accuracies. (2) We have empirically verified that the models trained with our approach are generic in the sense that they can be trained on one domain and applied to another, and they can be trained in one language and applied to another. (3) We have found that using the extracted titles we can significantly improve precision of document retrieval (by 10%). We conclude that we can indeed conduct reliable title extraction from general documents and use the extracted results to improve real applications. The rest of the paper is organized as follows. In section 2, we introduce related work, and in section 3, we explain the motivation and problem setting of our work. In section 4, we describe our method of title extraction, and in section 5, we describe our method of document retrieval using extracted titles. Section 6 gives our experimental results. We make concluding remarks in section 7. 2. RELATED WORK 2.1 Document Metadata Extraction Methods have been proposed for performing automatic metadata extraction from documents; however, the main focus was on extraction from research papers. The proposed methods fall into two categories: the rule based approach and the machine learning based approach. Giuffrida et al. [9], for instance, developed a rule-based system for automatically extracting metadata from research papers in Postscript. They used rules like titles are usually located on the upper portions of the first pages and they are usually in the largest font sizes. Liddy et al. [14] and Yilmazel el al. [23] performed metadata extraction from educational materials using rule-based natural language processing technologies. Mao et al. [16] also conducted automatic metadata extraction from research papers using rules on formatting information. The rule-based approach can achieve high performance. However, it also has disadvantages. It is less adaptive and robust when compared with the machine learning approach. Han et al. [10], for instance, conducted metadata extraction with the machine learning approach. They viewed the problem as that of classifying the lines in a document into the categories of metadata and proposed using Support Vector Machines as the classifier. They mainly used linguistic information as features. They reported high extraction accuracy from research papers in terms of precision and recall. 2.2 Information Extraction Metadata extraction can be viewed as an application of information extraction, in which given a sequence of instances, we identify a subsequence that represents information in which we are interested. Hidden Markov Model [6], Maximum Entropy Model [1, 4], Maximum Entropy Markov Model [17], Support Vector Machines [3], Conditional Random Field [12], and Voted Perceptron [2] are widely used information extraction models. Information extraction has been applied, for instance, to part-ofspeech tagging [20], named entity recognition [25] and table extraction [19]. 2.3 Search Using Title Information Title information is useful for document retrieval. In the system Citeseer, for instance, Giles et al. managed to extract titles from research papers and make use of the extracted titles in metadata search of papers [8]. In web search, the title fields (i.e., file properties) and anchor texts of web pages (HTML documents) can be viewed as ‘titles" of the pages [5]. Many search engines seem to utilize them for web page retrieval [7, 11, 18, 22]. Zhang et al., found that web pages with well-defined metadata are more easily retrieved than those without well-defined metadata [24]. To the best of our knowledge, no research has been conducted on using extracted titles from general documents (e.g., Office documents) for search of the documents. 146 3. MOTIVATION AND PROBLEM SETTING We consider the issue of automatically extracting titles from general documents. By general documents, we mean documents that belong to one of any number of specific genres. The documents can be presentations, books, book chapters, technical papers, brochures, reports, memos, specifications, letters, announcements, or resumes. General documents are more widely available in digital libraries, intranets, and internet, and thus investigation on title extraction from them is sorely needed. Figure 1 shows an estimate on distributions of file formats on intranet and internet [15]. Office and PDF are the main file formats on the intranet. Even on the internet, the documents in the formats are still not negligible, given its extremely large size. In this paper, without loss of generality, we take Office documents as an example. Figure 1. Distributions of file formats in internet and intranet. For Office documents, users can define titles as file properties using a feature provided by Office. We found in an experiment, however, that users seldom use the feature and thus titles in file properties are usually very inaccurate. That is to say, titles in file properties are usually inconsistent with the ‘true" titles in the file bodies that are created by the authors and are visible to readers. We collected 6,000 Word and 6,000 PowerPoint documents from an intranet and the internet and examined how many titles in the file properties are correct. We found that surprisingly the accuracy was only 0.265 (cf., Section 6.3 for details). A number of reasons can be considered. For example, if one creates a new file by copying an old file, then the file property of the new file will also be copied from the old file. In another experiment, we found that Google uses the titles in file properties of Office documents in search and browsing, but the titles are not very accurate. We created 50 queries to search Word and PowerPoint documents and examined the top 15 results of each query returned by Google. We found that nearly all the titles presented in the search results were from the file properties of the documents. However, only 0.272 of them were correct. Actually, ‘true" titles usually exist at the beginnings of the bodies of documents. If we can accurately extract the titles from the bodies of documents, then we can exploit reliable title information in document processing. This is exactly the problem we address in this paper. More specifically, given a Word document, we are to extract the title from the top region of the first page. Given a PowerPoint document, we are to extract the title from the first slide. A title sometimes consists of a main title and one or two subtitles. We only consider extraction of the main title. As baselines for title extraction, we use that of always using the first lines as titles and that of always using the lines with largest font sizes as titles. Figure 2. Title extraction from Word document. Figure 3. Title extraction from PowerPoint document. Next, we define a ‘specification" for human judgments in title data annotation. The annotated data will be used in training and testing of the title extraction methods. Summary of the specification: The title of a document should be identified on the basis of common sense, if there is no difficulty in the identification. However, there are many cases in which the identification is not easy. There are some rules defined in the specification that guide identification for such cases. The rules include a title is usually in consecutive lines in the same format, a document can have no title, titles in images are not considered, a title should not contain words like ‘draft", 147 ‘whitepaper", etc, if it is difficult to determine which is the title, select the one in the largest font size, and if it is still difficult to determine which is the title, select the first candidate. (The specification covers all the cases we have encountered in data annotation.) Figures 2 and 3 show examples of Office documents from which we conduct title extraction. In Figure 2, ‘Differences in Win32 API Implementations among Windows Operating Systems" is the title of the Word document. ‘Microsoft Windows" on the top of this page is a picture and thus is ignored. In Figure 3, ‘Building Competitive Advantages through an Agile Infrastructure" is the title of the PowerPoint document. We have developed a tool for annotation of titles by human annotators. Figure 4 shows a snapshot of the tool. Figure 4. Title annotation tool. 4. TITLE EXTRACTION METHOD 4.1 Outline Title extraction based on machine learning consists of training and extraction. The same pre-processing step occurs before training and extraction. During pre-processing, from the top region of the first page of a Word document or the first slide of a PowerPoint document a number of units for processing are extracted. If a line (lines are separated by ‘return" symbols) only has a single format, then the line will become a unit. If a line has several parts and each of them has its own format, then each part will become a unit. Each unit will be treated as an instance in learning. A unit contains not only content information (linguistic information) but also formatting information. The input to pre-processing is a document and the output of pre-processing is a sequence of units (instances). Figure 5 shows the units obtained from the document in Figure 2. Figure 5. Example of units. In learning, the input is sequences of units where each sequence corresponds to a document. We take labeled units (labeled as title_begin, title_end, or other) in the sequences as training data and construct models for identifying whether a unit is title_begin title_end, or other. We employ four types of models: Perceptron, Maximum Entropy (ME), Perceptron Markov Model (PMM), and Maximum Entropy Markov Model (MEMM). In extraction, the input is a sequence of units from one document. We employ one type of model to identify whether a unit is title_begin, title_end, or other. We then extract units from the unit labeled with ‘title_begin" to the unit labeled with ‘title_end". The result is the extracted title of the document. The unique characteristic of our approach is that we mainly utilize formatting information for title extraction. Our assumption is that although general documents vary in styles, their formats have certain patterns and we can learn and utilize the patterns for title extraction. This is in contrast to the work by Han et al., in which only linguistic features are used for extraction from research papers. 4.2 Models The four models actually can be considered in the same metadata extraction framework. That is why we apply them together to our current problem. Each input is a sequence of instances kxxx L21 together with a sequence of labels kyyy L21 . ix and iy represents an instance and its label, respectively ( ki ,,2,1 L= ). Recall that an instance here represents a unit. A label represents title_begin, title_end, or other. Here, k is the number of units in a document. In learning, we train a model which can be generally denoted as a conditional probability distribution )|( 11 kk XXYYP LL where iX and iY denote random variables taking instance ix and label iy as values, respectively ( ki ,,2,1 L= ). Learning Tool Extraction Tool 21121 2222122221 1121111211 nknnknn kk kk yyyxxx yyyxxx yyyxxx LL LL LL LL → → → )|(maxarg 11 mkmmkm xxyyP LL )|( 11 kk XXYYP LL Conditional Distribution mkmm xxx L21 Figure 6. Metadata extraction model. We can make assumptions about the general model in order to make it simple enough for training. 148 For example, we can assume that kYY ,,1 L are independent of each other given kXX ,,1 L . Thus, we have )|()|( )|( 11 11 kk kk XYPXYP XXYYP L LL = In this way, we decompose the model into a number of classifiers. We train the classifiers locally using the labeled data. As the classifier, we employ the Perceptron or Maximum Entropy model. We can also assume that the first order Markov property holds for kYY ,,1 L given kXX ,,1 L . Thus, we have )|()|( )|( 111 11 kkk kk XYYPXYP XXYYP −= L LL Again, we obtain a number of classifiers. However, the classifiers are conditioned on the previous label. When we employ the Percepton or Maximum Entropy model as a classifier, the models become a Percepton Markov Model or Maximum Entropy Markov Model, respectively. That is to say, the two models are more precise. In extraction, given a new sequence of instances, we resort to one of the constructed models to assign a sequence of labels to the sequence of instances, i.e., perform extraction. For Perceptron and ME, we assign labels locally and combine the results globally later using heuristics. Specifically, we first identify the most likely title_begin. Then we find the most likely title_end within three units after the title_begin. Finally, we extract as a title the units between the title_begin and the title_end. For PMM and MEMM, we employ the Viterbi algorithm to find the globally optimal label sequence. In this paper, for Perceptron, we actually employ an improved variant of it, called Perceptron with Uneven Margin [13]. This version of Perceptron can work well especially when the number of positive instances and the number of negative instances differ greatly, which is exactly the case in our problem. We also employ an improved version of Perceptron Markov Model in which the Perceptron model is the so-called Voted Perceptron [2]. In addition, in training, the parameters of the model are updated globally rather than locally. 4.3 Features There are two types of features: format features and linguistic features. We mainly use the former. The features are used for both the title-begin and the title-end classifiers. 4.3.1 Format Features Font Size: There are four binary features that represent the normalized font size of the unit (recall that a unit has only one type of font). If the font size of the unit is the largest in the document, then the first feature will be 1, otherwise 0. If the font size is the smallest in the document, then the fourth feature will be 1, otherwise 0. If the font size is above the average font size and not the largest in the document, then the second feature will be 1, otherwise 0. If the font size is below the average font size and not the smallest, the third feature will be 1, otherwise 0. It is necessary to conduct normalization on font sizes. For example, in one document the largest font size might be ‘12pt", while in another the smallest one might be ‘18pt". Boldface: This binary feature represents whether or not the current unit is in boldface. Alignment: There are four binary features that respectively represent the location of the current unit: ‘left", ‘center", ‘right", and ‘unknown alignment". The following format features with respect to ‘context" play an important role in title extraction. Empty Neighboring Unit: There are two binary features that represent, respectively, whether or not the previous unit and the current unit are blank lines. Font Size Change: There are two binary features that represent, respectively, whether or not the font size of the previous unit and the font size of the next unit differ from that of the current unit. Alignment Change: There are two binary features that represent, respectively, whether or not the alignment of the previous unit and the alignment of the next unit differ from that of the current one. Same Paragraph: There are two binary features that represent, respectively, whether or not the previous unit and the next unit are in the same paragraph as the current unit. 4.3.2 Linguistic Features The linguistic features are based on key words. Positive Word: This binary feature represents whether or not the current unit begins with one of the positive words. The positive words include ‘title:", ‘subject:", ‘subject line:" For example, in some documents the lines of titles and authors have the same formats. However, if lines begin with one of the positive words, then it is likely that they are title lines. Negative Word: This binary feature represents whether or not the current unit begins with one of the negative words. The negative words include ‘To", ‘By", ‘created by", ‘updated by", etc. There are more negative words than positive words. The above linguistic features are language dependent. Word Count: A title should not be too long. We heuristically create four intervals: [1, 2], [3, 6], [7, 9] and [9, ∞) and define one feature for each interval. If the number of words in a title falls into an interval, then the corresponding feature will be 1; otherwise 0. Ending Character: This feature represents whether the unit ends with ‘:", ‘-", or other special characters. A title usually does not end with such a character. 5. DOCUMENT RETRIEVAL METHOD We describe our method of document retrieval using extracted titles. Typically, in information retrieval a document is split into a number of fields including body, title, and anchor text. A ranking function in search can use different weights for different fields of 149 the document. Also, titles are typically assigned high weights, indicating that they are important for document retrieval. As explained previously, our experiment has shown that a significant number of documents actually have incorrect titles in the file properties, and thus in addition of using them we use the extracted titles as one more field of the document. By doing this, we attempt to improve the overall precision. In this paper, we employ a modification of BM25 that allows field weighting [21]. As fields, we make use of body, title, extracted title and anchor. First, for each term in the query we count the term frequency in each field of the document; each field frequency is then weighted according to the corresponding weight parameter: ∑= f tfft tfwwtf Similarly, we compute the document length as a weighted sum of lengths of each field. Average document length in the corpus becomes the average of all weighted document lengths. ∑= f ff dlwwdl In our experiments we used 75.0,8.11 == bk . Weight for content was 1.0, title was 10.0, anchor was 10.0, and extracted title was 5.0. 6. EXPERIMENTAL RESULTS 6.1 Data Sets and Evaluation Measures We used two data sets in our experiments. First, we downloaded and randomly selected 5,000 Word documents and 5,000 PowerPoint documents from an intranet of Microsoft. We call it MS hereafter. Second, we downloaded and randomly selected 500 Word and 500 PowerPoint documents from the DotGov and DotCom domains on the internet, respectively. Figure 7 shows the distributions of the genres of the documents. We see that the documents are indeed ‘general documents" as we define them. Figure 7. Distributions of document genres. Third, a data set in Chinese was also downloaded from the internet. It includes 500 Word documents and 500 PowerPoint documents in Chinese. We manually labeled the titles of all the documents, on the basis of our specification. Not all the documents in the two data sets have titles. Table 1 shows the percentages of the documents having titles. We see that DotCom and DotGov have more PowerPoint documents with titles than MS. This might be because PowerPoint documents published on the internet are more formal than those on the intranet. Table 1. The portion of documents with titles Domain Type MS DotCom DotGov Word 75.7% 77.8% 75.6% PowerPoint 82.1% 93.4% 96.4% In our experiments, we conducted evaluations on title extraction in terms of precision, recall, and F-measure. The evaluation measures are defined as follows: Precision: P = A / ( A + B ) Recall: R = A / ( A + C ) F-measure: F1 = 2PR / ( P + R ) Here, A, B, C, and D are numbers of documents as those defined in Table 2. Table 2. Contingence table with regard to title extraction Is title Is not title Extracted A B Not extracted C D 6.2 Baselines We test the accuracies of the two baselines described in section 4.2. They are denoted as ‘largest font size" and ‘first line" respectively. 6.3 Accuracy of Titles in File Properties We investigate how many titles in the file properties of the documents are reliable. We view the titles annotated by humans as true titles and test how many titles in the file properties can approximately match with the true titles. We use Edit Distance to conduct the approximate match. (Approximate match is only used in this evaluation). This is because sometimes human annotated titles can be slightly different from the titles in file properties on the surface, e.g., contain extra spaces). Given string A and string B: if ( (D == 0) or ( D / ( La + Lb ) < θ ) ) then string A = string B D: Edit Distance between string A and string B La: length of string A Lb: length of string B θ: 0.1 ∑ × ++− + = t t n N wtf avwdl wdl bbk kwtf FBM )log( ))1(( )1( 25 1 1 150 Table 3. Accuracies of titles in file properties File Type Domain Precision Recall F1 MS 0.299 0.311 0.305 DotCom 0.210 0.214 0.212Word DotGov 0.182 0.177 0.180 MS 0.229 0.245 0.237 DotCom 0.185 0.186 0.186PowerPoint DotGov 0.180 0.182 0.181 6.4 Comparison with Baselines We conducted title extraction from the first data set (Word and PowerPoint in MS). As the model, we used Perceptron. We conduct 4-fold cross validation. Thus, all the results reported here are those averaged over 4 trials. Tables 4 and 5 show the results. We see that Perceptron significantly outperforms the baselines. In the evaluation, we use exact matching between the true titles annotated by humans and the extracted titles. Table 4. Accuracies of title extraction with Word Precision Recall F1 Model Perceptron 0.810 0.837 0.823 Largest font size 0.700 0.758 0.727 Baselines First line 0.707 0.767 0.736 Table 5. Accuracies of title extraction with PowerPoint Precision Recall F1 Model Perceptron 0.875 0. 895 0.885 Largest font size 0.844 0.887 0.865 Baselines First line 0.639 0.671 0.655 We see that the machine learning approach can achieve good performance in title extraction. For Word documents both precision and recall of the approach are 8 percent higher than those of the baselines. For PowerPoint both precision and recall of the approach are 2 percent higher than those of the baselines. We conduct significance tests. The results are shown in Table 6. Here, ‘Largest" denotes the baseline of using the largest font size, ‘First" denotes the baseline of using the first line. The results indicate that the improvements of machine learning over baselines are statistically significant (in the sense p-value < 0.05) Table 6. Sign test results Documents Type Sign test between p-value Perceptron vs. Largest 3.59e-26 Word Perceptron vs. First 7.12e-10 Perceptron vs. Largest 0.010 PowerPoint Perceptron vs. First 5.13e-40 We see, from the results, that the two baselines can work well for title extraction, suggesting that font size and position information are most useful features for title extraction. However, it is also obvious that using only these two features is not enough. There are cases in which all the lines have the same font size (i.e., the largest font size), or cases in which the lines with the largest font size only contain general descriptions like ‘Confidential", ‘White paper", etc. For those cases, the ‘largest font size" method cannot work well. For similar reasons, the ‘first line" method alone cannot work well, either. With the combination of different features (evidence in title judgment), Perceptron can outperform Largest and First. We investigate the performance of solely using linguistic features. We found that it does not work well. It seems that the format features play important roles and the linguistic features are supplements.. Figure 8. An example Word document. Figure 9. An example PowerPoint document. We conducted an error analysis on the results of Perceptron. We found that the errors fell into three categories. (1) About one third of the errors were related to ‘hard cases". In these documents, the layouts of the first pages were difficult to understand, even for humans. Figure 8 and 9 shows examples. (2) Nearly one fourth of the errors were from the documents which do not have true titles but only contain bullets. Since we conduct extraction from the top regions, it is difficult to get rid of these errors with the current approach. (3). Confusions between main titles and subtitles were another type of error. Since we only labeled the main titles as titles, the extractions of both titles were considered incorrect. This type of error does little harm to document processing like search, however. 6.5 Comparison between Models To compare the performance of different machine learning models, we conducted another experiment. Again, we perform 4-fold cross 151 validation on the first data set (MS). Table 7, 8 shows the results of all the four models. It turns out that Perceptron and PMM perform the best, followed by MEMM, and ME performs the worst. In general, the Markovian models perform better than or as well as their classifier counterparts. This seems to be because the Markovian models are trained globally, while the classifiers are trained locally. The Perceptron based models perform better than the ME based counterparts. This seems to be because the Perceptron based models are created to make better classifications, while ME models are constructed for better prediction. Table 7. Comparison between different learning models for title extraction with Word Model Precision Recall F1 Perceptron 0.810 0.837 0.823 MEMM 0.797 0.824 0.810 PMM 0.827 0.823 0.825 ME 0.801 0.621 0.699 Table 8. Comparison between different learning models for title extraction with PowerPoint Model Precision Recall F1 Perceptron 0.875 0. 895 0. 885 MEMM 0.841 0.861 0.851 PMM 0.873 0.896 0.885 ME 0.753 0.766 0.759 6.6 Domain Adaptation We apply the model trained with the first data set (MS) to the second data set (DotCom and DotGov). Tables 9-12 show the results. Table 9. Accuracies of title extraction with Word in DotGov Precision Recall F1 Model Perceptron 0.716 0.759 0.737 Largest font size 0.549 0.619 0.582Baselines First line 0.462 0.521 0.490 Table 10. Accuracies of title extraction with PowerPoint in DotGov Precision Recall F1 Model Perceptron 0.900 0.906 0.903 Largest font size 0.871 0.888 0.879Baselines First line 0.554 0.564 0.559 Table 11. Accuracies of title extraction with Word in DotCom Precisio n Recall F1 Model Perceptron 0.832 0.880 0.855 Largest font size 0.676 0.753 0.712Baselines First line 0.577 0.643 0.608 Table 12. Performance of PowerPoint document title extraction in DotCom Precisio n Recall F1 Model Perceptron 0.910 0.903 0.907 Largest font size 0.864 0.886 0.875Baselines First line 0.570 0.585 0.577 From the results, we see that the models can be adapted to different domains well. There is almost no drop in accuracy. The results indicate that the patterns of title formats exist across different domains, and it is possible to construct a domain independent model by mainly using formatting information. 6.7 Language Adaptation We apply the model trained with the data in English (MS) to the data set in Chinese. Tables 13-14 show the results. Table 13. Accuracies of title extraction with Word in Chinese Precision Recall F1 Model Perceptron 0.817 0.805 0.811 Largest font size 0.722 0.755 0.738Baselines First line 0.743 0.777 0.760 Table 14. Accuracies of title extraction with PowerPoint in Chinese Precision Recall F1 Model Perceptron 0.766 0.812 0.789 Largest font size 0.753 0.813 0.782Baselines First line 0.627 0.676 0.650 We see that the models can be adapted to a different language. There are only small drops in accuracy. Obviously, the linguistic features do not work for Chinese, but the effect of not using them is negligible. The results indicate that the patterns of title formats exist across different languages. From the domain adaptation and language adaptation results, we conclude that the use of formatting information is the key to a successful extraction from general documents. 6.8 Search with Extracted Titles We performed experiments on using title extraction for document retrieval. As a baseline, we employed BM25 without using extracted titles. The ranking mechanism was as described in Section 5. The weights were heuristically set. We did not conduct optimization on the weights. The evaluation was conducted on a corpus of 1.3 M documents crawled from the intranet of Microsoft using 100 evaluation queries obtained from this intranet"s search engine query logs. 50 queries were from the most popular set, while 50 queries other were chosen randomly. Users were asked to provide judgments of the degree of document relevance from a scale of 1to 5 (1 meaning detrimental, 2 - bad, 3 - fair, 4 - good and 5 - excellent). 152 Figure 10 shows the results. In the chart two sets of precision results were obtained by either considering good or excellent documents as relevant (left 3 bars with relevance threshold 0.5), or by considering only excellent documents as relevant (right 3 bars with relevance threshold 1.0) 0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 P@10 P@5 Reciprocal P@10 P@5 Reciprocal 0.5 1 BM25 Anchor, Title, Body BM25 Anchor, Title, Body, ExtractedTitle Name All RelevanceThreshold Data Description Figure 10. Search ranking results. Figure 10 shows different document retrieval results with different ranking functions in terms of precision @10, precision @5 and reciprocal rank: • Blue bar - BM25 including the fields body, title (file property), and anchor text. • Purple bar - BM25 including the fields body, title (file property), anchor text, and extracted title. With the additional field of extracted title included in BM25 the precision @10 increased from 0.132 to 0.145, or by ~10%. Thus, it is safe to say that the use of extracted title can indeed improve the precision of document retrieval. 7. CONCLUSION In this paper, we have investigated the problem of automatically extracting titles from general documents. We have tried using a machine learning approach to address the problem. Previous work showed that the machine learning approach can work well for metadata extraction from research papers. In this paper, we showed that the approach can work for extraction from general documents as well. Our experimental results indicated that the machine learning approach can work significantly better than the baselines in title extraction from Office documents. Previous work on metadata extraction mainly used linguistic features in documents, while we mainly used formatting information. It appeared that using formatting information is a key for successfully conducting title extraction from general documents. We tried different machine learning models including Perceptron, Maximum Entropy, Maximum Entropy Markov Model, and Voted Perceptron. We found that the performance of the Perceptorn models was the best. We applied models constructed in one domain to another domain and applied models trained in one language to another language. We found that the accuracies did not drop substantially across different domains and across different languages, indicating that the models were generic. We also attempted to use the extracted titles in document retrieval. We observed a significant improvement in document ranking performance for search when using extracted title information. All the above investigations were not conducted in previous work, and through our investigations we verified the generality and the significance of the title extraction approach. 8. ACKNOWLEDGEMENTS We thank Chunyu Wei and Bojuan Zhao for their work on data annotation. We acknowledge Jinzhu Li for his assistance in conducting the experiments. We thank Ming Zhou, John Chen, Jun Xu, and the anonymous reviewers of JCDL"05 for their valuable comments on this paper. 9. REFERENCES [1] Berger, A. L., Della Pietra, S. A., and Della Pietra, V. J. A maximum entropy approach to natural language processing. Computational Linguistics, 22:39-71, 1996. [2] Collins, M. Discriminative training methods for hidden markov models: theory and experiments with perceptron algorithms. In Proceedings of Conference on Empirical Methods in Natural Language Processing, 1-8, 2002. [3] Cortes, C. and Vapnik, V. Support-vector networks. Machine Learning, 20:273-297, 1995. [4] Chieu, H. L. and Ng, H. T. A maximum entropy approach to information extraction from semi-structured and free text. In Proceedings of the Eighteenth National Conference on Artificial Intelligence, 768-791, 2002. [5] Evans, D. K., Klavans, J. L., and McKeown, K. R. Columbia newsblaster: multilingual news summarization on the Web. In Proceedings of Human Language Technology conference / North American chapter of the Association for Computational Linguistics annual meeting, 1-4, 2004. [6] Ghahramani, Z. and Jordan, M. I. Factorial hidden markov models. Machine Learning, 29:245-273, 1997. [7] Gheel, J. and Anderson, T. Data and metadata for finding and reminding, In Proceedings of the 1999 International Conference on Information Visualization, 446-451,1999. [8] Giles, C. L., Petinot, Y., Teregowda P. B., Han, H., Lawrence, S., Rangaswamy, A., and Pal, N. eBizSearch: a niche search engine for e-Business. In Proceedings of the 26th Annual International ACM SIGIR Conference on Research and Development in Information Retrieval, 413414, 2003. [9] Giuffrida, G., Shek, E. C., and Yang, J. Knowledge-based metadata extraction from PostScript files. In Proceedings of the Fifth ACM Conference on Digital Libraries, 77-84, 2000. [10] Han, H., Giles, C. L., Manavoglu, E., Zha, H., Zhang, Z., and Fox, E. A. Automatic document metadata extraction using support vector machines. In Proceedings of the Third ACM/IEEE-CS Joint Conference on Digital Libraries, 37-48, 2003. [11] Kobayashi, M., and Takeda, K. Information retrieval on the Web. ACM Computing Surveys, 32:144-173, 2000. [12] Lafferty, J., McCallum, A., and Pereira, F. Conditional random fields: probabilistic models for segmenting and 153 labeling sequence data. In Proceedings of the Eighteenth International Conference on Machine Learning, 282-289, 2001. [13] Li, Y., Zaragoza, H., Herbrich, R., Shawe-Taylor J., and Kandola, J. S. The perceptron algorithm with uneven margins. In Proceedings of the Nineteenth International Conference on Machine Learning, 379-386, 2002. [14] Liddy, E. D., Sutton, S., Allen, E., Harwell, S., Corieri, S., Yilmazel, O., Ozgencil, N. E., Diekema, A., McCracken, N., and Silverstein, J. Automatic Metadata generation & evaluation. In Proceedings of the 25th Annual International ACM SIGIR Conference on Research and Development in Information Retrieval, 401-402, 2002. [15] Littlefield, A. Effective enterprise information retrieval across new content formats. In Proceedings of the Seventh Search Engine Conference, http://www.infonortics.com/searchengines/sh02/02prog.html, 2002. [16] Mao, S., Kim, J. W., and Thoma, G. R. A dynamic feature generation system for automated metadata extraction in preservation of digital materials. In Proceedings of the First International Workshop on Document Image Analysis for Libraries, 225-232, 2004. [17] McCallum, A., Freitag, D., and Pereira, F. Maximum entropy markov models for information extraction and segmentation. In Proceedings of the Seventeenth International Conference on Machine Learning, 591-598, 2000. [18] Murphy, L. D. Digital document metadata in organizations: roles, analytical approaches, and future research directions. In Proceedings of the Thirty-First Annual Hawaii International Conference on System Sciences, 267-276, 1998. [19] Pinto, D., McCallum, A., Wei, X., and Croft, W. B. Table extraction using conditional random fields. In Proceedings of the 26th Annual International ACM SIGIR Conference on Research and Development in Information Retrieval, 235242, 2003. [20] Ratnaparkhi, A. Unsupervised statistical models for prepositional phrase attachment. In Proceedings of the Seventeenth International Conference on Computational Linguistics. 1079-1085, 1998. [21] Robertson, S., Zaragoza, H., and Taylor, M. Simple BM25 extension to multiple weighted fields, In Proceedings of ACM Thirteenth Conference on Information and Knowledge Management, 42-49, 2004. [22] Yi, J. and Sundaresan, N. Metadata based Web mining for relevance, In Proceedings of the 2000 International Symposium on Database Engineering & Applications, 113121, 2000. [23] Yilmazel, O., Finneran, C. M., and Liddy, E. D. MetaExtract: An NLP system to automatically assign metadata. In Proceedings of the 2004 Joint ACM/IEEE Conference on Digital Libraries, 241-242, 2004. [24] Zhang, J. and Dimitroff, A. Internet search engines' response to metadata Dublin Core implementation. Journal of Information Science, 30:310-320, 2004. [25] Zhang, L., Pan, Y., and Zhang, T. Recognising and using named entities: focused named entity recognition using machine learning. In Proceedings of the 27th Annual International ACM SIGIR Conference on Research and Development in Information Retrieval, 281-288, 2004. [26] http://dublincore.org/groups/corporate/Seattle/ 154

machine learn;formatting information;metada of document;title extraction;language independence;linguistic feature;usefulness of extracted title;extracted title usefulness;genre;classifier;automatic title extraction;information extraction;model generality;metada extraction;search;document retrieval;comparison between model;generality of model

train_H-79

Beyond PageRank: Machine Learning for Static Ranking

Since the publication of Brin and Page"s paper on PageRank, many in the Web community have depended on PageRank for the static (query-independent) ordering of Web pages. We show that we can significantly outperform PageRank using features that are independent of the link structure of the Web. We gain a further boost in accuracy by using data on the frequency at which users visit Web pages. We use RankNet, a ranking machine learning algorithm, to combine these and other static features based on anchor text and domain characteristics. The resulting model achieves a static ranking pairwise accuracy of 67.3% (vs. 56.7% for PageRank or 50% for random).

1. INTRODUCTION Over the past decade, the Web has grown exponentially in size. Unfortunately, this growth has not been isolated to good-quality pages. The number of incorrect, spamming, and malicious (e.g., phishing) sites has also grown rapidly. The sheer number of both good and bad pages on the Web has led to an increasing reliance on search engines for the discovery of useful information. Users rely on search engines not only to return pages related to their search query, but also to separate the good from the bad, and order results so that the best pages are suggested first. To date, most work on Web page ranking has focused on improving the ordering of the results returned to the user (querydependent ranking, or dynamic ranking). However, having a good query-independent ranking (static ranking) is also crucially important for a search engine. A good static ranking algorithm provides numerous benefits: • Relevance: The static rank of a page provides a general indicator to the overall quality of the page. This is a useful input to the dynamic ranking algorithm. • Efficiency: Typically, the search engine"s index is ordered by static rank. By traversing the index from highquality to low-quality pages, the dynamic ranker may abort the search when it determines that no later page will have as high of a dynamic rank as those already found. The more accurate the static rank, the better this early-stopping ability, and hence the quicker the search engine may respond to queries. • Crawl Priority: The Web grows and changes as quickly as search engines can crawl it. Search engines need a way to prioritize their crawl-to determine which pages to recrawl, how frequently, and how often to seek out new pages. Among other factors, the static rank of a page is used to determine this prioritization. A better static rank thus provides the engine with a higher quality, more upto-date index. Google is often regarded as the first commercially successful search engine. Their ranking was originally based on the PageRank algorithm [5][27]. Due to this (and possibly due to Google"s promotion of PageRank to the public), PageRank is widely regarded as the best method for the static ranking of Web pages. Though PageRank has historically been thought to perform quite well, there has yet been little academic evidence to support this claim. Even worse, there has recently been work showing that PageRank may not perform any better than other simple measures on certain tasks. Upstill et al. have found that for the task of finding home pages, the number of pages linking to a page and the type of URL were as, or more, effective than PageRank [32]. They found similar results for the task of finding high quality companies [31]. PageRank has also been used in systems for TREC"s very large collection and Web track competitions, but with much less success than had been expected [17]. Finally, Amento et al. [1] found that simple features, such as the number of pages on a site, performed as well as PageRank. Despite these, the general belief remains among many, both academic and in the public, that PageRank is an essential factor for a good static rank. Failing this, it is still assumed that using the link structure is crucial, in the form of the number of inlinks or the amount of anchor text. In this paper, we show there are a number of simple url- or pagebased features that significantly outperform PageRank (for the purposes of statically ranking Web pages) despite ignoring the structure of the Web. We combine these and other static features using machine learning to achieve a ranking system that is significantly better than PageRank (in pairwise agreement with human labels). A machine learning approach for static ranking has other advantages besides the quality of the ranking. Because the measure consists of many features, it is harder for malicious users to manipulate it (i.e., to raise their page"s static rank to an undeserved level through questionable techniques, also known as Web spamming). This is particularly true if the feature set is not known. In contrast, a single measure like PageRank can be easier to manipulate because spammers need only concentrate on one goal: how to cause more pages to point to their page. With an algorithm that learns, a feature that becomes unusable due to spammer manipulation will simply be reduced or removed from the final computation of rank. This flexibility allows a ranking system to rapidly react to new spamming techniques. A machine learning approach to static ranking is also able to take advantage of any advances in the machine learning field. For example, recent work on adversarial classification [12] suggests that it may be possible to explicitly model the Web page spammer"s (the adversary) actions, adjusting the ranking model in advance of the spammer"s attempts to circumvent it. Another example is the elimination of outliers in constructing the model, which helps reduce the effect that unique sites may have on the overall quality of the static rank. By moving static ranking to a machine learning framework, we not only gain in accuracy, but also gain in the ability to react to spammer"s actions, to rapidly add new features to the ranking algorithm, and to leverage advances in the rapidly growing field of machine learning. Finally, we believe there will be significant advantages to using this technique for other domains, such as searching a local hard drive or a corporation"s intranet. These are domains where the link structure is particularly weak (or non-existent), but there are other domain-specific features that could be just as powerful. For example, the author of an intranet page and his/her position in the organization (e.g., CEO, manager, or developer) could provide significant clues as to the importance of that page. A machine learning approach thus allows rapid development of a good static algorithm in new domains. This paper"s contribution is a systematic study of static features, including PageRank, for the purposes of (statically) ranking Web pages. Previous studies on PageRank typically used subsets of the Web that are significantly smaller (e.g., the TREC VLC2 corpus, used by many, contains only 19 million pages). Also, the performance of PageRank and other static features has typically been evaluated in the context of a complete system for dynamic ranking, or for other tasks such as question answering. In contrast, we explore the use of PageRank and other features for the direct task of statically ranking Web pages. We first briefly describe the PageRank algorithm. In Section 3 we introduce RankNet, the machine learning technique used to combine static features into a final ranking. Section 4 describes the static features. The heart of the paper is in Section 5, which presents our experiments and results. We conclude with a discussion of related and future work. 2. PAGERANK The basic idea behind PageRank is simple: a link from a Web page to another can be seen as an endorsement of that page. In general, links are made by people. As such, they are indicative of the quality of the pages to which they point - when creating a page, an author presumably chooses to link to pages deemed to be of good quality. We can take advantage of this linkage information to order Web pages according to their perceived quality. Imagine a Web surfer who jumps from Web page to Web page, choosing with uniform probability which link to follow at each step. In order to reduce the effect of dead-ends or endless cycles the surfer will occasionally jump to a random page with some small probability α, or when on a page with no out-links. If averaged over a sufficient number of steps, the probability the surfer is on page j at some point in time is given by the formula: ∑∈ + − = ji i iP N jP B F )()1( )( α α (1) Where Fi is the set of pages that page i links to, and Bj is the set of pages that link to page j. The PageRank score for node j is defined as this probability: PR(j)=P(j). Because equation (1) is recursive, it must be iteratively evaluated until P(j) converges (typically, the initial distribution for P(j) is uniform). The intuition is, because a random surfer would end up at the page more frequently, it is likely a better page. An alternative view for equation (1) is that each page is assigned a quality, P(j). A page gives an equal share of its quality to each page it points to. PageRank is computationally expensive. Our collection of 5 billion pages contains approximately 370 billion links. Computing PageRank requires iterating over these billions of links multiple times (until convergence). It requires large amounts of memory (or very smart caching schemes that slow the computation down even further), and if spread across multiple machines, requires significant communication between them. Though much work has been done on optimizing the PageRank computation (see e.g., [25] and [6]), it remains a relatively slow, computationally expensive property to compute. 3. RANKNET Much work in machine learning has been done on the problems of classification and regression. Let X={xi} be a collection of feature vectors (typically, a feature is any real valued number), and Y={yi} be a collection of associated classes, where yi is the class of the object described by feature vector xi. The classification problem is to learn a function f that maps yi=f(xi), for all i. When yi is real-valued as well, this is called regression. Static ranking can be seen as a regression problem. If we let xi represent features of page i, and yi be a value (say, the rank) for each page, we could learn a regression function that mapped each page"s features to their rank. However, this over-constrains the problem we wish to solve. All we really care about is the order of the pages, not the actual value assigned to them. Recent work on this ranking problem [7][13][18] directly attempts to optimize the ordering of the objects, rather than the value assigned to them. For these, let Z={<i,j>} be a collection of pairs of items, where item i should be assigned a higher value than item j. The goal of the ranking problem, then, is to learn a function f such that, )()(,, ji ffji xxZ >∈∀ 708 Note that, as with learning a regression function, the result of this process is a function (f) that maps feature vectors to real values. This function can still be applied anywhere that a regressionlearned function could be applied. The only difference is the technique used to learn the function. By directly optimizing the ordering of objects, these methods are able to learn a function that does a better job of ranking than do regression techniques. We used RankNet [7], one of the aforementioned techniques for learning ranking functions, to learn our static rank function. RankNet is a straightforward modification to the standard neural network back-prop algorithm. As with back-prop, RankNet attempts to minimize the value of a cost function by adjusting each weight in the network according to the gradient of the cost function with respect to that weight. The difference is that, while a typical neural network cost function is based on the difference between the network output and the desired output, the RankNet cost function is based on the difference between a pair of network outputs. That is, for each pair of feature vectors <i,j> in the training set, RankNet computes the network outputs oi and oj. Since vector i is supposed to be ranked higher than vector j, the larger is oj-oi, the larger the cost. RankNet also allows the pairs in Z to be weighted with a confidence (posed as the probability that the pair satisfies the ordering induced by the ranking function). In this paper, we used a probability of one for all pairs. In the next section, we will discuss the features used in our feature vectors, xi. 4. FEATURES To apply RankNet (or other machine learning techniques) to the ranking problem, we needed to extract a set of features from each page. We divided our feature set into four, mutually exclusive, categories: page-level (Page), domain-level (Domain), anchor text and inlinks (Anchor), and popularity (Popularity). We also optionally used the PageRank of a page as a feature. Below, we describe each of these feature categories in more detail. PageRank We computed PageRank on a Web graph of 5 billion crawled pages (and 20 billion known URLs linked to by these pages). This represents a significant portion of the Web, and is approximately the same number of pages as are used by Google, Yahoo, and MSN for their search engines. Because PageRank is a graph-based algorithm, it is important that it be run on as large a subset of the Web as possible. Most previous studies on PageRank used subsets of the Web that are significantly smaller (e.g. the TREC VLC2 corpus, used by many, contains only 19 million pages) We computed PageRank using the standard value of 0.85 for α. Popularity Another feature we used is the actual popularity of a Web page, measured as the number of times that it has been visited by users over some period of time. We have access to such data from users who have installed the MSN toolbar and have opted to provide it to MSN. The data is aggregated into a count, for each Web page, of the number of users who viewed that page. Though popularity data is generally unavailable, there are two other sources for it. The first is from proxy logs. For example, a university that requires its students to use a proxy has a record of all the pages they have visited while on campus. Unfortunately, proxy data is quite biased and relatively small. Another source, internal to search engines, are records of which results their users clicked on. Such data was used by the search engine Direct Hit, and has recently been explored for dynamic ranking purposes [20]. An advantage of the toolbar data over this is that it contains information about URL visits that are not just the result of a search. The raw popularity is processed into a number of features such as the number of times a page was viewed and the number of times any page in the domain was viewed. More details are provided in section 5.5. Anchor text and inlinks These features are based on the information associated with links to the page in question. It includes features such as the total amount of text in links pointing to the page (anchor text), the number of unique words in that text, etc. Page This category consists of features which may be determined by looking at the page (and its URL) alone. We used only eight, simple features such as the number of words in the body, the frequency of the most common term, etc. Domain This category contains features that are computed as averages across all pages in the domain. For example, the average number of outlinks on any page and the average PageRank. Many of these features have been used by others for ranking Web pages, particularly the anchor and page features. As mentioned, the evaluation is typically for dynamic ranking, and we wish to evaluate the use of them for static ranking. Also, to our knowledge, this is the first study on the use of actual page visitation popularity for static ranking. The closest similar work is on using click-through behavior (that is, which search engine results the users click on) to affect dynamic ranking (see e.g., [20]). Because we use a wide variety of features to come up with a static ranking, we refer to this as fRank (for feature-based ranking). fRank uses RankNet and the set of features described in this section to learn a ranking function for Web pages. Unless otherwise specified, fRank was trained with all of the features. 5. EXPERIMENTS In this section, we will demonstrate that we can out perform PageRank by applying machine learning to a straightforward set of features. Before the results, we first discuss the data, the performance metric, and the training method. 5.1 Data In order to evaluate the quality of a static ranking, we needed a gold standard defining the correct ordering for a set of pages. For this, we employed a dataset which contains human judgments for 28000 queries. For each query, a number of results are manually assigned a rating, from 0 to 4, by human judges. The rating is meant to be a measure of how relevant the result is for the query, where 0 means poor and 4 means excellent. There are approximately 500k judgments in all, or an average of 18 ratings per query. The queries are selected by randomly choosing queries from among those issued to the MSN search engine. The probability that a query is selected is proportional to its frequency among all 709 of the queries. As a result, common queries are more likely to be judged than uncommon queries. As an example of how diverse the queries are, the first four queries in the training set are chef schools, chicagoland speedway, eagles fan club, and Turkish culture. The documents selected for judging are those that we expected would, on average, be reasonably relevant (for example, the top ten documents returned by MSN"s search engine). This provides significantly more information than randomly selecting documents on the Web, the vast majority of which would be irrelevant to a given query. Because of this process, the judged pages tend to be of higher quality than the average page on the Web, and tend to be pages that will be returned for common search queries. This bias is good when evaluating the quality of static ranking for the purposes of index ordering and returning relevant documents. This is because the most important portion of the index to be well-ordered and relevant is the portion that is frequently returned for search queries. Because of this bias, however, the results in this paper are not applicable to crawl prioritization. In order to obtain experimental results on crawl prioritization, we would need ratings on a random sample of Web pages. To convert the data from query-dependent to query-independent, we simply removed the query, taking the maximum over judgments for a URL that appears in more than one query. The reasoning behind this is that a page that is relevant for some query and irrelevant for another is probably a decent page and should have a high static rank. Because we evaluated the pages on queries that occur frequently, our data indicates the correct index ordering, and assigns high value to pages that are likely to be relevant to a common query. We randomly assigned queries to a training, validation, or test set, such that they contained 84%, 8%, and 8% of the queries, respectively. Each set contains all of the ratings for a given query, and no query appears in more than one set. The training set was used to train fRank. The validation set was used to select the model that had the highest performance. The test set was used for the final results. This data gives us a query-independent ordering of pages. The goal for a static ranking algorithm will be to reproduce this ordering as closely as possible. In the next section, we describe the measure we used to evaluate this. 5.2 Measure We chose to use pairwise accuracy to evaluate the quality of a static ranking. The pairwise accuracy is the fraction of time that the ranking algorithm and human judges agree on the ordering of a pair of Web pages. If S(x) is the static ranking assigned to page x, and H(x) is the human judgment of relevance for x, then consider the following sets: )}()(:,{ yHxHyx >=pH and )}()(:,{ ySxSyx >=pS The pairwise accuracy is the portion of Hp that is also contained in Sp: p pp H SH ∩ =accuracypairwise This measure was chosen for two reasons. First, the discrete human judgments provide only a partial ordering over Web pages, making it difficult to apply a measure such as the Spearman rank order correlation coefficient (in the pairwise accuracy measure, a pair of documents with the same human judgment does not affect the score). Second, the pairwise accuracy has an intuitive meaning: it is the fraction of pairs of documents that, when the humans claim one is better than the other, the static rank algorithm orders them correctly. 5.3 Method We trained fRank (a RankNet based neural network) using the following parameters. We used a fully connected 2 layer network. The hidden layer had 10 hidden nodes. The input weights to this layer were all initialized to be zero. The output layer (just a single node) weights were initialized using a uniform random distribution in the range [-0.1, 0.1]. We used tanh as the transfer function from the inputs to the hidden layer, and a linear function from the hidden layer to the output. The cost function is the pairwise cross entropy cost function as discussed in section 3. The features in the training set were normalized to have zero mean and unit standard deviation. The same linear transformation was then applied to the features in the validation and test sets. For training, we presented the network with 5 million pairings of pages, where one page had a higher rating than the other. The pairings were chosen uniformly at random (with replacement) from all possible pairings. When forming the pairs, we ignored the magnitude of the difference between the ratings (the rating spread) for the two URLs. Hence, the weight for each pair was constant (one), and the probability of a pair being selected was independent of its rating spread. We trained the network for 30 epochs. On each epoch, the training pairs were randomly shuffled. The initial training rate was 0.001. At each epoch, we checked the error on the training set. If the error had increased, then we decreased the training rate, under the hypothesis that the network had probably overshot. The training rate at each epoch was thus set to: Training rate = 1+ε κ Where κ is the initial rate (0.001), and ε is the number of times the training set error has increased. After each epoch, we measured the performance of the neural network on the validation set, using 1 million pairs (chosen randomly with replacement). The network with the highest pairwise accuracy on the validation set was selected, and then tested on the test set. We report the pairwise accuracy on the test set, calculated using all possible pairs. These parameters were determined and fixed before the static rank experiments in this paper. In particular, the choice of initial training rate, number of epochs, and training rate decay function were taken directly from Burges et al [7]. Though we had the option of preprocessing any of the features before they were input to the neural network, we refrained from doing so on most of them. The only exception was the popularity features. As with most Web phenomenon, we found that the distribution of site popularity is Zipfian. To reduce the dynamic range, and hopefully make the feature more useful, we presented the network with both the unpreprocessed, as well as the logarithm, of the popularity features (As with the others, the logarithmic feature values were also normalized to have zero mean and unit standard deviation). 710 Applying fRank to a document is computationally efficient, taking time that is only linear in the number of input features; it is thus within a constant factor of other simple machine learning methods such as naïve Bayes. In our experiments, computing the fRank for all five billion Web pages was approximately 100 times faster than computing the PageRank for the same set. 5.4 Results As Table 1 shows, fRank significantly outperforms PageRank for the purposes of static ranking. With a pairwise accuracy of 67.4%, fRank more than doubles the accuracy of PageRank (relative to the baseline of 50%, which is the accuracy that would be achieved by a random ordering of Web pages). Note that one of fRank"s input features is the PageRank of the page, so we would expect it to perform no worse than PageRank. The significant increase in accuracy implies that the other features (anchor, popularity, etc.) do in fact contain useful information regarding the overall quality of a page. Table 1: Basic Results Technique Accuracy (%) None (Baseline) 50.00 PageRank 56.70 fRank 67.43 There are a number of decisions that go into the computation of PageRank, such as how to deal with pages that have no outlinks, the choice of α, numeric precision, convergence threshold, etc. We were able to obtain a computation of PageRank from a completely independent implementation (provided by Marc Najork) that varied somewhat in these parameters. It achieved a pairwise accuracy of 56.52%, nearly identical to that obtained by our implementation. We thus concluded that the quality of the PageRank is not sensitive to these minor variations in algorithm, nor was PageRank"s low accuracy due to problems with our implementation of it. We also wanted to find how well each feature set performed. To answer this, for each feature set, we trained and tested fRank using only that set of features. The results are shown in Table 2. As can be seen, every single feature set individually outperformed PageRank on this test. Perhaps the most interesting result is that the Page-level features had the highest performance out of all the feature sets. This is surprising because these are features that do not depend on the overall graph structure of the Web, nor even on what pages point to a given page. This is contrary to the common belief that the Web graph structure is the key to finding a good static ranking of Web pages. Table 2: Results for individual feature sets. Feature Set Accuracy (%) PageRank 56.70 Popularity 60.82 Anchor 59.09 Page 63.93 Domain 59.03 All Features 67.43 Because we are using a two-layer neural network, the features in the learned network can interact with each other in interesting, nonlinear ways. This means that a particular feature that appears to have little value in isolation could actually be very important when used in combination with other features. To measure the final contribution of a feature set, in the context of all the other features, we performed an ablation study. That is, for each set of features, we trained a network to contain all of the features except that set. We then compared the performance of the resulting network to the performance of the network with all of the features. Table 3 shows the results of this experiment, where the decrease in accuracy is the difference in pairwise accuracy between the network trained with all of the features, and the network missing the given feature set. Table 3: Ablation study. Shown is the decrease in accuracy when we train a network that has all but the given set of features. The last line is shows the effect of removing the anchor, PageRank, and domain features, hence a model containing no network or link-based information whatsoever. Feature Set Decrease in Accuracy PageRank 0.18 Popularity 0.78 Anchor 0.47 Page 5.42 Domain Anchor, PageRank & Domain 0.10 0.60 The results of the ablation study are consistent with the individual feature set study. Both show that the most important feature set is the Page-level feature set, and the second most important is the popularity feature set. Finally, we wished to see how the performance of fRank improved as we added features; we wanted to find at what point adding more feature sets became relatively useless. Beginning with no features, we greedily added the feature set that improved performance the most. The results are shown in Table 4. For example, the fourth line of the table shows that fRank using the page, popularity, and anchor features outperformed any network that used the page, popularity, and some other feature set, and that the performance of this network was 67.25%. Table 4: fRank performance as feature sets are added. At each row, the feature set that gave the greatest increase in accuracy was added to the list of features (i.e., we conducted a greedy search over feature sets). Feature Set Accuracy (%) None 50.00 +Page 63.93 +Popularity 66.83 +Anchor 67.25 +PageRank 67.31 +Domain 67.43 711 Finally, we present a qualitative comparison of PageRank vs. fRank. In Table 5 are the top ten URLs returned for PageRank and for fRank. PageRank"s results are heavily weighted towards technology sites. It contains two QuickTime URLs (Apple"s video playback software), as well as Internet Explorer and FireFox URLs (both of which are Web browsers). fRank, on the other hand, contains more consumer-oriented sites such as American Express, Target, Dell, etc. PageRank"s bias toward technology can be explained through two processes. First, there are many pages with buttons at the bottom suggesting that the site is optimized for Internet Explorer, or that the visitor needs QuickTime. These generally link back to, in these examples, the Internet Explorer and QuickTime download sites. Consequently, PageRank ranks those pages highly. Though these pages are important, they are not as important as it may seem by looking at the link structure alone. One fix for this is to add information about the link to the PageRank computation, such as the size of the text, whether it was at the bottom of the page, etc. The other bias comes from the fact that the population of Web site authors is different than the population of Web users. Web authors tend to be technologically-oriented, and thus their linking behavior reflects those interests. fRank, by knowing the actual visitation popularity of a site (the popularity feature set), is able to eliminate some of that bias. It has the ability to depend more on where actual Web users visit rather than where the Web site authors have linked. The results confirm that fRank outperforms PageRank in pairwise accuracy. The two most important feature sets are the page and popularity features. This is surprising, as the page features consisted only of a few (8) simple features. Further experiments found that, of the page features, those based on the text of the page (as opposed to the URL) performed the best. In the next section, we explore the popularity feature in more detail. 5.5 Popularity Data As mentioned in section 4, our popularity data came from MSN toolbar users. For privacy reasons, we had access only to an aggregate count of, for each URL, how many times it was visited by any toolbar user. This limited the possible features we could derive from this data. For possible extensions, see section 6.3, future work. For each URL in our train and test sets, we provided a feature to fRank which was how many times it had been visited by a toolbar user. However, this feature was quite noisy and sparse, particularly for URLs with query parameters (e.g., http://search.msn.com/results.aspx?q=machine+learning&form=QBHP). One solution was to provide an additional feature which was the number of times any URL at the given domain was visited by a toolbar user. Adding this feature dramatically improved the performance of fRank. We took this one step further and used the built-in hierarchical structure of URLs to construct many levels of backoff between the full URL and the domain. We did this by using the set of features shown in Table 6. Table 6: URL functions used to compute the Popularity feature set. Function Example Exact URL cnn.com/2005/tech/wikipedia.html?v=mobile No Params cnn.com/2005/tech/wikipedia.html Page wikipedia.html URL-1 cnn.com/2005/tech URL-2 cnn.com/2005 … Domain cnn.com Domain+1 cnn.com/2005 … Each URL was assigned one feature for each function shown in the table. The value of the feature was the count of the number of times a toolbar user visited a URL, where the function applied to that URL matches the function applied to the URL in question. For example, a user"s visit to cnn.com/2005/sports.html would increment the Domain and Domain+1 features for the URL cnn.com/2005/tech/wikipedia.html. As seen in Table 7, adding the domain counts significantly improved the quality of the popularity feature, and adding the numerous backoff functions listed in Table 6 improved the accuracy even further. Table 7: Effect of adding backoff to the popularity feature set Features Accuracy (%) URL count 58.15 URL and Domain counts 59.31 All backoff functions (Table 6) 60.82 Table 5: Top ten URLs for PageRank vs. fRank PageRank fRank google.com google.com apple.com/quicktime/download yahoo.com amazon.com americanexpress.com yahoo.com hp.com microsoft.com/windows/ie target.com apple.com/quicktime bestbuy.com mapquest.com dell.com ebay.com autotrader.com mozilla.org/products/firefox dogpile.com ftc.gov bankofamerica.com 712 Backing off to subsets of the URL is one technique for dealing with the sparsity of data. It is also informative to see how the performance of fRank depends on the amount of popularity data that we have collected. In Figure 1 we show the performance of fRank trained with only the popularity feature set vs. the amount of data we have for the popularity feature set. Each day, we receive additional popularity data, and as can be seen in the plot, this increases the performance of fRank. The relation is logarithmic: doubling the amount of popularity data provides a constant improvement in pairwise accuracy. In summary, we have found that the popularity features provide a useful boost to the overall fRank accuracy. Gathering more popularity data, as well as employing simple backoff strategies, improve this boost even further. 5.6 Summary of Results The experiments provide a number of conclusions. First, fRank performs significantly better than PageRank, even without any information about the Web graph. Second, the page level and popularity features were the most significant contributors to pairwise accuracy. Third, by collecting more popularity data, we can continue to improve fRank"s performance. The popularity data provides two benefits to fRank. First, we see that qualitatively, fRank"s ordering of Web pages has a more favorable bias than PageRank"s. fRank"s ordering seems to correspond to what Web users, rather than Web page authors, prefer. Second, the popularity data is more timely than PageRank"s link information. The toolbar provides information about which Web pages people find interesting right now, whereas links are added to pages more slowly, as authors find the time and interest. 6. RELATED AND FUTURE WORK 6.1 Improvements to PageRank Since the original PageRank paper, there has been work on improving it. Much of that work centers on speeding up and parallelizing the computation [15][25]. One recognized problem with PageRank is that of topic drift: A page about dogs will have high PageRank if it is linked to by many pages that themselves have high rank, regardless of their topic. In contrast, a search engine user looking for good pages about dogs would likely prefer to find pages that are pointed to by many pages that are themselves about dogs. Hence, a link that is on topic should have higher weight than a link that is not. Richardson and Domingos"s Query Dependent PageRank [29] and Haveliwala"s Topic-Sensitive PageRank [16] are two approaches that tackle this problem. Other variations to PageRank include differently weighting links for inter- vs. intra-domain links, adding a backwards step to the random surfer to simulate the back button on most browsers [24] and modifying the jump probability (α) [3]. See Langville and Meyer [23] for a good survey of these, and other modifications to PageRank. 6.2 Other related work PageRank is not the only link analysis algorithm used for ranking Web pages. The most well-known other is HITS [22], which is used by the Teoma search engine [30]. HITS produces a list of hubs and authorities, where hubs are pages that point to many authority pages, and authorities are pages that are pointed to by many hubs. Previous work has shown HITS to perform comparably to PageRank [1]. One field of interest is that of static index pruning (see e.g., Carmel et al. [8]). Static index pruning methods reduce the size of the search engine"s index by removing documents that are unlikely to be returned by a search query. The pruning is typically done based on the frequency of query terms. Similarly, Pandey and Olston [28] suggest crawling pages frequently if they are likely to incorrectly appear (or not appear) as a result of a search. Similar methods could be incorporated into the static rank (e.g., how many frequent queries contain words found on this page). Others have investigated the effect that PageRank has on the Web at large [9]. They argue that pages with high PageRank are more likely to be found by Web users, thus more likely to be linked to, and thus more likely to maintain a higher PageRank than other pages. The same may occur for the popularity data. If we increase the ranking for popular pages, they are more likely to be clicked on, thus further increasing their popularity. Cho et al. [10] argue that a more appropriate measure of Web page quality would depend on not only the current link structure of the Web, but also on the change in that link structure. The same technique may be applicable to popularity data: the change in popularity of a page may be more informative than the absolute popularity. One interesting related work is that of Ivory and Hearst [19]. Their goal was to build a model of Web sites that are considered high quality from the perspective of content, structure and navigation, visual design, functionality, interactivity, and overall experience. They used over 100 page level features, as well as features encompassing the performance and structure of the site. This let them qualitatively describe the qualities of a page that make it appear attractive (e.g., rare use of italics, at least 9 point font, …), and (in later work) to build a system that assists novel Web page authors in creating quality pages by evaluating it according to these features. The primary differences between this work and ours are the goal (discovering what constitutes a good Web page vs. ordering Web pages for the purposes of Web search), the size of the study (they used a dataset of less than 6000 pages vs. our set of 468,000), and our comparison with PageRank. y = 0.577Ln(x) + 58.283 R 2 = 0.9822 58 58.5 59 59.5 60 60.5 61 1 10 100 Days of Toolbar Data PairwiseAccuracy Figure 1: Relation between the amount of popularity data and the performance of the popularity feature set. Note the x-axis is a logarithmic scale. 713 Nevertheless, their work provides insights to additional useful static features that we could incorporate into fRank in the future. Recent work on incorporating novel features into dynamic ranking includes that by Joachims et al. [21], who investigate the use of implicit feedback from users, in the form of which search engine results are clicked on. Craswell et al. [11] present a method for determining the best transformation to apply to query independent features (such as those used in this paper) for the purposes of improving dynamic ranking. Other work, such as Boyan et al. [4] and Bartell et al. [2] apply machine learning for the purposes of improving the overall relevance of a search engine (i.e., the dynamic ranking). They do not apply their techniques to the problem of static ranking. 6.3 Future work There are many ways in which we would like to extend this work. First, fRank uses only a small number of features. We believe we could achieve even more significant results with more features. In particular the existence, or lack thereof, of certain words could prove very significant (for instance, under construction probably signifies a low quality page). Other features could include the number of images on a page, size of those images, number of layout elements (tables, divs, and spans), use of style sheets, conforming to W3C standards (like XHTML 1.0 Strict), background color of a page, etc. Many pages are generated dynamically, the contents of which may depend on parameters in the URL, the time of day, the user visiting the site, or other variables. For such pages, it may be useful to apply the techniques found in [26] to form a static approximation for the purposes of extracting features. The resulting grammar describing the page could itself be a source of additional features describing the complexity of the page, such as how many non-terminal nodes it has, the depth of the grammar tree, etc. fRank allows one to specify a confidence in each pairing of documents. In the future, we will experiment with probabilities that depend on the difference in human judgments between the two items in the pair. For example, a pair of documents where one was rated 4 and the other 0 should have a higher confidence than a pair of documents rated 3 and 2. The experiments in this paper are biased toward pages that have higher than average quality. Also, fRank with all of the features can only be applied to pages that have already been crawled. Thus, fRank is primarily useful for index ordering and improving relevance, not for directing the crawl. We would like to investigate a machine learning approach for crawl prioritization as well. It may be that a combination of methods is best: for example, using PageRank to select the best 5 billion of the 20 billion pages on the Web, then using fRank to order the index and affect search relevancy. Another interesting direction for exploration is to incorporate fRank and page-level features directly into the PageRank computation itself. Work on biasing the PageRank jump vector [16], and transition matrix [29], have demonstrated the feasibility and advantages of such an approach. There is reason to believe that a direct application of [29], using the fRank of a page for its relevance, could lead to an improved overall static rank. Finally, the popularity data can be used in other interesting ways. The general surfing and searching habits of Web users varies by time of day. Activity in the morning, daytime, and evening are often quite different (e.g., reading the news, solving problems, and accessing entertainment, respectively). We can gain insight into these differences by using the popularity data, divided into segments of the day. When a query is issued, we would then use the popularity data matching the time of query in order to do the ranking of Web pages. We also plan to explore popularity features that use more than just the counts of how often a page was visited. For example, how long users tended to dwell on a page, did they leave the page by clicking a link or by hitting the back button, etc. Fox et al. did a study that showed that features such as this can be valuable for the purposes of dynamic ranking [14]. Finally, the popularity data could be used as the label rather than as a feature. Using fRank in this way to predict the popularity of a page may useful for the tasks of relevance, efficiency, and crawl priority. There is also significantly more popularity data than human labeled data, potentially enabling more complex machine learning methods, and significantly more features. 7. CONCLUSIONS A good static ranking is an important component for today"s search engines and information retrieval systems. We have demonstrated that PageRank does not provide a very good static ranking; there are many simple features that individually out perform PageRank. By combining many static features, fRank achieves a ranking that has a significantly higher pairwise accuracy than PageRank alone. A qualitative evaluation of the top documents shows that fRank is less technology-biased than PageRank; by using popularity data, it is biased toward pages that Web users, rather than Web authors, visit. The machine learning component of fRank gives it the additional benefit of being more robust against spammers, and allows it to leverage further developments in the machine learning community in areas such as adversarial classification. We have only begun to explore the options, and believe that significant strides can be made in the area of static ranking by further experimentation with additional features, other machine learning techniques, and additional sources of data. 8. ACKNOWLEDGMENTS Thank you to Marc Najork for providing us with additional PageRank computations and to Timo Burkard for assistance with the popularity data. Many thanks to Chris Burges for providing code and significant support in using training RankNets. Also, we thank Susan Dumais and Nick Craswell for their edits and suggestions. 9. REFERENCES [1] B. Amento, L. Terveen, and W. Hill. Does authority mean quality? Predicting expert quality ratings of Web documents. In Proceedings of the 23rd Annual International ACM SIGIR Conference on Research and Development in Information Retrieval, 2000. [2] B. Bartell, G. Cottrell, and R. Belew. Automatic combination of multiple ranked retrieval systems. In Proceedings of the 17th Annual International ACM SIGIR Conference on Research and Development in Information Retrieval, 1994. [3] P. Boldi, M. Santini, and S. Vigna. PageRank as a function of the damping factor. In Proceedings of the International World Wide Web Conference, May 2005. 714 [4] J. Boyan, D. Freitag, and T. Joachims. A machine learning architecture for optimizing web search engines. In AAAI Workshop on Internet Based Information Systems, August 1996. [5] S. Brin and L. Page. The anatomy of a large-scale hypertextual web search engine. In Proceedings of the Seventh International Wide Web Conference, Brisbane, Australia, 1998. Elsevier. [6] A. Broder, R. Lempel, F. Maghoul, and J. Pederson. Efficient PageRank approximation via graph aggregation. In Proceedings of the International World Wide Web Conference, May 2004. [7] C. Burges, T. Shaked, E. Renshaw, A. Lazier, M. Deeds, N. Hamilton, G. Hullender. Learning to rank using gradient descent. In Proceedings of the 22nd International Conference on Machine Learning, Bonn, Germany, 2005. [8] D. Carmel, D. Cohen, R. Fagin, E. Farchi, M. Herscovici, Y. S. Maarek, and A. Soffer. Static index pruning for information retrieval systems. In Proceedings of the 24th Annual International ACM SIGIR Conference on Research and Development in Information Retrieval, pages 43-50, New Orleans, Louisiana, USA, September 2001. [9] J. Cho and S. Roy. Impact of search engines on page popularity. In Proceedings of the International World Wide Web Conference, May 2004. [10]J. Cho, S. Roy, R. Adams. Page Quality: In search of an unbiased web ranking. In Proceedings of the ACM SIGMOD 2005 Conference. Baltimore, Maryland. June 2005. [11]N. Craswell, S. Robertson, H. Zaragoza, and M. Taylor. Relevance weighting for query independent evidence. In Proceedings of the 28th Annual Conference on Research and Development in Information Retrieval (SIGIR), August, 2005. [12]N. Dalvi, P. Domingos, Mausam, S. Sanghai, D. Verma. Adversarial Classification. In Proceedings of the Tenth International Conference on Knowledge Discovery and Data Mining (pp. 99-108), Seattle, WA, 2004. [13]O. Dekel, C. Manning, and Y. Singer. Log-linear models for label-ranking. In Advances in Neural Information Processing Systems 16. Cambridge, MA: MIT Press, 2003. [14]S. Fox, K S. Fox, K. Karnawat, M. Mydland, S. T. Dumais and T. White (2005). Evaluating implicit measures to improve the search experiences. In the ACM Transactions on Information Systems, 23(2), pp. 147-168. April 2005. [15]T. Haveliwala. Efficient computation of PageRank. Stanford University Technical Report, 1999. [16]T. Haveliwala. Topic-sensitive PageRank. In Proceedings of the International World Wide Web Conference, May 2002. [17]D. Hawking and N. Craswell. Very large scale retrieval and Web search. In D. Harman and E. Voorhees (eds), The TREC Book. MIT Press. [18]R. Herbrich, T. Graepel, and K. Obermayer. Support vector learning for ordinal regression. In Proceedings of the Ninth International Conference on Artificial Neural Networks, pp. 97-102. 1999. [19]M. Ivory and M. Hearst. Statistical profiles of highly-rated Web sites. In Proceedings of the ACM SIGCHI Conference on Human Factors in Computing Systems, 2002. [20]T. Joachims. Optimizing search engines using clickthrough data. In Proceedings of the ACM Conference on Knowledge Discovery and Data Mining (KDD), 2002. [21]T. Joachims, L. Granka, B. Pang, H. Hembrooke, and G. Gay. Accurately Interpreting Clickthrough Data as Implicit Feedback. In Proceedings of the Conference on Research and Development in Information Retrieval (SIGIR), 2005. [22]J. Kleinberg. Authoritative sources in a hyperlinked environment. Journal of the ACM 46:5, pp. 604-32. 1999. [23]A. Langville and C. Meyer. Deeper inside PageRank. Internet Mathematics 1(3):335-380, 2004. [24]F. Matthieu and M. Bouklit. The effect of the back button in a random walk: application for PageRank. In Alternate track papers and posters of the Thirteenth International World Wide Web Conference, 2004. [25]F. McSherry. A uniform approach to accelerated PageRank computation. In Proceedings of the International World Wide Web Conference, May 2005. [26]Y. Minamide. Static approximation of dynamically generated Web pages. In Proceedings of the International World Wide Web Conference, May 2005. [27]L. Page, S. Brin, R. Motwani, and T. Winograd. The PageRank citation ranking: Bringing order to the web. Technical report, Stanford University, Stanford, CA, 1998. [28]S. Pandey and C. Olston. User-centric Web crawling. In Proceedings of the International World Wide Web Conference, May 2005. [29]M. Richardson and P. Domingos. The intelligent surfer: probabilistic combination of link and content information in PageRank. In Advances in Neural Information Processing Systems 14, pp. 1441-1448. Cambridge, MA: MIT Press, 2002. [30]C. Sherman. Teoma vs. Google, Round 2. Available from World Wide Web (http://dc.internet.com/news/article.php/ 1002061), 2002. [31]T. Upstill, N. Craswell, and D. Hawking. Predicting fame and fortune: PageRank or indegree?. In the Eighth Australasian Document Computing Symposium. 2003. [32]T. Upstill, N. Craswell, and D. Hawking. Query-independent evidence in home page finding. In ACM Transactions on Information Systems. 2003. 715

ranknet;relevance;static rank;visitation popularity;search engine;pagerank;regression;adversarial classification;feature-based ranking;information retrieval;machine learning;static ranking

train_H-81

Distance Measures for MPEG-7-based Retrieval

In visual information retrieval the careful choice of suitable proximity measures is a crucial success factor. The evaluation presented in this paper aims at showing that the distance measures suggested by the MPEG-7 group for the visual descriptors can be beaten by general-purpose measures. Eight visual MPEG-7 descriptors were selected and 38 distance measures implemented. Three media collections were created and assessed, performance indicators developed and more than 22500 tests performed. Additionally, a quantisation model was developed to be able to use predicate-based distance measures on continuous data as well. The evaluation shows that the distance measures recommended in the MPEG-7-standard are among the best but that other measures perform even better.

1. INTRODUCTION The MPEG-7 standard defines - among others - a set of descriptors for visual media. Each descriptor consists of a feature extraction mechanism, a description (in binary and XML format) and guidelines that define how to apply the descriptor on different kinds of media (e.g. on temporal media). The MPEG-7 descriptors have been carefully designed to meet - partially complementaryrequirements of different application domains: archival, browsing, retrieval, etc. [9]. In the following, we will exclusively deal with the visual MPEG-7 descriptors in the context of media retrieval. The visual MPEG-7 descriptors fall in five groups: colour, texture, shape, motion and others (e.g. face description) and sum up to 16 basic descriptors. For retrieval applications, a rule for each descriptor is mandatory that defines how to measure the similarity of two descriptions. Common rules are distance functions, like the Euclidean distance and the Mahalanobis distance. Unfortunately, the MPEG-7 standard does not include distance measures in the normative part, because it was not designed to be (and should not exclusively understood to be) retrieval-specific. However, the MPEG-7 authors give recommendations, which distance measure to use on a particular descriptor. These recommendations are based on accurate knowledge of the descriptors' behaviour and the description structures. In the present study a large number of successful distance measures from different areas (statistics, psychology, medicine, social and economic sciences, etc.) were implemented and applied on MPEG-7 data vectors to verify whether or not the recommended MPEG-7 distance measures are really the best for any reasonable class of media objects. From the MPEG-7 tests and the recommendations it does not become clear, how many and which distance measures have been tested on the visual descriptors and the MPEG-7 test datasets. The hypothesis is that analytically derived distance measures may be good in general but only a quantitative analysis is capable to identify the best distance measure for a specific feature extraction method. The paper is organised as follows. Section 2 gives a minimum of background information on the MPEG-7 descriptors and distance measurement in visual information retrieval (VIR, see [3], [16]). Section 3 gives an overview over the implemented distance measures. Section 4 describes the test setup, including the test data and the implemented evaluation methods. Finally, Section 5 presents the results per descriptor and over all descriptors. 2. BACKGROUND 2.1 MPEG-7: visual descriptors The visual part of the MPEG-7 standard defines several descriptors. Not all of them are really descriptors in the sense that they extract properties from visual media. Some of them are just structures for descriptor aggregation or localisation. The basic descriptors are Color Layout, Color Structure, Dominant Color, Scalable Color, Edge Histogram, Homogeneous Texture, Texture Browsing, Region-based Shape, Contour-based Shape, Camera Motion, Parametric Motion and Motion Activity. Other descriptors are based on low-level descriptors or semantic information: Group-of-Frames/Group-of-Pictures Color (based on Scalable Color), Shape 3D (based on 3D mesh information), Motion Trajectory (based on object segmentation) and Face Recognition (based on face extraction). Descriptors for spatiotemporal aggregation and localisation are: Spatial 2D Coordinates, Grid Layout, Region Locator (spatial), Time Series, Temporal Interpolation (temporal) and SpatioTemporal Locator (combined). Finally, other structures exist for colour spaces, colour quantisation and multiple 2D views of 3D objects. These additional structures allow combining the basic descriptors in multiple ways and on different levels. But they do not change the characteristics of the extracted information. Consequently, structures for aggregation and localisation were not considered in the work described in this paper. 2.2 Similarity measurement on visual data Generally, similarity measurement on visual information aims at imitating human visual similarity perception. Unfortunately, human perception is much more complex than any of the existing similarity models (it includes perception, recognition and subjectivity). The common approach in visual information retrieval is measuring dis-similarity as distance. Both, query object and candidate object are represented by their corresponding feature vectors. The distance between these objects is measured by computing the distance between the two vectors. Consequently, the process is independent of the employed querying paradigm (e.g. query by example). The query object may be natural (e.g. a real object) or artificial (e.g. properties of a group of objects). Goal of the measurement process is to express a relationship between the two objects by their distance. Iteration for multiple candidates allows then to define a partial order over the candidates and to address those in a (to be defined) neighbourhood being similar to the query object. At this point, it has to be mentioned that in a multi-descriptor environmentespecially in MPEG-7 - we are only half way towards a statement on similarity. If multiple descriptors are used (e.g. a descriptor scheme), a rule has to be defined how to combine all distances to a global value for each object. Still, distance measurement is the most important first step in similarity measurement. Obviously, the main task of good distance measures is to reorganise descriptor space in a way that media objects with the highest similarity are nearest to the query object. If distance is defined minimal, the query object is always in the origin of distance space and similar candidates should form clusters around the origin that are as large as possible. Consequently, many well known distance measures are based on geometric assumptions of descriptor space (e.g. Euclidean distance is based on the metric axioms). Unfortunately, these measures do not fit ideally with human similarity perception (e.g. due to human subjectivity). To overcome this shortage, researchers from different areas have developed alternative models that are mostly predicate-based (descriptors are assumed to contain just binary elements, e.g. Tversky's Feature Contrast Model [17]) and fit better with human perception. In the following distance measures of both groups of approaches will be considered. 3. DISTANCE MEASURES The distance measures used in this work have been collected from various areas (Subsection 3.1). Because they work on differently quantised data, Subsection 3.2 sketches a model for unification on the basis of quantitative descriptions. Finally, Subsection 3.3 introduces the distance measures as well as their origin and the idea they implement. 3.1 Sources Distance measurement is used in many research areas such as psychology, sociology (e.g. comparing test results), medicine (e.g. comparing parameters of test persons), economics (e.g. comparing balance sheet ratios), etc. Naturally, the character of data available in these areas differs significantly. Essentially, there are two extreme cases of data vectors (and distance measures): predicatebased (all vector elements are binary, e.g. {0, 1}) and quantitative (all vector elements are continuous, e.g. [0, 1]). Predicates express the existence of properties and represent highlevel information while quantitative values can be used to measure and mostly represent low-level information. Predicates are often employed in psychology, sociology and other human-related sciences and most predicate-based distance measures were therefore developed in these areas. Descriptions in visual information retrieval are nearly ever (if they do not integrate semantic information) quantitative. Consequently, mostly quantitative distance measures are used in visual information retrieval. The goal of this work is to compare the MPEG-7 distance measures with the most powerful distance measures developed in other areas. Since MPEG-7 descriptions are purely quantitative but some of the most sophisticated distance measures are defined exclusively on predicates, a model is mandatory that allows the application of predicate-based distance measures on quantitative data. The model developed for this purpose is presented in the next section. 3.2 Quantisation model The goal of the quantisation model is to redefine the set operators that are usually used in predicate-based distance measures on continuous data. The first in visual information retrieval to follow this approach were Santini and Jain, who tried to apply Tversky's Feature Contrast Model [17] to content-based image retrieval [12], [13]. They interpreted continuous data as fuzzy predicates and used fuzzy set operators. Unfortunately, their model suffered from several shortcomings they described in [12], [13] (for example, the quantitative model worked only for one specific version of the original predicate-based measure). The main idea of the presented quantisation model is that set operators are replaced by statistical functions. In [5] the authors could show that this interpretation of set operators is reasonable. The model offers a solution for the descriptors considered in the evaluation. It is not specific to one distance measure, but can be applied to any predicate-based measure. Below, it will be shown that the model does not only work for predicate data but for quantitative data as well. Each measure implementing the model can be used as a substitute for the original predicate-based measure. Generally, binary properties of two objects (e.g. media objects) can exist in both objects (denoted as a), in just one (b, c) or in none of them (d). The operator needed for these relationships are UNION, MINUS and NOT. In the quantisation model they are replaced as follows (see [5] for further details). 131 ∑     ≤ + − + ==∩= k jkikjkik kkji else xx Mif xx ssXXa 0 22, 1ε ( ) ( ) ∑ ∑ ∑     ≤ ++ −==¬∩¬=    ≤−−− ==−=    ≤−−− ==−= k jkikjkik kkji k ikjkikjk kkij k jkikjkik kkji else xx if xx MssXXd else xxMifxx ssXXc else xxMifxx ssXXb 0 22, 0 , 0 , 1 2 2 ε ε ε with: ( ) [ ] ( ) { }0\ . 0 1 . 0 1 , 2 2 1 minmax maxmin + ∈ − =     ≥      − = =     ≥      − = −= ∈= ∑ ∑ ∑ ∑ Rp ki x where else pif p M ki x where else pif p M xxM xxxwithxX i k ik i k ik ikiki µ σ σ σ ε µ µ µ ε a selects properties that are present in both data vectors (Xi, Xj representing media objects), b and c select properties that are present in just one of them and d selects properties that are present in neither of the two data vectors. Every property is selected by the extent to which it is present (a and d: mean, b and c: difference) and only if the amount to which it is present exceeds a certain threshold (depending on the mean and standard deviation over all elements of descriptor space). The implementation of these operators is based on one assumption. It is assumed that vector elements measure on interval scale. That means, each element expresses that the measured property is "more or less" present ("0": not at all, "M": fully present). This is true for most visual descriptors and all MPEG-7 descriptors. A natural origin as it is assumed here ("0") is not needed. Introducing p (called discriminance-defining parameter) for the thresholds 21 ,εε has the positive consequence that a, b, c, d can then be controlled through a single parameter. p is an additional criterion for the behaviour of a distance measure and determines the thresholds used in the operators. It expresses how accurate data items are present (quantisation) and consequently, how accurate they should be investigated. p can be set by the user or automatically. Interesting are the limits: 1. Mp →⇒∞→ 21 ,εε In this case, all elements (=properties) are assumed to be continuous (high quantisation). In consequence, all properties of a descriptor are used by the operators. Then, the distance measure is not discriminant for properties. 2. 0,0 21 →⇒→ εεp In this case, all properties are assumed to be predicates. In consequence, only binary elements (=predicates) are used by the operators (1-bit quantisation). The distance measure is then highly discriminant for properties. Between these limits, a distance measure that uses the quantisation model is - depending on p - more or less discriminant for properties. This means, it selects a subset of all available description vector elements for distance measurement. For both predicate data and quantitative data it can be shown that the quantisation model is reasonable. If description vectors consist of binary elements only, p should be used as follows (for example, p can easily be set automatically): ( )σµεε ,min..,0,0 21 ==⇒→ pgep In this case, a, b, c, d measure like the set operators they replace. For example, Table 1 shows their behaviour for two onedimensional feature vectors Xi and Xj. As can be seen, the statistical measures work like set operators. Actually, the quantisation model works accurate on predicate data for any p≠∞. To show that the model is reasonable for quantitative data the following fact is used. It is easy to show that for predicate data some quantitative distance measures degenerate to predicatebased measures. For example, the L1 metric (Manhattan metric) degenerates to the Hamming distance (from [9], without weights): distanceHammingcbxxL k jkik =+≡−= ∑1 If it can be shown that the quantisation model is able to reconstruct the quantitative measure from the degenerated predicate-based measure, the model is obviously able to extend predicate-based measures to the quantitative domain. This is easy to illustrate. For purely quantitative feature vectors, p should be used as follows (again, p can easily be set automatically): 1, 21 =⇒∞→ εεp Then, a and d become continuous functions: ∑ ∑ + −==⇒≡≤ + + ==⇒≡≤ + − k jkik kk jkik k jkik kk jkik xx MswheresdtrueM xx xx swheresatrueM xx M 22 22 b and c can be made continuous for the following expressions: ( ) ( ) ∑ ∑ ∑ −==+⇒    ≥−− ==⇒ ≥−≡≤−−    ≥−− ==⇒ ≥−≡≤−− k jkikkk k ikjkikjk kk ikjkikjk k jkikjkik kk jkikjkik xxswherescb else xxifxx swheresc xxMxxM else xxifxx swheresb xxMxxM 0 0 0 0 0 0 Table 1. Quantisation model on predicate vectors. Xi Xj a b c d (1) (1) 1 0 0 0 (1) (0) 0 1 0 0 (0) (1) 0 0 1 0 (0) (0) 0 0 0 1 132 ∑ ∑ −==− −==− k ikjkkk k jkikkk xxswheresbc xxswherescb This means, for sufficiently high p every predicate-based distance measure that is either not using b and c or just as b+c, b-c or c-b, can be transformed into a continuous quantitative distance measure. For example, the Hamming distance (again, without weights): 1 Lxxxxswherescb k jkik k jkikkk =−=−==+ ∑∑ The quantisation model successfully reconstructs the L1 metric and no distance measure-specific modification has to be made to the model. This demonstrates that the model is reasonable. In the following it will be used to extend successful predicate-based distance measures on the quantitative domain. The major advantages of the quantisation model are: (1) it is application domain independent, (2) the implementation is straightforward, (3) the model is easy to use and finally, (4) the new parameter p allows to control the similarity measurement process in a new way (discriminance on property level). 3.3 Implemented measures For the evaluation described in this work next to predicate-based (based on the quantisation model) and quantitative measures, the distance measures recommended in the MPEG-7 standard were implemented (all together 38 different distance measures). Table 2 summarises those predicate-based measures that performed best in the evaluation (in sum 20 predicate-based measures were investigated). For these measures, K is the number of predicates in the data vectors Xi and Xj. In P1, the sum is used for Tversky's f() (as Tversky himself does in [17]) and α, β are weights for element b and c. In [5] the author's investigated Tversky's Feature Contrast Model and found α=1, β=0 to be the optimum parameters. Some of the predicate-based measures are very simple (e.g. P2, P4) but have been heavily exploited in psychological research. Pattern difference (P6) - a very powerful measure - is used in the statistics package SPSS for cluster analysis. P7 is a correlation coefficient for predicates developed by Pearson. Table 3 shows the best quantitative distance measures that were used. Q1 and Q2 are metric-based and were implemented as representatives for the entire group of Minkowski distances. The wi are weights. In Q5, ii σµ , are mean and standard deviation for the elements of descriptor Xi. In Q6, m is 2 M (=0.5). Q3, the Canberra metric, is a normalised form of Q1. Similarly, Q4, Clark's divergence coefficient is a normalised version of Q2. Q6 is a further-developed correlation coefficient that is invariant against sign changes. This measure is used even though its particular properties are of minor importance for this application domain. Finally, Q8 is a measure that takes the differences between adjacent vector elements into account. This makes it structurally different from all other measures. Obviously, one important distance measure is missing. The Mahalanobis distance was not considered, because different descriptors would require different covariance matrices and for some descriptors it is simply impossible to define a covariance matrix. If the identity matrix was used in this case, the Mahalanobis distance would degenerate to a Minkowski distance. Additionally, the recommended MPEG-7 distances were implemented with the following parameters: In the distance measure of the Color Layout descriptor all weights were set to "1" (as in all other implemented measures). In the distance measure of the Dominant Color descriptor the following parameters were used: 20,1,3.0,7.0 21 ==== dTww α (as recommended). In the Homogeneous Texture descriptor's distance all ( )kα were set to "1" and matching was done rotation- and scale-invariant. Important! Some of the measures presented in this section are distance measures while others are similarity measures. For the tests, it is important to notice, that all similarity measures were inverted to distance measures. 4. TEST SETUP Subsection 4.1 describes the descriptors (including parameters) and the collections (including ground truth information) that were used in the evaluation. Subsection 4.2 discusses the evaluation method that was implemented and Subsection 4.3 sketches the test environment used for the evaluation process. 4.1 Test data For the evaluation eight MPEG-7 descriptors were used. All colour descriptors: Color Layout, Color Structure, Dominant Color, Scalable Color, all texture descriptors: Edge Histogram, Homogeneous Texture, Texture Browsing and one shape descriptor: Region-based Shape. Texture Browsing was used even though the MPEG-7 standard suggests that it is not suitable for retrieval. The other basic shape descriptor, Contour-based Shape, was not used, because it produces structurally different descriptions that cannot be transformed to data vectors with elements measuring on interval-scales. The motion descriptors were not used, because they integrate the temporal dimension of visual media and would only be comparable, if the basic colour, texture and shape descriptors would be aggregated over time. This was not done. Finally, no high-level descriptors were used (Localisation, Face Recognition, etc., see Subsection 2.1), because - to the author's opinion - the behaviour of the basic descriptors on elementary media objects should be evaluated before conclusions on aggregated structures can be drawn. Table 2. Predicate-based distance measures. No. Measure Comment P1 cba .. βα −− Feature Contrast Model, Tversky 1977 [17] P2 a No. of co-occurrences P3 cb + Hamming distance P4 K a Russel 1940 [14] P5 cb a + Kulczvnski 1927 [14] P6 2 K bc Pattern difference [14] P7 ( )( )( )( )dcdbcaba bcad ++++ − Pearson 1926 [11] 133 The Texture Browsing descriptions had to be transformed from five bins to an eight bin representation in order that all elements of the descriptor measure on an interval scale. A Manhattan metric was used to measure proximity (see [6] for details). Descriptor extraction was performed using the MPEG-7 reference implementation. In the extraction process each descriptor was applied on the entire content of each media object and the following extraction parameters were used. Colour in Color Structure was quantised to 32 bins. For Dominant Color colour space was set to YCrCb, 5-bit default quantisation was used and the default value for spatial coherency was used. Homogeneous Texture was quantised to 32 components. Scalable Color values were quantised to sizeof(int)-3 bits and 64 bins were used. Finally, Texture Browsing was used with five components. These descriptors were applied on three media collections with image content: the Brodatz dataset (112 images, 512x512 pixel), a subset of the Corel dataset (260 images, 460x300 pixel, portrait and landscape) and a dataset with coats-of-arms images (426 images, 200x200 pixel). Figure 1 shows examples from the three collections. Designing appropriate test sets for a visual evaluation is a highly difficult task (for example, see the TREC video 2002 report [15]). Of course, for identifying the best distance measure for a descriptor, it should be tested on an infinite number of media objects. But this is not the aim of this study. It is just evaluated if - for likely image collections - better proximity measures than those suggested by the MPEG-7 group can be found. Collections of this relatively small size were used in the evaluation, because the applied evaluation methods are above a certain minimum size invariant against collection size and for smaller collections it is easier to define a high-quality ground truth. Still, the average ratio of ground truth size to collection size is at least 1:7. Especially, no collection from the MPEG-7 dataset was used in the evaluation because the evaluations should show, how well the descriptors and the recommended distance measures perform on "unknown" material. When the descriptor extraction was finished, the resulting XML descriptions were transformed into a data matrix with 798 lines (media objects) and 314 columns (descriptor elements). To be usable with distance measures that do not integrate domain knowledge, the elements of this data matrix were normalised to [0, 1]. For the distance evaluation - next to the normalised data matrixhuman similarity judgement is needed. In this work, the ground truth is built of twelve groups of similar images (four for each dataset). Group membership was rated by humans based on semantic criterions. Table 4 summarises the twelve groups and the underlying descriptions. It has to be noticed, that some of these groups (especially 5, 7 and 10) are much harder to find with lowlevel descriptors than others. 4.2 Evaluation method Usually, retrieval evaluation is performed based on a ground truth with recall and precision (see, for example, [3], [16]). In multidescriptor environments this leads to a problem, because the resulting recall and precision values are strongly influenced by the method used to merge the distance values for one media object. Even though it is nearly impossible to say, how big the influence of a single distance measure was on the resulting recall and precision values, this problem has been almost ignored so far. In Subsection 2.2 it was stated that the major task of a distance measure is to bring the relevant media objects as close to the origin (where the query object lies) as possible. Even in a multidescriptor environment it is then simple to identify the similar objects in a large distance space. Consequently, it was decided to Table 3. Quantitative distance measures. No. Measure Comment No. Measure Comment Q1 ∑ − k jkiki xxw City block distance (L1 ) Q2 ( )∑ − k jkiki xxw 2 Euclidean distance (L2 ) Q3 ∑ + − k jkik jkik xx xx Canberra metric, Lance, Williams 1967 [8] Q4 ( ) ∑ + − k jkik jkik xx xx K 2 1 Divergence coefficient, Clark 1952 [1] Q5 ( )( ) ( ) ( )∑ ∑ ∑ −− −− k k jjkiik k jjkiik xx xx 22 µµ µµ Correlation coefficient Q6       −+      −−       +−− ∑∑∑ ∑ ∑∑ k ik k jkik k ik k k jk k ikjkik xmKmxxmKmx xxmKmxx 2..2 2222 Cohen 1969 [2] Q7 ∑ ∑ ∑ k k jkik k jkik xx xx 22 Angular distance, Gower 1967 [7] Q8 ( ) ( )( )∑ − ++ −−− 1 2 11 K k jkjkikik xxxx Meehl Index [10] Table 4. Ground truth information. Coll. No Images Description 1 19 Regular, chequered patterns 2 38 Dark white noise 3 33 Moon-like surfaces Brodatz 4 35 Water-like surfaces 5 73 Humans in nature (difficult) 6 17 Images with snow (mountains, skiing) 7 76 Animals in nature (difficult) Corel 8 27 Large coloured flowers 9 12 Bavarian communal arms 10 10 All Bavarian arms (difficult) 11 18 Dark objects / light unsegmented shield Arms 12 14 Major charges on blue or red shield 134 use indicators measuring the distribution in distance space of candidates similar to the query object for this evaluation instead of recall and precision. Identifying clusters of similar objects (based on the given ground truth) is relatively easy, because the resulting distance space for one descriptor and any distance measure is always one-dimensional. Clusters are found by searching from the origin of distance space to the first similar object, grouping all following similar objects in the cluster, breaking off the cluster with the first un-similar object and so forth. For the evaluation two indicators were defined. The first measures the average distance of all cluster means to the origin: distanceavgclustersno sizecluster distanceclustersno i i sizecluster j ij d i _._ _ _ _ ∑ ∑ =µ where distanceij is the distance value of the j-th element in the i-th cluster, ∑ ∑ ∑ = CLUSTERS i i CLUSTERS i sizecluster j ij sizecluster distance distanceavg i _ _ _ , no_clusters is the number of found clusters and cluster_sizei is the size of the i-th cluster. The resulting indicator is normalised by the distribution characteristics of the distance measure (avg_distance). Additionally, the standard deviation is used. In the evaluation process this measure turned out to produce valuable results and to be relatively robust against parameter p of the quantisation model. In Subsection 3.2 we noted that p affects the discriminance of a predicate-based distance measure: The smaller p is set the larger are the resulting clusters because the quantisation model is then more discriminant against properties and less elements of the data matrix are used. This causes a side-effect that is measured by the second indicator: more and more un-similar objects come out with exactly the same distance value as similar objects (a problem that does not exist for large p's) and become indiscernible from similar objects. Consequently, they are (false) cluster members. This phenomenon (conceptually similar to the "false negatives" indicator) was named "cluster pollution" and the indicator measures the average cluster pollution over all clusters: clustersno doublesno cp clustersno i sizecluster j ij i _ _ _ _ ∑ ∑ = where no_doublesij is the number of indiscernible un-similar objects associated with the j-th element of cluster i. Remark: Even though there is a certain influence, it could be proven in [5] that no significant correlation exists between parameter p of the quantisation model and cluster pollution. 4.3 Test environment As pointed out above, to generate the descriptors, the MPEG-7 reference implementation in version 5.6 was used (provided by TU Munich). Image processing was done with Adobe Photoshop and normalisation and all evaluations were done with Perl. The querying process was performed in the following steps: (1) random selection of a ground truth group, (2) random selection of a query object from this group, (3) distance comparison for all other objects in the dataset, (4) clustering of the resulting distance space based on the ground truth and finally, (5) evaluation. For each combination of dataset and distance measure 250 queries were issued and evaluations were aggregated over all datasets and descriptors. The next section shows the - partially surprisingresults. 5. RESULTS In the results presented below the first indicator from Subsection 4.2 was used to evaluate distance measures. In a first step parameter p had to be set in a way that all measures are equally discriminant. Distance measurement is fair if the following condition holds true for any predicate-based measure dP and any continuous measure dC: ( ) ( )CP dcppdcp ≈, Then, it is guaranteed that predicate-based measures do not create larger clusters (with a higher number of similar objects) for the price of higher cluster pollution. In more than 1000 test queries the optimum value was found to be p=1. Results are organised as follows: Subsection 5.1 summarises the Figure 1. Test datasets. Left: Brodatz dataset, middle: Corel dataset, right: coats-of-arms dataset. 135 best distance measures per descriptor, Section 5.2 shows the best overall distance measures and Section 5.3 points out other interesting results (for example, distance measures that work particularly good on specific ground truth groups). 5.1 Best measure per descriptor Figure 2 shows the evaluation results for the first indicator. For each descriptor the best measure and the performance of the MPEG-7 recommendation are shown. The results are aggregated over the tested datasets. On first sight, it becomes clear that the MPEG-7 recommendations are mostly relatively good but never the best. For Color Layout the difference between MP7 and the best measure, the Meehl index (Q8), is just 4% and the MPEG-7 measure has a smaller standard deviation. The reason why the Meehl index is better may be that this descriptors generates descriptions with elements that have very similar variance. Statistical analysis confirmed that (see [6]). For Color Structure, Edge Histogram, Homogeneous Texture, Region-based Shape and Scalable Color by far the best measure is pattern difference (P6). Psychological research on human visual perception has revealed that in many situation differences between the query object and a candidate weigh much stronger than common properties. The pattern difference measure implements this insight in the most consequent way. In the author's opinion, the reason why pattern difference performs so extremely well on many descriptors is due to this fact. Additional advantages of pattern difference are that it usually has a very low variance andbecause it is a predicate-based measure - its discriminance (and cluster structure) can be tuned with parameter p. The best measure for Dominant Color turned out to be Clark's Divergence coefficient (Q4). This is a similar measure to pattern difference on the continuous domain. The Texture Browsing descriptor is a special problem. In the MPEG-7 standard it is recommended to use it exclusively for browsing. After testing it for retrieval on various distance measures the author supports this opinion. It is very difficult to find a good distance measure for Texture Browsing. The proposed Manhattan metric, for example, performs very bad. The best measure is predicate-based (P7). It works on common properties (a, d) but produces clusters with very high cluster pollution. For this descriptor the second indicator is up to eight times higher than for predicate-based measures on other descriptors. 5.2 Best overall measures Figure 3 summarises the results over all descriptors and media collections. The diagram should give an indication on the general potential of the investigated distance measures for visual information retrieval. It can be seen that the best overall measure is a predicate-based one. The top performance of pattern difference (P6) proves that the quantisation model is a reasonable method to extend predicate-based distance measures on the continuous domain. The second best group of measures are the MPEG-7 recommendations, which have a slightly higher mean but a lower standard deviation than pattern difference. The third best measure is the Meehl index (Q8), a measure developed for psychological applications but because of its characteristic properties tailormade for certain (homogeneous) descriptors. Minkowski metrics are also among the best measures: the average mean and variance of the Manhattan metric (Q1) and the Euclidean metric (Q2) are in the range of Q8. Of course, these measures do not perform particularly well for any of the descriptors. Remarkably for a predicate-based measure, Tversky's Feature Contrast Model (P1) is also in the group of very good measures (even though it is not among the best) that ends with Q5, the correlation coefficient. The other measures either have a significantly higher mean or a very large standard deviation. 5.3 Other interesting results Distance measures that perform in average worse than others may in certain situations (e.g. on specific content) still perform better. For Color Layout, for example, Q7 is a very good measure on colour photos. It performs as good as Q8 and has a lower standard deviation. For artificial images the pattern difference and the Hamming distance produce comparable results as well. If colour information is available in media objects, pattern difference performs well on Dominant Color (just 20% worse Q4) and in case of difficult ground truth (group 5, 7, 10) the Meehl index is as strong as P6. 0,000 0,001 0,002 0,003 0,004 0,005 0,006 0,007 0,008 Q8 MP7 P6 MP7 Q4 MP7 P6 MP7 P6 MP7 P6 MP7 P6 MP7 P7 Q2 Color Layout Color Structure Dominant Color Edge Histogram Homog. Texture Region Shape Scalable Color Texture Browsing Figure 2. Results per measure and descriptor. The horizontal axis shows the best measure and the performance of the MPEG-7 recommendation for each descriptor. The vertical axis shows the values for the first indicator (smaller value = better cluster structure). Shades have the following meaning: black=µ-σ (good cases), black + dark grey=µ (average) and black + dark grey + light grey=µ+σ (bad). 136 6. CONCLUSION The evaluation presented in this paper aims at testing the recommended distance measures and finding better ones for the basic visual MPEG-7 descriptors. Eight descriptors were selected, 38 distance measures were implemented, media collections were created and assessed, performance indicators were defined and more than 22500 tests were performed. To be able to use predicate-based distance measures next to quantitative measures a quantisation model was defined that allows the application of predicate-based measures on continuous data. In the evaluation the best overall distance measures for visual content - as extracted by the visual MPEG-7 descriptors - turned out to be the pattern difference measure and the Meehl index (for homogeneous descriptions). Since these two measures perform significantly better than the MPEG-7 recommendations they should be further tested on large collections of image and video content (e.g. from [15]). The choice of the right distance function for similarity measurement depends on the descriptor, the queried media collection and the semantic level of the user's idea of similarity. This work offers suitable distance measures for various situations. In consequence, the distance measures identified as the best will be implemented in the open MPEG-7 based visual information retrieval framework VizIR [4]. ACKNOWLEDGEMENTS The author would like to thank Christian Breiteneder for his valuable comments and suggestions for improvement. The work presented in this paper is part of the VizIR project funded by the Austrian Scientific Research Fund FWF under grant no. P16111. REFERENCES [1] Clark, P.S. An extension of the coefficient of divergence for use with multiple characters. Copeia, 2 (1952), 61-64. [2] Cohen, J. A profile similarity coefficient invariant over variable reflection. Psychological Bulletin, 71 (1969), 281284. [3] Del Bimbo, A. Visual information retrieval. Morgan Kaufmann Publishers, San Francisco CA, 1999. [4] Eidenberger, H., and Breiteneder, C. A framework for visual information retrieval. In Proceedings Visual Information Systems Conference (HSinChu Taiwan, March 2002), LNCS 2314, Springer Verlag, 105-116. [5] Eidenberger, H., and Breiteneder, C. Visual similarity measurement with the Feature Contrast Model. In Proceedings SPIE Storage and Retrieval for Media Databases Conference (Santa Clara CA, January 2003), SPIE Vol. 5021, 64-76. [6] Eidenberger, H., How good are the visual MPEG-7 features? In Proceedings SPIE Visual Communications and Image Processing Conference (Lugano Switzerland, July 2003), SPIE Vol. 5150, 476-488. [7] Gower, J.G. Multivariate analysis and multidimensional geometry. The Statistician, 17 (1967),13-25. [8] Lance, G.N., and Williams, W.T. Mixed data classificatory programs. Agglomerative Systems Australian Comp. Journal, 9 (1967), 373-380. [9] Manjunath, B.S., Ohm, J.R., Vasudevan, V.V., and Yamada, A. Color and texture descriptors. In Special Issue on MPEG7. IEEE Transactions on Circuits and Systems for Video Technology, 11/6 (June 2001), 703-715. [10] Meehl, P. E. The problem is epistemology, not statistics: Replace significance tests by confidence intervals and quantify accuracy of risky numerical predictions. In Harlow, L.L., Mulaik, S.A., and Steiger, J.H. (Eds.). What if there were no significance tests? Erlbaum, Mahwah NJ, 393-425. [11] Pearson, K. On the coefficients of racial likeness. Biometrica, 18 (1926), 105-117. [12] Santini, S., and Jain, R. Similarity is a geometer. Multimedia Tools and Application, 5/3 (1997), 277-306. [13] Santini, S., and Jain, R. Similarity measures. IEEE Transactions on Pattern Analysis and Machine Intelligence, 21/9 (September 1999), 871-883. [14] Sint, P.P. Similarity structures and similarity measures. Austrian Academy of Sciences Press, Vienna Austria, 1975 (in German). [15] Smeaton, A.F., and Over, P. The TREC-2002 video track report. NIST Special Publication SP 500-251 (March 2003), available from: http://trec.nist.gov/pubs/trec11/papers/ VIDEO.OVER.pdf (last visited: 2003-07-29) [16] Smeulders, A.W.M., Worring, M., Santini, S., Gupta, A., and Jain, R. Content-based image retrieval at the end of the early years. IEEE Transactions on Pattern Analysis and Machine Intelligence, 22/12 (December 2000), 1349-1380. [17] Tversky, A. Features of similarity. Psychological Review, 84/4 (July 1977), 327-351. 0,000 0,002 0,004 0,006 0,008 0,010 0,012 0,014 0,016 0,018 0,020 P6 MP7 Q8 Q1 Q4 Q2 P2 P4 Q6 Q3 Q7 P1 Q5 P3 P5 P7 Figure 3. Overall results (ordered by the first indicator). The vertical axis shows the values for the first indicator (smaller value = better cluster structure). Shades have the following meaning: black=µ-σ, black + dark grey=µ and black + dark grey + light grey=µ+σ. 137

similarity measurement;performance indicator;visual media;content-base video retrieval;media collection;distance measurement;distance measure;mpeg-7;visual descriptor;mpeg-7-based retrieval;meehl index;visual information retrieval;similarity perception;human similarity perception;content-base image retrieval;predicate-based model

train_H-82

Downloading Textual Hidden Web Content Through Keyword Queries

An ever-increasing amount of information on the Web today is available only through search interfaces: the users have to type in a set of keywords in a search form in order to access the pages from certain Web sites. These pages are often referred to as the Hidden Web or the Deep Web. Since there are no static links to the Hidden Web pages, search engines cannot discover and index such pages and thus do not return them in the results. However, according to recent studies, the content provided by many Hidden Web sites is often of very high quality and can be extremely valuable to many users. In this paper, we study how we can build an effective Hidden Web crawler that can autonomously discover and download pages from the Hidden Web. Since the only entry point to a Hidden Web site is a query interface, the main challenge that a Hidden Web crawler has to face is how to automatically generate meaningful queries to issue to the site. Here, we provide a theoretical framework to investigate the query generation problem for the Hidden Web and we propose effective policies for generating queries automatically. Our policies proceed iteratively, issuing a different query in every iteration. We experimentally evaluate the effectiveness of these policies on 4 real Hidden Web sites and our results are very promising. For instance, in one experiment, one of our policies downloaded more than 90% of a Hidden Web site (that contains 14 million documents) after issuing fewer than 100 queries.

1. INTRODUCTION Recent studies show that a significant fraction of Web content cannot be reached by following links [7, 12]. In particular, a large part of the Web is hidden behind search forms and is reachable only when users type in a set of keywords, or queries, to the forms. These pages are often referred to as the Hidden Web [17] or the Deep Web [7], because search engines typically cannot index the pages and do not return them in their results (thus, the pages are essentially hidden from a typical Web user). According to many studies, the size of the Hidden Web increases rapidly as more organizations put their valuable content online through an easy-to-use Web interface [7]. In [12], Chang et al. estimate that well over 100,000 Hidden-Web sites currently exist on the Web. Moreover, the content provided by many Hidden-Web sites is often of very high quality and can be extremely valuable to many users [7]. For example, PubMed hosts many high-quality papers on medical research that were selected from careful peerreview processes, while the site of the US Patent and Trademarks Office1 makes existing patent documents available, helping potential inventors examine prior art. In this paper, we study how we can build a Hidden-Web crawler2 that can automatically download pages from the Hidden Web, so that search engines can index them. Conventional crawlers rely on the hyperlinks on the Web to discover pages, so current search engines cannot index the Hidden-Web pages (due to the lack of links). We believe that an effective Hidden-Web crawler can have a tremendous impact on how users search information on the Web: • Tapping into unexplored information: The Hidden-Web crawler will allow an average Web user to easily explore the vast amount of information that is mostly hidden at present. Since a majority of Web users rely on search engines to discover pages, when pages are not indexed by search engines, they are unlikely to be viewed by many Web users. Unless users go directly to Hidden-Web sites and issue queries there, they cannot access the pages at the sites. • Improving user experience: Even if a user is aware of a number of Hidden-Web sites, the user still has to waste a significant amount of time and effort, visiting all of the potentially relevant sites, querying each of them and exploring the result. By making the Hidden-Web pages searchable at a central location, we can significantly reduce the user"s wasted time and effort in searching the Hidden Web. • Reducing potential bias: Due to the heavy reliance of many Web users on search engines for locating information, search engines influence how the users perceive the Web [28]. Users do not necessarily perceive what actually exists on the Web, but what is indexed by search engines [28]. According to a recent article [5], several organizations have recognized the importance of bringing information of their Hidden Web sites onto the surface, and committed considerable resources towards this effort. Our 1 US Patent Office: http://www.uspto.gov 2 Crawlers are the programs that traverse the Web automatically and download pages for search engines. 100 Figure 1: A single-attribute search interface Hidden-Web crawler attempts to automate this process for Hidden Web sites with textual content, thus minimizing the associated costs and effort required. Given that the only entry to Hidden Web pages is through querying a search form, there are two core challenges to implementing an effective Hidden Web crawler: (a) The crawler has to be able to understand and model a query interface, and (b) The crawler has to come up with meaningful queries to issue to the query interface. The first challenge was addressed by Raghavan and Garcia-Molina in [29], where a method for learning search interfaces was presented. Here, we present a solution to the second challenge, i.e. how a crawler can automatically generate queries so that it can discover and download the Hidden Web pages. Clearly, when the search forms list all possible values for a query (e.g., through a drop-down list), the solution is straightforward. We exhaustively issue all possible queries, one query at a time. When the query forms have a free text input, however, an infinite number of queries are possible, so we cannot exhaustively issue all possible queries. In this case, what queries should we pick? Can the crawler automatically come up with meaningful queries without understanding the semantics of the search form? In this paper, we provide a theoretical framework to investigate the Hidden-Web crawling problem and propose effective ways of generating queries automatically. We also evaluate our proposed solutions through experiments conducted on real Hidden-Web sites. In summary, this paper makes the following contributions: • We present a formal framework to study the problem of HiddenWeb crawling. (Section 2). • We investigate a number of crawling policies for the Hidden Web, including the optimal policy that can potentially download the maximum number of pages through the minimum number of interactions. Unfortunately, we show that the optimal policy is NP-hard and cannot be implemented in practice (Section 2.2). • We propose a new adaptive policy that approximates the optimal policy. Our adaptive policy examines the pages returned from previous queries and adapts its query-selection policy automatically based on them (Section 3). • We evaluate various crawling policies through experiments on real Web sites. Our experiments will show the relative advantages of various crawling policies and demonstrate their potential. The results from our experiments are very promising. In one experiment, for example, our adaptive policy downloaded more than 90% of the pages within PubMed (that contains 14 million documents) after it issued fewer than 100 queries. 2. FRAMEWORK In this section, we present a formal framework for the study of the Hidden-Web crawling problem. In Section 2.1, we describe our assumptions on Hidden-Web sites and explain how users interact with the sites. Based on this interaction model, we present a highlevel algorithm for a Hidden-Web crawler in Section 2.2. Finally in Section 2.3, we formalize the Hidden-Web crawling problem. 2.1 Hidden-Web database model There exists a variety of Hidden Web sources that provide information on a multitude of topics. Depending on the type of information, we may categorize a Hidden-Web site either as a textual database or a structured database. A textual database is a site that Figure 2: A multi-attribute search interface mainly contains plain-text documents, such as PubMed and LexisNexis (an online database of legal documents [1]). Since plaintext documents do not usually have well-defined structure, most textual databases provide a simple search interface where users type a list of keywords in a single search box (Figure 1). In contrast, a structured database often contains multi-attribute relational data (e.g., a book on the Amazon Web site may have the fields title=‘Harry Potter", author=‘J.K. Rowling" and isbn=‘0590353403") and supports multi-attribute search interfaces (Figure 2). In this paper, we will mainly focus on textual databases that support single-attribute keyword queries. We discuss how we can extend our ideas for the textual databases to multi-attribute structured databases in Section 6.1. Typically, the users need to take the following steps in order to access pages in a Hidden-Web database: 1. Step 1. First, the user issues a query, say liver, through the search interface provided by the Web site (such as the one shown in Figure 1). 2. Step 2. Shortly after the user issues the query, she is presented with a result index page. That is, the Web site returns a list of links to potentially relevant Web pages, as shown in Figure 3(a). 3. Step 3. From the list in the result index page, the user identifies the pages that look interesting and follows the links. Clicking on a link leads the user to the actual Web page, such as the one shown in Figure 3(b), that the user wants to look at. 2.2 A generic Hidden Web crawling algorithm Given that the only entry to the pages in a Hidden-Web site is its search from, a Hidden-Web crawler should follow the three steps described in the previous section. That is, the crawler has to generate a query, issue it to the Web site, download the result index page, and follow the links to download the actual pages. In most cases, a crawler has limited time and network resources, so the crawler repeats these steps until it uses up its resources. In Figure 4 we show the generic algorithm for a Hidden-Web crawler. For simplicity, we assume that the Hidden-Web crawler issues single-term queries only.3 The crawler first decides which query term it is going to use (Step (2)), issues the query, and retrieves the result index page (Step (3)). Finally, based on the links found on the result index page, it downloads the Hidden Web pages from the site (Step (4)). This same process is repeated until all the available resources are used up (Step (1)). Given this algorithm, we can see that the most critical decision that a crawler has to make is what query to issue next. If the crawler can issue successful queries that will return many matching pages, the crawler can finish its crawling early on using minimum resources. In contrast, if the crawler issues completely irrelevant queries that do not return any matching pages, it may waste all of its resources simply issuing queries without ever retrieving actual pages. Therefore, how the crawler selects the next query can greatly affect its effectiveness. In the next section, we formalize this query selection problem. 3 For most Web sites that assume AND for multi-keyword queries, single-term queries return the maximum number of results. Extending our work to multi-keyword queries is straightforward. 101 (a) List of matching pages for query liver. (b) The first matching page for liver. Figure 3: Pages from the PubMed Web site. ALGORITHM 2.1. Crawling a Hidden Web site Procedure (1) while ( there are available resources ) do // select a term to send to the site (2) qi = SelectTerm() // send query and acquire result index page (3) R(qi) = QueryWebSite( qi ) // download the pages of interest (4) Download( R(qi) ) (5) done Figure 4: Algorithm for crawling a Hidden Web site. S q1 q qq 2 34 Figure 5: A set-formalization of the optimal query selection problem. 2.3 Problem formalization Theoretically, the problem of query selection can be formalized as follows: We assume that the crawler downloads pages from a Web site that has a set of pages S (the rectangle in Figure 5). We represent each Web page in S as a point (dots in Figure 5). Every potential query qi that we may issue can be viewed as a subset of S, containing all the points (pages) that are returned when we issue qi to the site. Each subset is associated with a weight that represents the cost of issuing the query. Under this formalization, our goal is to find which subsets (queries) cover the maximum number of points (Web pages) with the minimum total weight (cost). This problem is equivalent to the set-covering problem in graph theory [16]. There are two main difficulties that we need to address in this formalization. First, in a practical situation, the crawler does not know which Web pages will be returned by which queries, so the subsets of S are not known in advance. Without knowing these subsets the crawler cannot decide which queries to pick to maximize the coverage. Second, the set-covering problem is known to be NP-Hard [16], so an efficient algorithm to solve this problem optimally in polynomial time has yet to be found. In this paper, we will present an approximation algorithm that can find a near-optimal solution at a reasonable computational cost. Our algorithm leverages the observation that although we do not know which pages will be returned by each query qi that we issue, we can predict how many pages will be returned. Based on this information our query selection algorithm can then select the best queries that cover the content of the Web site. We present our prediction method and our query selection algorithm in Section 3. 2.3.1 Performance Metric Before we present our ideas for the query selection problem, we briefly discuss some of our notation and the cost/performance metrics. Given a query qi, we use P(qi) to denote the fraction of pages that we will get back if we issue query qi to the site. For example, if a Web site has 10,000 pages in total, and if 3,000 pages are returned for the query qi = medicine, then P(qi) = 0.3. We use P(q1 ∧ q2) to represent the fraction of pages that are returned from both q1 and q2 (i.e., the intersection of P(q1) and P(q2)). Similarly, we use P(q1 ∨ q2) to represent the fraction of pages that are returned from either q1 or q2 (i.e., the union of P(q1) and P(q2)). We also use Cost(qi) to represent the cost of issuing the query qi. Depending on the scenario, the cost can be measured either in time, network bandwidth, the number of interactions with the site, or it can be a function of all of these. As we will see later, our proposed algorithms are independent of the exact cost function. In the most common case, the query cost consists of a number of factors, including the cost for submitting the query to the site, retrieving the result index page (Figure 3(a)) and downloading the actual pages (Figure 3(b)). We assume that submitting a query incurs a fixed cost of cq. The cost for downloading the result index page is proportional to the number of matching documents to the query, while the cost cd for downloading a matching document is also fixed. Then the overall cost of query qi is Cost(qi) = cq + crP(qi) + cdP(qi). (1) In certain cases, some of the documents from qi may have already been downloaded from previous queries. In this case, the crawler may skip downloading these documents and the cost of qi can be Cost(qi) = cq + crP(qi) + cdPnew(qi). (2) Here, we use Pnew(qi) to represent the fraction of the new documents from qi that have not been retrieved from previous queries. Later in Section 3.1 we will study how we can estimate P(qi) and Pnew(qi) to estimate the cost of qi. Since our algorithms are independent of the exact cost function, we will assume a generic cost function Cost(qi) in this paper. When we need a concrete cost function, however, we will use Equation 2. Given the notation, we can formalize the goal of a Hidden-Web crawler as follows: 102 PROBLEM 1. Find the set of queries q1, . . . , qn that maximizes P(q1 ∨ · · · ∨ qn) under the constraint n i=1 Cost(qi) ≤ t. Here, t is the maximum download resource that the crawler has. 3. KEYWORD SELECTION How should a crawler select the queries to issue? Given that the goal is to download the maximum number of unique documents from a textual database, we may consider one of the following options: • Random: We select random keywords from, say, an English dictionary and issue them to the database. The hope is that a random query will return a reasonable number of matching documents. • Generic-frequency: We analyze a generic document corpus collected elsewhere (say, from the Web) and obtain the generic frequency distribution of each keyword. Based on this generic distribution, we start with the most frequent keyword, issue it to the Hidden-Web database and retrieve the result. We then continue to the second-most frequent keyword and repeat this process until we exhaust all download resources. The hope is that the frequent keywords in a generic corpus will also be frequent in the Hidden-Web database, returning many matching documents. • Adaptive: We analyze the documents returned from the previous queries issued to the Hidden-Web database and estimate which keyword is most likely to return the most documents. Based on this analysis, we issue the most promising query, and repeat the process. Among these three general policies, we may consider the random policy as the base comparison point since it is expected to perform the worst. Between the generic-frequency and the adaptive policies, both policies may show similar performance if the crawled database has a generic document collection without a specialized topic. The adaptive policy, however, may perform significantly better than the generic-frequency policy if the database has a very specialized collection that is different from the generic corpus. We will experimentally compare these three policies in Section 4. While the first two policies (random and generic-frequency policies) are easy to implement, we need to understand how we can analyze the downloaded pages to identify the most promising query in order to implement the adaptive policy. We address this issue in the rest of this section. 3.1 Estimating the number of matching pages In order to identify the most promising query, we need to estimate how many new documents we will download if we issue the query qi as the next query. That is, assuming that we have issued queries q1, . . . , qi−1 we need to estimate P(q1∨· · ·∨qi−1∨qi), for every potential next query qi and compare this value. In estimating this number, we note that we can rewrite P(q1 ∨ · · · ∨ qi−1 ∨ qi) as: P((q1 ∨ · · · ∨ qi−1) ∨ qi) = P(q1 ∨ · · · ∨ qi−1) + P(qi) − P((q1 ∨ · · · ∨ qi−1) ∧ qi) = P(q1 ∨ · · · ∨ qi−1) + P(qi) − P(q1 ∨ · · · ∨ qi−1)P(qi|q1 ∨ · · · ∨ qi−1) (3) In the above formula, note that we can precisely measure P(q1 ∨ · · · ∨ qi−1) and P(qi | q1 ∨ · · · ∨ qi−1) by analyzing previouslydownloaded pages: We know P(q1 ∨ · · · ∨ qi−1), the union of all pages downloaded from q1, . . . , qi−1, since we have already issued q1, . . . , qi−1 and downloaded the matching pages.4 We can also measure P(qi | q1 ∨ · · · ∨ qi−1), the probability that qi appears in the pages from q1, . . . , qi−1, by counting how many times qi appears in the pages from q1, . . . , qi−1. Therefore, we only need to estimate P(qi) to evaluate P(q1 ∨ · · · ∨ qi). We may consider a number of different ways to estimate P(qi), including the following: 1. Independence estimator: We assume that the appearance of the term qi is independent of the terms q1, . . . , qi−1. That is, we assume that P(qi) = P(qi|q1 ∨ · · · ∨ qi−1). 2. Zipf estimator: In [19], Ipeirotis et al. proposed a method to estimate how many times a particular term occurs in the entire corpus based on a subset of documents from the corpus. Their method exploits the fact that the frequency of terms inside text collections follows a power law distribution [30, 25]. That is, if we rank all terms based on their occurrence frequency (with the most frequent term having a rank of 1, second most frequent a rank of 2 etc.), then the frequency f of a term inside the text collection is given by: f = α(r + β)−γ (4) where r is the rank of the term and α, β, and γ are constants that depend on the text collection. Their main idea is (1) to estimate the three parameters, α, β and γ, based on the subset of documents that we have downloaded from previous queries, and (2) use the estimated parameters to predict f given the ranking r of a term within the subset. For a more detailed description on how we can use this method to estimate P(qi), we refer the reader to the extended version of this paper [27]. After we estimate P(qi) and P(qi|q1 ∨ · · · ∨ qi−1) values, we can calculate P(q1 ∨ · · · ∨ qi). In Section 3.3, we explain how we can efficiently compute P(qi|q1 ∨ · · · ∨ qi−1) by maintaining a succinct summary table. In the next section, we first examine how we can use this value to decide which query we should issue next to the Hidden Web site. 3.2 Query selection algorithm The goal of the Hidden-Web crawler is to download the maximum number of unique documents from a database using its limited download resources. Given this goal, the Hidden-Web crawler has to take two factors into account. (1) the number of new documents that can be obtained from the query qi and (2) the cost of issuing the query qi. For example, if two queries, qi and qj, incur the same cost, but qi returns more new pages than qj, qi is more desirable than qj. Similarly, if qi and qj return the same number of new documents, but qi incurs less cost then qj, qi is more desirable. Based on this observation, the Hidden-Web crawler may use the following efficiency metric to quantify the desirability of the query qi: Efficiency(qi) = Pnew(qi) Cost(qi) Here, Pnew(qi) represents the amount of new documents returned for qi (the pages that have not been returned for previous queries). Cost(qi) represents the cost of issuing the query qi. Intuitively, the efficiency of qi measures how many new documents are retrieved per unit cost, and can be used as an indicator of 4 For exact estimation, we need to know the total number of pages in the site. However, in order to compare only relative values among queries, this information is not actually needed. 103 ALGORITHM 3.1. Greedy SelectTerm() Parameters: T: The list of potential query keywords Procedure (1) Foreach tk in T do (2) Estimate Efficiency(tk) = Pnew(tk) Cost(tk) (3) done (4) return tk with maximum Efficiency(tk) Figure 6: Algorithm for selecting the next query term. how well our resources are spent when issuing qi. Thus, the Hidden Web crawler can estimate the efficiency of every candidate qi, and select the one with the highest value. By using its resources more efficiently, the crawler may eventually download the maximum number of unique documents. In Figure 6, we show the query selection function that uses the concept of efficiency. In principle, this algorithm takes a greedy approach and tries to maximize the potential gain in every step. We can estimate the efficiency of every query using the estimation method described in Section 3.1. That is, the size of the new documents from the query qi, Pnew(qi), is Pnew(qi) = P(q1 ∨ · · · ∨ qi−1 ∨ qi) − P(q1 ∨ · · · ∨ qi−1) = P(qi) − P(q1 ∨ · · · ∨ qi−1)P(qi|q1 ∨ · · · ∨ qi−1) from Equation 3, where P(qi) can be estimated using one of the methods described in section 3. We can also estimate Cost(qi) similarly. For example, if Cost(qi) is Cost(qi) = cq + crP(qi) + cdPnew(qi) (Equation 2), we can estimate Cost(qi) by estimating P(qi) and Pnew(qi). 3.3 Efficient calculation of query statistics In estimating the efficiency of queries, we found that we need to measure P(qi|q1∨· · ·∨qi−1) for every potential query qi. This calculation can be very time-consuming if we repeat it from scratch for every query qi in every iteration of our algorithm. In this section, we explain how we can compute P(qi|q1 ∨ · · · ∨ qi−1) efficiently by maintaining a small table that we call a query statistics table. The main idea for the query statistics table is that P(qi|q1 ∨· · ·∨ qi−1) can be measured by counting how many times the keyword qi appears within the documents downloaded from q1, . . . , qi−1. We record these counts in a table, as shown in Figure 7(a). The left column of the table contains all potential query terms and the right column contains the number of previously-downloaded documents containing the respective term. For example, the table in Figure 7(a) shows that we have downloaded 50 documents so far, and the term model appears in 10 of these documents. Given this number, we can compute that P(model|q1 ∨ · · · ∨ qi−1) = 10 50 = 0.2. We note that the query statistics table needs to be updated whenever we issue a new query qi and download more documents. This update can be done efficiently as we illustrate in the following example. EXAMPLE 1. After examining the query statistics table of Figure 7(a), we have decided to use the term computer as our next query qi. From the new query qi = computer, we downloaded 20 more new pages. Out of these, 12 contain the keyword model Term tk N(tk) model 10 computer 38 digital 50 Term tk N(tk) model 12 computer 20 disk 18 Total pages: 50 New pages: 20 (a) After q1, . . . , qi−1 (b) New from qi = computer Term tk N(tk) model 10+12 = 22 computer 38+20 = 58 disk 0+18 = 18 digital 50+0 = 50 Total pages: 50 + 20 = 70 (c) After q1, . . . , qi Figure 7: Updating the query statistics table. q i1 i−1 q\/ ... \/q q i / S Figure 8: A Web site that does not return all the results. and 18 the keyword disk. The table in Figure 7(b) shows the frequency of each term in the newly-downloaded pages. We can update the old table (Figure 7(a)) to include this new information by simply adding corresponding entries in Figures 7(a) and (b). The result is shown on Figure 7(c). For example, keyword model exists in 10 + 12 = 22 pages within the pages retrieved from q1, . . . , qi. According to this new table, P(model|q1∨· · ·∨qi) is now 22 70 = 0.3. 3.4 Crawling sites that limit the number of results In certain cases, when a query matches a large number of pages, the Hidden Web site returns only a portion of those pages. For example, the Open Directory Project [2] allows the users to see only up to 10, 000 results after they issue a query. Obviously, this kind of limitation has an immediate effect on our Hidden Web crawler. First, since we can only retrieve up to a specific number of pages per query, our crawler will need to issue more queries (and potentially will use up more resources) in order to download all the pages. Second, the query selection method that we presented in Section 3.2 assumes that for every potential query qi, we can find P(qi|q1 ∨ · · · ∨ qi−1). That is, for every query qi we can find the fraction of documents in the whole text database that contains qi with at least one of q1, . . . , qi−1. However, if the text database returned only a portion of the results for any of the q1, . . . , qi−1 then the value P(qi|q1 ∨ · · · ∨ qi−1) is not accurate and may affect our decision for the next query qi, and potentially the performance of our crawler. Since we cannot retrieve more results per query than the maximum number the Web site allows, our crawler has no other choice besides submitting more queries. However, there is a way to estimate the correct value for P(qi|q1 ∨ · · · ∨ qi−1) in the case where the Web site returns only a portion of the results. 104 Again, assume that the Hidden Web site we are currently crawling is represented as the rectangle on Figure 8 and its pages as points in the figure. Assume that we have already issued queries q1, . . . , qi−1 which returned a number of results less than the maximum number than the site allows, and therefore we have downloaded all the pages for these queries (big circle in Figure 8). That is, at this point, our estimation for P(qi|q1 ∨· · ·∨qi−1) is accurate. Now assume that we submit query qi to the Web site, but due to a limitation in the number of results that we get back, we retrieve the set qi (small circle in Figure 8) instead of the set qi (dashed circle in Figure 8). Now we need to update our query statistics table so that it has accurate information for the next step. That is, although we got the set qi back, for every potential query qi+1 we need to find P(qi+1|q1 ∨ · · · ∨ qi): P(qi+1|q1 ∨ · · · ∨ qi) = 1 P(q1 ∨ · · · ∨ qi) · [P(qi+1 ∧ (q1 ∨ · · · ∨ qi−1))+ P(qi+1 ∧ qi) − P(qi+1 ∧ qi ∧ (q1 ∨ · · · ∨ qi−1))] (5) In the previous equation, we can find P(q1 ∨· · ·∨qi) by estimating P(qi) with the method shown in Section 3. Additionally, we can calculate P(qi+1 ∧ (q1 ∨ · · · ∨ qi−1)) and P(qi+1 ∧ qi ∧ (q1 ∨ · · · ∨ qi−1)) by directly examining the documents that we have downloaded from queries q1, . . . , qi−1. The term P(qi+1 ∧ qi) however is unknown and we need to estimate it. Assuming that qi is a random sample of qi, then: P(qi+1 ∧ qi) P(qi+1 ∧ qi) = P(qi) P(qi) (6) From Equation 6 we can calculate P(qi+1 ∧ qi) and after we replace this value to Equation 5 we can find P(qi+1|q1 ∨ · · · ∨ qi). 4. EXPERIMENTAL EVALUATION In this section we experimentally evaluate the performance of the various algorithms for Hidden Web crawling presented in this paper. Our goal is to validate our theoretical analysis through realworld experiments, by crawling popular Hidden Web sites of textual databases. Since the number of documents that are discovered and downloaded from a textual database depends on the selection of the words that will be issued as queries5 to the search interface of each site, we compare the various selection policies that were described in section 3, namely the random, generic-frequency, and adaptive algorithms. The adaptive algorithm learns new keywords and terms from the documents that it downloads, and its selection process is driven by a cost model as described in Section 3.2. To keep our experiment and its analysis simple at this point, we will assume that the cost for every query is constant. That is, our goal is to maximize the number of downloaded pages by issuing the least number of queries. Later, in Section 4.4 we will present a comparison of our policies based on a more elaborate cost model. In addition, we use the independence estimator (Section 3.1) to estimate P(qi) from downloaded pages. Although the independence estimator is a simple estimator, our experiments will show that it can work very well in practice.6 For the generic-frequency policy, we compute the frequency distribution of words that appear in a 5.5-million-Web-page corpus 5 Throughout our experiments, once an algorithm has submitted a query to a database, we exclude the query from subsequent submissions to the same database from the same algorithm. 6 We defer the reporting of results based on the Zipf estimation to a future work. downloaded from 154 Web sites of various topics [26]. Keywords are selected based on their decreasing frequency with which they appear in this document set, with the most frequent one being selected first, followed by the second-most frequent keyword, etc.7 Regarding the random policy, we use the same set of words collected from the Web corpus, but in this case, instead of selecting keywords based on their relative frequency, we choose them randomly (uniform distribution). In order to further investigate how the quality of the potential query-term list affects the random-based algorithm, we construct two sets: one with the 16, 000 most frequent words of the term collection used in the generic-frequency policy (hereafter, the random policy with the set of 16,000 words will be referred to as random-16K), and another set with the 1 million most frequent words of the same collection as above (hereafter, referred to as random-1M). The former set has frequent words that appear in a large number of documents (at least 10, 000 in our collection), and therefore can be considered of high-quality terms. The latter set though contains a much larger collection of words, among which some might be bogus, and meaningless. The experiments were conducted by employing each one of the aforementioned algorithms (adaptive, generic-frequency, random16K, and random-1M) to crawl and download contents from three Hidden Web sites: The PubMed Medical Library,8 Amazon,9 and the Open Directory Project[2]. According to the information on PubMed"s Web site, its collection contains approximately 14 million abstracts of biomedical articles. We consider these abstracts as the documents in the site, and in each iteration of the adaptive policy, we use these abstracts as input to the algorithm. Thus our goal is to discover as many unique abstracts as possible by repeatedly querying the Web query interface provided by PubMed. The Hidden Web crawling on the PubMed Web site can be considered as topic-specific, due to the fact that all abstracts within PubMed are related to the fields of medicine and biology. In the case of the Amazon Web site, we are interested in downloading all the hidden pages that contain information on books. The querying to Amazon is performed through the Software Developer"s Kit that Amazon provides for interfacing to its Web site, and which returns results in XML form. The generic keyword field is used for searching, and as input to the adaptive policy we extract the product description and the text of customer reviews when present in the XML reply. Since Amazon does not provide any information on how many books it has in its catalogue, we use random sampling on the 10-digit ISBN number of the books to estimate the size of the collection. Out of the 10, 000 random ISBN numbers queried, 46 are found in the Amazon catalogue, therefore the size of its book collection is estimated to be 46 10000 · 1010 = 4.6 million books. It"s also worth noting here that Amazon poses an upper limit on the number of results (books in our case) returned by each query, which is set to 32, 000. As for the third Hidden Web site, the Open Directory Project (hereafter also referred to as dmoz), the site maintains the links to 3.8 million sites together with a brief summary of each listed site. The links are searchable through a keyword-search interface. We consider each indexed link together with its brief summary as the document of the dmoz site, and we provide the short summaries to the adaptive algorithm to drive the selection of new keywords for querying. On the dmoz Web site, we perform two Hidden Web crawls: the first is on its generic collection of 3.8-million indexed 7 We did not manually exclude stop words (e.g., the, is, of, etc.) from the keyword list. As it turns out, all Web sites except PubMed return matching documents for the stop words, such as the. 8 PubMed Medical Library: http://www.pubmed.org 9 Amazon Inc.: http://www.amazon.com 105 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0 50 100 150 200 fractionofdocuments query number Cumulative fraction of unique documents - PubMed website adaptive generic-frequency random-16K random-1M Figure 9: Coverage of policies for Pubmed 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0 100 200 300 400 500 600 700 fractionofdocuments query number Cumulative fraction of unique documents - Amazon website adaptive generic-frequency random-16K random-1M Figure 10: Coverage of policies for Amazon sites, regardless of the category that they fall into. The other crawl is performed specifically on the Arts section of dmoz (http:// dmoz.org/Arts), which comprises of approximately 429, 000 indexed sites that are relevant to Arts, making this crawl topicspecific, as in PubMed. Like Amazon, dmoz also enforces an upper limit on the number of returned results, which is 10, 000 links with their summaries. 4.1 Comparison of policies The first question that we seek to answer is the evolution of the coverage metric as we submit queries to the sites. That is, what fraction of the collection of documents stored in the Hidden Web site can we download as we continuously query for new words selected using the policies described above? More formally, we are interested in the value of P(q1 ∨ · · · ∨ qi−1 ∨ qi), after we submit q1, . . . , qi queries, and as i increases. In Figures 9, 10, 11, and 12 we present the coverage metric for each policy, as a function of the query number, for the Web sites of PubMed, Amazon, general dmoz and the art-specific dmoz, respectively. On the y-axis the fraction of the total documents downloaded from the website is plotted, while the x-axis represents the query number. A first observation from these graphs is that in general, the generic-frequency and the adaptive policies perform much better than the random-based algorithms. In all of the figures, the graphs for the random-1M and the random-16K are significantly below those of other policies. Between the generic-frequency and the adaptive policies, we can see that the latter outperforms the former when the site is topic spe0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0 100 200 300 400 500 600 700 fractionofdocuments query number Cumulative fraction of unique documents - dmoz website adaptive generic-frequency random-16K random-1M Figure 11: Coverage of policies for general dmoz 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0 50 100 150 200 250 300 350 400 450 fractionofdocuments query number Cumulative fraction of unique documents - dmoz/Arts website adaptive generic-frequency random-16K random-1M Figure 12: Coverage of policies for the Arts section of dmoz cific. For example, for the PubMed site (Figure 9), the adaptive algorithm issues only 83 queries to download almost 80% of the documents stored in PubMed, but the generic-frequency algorithm requires 106 queries for the same coverage,. For the dmoz/Arts crawl (Figure 12), the difference is even more substantial: the adaptive policy is able to download 99.98% of the total sites indexed in the Directory by issuing 471 queries, while the frequency-based algorithm is much less effective using the same number of queries, and discovers only 72% of the total number of indexed sites. The adaptive algorithm, by examining the contents of the pages that it downloads at each iteration, is able to identify the topic of the site as expressed by the words that appear most frequently in the result-set. Consequently, it is able to select words for subsequent queries that are more relevant to the site, than those preferred by the genericfrequency policy, which are drawn from a large, generic collection. Table 1 shows a sample of 10 keywords out of 211 chosen and submitted to the PubMed Web site by the adaptive algorithm, but not by the other policies. For each keyword, we present the number of the iteration, along with the number of results that it returned. As one can see from the table, these keywords are highly relevant to the topics of medicine and biology of the Public Medical Library, and match against numerous articles stored in its Web site. In both cases examined in Figures 9, and 12, the random-based policies perform much worse than the adaptive algorithm, and the generic-frequency. It is worthy noting however, that the randombased policy with the small, carefully selected set of 16, 000 quality words manages to download a considerable fraction of 42.5% 106 Iteration Keyword Number of Results 23 department 2, 719, 031 34 patients 1, 934, 428 53 clinical 1, 198, 322 67 treatment 4, 034, 565 69 medical 1, 368, 200 70 hospital 503, 307 146 disease 1, 520, 908 172 protein 2, 620, 938 Table 1: Sample of keywords queried to PubMed exclusively by the adaptive policy from the PubMed Web site after 200 queries, while the coverage for the Arts section of dmoz reaches 22.7%, after 471 queried keywords. On the other hand, the random-based approach that makes use of the vast collection of 1 million words, among which a large number is bogus keywords, fails to download even a mere 1% of the total collection, after submitting the same number of query words. For the generic collections of Amazon and the dmoz sites, shown in Figures 10 and 11 respectively, we get mixed results: The genericfrequency policy shows slightly better performance than the adaptive policy for the Amazon site (Figure 10), and the adaptive method clearly outperforms the generic-frequency for the general dmoz site (Figure 11). A closer look at the log files of the two Hidden Web crawlers reveals the main reason: Amazon was functioning in a very flaky way when the adaptive crawler visited it, resulting in a large number of lost results. Thus, we suspect that the slightly poor performance of the adaptive policy is due to this experimental variance. We are currently running another experiment to verify whether this is indeed the case. Aside from this experimental variance, the Amazon result indicates that if the collection and the words that a Hidden Web site contains are generic enough, then the generic-frequency approach may be a good candidate algorithm for effective crawling. As in the case of topic-specific Hidden Web sites, the randombased policies also exhibit poor performance compared to the other two algorithms when crawling generic sites: for the Amazon Web site, random-16K succeeds in downloading almost 36.7% after issuing 775 queries, alas for the generic collection of dmoz, the fraction of the collection of links downloaded is 13.5% after the 770th query. Finally, as expected, random-1M is even worse than random16K, downloading only 14.5% of Amazon and 0.3% of the generic dmoz. In summary, the adaptive algorithm performs remarkably well in all cases: it is able to discover and download most of the documents stored in Hidden Web sites by issuing the least number of queries. When the collection refers to a specific topic, it is able to identify the keywords most relevant to the topic of the site and consequently ask for terms that is most likely that will return a large number of results . On the other hand, the generic-frequency policy proves to be quite effective too, though less than the adaptive: it is able to retrieve relatively fast a large portion of the collection, and when the site is not topic-specific, its effectiveness can reach that of adaptive (e.g. Amazon). Finally, the random policy performs poorly in general, and should not be preferred. 4.2 Impact of the initial query An interesting issue that deserves further examination is whether the initial choice of the keyword used as the first query issued by the adaptive algorithm affects its effectiveness in subsequent iterations. The choice of this keyword is not done by the selection of the 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0 10 20 30 40 50 60 fractionofdocuments query number Convergence of adaptive under different initial queries - PubMed website pubmed data information return Figure 13: Convergence of the adaptive algorithm using different initial queries for crawling the PubMed Web site adaptive algorithm itself and has to be manually set, since its query statistics tables have not been populated yet. Thus, the selection is generally arbitrary, so for purposes of fully automating the whole process, some additional investigation seems necessary. For this reason, we initiated three adaptive Hidden Web crawlers targeting the PubMed Web site with different seed-words: the word data, which returns 1,344,999 results, the word information that reports 308, 474 documents, and the word return that retrieves 29, 707 pages, out of 14 million. These keywords represent varying degrees of term popularity in PubMed, with the first one being of high popularity, the second of medium, and the third of low. We also show results for the keyword pubmed, used in the experiments for coverage of Section 4.1, and which returns 695 articles. As we can see from Figure 13, after a small number of queries, all four crawlers roughly download the same fraction of the collection, regardless of their starting point: Their coverages are roughly equivalent from the 25th query. Eventually, all four crawlers use the same set of terms for their queries, regardless of the initial query. In the specific experiment, from the 36th query onward, all four crawlers use the same terms for their queries in each iteration, or the same terms are used off by one or two query numbers. Our result confirms the observation of [11] that the choice of the initial query has minimal effect on the final performance. We can explain this intuitively as follows: Our algorithm approximates the optimal set of queries to use for a particular Web site. Once the algorithm has issued a significant number of queries, it has an accurate estimation of the content of the Web site, regardless of the initial query. Since this estimation is similar for all runs of the algorithm, the crawlers will use roughly the same queries. 4.3 Impact of the limit in the number of results While the Amazon and dmoz sites have the respective limit of 32,000 and 10,000 in their result sizes, these limits may be larger than those imposed by other Hidden Web sites. In order to investigate how a tighter limit in the result size affects the performance of our algorithms, we performed two additional crawls to the generic-dmoz site: we ran the generic-frequency and adaptive policies but we retrieved only up to the top 1,000 results for every query. In Figure 14 we plot the coverage for the two policies as a function of the number of queries. As one might expect, by comparing the new result in Figure 14 to that of Figure 11 where the result limit was 10,000, we conclude that the tighter limit requires a higher number of queries to achieve the same coverage. For example, when the result limit was 10,000, the adaptive pol107 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0 500 1000 1500 2000 2500 3000 3500 FractionofUniquePages Query Number Cumulative fraction of unique pages downloaded per query - Dmoz Web site (cap limit 1000) adaptive generic-frequency Figure 14: Coverage of general dmoz after limiting the number of results to 1,000 icy could download 70% of the site after issuing 630 queries, while it had to issue 2,600 queries to download 70% of the site when the limit was 1,000. On the other hand, our new result shows that even with a tight result limit, it is still possible to download most of a Hidden Web site after issuing a reasonable number of queries. The adaptive policy could download more than 85% of the site after issuing 3,500 queries when the limit was 1,000. Finally, our result shows that our adaptive policy consistently outperforms the generic-frequency policy regardless of the result limit. In both Figure 14 and Figure 11, our adaptive policy shows significantly larger coverage than the generic-frequency policy for the same number of queries. 4.4 Incorporating the document download cost For brevity of presentation, the performance evaluation results provided so far assumed a simplified cost-model where every query involved a constant cost. In this section we present results regarding the performance of the adaptive and generic-frequency algorithms using Equation 2 to drive our query selection process. As we discussed in Section 2.3.1, this query cost model includes the cost for submitting the query to the site, retrieving the result index page, and also downloading the actual pages. For these costs, we examined the size of every result in the index page and the sizes of the documents, and we chose cq = 100, cr = 100, and cd = 10000, as values for the parameters of Equation 2, and for the particular experiment that we ran on the PubMed website. The values that we selected imply that the cost for issuing one query and retrieving one result from the result index page are roughly the same, while the cost for downloading an actual page is 100 times larger. We believe that these values are reasonable for the PubMed Web site. Figure 15 shows the coverage of the adaptive and genericfrequency algorithms as a function of the resource units used during the download process. The horizontal axis is the amount of resources used, and the vertical axis is the coverage. As it is evident from the graph, the adaptive policy makes more efficient use of the available resources, as it is able to download more articles than the generic-frequency, using the same amount of resource units. However, the difference in coverage is less dramatic in this case, compared to the graph of Figure 9. The smaller difference is due to the fact that under the current cost metric, the download cost of documents constitutes a significant portion of the cost. Therefore, when both policies downloaded the same number of documents, the saving of the adaptive policy is not as dramatic as before. That 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0 5000 10000 15000 20000 25000 30000 FractionofUniquePages Total Cost (cq=100, cr=100, cd=10000) Cumulative fraction of unique pages downloaded per cost unit - PubMed Web site adaptive frequency Figure 15: Coverage of PubMed after incorporating the document download cost is, the savings in the query cost and the result index download cost is only a relatively small portion of the overall cost. Still, we observe noticeable savings from the adaptive policy. At the total cost of 8000, for example, the coverage of the adaptive policy is roughly 0.5 while the coverage of the frequency policy is only 0.3. 5. RELATED WORK In a recent study, Raghavan and Garcia-Molina [29] present an architectural model for a Hidden Web crawler. The main focus of this work is to learn Hidden-Web query interfaces, not to generate queries automatically. The potential queries are either provided manually by users or collected from the query interfaces. In contrast, our main focus is to generate queries automatically without any human intervention. The idea of automatically issuing queries to a database and examining the results has been previously used in different contexts. For example, in [10, 11], Callan and Connel try to acquire an accurate language model by collecting a uniform random sample from the database. In [22] Lawrence and Giles issue random queries to a number of Web Search Engines in order to estimate the fraction of the Web that has been indexed by each of them. In a similar fashion, Bharat and Broder [8] issue random queries to a set of Search Engines in order to estimate the relative size and overlap of their indexes. In [6], Barbosa and Freire experimentally evaluate methods for building multi-keyword queries that can return a large fraction of a document collection. Our work differs from the previous studies in two ways. First, it provides a theoretical framework for analyzing the process of generating queries for a database and examining the results, which can help us better understand the effectiveness of the methods presented in the previous work. Second, we apply our framework to the problem of Hidden Web crawling and demonstrate the efficiency of our algorithms. Cope et al. [15] propose a method to automatically detect whether a particular Web page contains a search form. This work is complementary to ours; once we detect search interfaces on the Web using the method in [15], we may use our proposed algorithms to download pages automatically from those Web sites. Reference [4] reports methods to estimate what fraction of a text database can be eventually acquired by issuing queries to the database. In [3] the authors study query-based techniques that can extract relational data from large text databases. Again, these works study orthogonal issues and are complementary to our work. In order to make documents in multiple textual databases searchable at a central place, a number of harvesting approaches have 108 been proposed (e.g., OAI [21], DP9 [24]). These approaches essentially assume cooperative document databases that willingly share some of their metadata and/or documents to help a third-party search engine to index the documents. Our approach assumes uncooperative databases that do not share their data publicly and whose documents are accessible only through search interfaces. There exists a large body of work studying how to identify the most relevant database given a user query [20, 19, 14, 23, 18]. This body of work is often referred to as meta-searching or database selection problem over the Hidden Web. For example, [19] suggests the use of focused probing to classify databases into a topical category, so that given a query, a relevant database can be selected based on its topical category. Our vision is different from this body of work in that we intend to download and index the Hidden pages at a central location in advance, so that users can access all the information at their convenience from one single location. 6. CONCLUSION AND FUTURE WORK Traditional crawlers normally follow links on the Web to discover and download pages. Therefore they cannot get to the Hidden Web pages which are only accessible through query interfaces. In this paper, we studied how we can build a Hidden Web crawler that can automatically query a Hidden Web site and download pages from it. We proposed three different query generation policies for the Hidden Web: a policy that picks queries at random from a list of keywords, a policy that picks queries based on their frequency in a generic text collection, and a policy which adaptively picks a good query based on the content of the pages downloaded from the Hidden Web site. Experimental evaluation on 4 real Hidden Web sites shows that our policies have a great potential. In particular, in certain cases the adaptive policy can download more than 90% of a Hidden Web site after issuing approximately 100 queries. Given these results, we believe that our work provides a potential mechanism to improve the search-engine coverage of the Web and the user experience of Web search. 6.1 Future Work We briefly discuss some future-research avenues. Multi-attribute Databases We are currently investigating how to extend our ideas to structured multi-attribute databases. While generating queries for multi-attribute databases is clearly a more difficult problem, we may exploit the following observation to address this problem: When a site supports multi-attribute queries, the site often returns pages that contain values for each of the query attributes. For example, when an online bookstore supports queries on title, author and isbn, the pages returned from a query typically contain the title, author and ISBN of corresponding books. Thus, if we can analyze the returned pages and extract the values for each field (e.g, title = ‘Harry Potter", author = ‘J.K. Rowling", etc), we can apply the same idea that we used for the textual database: estimate the frequency of each attribute value and pick the most promising one. The main challenge is to automatically segment the returned pages so that we can identify the sections of the pages that present the values corresponding to each attribute. Since many Web sites follow limited formatting styles in presenting multiple attributes - for example, most book titles are preceded by the label Title: - we believe we may learn page-segmentation rules automatically from a small set of training examples. Other Practical Issues In addition to the automatic query generation problem, there are many practical issues to be addressed to build a fully automatic Hidden-Web crawler. For example, in this paper we assumed that the crawler already knows all query interfaces for Hidden-Web sites. But how can the crawler discover the query interfaces? The method proposed in [15] may be a good starting point. In addition, some Hidden-Web sites return their results in batches of, say, 20 pages, so the user has to click on a next button in order to see more results. In this case, a fully automatic Hidden-Web crawler should know that the first result index page contains only a partial result and press the next button automatically. Finally, some Hidden Web sites may contain an infinite number of Hidden Web pages which do not contribute much significant content (e.g. a calendar with links for every day). In this case the Hidden-Web crawler should be able to detect that the site does not have much more new content and stop downloading pages from the site. Page similarity detection algorithms may be useful for this purpose [9, 13]. 7. REFERENCES [1] Lexisnexis http://www.lexisnexis.com. [2] The Open Directory Project, http://www.dmoz.org. [3] E. Agichtein and L. Gravano. Querying text databases for efficient information extraction. In ICDE, 2003. [4] E. Agichtein, P. Ipeirotis, and L. Gravano. Modeling query-based access to text databases. In WebDB, 2003. [5] Article on New York Times. Old Search Engine, the Library, Tries to Fit Into a Google World. Available at: http: //www.nytimes.com/2004/06/21/technology/21LIBR.html, June 2004. [6] L. Barbosa and J. Freire. Siphoning hidden-web data through keyword-based interfaces. In SBBD, 2004. [7] M. K. Bergman. The deep web: Surfacing hidden value,http: //www.press.umich.edu/jep/07-01/bergman.html. [8] K. Bharat and A. Broder. A technique for measuring the relative size and overlap of public web search engines. In WWW, 1998. [9] A. Z. Broder, S. C. Glassman, M. S. Manasse, and G. Zweig. Syntactic clustering of the web. In WWW, 1997. [10] J. Callan, M. Connell, and A. Du. Automatic discovery of language models for text databases. In SIGMOD, 1999. [11] J. P. Callan and M. E. Connell. Query-based sampling of text databases. Information Systems, 19(2):97-130, 2001. [12] K. C.-C. Chang, B. He, C. Li, and Z. Zhang. Structured databases on the web: Observations and implications. Technical report, UIUC. [13] J. Cho, N. Shivakumar, and H. Garcia-Molina. Finding replicated web collections. In SIGMOD, 2000. [14] W. Cohen and Y. Singer. Learning to query the web. In AAAI Workshop on Internet-Based Information Systems, 1996. [15] J. Cope, N. Craswell, and D. Hawking. Automated discovery of search interfaces on the web. In 14th Australasian conference on Database technologies, 2003. [16] T. H. Cormen, C. E. Leiserson, and R. L. Rivest. Introduction to Algorithms, 2nd Edition. MIT Press/McGraw Hill, 2001. [17] D. Florescu, A. Y. Levy, and A. O. Mendelzon. Database techniques for the world-wide web: A survey. SIGMOD Record, 27(3):59-74, 1998. [18] B. He and K. C.-C. Chang. Statistical schema matching across web query interfaces. In SIGMOD Conference, 2003. [19] P. Ipeirotis and L. Gravano. Distributed search over the hidden web: Hierarchical database sampling and selection. In VLDB, 2002. [20] P. G. Ipeirotis, L. Gravano, and M. Sahami. Probe, count, and classify: Categorizing hidden web databases. In SIGMOD, 2001. [21] C. Lagoze and H. V. Sompel. The Open Archives Initiative: Building a low-barrier interoperability framework In JCDL, 2001. [22] S. Lawrence and C. L. Giles. Searching the World Wide Web. Science, 280(5360):98-100, 1998. [23] V. Z. Liu, J. C. Richard C. Luo and, and W. W. Chu. Dpro: A probabilistic approach for hidden web database selection using dynamic probing. In ICDE, 2004. [24] X. Liu, K. Maly, M. Zubair and M. L. Nelson. DP9-An OAI Gateway Service for Web Crawlers. In JCDL, 2002. [25] B. B. Mandelbrot. Fractal Geometry of Nature. W. H. Freeman & Co. [26] A. Ntoulas, J. Cho, and C. Olston. What"s new on the web? the evolution of the web from a search engine perspective. In WWW, 2004. [27] A. Ntoulas, P. Zerfos, and J. Cho. Downloading hidden web content. Technical report, UCLA, 2004. [28] S. Olsen. Does search engine"s power threaten web"s independence? http://news.com.com/2009-1023-963618.html. [29] S. Raghavan and H. Garcia-Molina. Crawling the hidden web. In VLDB, 2001. [30] G. K. Zipf. Human Behavior and the Principle of Least-Effort. Addison-Wesley, Cambridge, MA, 1949. 109

textual database;independence estimator;hidden web;hide web crawl;generic-frequency policy;deep web crawler;query selection;query-selection policy;adaptive algorithm;crawling policy;adaptive policy;keyword query;deep web;hidden web crawler

train_H-83

Estimating the Global PageRank of Web Communities

Localized search engines are small-scale systems that index a particular community on the web. They offer several benefits over their large-scale counterparts in that they are relatively inexpensive to build, and can provide more precise and complete search capability over their relevant domains. One disadvantage such systems have over large-scale search engines is the lack of global PageRank values. Such information is needed to assess the value of pages in the localized search domain within the context of the web as a whole. In this paper, we present well-motivated algorithms to estimate the global PageRank values of a local domain. The algorithms are all highly scalable in that, given a local domain of size n, they use O(n) resources that include computation time, bandwidth, and storage. We test our methods across a variety of localized domains, including site-specific domains and topic-specific domains. We demonstrate that by crawling as few as n or 2n additional pages, our methods can give excellent global PageRank estimates.

1. INTRODUCTION Localized search engines are small-scale search engines that index only a single community of the web. Such communities can be site-specific domains, such as pages within the cs.utexas.edu domain, or topic-related communitiesfor example, political websites. Compared to the web graph crawled and indexed by large-scale search engines, the size of such local communities is typically orders of magnitude smaller. Consequently, the computational resources needed to build such a search engine are also similarly lighter. By restricting themselves to smaller, more manageable sections of the web, localized search engines can also provide more precise and complete search capabilities over their respective domains. One drawback of localized indexes is the lack of global information needed to compute link-based rankings. The PageRank algorithm [3], has proven to be an effective such measure. In general, the PageRank of a given page is dependent on pages throughout the entire web graph. In the context of a localized search engine, if the PageRanks are computed using only the local subgraph, then we would expect the resulting PageRanks to reflect the perceived popularity within the local community and not of the web as a whole. For example, consider a localized search engine that indexes political pages with conservative views. A person wishing to research the opinions on global warming within the conservative political community may encounter numerous such opinions across various websites. If only local PageRank values are available, then the search results will reflect only strongly held beliefs within the community. However, if global PageRanks are also available, then the results can additionally reflect outsiders" views of the conservative community (those documents that liberals most often access within the conservative community). Thus, for many localized search engines, incorporating global PageRanks can improve the quality of search results. However, the number of pages a local search engine indexes is typically orders of magnitude smaller than the number of pages indexed by their large-scale counterparts. Localized search engines do not have the bandwidth, storage capacity, or computational power to crawl, download, and compute the global PageRanks of the entire web. In this work, we present a method of approximating the global PageRanks of a local domain while only using resources of the same order as those needed to compute the PageRanks of the local subgraph. Our proposed method looks for a supergraph of our local subgraph such that the local PageRanks within this supergraph are close to the true global PageRanks. We construct this supergraph by iteratively crawling global pages on the current web frontier-i.e., global pages with inlinks from pages that have already been crawled. In order to provide 116 Research Track Paper a good approximation to the global PageRanks, care must be taken when choosing which pages to crawl next; in this paper, we present a well-motivated page selection algorithm that also performs well empirically. This algorithm is derived from a well-defined problem objective and has a running time linear in the number of local nodes. We experiment across several types of local subgraphs, including four topic related communities and several sitespecific domains. To evaluate performance, we measure the difference between the current global PageRank estimate and the global PageRank, as a function of the number of pages crawled. We compare our algorithm against several heuristics and also against a baseline algorithm that chooses pages at random, and we show that our method outperforms these other methods. Finally, we empirically demonstrate that, given a local domain of size n, we can provide good approximations to the global PageRank values by crawling at most n or 2n additional pages. The paper is organized as follows. Section 2 gives an overview of localized search engines and outlines their advantages over global search. Section 3 provides background on the PageRank algorithm. Section 4 formally defines our problem, and section 5 presents our page selection criteria and derives our algorithms. Section 6 provides experimental results, section 7 gives an overview of related work, and, finally, conclusions are given in section 8. 2. LOCALIZED SEARCH ENGINES Localized search engines index a single community of the web, typically either a site-specific community, or a topicspecific community. Localized search engines enjoy three major advantages over their large-scale counterparts: they are relatively inexpensive to build, they can offer more precise search capability over their local domain, and they can provide a more complete index. The resources needed to build a global search engine are enormous. A 2003 study by Lyman et al. [13] found that the ‘surface web" (publicly available static sites) consists of 8.9 billion pages, and that the average size of these pages is approximately 18.7 kilobytes. To download a crawl of this size, approximately 167 terabytes of space is needed. For a researcher who wishes to build a search engine with access to a couple of workstations or a small server, storage of this magnitude is simply not available. However, building a localized search engine over a web community of a hundred thousand pages would only require a few gigabytes of storage. The computational burden required to support search queries over a database this size is more manageable as well. We note that, for topic-specific search engines, the relevant community can be efficiently identified and downloaded by using a focused crawler [21, 4]. For site-specific domains, the local domain is readily available on their own web server. This obviates the need for crawling or spidering, and a complete and up-to-date index of the domain can thus be guaranteed. This is in contrast to their large-scale counterparts, which suffer from several shortcomings. First, crawling dynamically generated pages-pages in the ‘hidden web"-has been the subject of research [20] and is a non-trivial task for an external crawler. Second, site-specific domains can enable the robots exclusion policy. This prohibits external search engines" crawlers from downloading content from the domain, and an external search engine must instead rely on outside links and anchor text to index these restricted pages. By restricting itself to only a specific domain of the internet, a localized search engine can provide more precise search results. Consider the canonical ambiguous search query, ‘jaguar", which can refer to either the car manufacturer or the animal. A scientist trying to research the habitat and evolutionary history of a jaguar may have better success using a finely tuned zoology-specific search engine than querying Google with multiple keyword searches and wading through irrelevant results. A method to learn better ranking functions for retrieval was recently proposed by Radlinski and Joachims [19] and has been applied to various local domains, including Cornell University"s website [8]. 3. PAGERANK OVERVIEW The PageRank algorithm defines the importance of web pages by analyzing the underlying hyperlink structure of a web graph. The algorithm works by building a Markov chain from the link structure of the web graph and computing its stationary distribution. One way to compute the stationary distribution of a Markov chain is to find the limiting distribution of a random walk over the chain. Thus, the PageRank algorithm uses what is sometimes referred to as the ‘random surfer" model. In each step of the random walk, the ‘surfer" either follows an outlink from the current page (i.e. the current node in the chain), or jumps to a random page on the web. We now precisely define the PageRank problem. Let U be an m × m adjacency matrix for a given web graph such that Uji = 1 if page i links to page j and Uji = 0 otherwise. We define the PageRank matrix PU to be: PU = αUD−1 U + (1 − α)veT , (1) where DU is the (unique) diagonal matrix such that UD−1 U is column stochastic, α is a given scalar such that 0 ≤ α ≤ 1, e is the vector of all ones, and v is a non-negative, L1normalized vector, sometimes called the ‘random surfer" vector. Note that the matrix D−1 U is well-defined only if each column of U has at least one non-zero entry-i.e., each page in the webgraph has at least one outlink. In the presence of such ‘dangling nodes" that have no outlinks, one commonly used solution, proposed by Brin et al. [3], is to replace each zero column of U by a non-negative, L1-normalized vector. The PageRank vector r is the dominant eigenvector of the PageRank matrix, r = PU r. We will assume, without loss of generality, that r has an L1-norm of one. Computationally, r can be computed using the power method. This method first chooses a random starting vector r(0) , and iteratively multiplies the current vector by the PageRank matrix PU ; see Algorithm 1. In general, each iteration of the power method can take O(m2 ) operations when PU is a dense matrix. However, in practice, the number of links in a web graph will be of the order of the number of pages. By exploiting the sparsity of the PageRank matrix, the work per iteration can be reduced to O(km), where k is the average number of links per web page. It has also been shown that the total number of iterations needed for convergence is proportional to α and does not depend on the size of the web graph [11, 7]. Finally, the total space needed is also O(km), mainly to store the matrix U. 117 Research Track Paper Algorithm 1: A linear time (per iteration) algorithm for computing PageRank. ComputePR(U) Input: U: Adjacency matrix. Output: r: PageRank vector. Choose (randomly) an initial non-negative vector r(0) such that r(0) 1 = 1. i ← 0 repeat i ← i + 1 ν ← αUD−1 U r(i−1) {α is the random surfing probability} r(i) ← ν + (1 − α)v {v is the random surfer vector.} until r(i) − r(i−1) < δ {δ is the convergence threshold.} r ← r(i) 4. PROBLEM DEFINITION Given a local domain L, let G be an N × N adjacency matrix for the entire connected component of the web that contains L, such that Gji = 1 if page i links to page j and Gji = 0 otherwise. Without loss of generality, we will partition G as: G = L Gout Lout Gwithin , (2) where L is the n × n local subgraph corresponding to links inside the local domain, Lout is the subgraph that corresponds to links from the local domain pointing out to the global domain, Gout is the subgraph containing links from the global domain into the local domain, and Gwithin contains links within the global domain. We assume that when building a localized search engine, only pages inside the local domain are crawled, and the links between these pages are represented by the subgraph L. The links in Lout are also known, as these point from crawled pages in the local domain to uncrawled pages in the global domain. As defined in equation (1), PG is the PageRank matrix formed from the global graph G, and we define the global PageRank vector of this graph to be g. Let the n-length vector p∗ be the L1-normalized vector corresponding to the global PageRank of the pages in the local domain L: p∗ = EL g ELg 1 , where EL = [ I | 0 ] is the restriction matrix that selects the components from g corresponding to nodes in L. Let p denote the PageRank vector constructed from the local domain subgraph L. In practice, the observed local PageRank p and the global PageRank p∗ will be quite different. One would expect that as the size of local matrix L approaches the size of global matrix G, the global PageRank and the observed local PageRank will become more similar. Thus, one approach to estimating the global PageRank is to crawl the entire global domain, compute its PageRank, and extract the PageRanks of the local domain. Typically, however, n N , i.e., the number of global pages is much larger than the number of local pages. Therefore, crawling all global pages will quickly exhaust all local resources (computational, storage, and bandwidth) available to create the local search engine. We instead seek a supergraph ˆF of our local subgraph L with size O(n). Our goal Algorithm 2: The FindGlobalPR algorithm. FindGlobalPR(L, Lout, T, k) Input: L: zero-one adjacency matrix for the local domain, Lout: zero-one outlink matrix from L to global subgraph as in (2), T: number of iterations, k: number of pages to crawl per iteration. Output: ˆp: an improved estimate of the global PageRank of L. F ← L Fout ← Lout f ← ComputePR(F ) for (i = 1 to T) {Determine which pages to crawl next} pages ← SelectNodes(F , Fout, f, k) Crawl pages, augment F and modify Fout {Update PageRanks for new local domain} f ← ComputePR(F ) end {Extract PageRanks of original local domain & normalize} ˆp ← ELf ELf 1 is to find such a supergraph ˆF with PageRank ˆf, so that ˆf when restricted to L is close to p∗ . Formally, we seek to minimize GlobalDiff( ˆf) = EL ˆf EL ˆf 1 − p∗ 1 . (3) We choose the L1 norm for measuring the error as it does not place excessive weight on outliers (as the L2 norm does, for example), and also because it is the most commonly used distance measure in the literature for comparing PageRank vectors, as well as for detecting convergence of the algorithm [3]. We propose a greedy framework, given in Algorithm 2, for constructing ˆF . Initially, F is set to the local subgraph L, and the PageRank f of this graph is computed. The algorithm then proceeds as follows. First, the SelectNodes algorithm (which we discuss in the next section) is called and it returns a set of k nodes to crawl next from the set of nodes in the current crawl frontier, Fout. These selected nodes are then crawled to expand the local subgraph, F , and the PageRanks of this expanded graph are then recomputed. These steps are repeated for each of T iterations. Finally, the PageRank vector ˆp, which is restricted to pages within the original local domain, is returned. Given our computation, bandwidth, and memory restrictions, we will assume that the algorithm will crawl at most O(n) pages. Since the PageRanks are computed in each iteration of the algorithm, which is an O(n) operation, we will also assume that the number of iterations T is a constant. Of course, the main challenge here is in selecting which set of k nodes to crawl next. In the next section, we formally define the problem and give efficient algorithms. 5. NODE SELECTION In this section, we present node selection algorithms that operate within the greedy framework presented in the previous section. We first give a well-defined criteria for the page selection problem and provide experimental evidence that this criteria can effectively identify pages that optimize our problem objective (3). We then present our main al118 Research Track Paper gorithmic contribution of the paper, a method with linear running time that is derived from this page selection criteria. Finally, we give an intuitive analysis of our algorithm in terms of ‘leaks" and ‘flows". We show that if only the ‘flow" is considered, then the resulting method is very similar to a widely used page selection heuristic [6]. 5.1 Formulation For a given page j in the global domain, we define the expanded local graph Fj: Fj = F s uT j 0 , (4) where uj is the zero-one vector containing the outlinks from F into page j, and s contains the inlinks from page j into the local domain. Note that we do not allow self-links in this framework. In practice, self-links are often removed, as they only serve to inflate a given page"s PageRank. Observe that the inlinks into F from node j are not known until after node j is crawled. Therefore, we estimate this inlink vector as the expectation over inlink counts among the set of already crawled pages, s = F T e F T e 1 . (5) In practice, for any given page, this estimate may not reflect the true inlinks from that page. Furthermore, this expectation is sampled from the set of links within the crawled domain, whereas a better estimate would also use links from the global domain. However, the latter distribution is not known to a localized search engine, and we contend that the above estimate will, on average, be a better estimate than the uniform distribution, for example. Let the PageRank of F be f. We express the PageRank f+ j of the expanded local graph Fj as f+ j = (1 − xj)fj xj , (6) where xj is the PageRank of the candidate global node j, and fj is the L1-normalized PageRank vector restricted to the pages in F . Since directly optimizing our problem goal requires knowing the global PageRank p∗ , we instead propose to crawl those nodes that will have the greatest influence on the PageRanks of pages in the original local domain L: influence(j) = k∈L |fj[k] − f[k]| (7) = EL (fj − f) 1. Experimentally, the influence score is a very good predictor of our problem objective (3). For each candidate global node j, figure 1(a) shows the objective function value Global Diff(fj) as a function of the influence of page j. The local domain used here is a crawl of conservative political pages (we will provide more details about this dataset in section 6); we observed similar results in other domains. The correlation is quite strong, implying that the influence criteria can effectively identify pages that improve the global PageRank estimate. As a baseline, figure 1(b) compares our objective with an alternative criteria, outlink count. The outlink count is defined as the number of outlinks from the local domain to page j. The correlation here is much weaker. .00001 .0001 .001 .01 0.26 0.262 0.264 0.266 Influence Objective 1 10 100 1000 0.266 0.264 0.262 0.26 Outlink Count Objective (a) (b) Figure 1: (a) The correlation between our influence page selection criteria (7) and the actual objective function (3) value is quite strong. (b) This is in contrast to other criteria, such as outlink count, which exhibit a much weaker correlation. 5.2 Computation As described, for each candidate global page j, the influence score (7) must be computed. If fj is computed exactly for each global page j, then the PageRank algorithm would need to be run for each of the O(n) such global pages j we consider, resulting in an O(n2 ) computational cost for the node selection method. Thus, computing the exact value of fj will lead to a quadratic algorithm, and we must instead turn to methods of approximating this vector. The algorithm we present works by performing one power method iteration used by the PageRank algorithm (Algorithm 1). The convergence rate for the PageRank algorithm has been shown to equal the random surfer probability α [7, 11]. Given a starting vector x(0) , if k PageRank iterations are performed, the current PageRank solution x(k) satisfies: x(k) − x∗ 1 = O(αk x(0) − x∗ 1), (8) where x∗ is the desired PageRank vector. Therefore, if only one iteration is performed, choosing a good starting vector is necessary to achieve an accurate approximation. We partition the PageRank matrix PFj , corresponding to the × subgraph Fj as: PFj = ˜F ˜s ˜uT j w , (9) where ˜F = αF (DF + diag(uj))−1 + (1 − α) e + 1 eT , ˜s = αs + (1 − α) e + 1 , ˜uj = α(DF + diag(uj))−1 uj + (1 − α) e + 1 , w = 1 − α + 1 , and diag(uj) is the diagonal matrix with the (i, i)th entry equal to one if the ith element of uj equals one, and is zero otherwise. We have assumed here that the random surfer vector is the uniform vector, and that L has no ‘dangling links". These assumptions are not necessary and serve only to simplify discussion and analysis. A simple approach for estimating fj is the following. First, estimate the PageRank f+ j of Fj by computing one PageRank iteration over the matrix PFj , using the starting vector ν = f 0 . Then, estimate fj by removing the last 119 Research Track Paper component from our estimate of f+ j (i.e., the component corresponding to the added node j), and renormalizing. The problem with this approach is in the starting vector. Recall from (6) that xj is the PageRank of the added node j. The difference between the actual PageRank f+ j of PFj and the starting vector ν is ν − f+ j 1 = xj + f − (1 − xj)fj 1 ≥ xj + | f 1 − (1 − xj) fj 1| = xj + |xj| = 2xj. Thus, by (8), after one PageRank iteration, we expect our estimate of f+ j to still have an error of about 2αxj. In particular, for candidate nodes j with relatively high PageRank xj, this method will yield more inaccurate results. We will next present a method that eliminates this bias and runs in O(n) time. 5.2.1 Stochastic Complementation Since f+ j , as given in (6) is the PageRank of the matrix PFj , we have: fj(1 − xj) xj = ˜F ˜s ˜uT j w fj(1 − xj) xj = ˜F fj(1 − xj) + ˜sxj ˜uT j fj(1 − xj) + wxj . Solving the above system for fj can be shown to yield fj = ( ˜F + (1 − w)−1 ˜s˜uT j )fj. (10) The matrix S = ˜F +(1−w)−1 ˜s˜uT j is known as the stochastic complement of the column stochastic matrix PFj with respect to the sub matrix ˜F . The theory of stochastic complementation is well studied, and it can be shown the stochastic complement of an irreducible matrix (such as the PageRank matrix) is unique. Furthermore, the stochastic complement is also irreducible and therefore has a unique stationary distribution as well. For an extensive study, see [15]. It can be easily shown that the sub-dominant eigenvalue of S is at most +1 α, where is the size of F . For sufficiently large , this value will be very close to α. This is important, as other properties of the PageRank algorithm, notably the algorithm"s sensitivity, are dependent on this value [11]. In this method, we estimate the length vector fj by computing one PageRank iteration over the × stochastic complement S, starting at the vector f: fj ≈ Sf. (11) This is in contrast to the simple method outlined in the previous section, which first iterates over the ( + 1) × ( + 1) matrix PFj to estimate f+ j , and then removes the last component from the estimate and renormalizes to approximate fj. The problem with the latter method is in the choice of the ( + 1) length starting vector, ν. Consequently, the PageRank estimate given by the simple method differs from the true PageRank by at least 2αxj, where xj is the PageRank of page j. By using the stochastic complement, we can establish a tight lower bound of zero for this difference. To see this, consider the case in which a node k is added to F to form the augmented local subgraph Fk, and that the PageRank of this new graph is (1 − xk)f xk . Specifically, the addition of page k does not change the PageRanks of the pages in F , and thus fk = f. By construction of the stochastic complement, fk = Sfk, so the approximation given in equation (11) will yield the exact solution. Next, we present the computational details needed to efficiently compute the quantity fj −f 1 over all known global pages j. We begin by expanding the difference fj −f, where the vector fj is estimated as in (11), fj − f ≈ Sf − f = αF (DF + diag(uj))−1 f + (1 − α) e + 1 eT f +(1 − w)−1 (˜uT j f)˜s − f. (12) Note that the matrix (DF +diag(uj))−1 is diagonal. Letting o[k] be the outlink count for page k in F , we can express the kth diagonal element as: (DF + diag(uj))−1 [k, k] = 1 o[k]+1 if uj[k] = 1 1 o[k] if uj[k] = 0 Noting that (o[k] + 1)−1 = o[k]−1 − (o[k](o[k] + 1))−1 and rewriting this in matrix form yields (DF +diag(uj))−1 = D−1 F −D−1 F (DF +diag(uj))−1 diag(uj). (13) We use the same identity to express e + 1 = e − e ( + 1) . (14) Recall that, by definition, we have PF = αF D−1 F +(1−α)e . Substituting (13) and (14) in (12) yields fj − f ≈ (PF f − f) −αF D−1 F (DF + diag(uj))−1 diag(uj)f −(1 − α) e ( + 1) + (1 − w)−1 (˜uT j f)˜s = x + y + (˜uT j f)z, (15) noting that by definition, f = PF f, and defining the vectors x, y, and z to be x = −αF D−1 F (DF + diag(uj))−1 diag(uj)f (16) y = −(1 − α) e ( + 1) (17) z = (1 − w)−1 ˜s. (18) The first term x is a sparse vector, and takes non-zero values only for local pages k that are siblings of the global page j. We define (i, j) ∈ F if and only if F [j, i] = 1 (equivalently, page i links to page j) and express the value of the component x[k ] as: x[k ] = −α k:(k,k )∈F ,uj [k]=1 f[k] o[k](o[k] + 1) , (19) where o[k], as before, is the number of outlinks from page k in the local domain. Note that the last two terms, y and z are not dependent on the current global node j. Given the function hj(f) = y + (˜uT j f)z 1, the quantity fj − f 1 120 Research Track Paper can be expressed as fj − f 1 = k x[k] + y[k] + (˜uT j f)z[k] = k:x[k]=0 y[k] + (˜uT j f)z[k] + k:x[k]=0 x[k] + y[k] + (˜uT j f)z[k] = hj(f) − k:x[k]=0 y[k] + (˜uT j f)z[k] + k:x[k]=0 x[k] + y[k] + (˜uT j f)z[k] .(20) If we can compute the function hj in linear time, then we can compute each value of fj − f 1 using an additional amount of time that is proportional to the number of nonzero components in x. These optimizations are carried out in Algorithm 3. Note that (20) computes the difference between all components of f and fj, whereas our node selection criteria, given in (7), is restricted to the components corresponding to nodes in the original local domain L. Let us examine Algorithm 3 in more detail. First, the algorithm computes the outlink counts for each page in the local domain. The algorithm then computes the quantity ˜uT j f for each known global page j. This inner product can be written as (1 − α) 1 + 1 + α k:(k,j)∈Fout f[k] o[k] + 1 , where the second term sums over the set of local pages that link to page j. Since the total number of edges in Fout was assumed to have size O( ) (recall that is the number of pages in F ), the running time of this step is also O( ). The algorithm then computes the vectors y and z, as given in (17) and (18), respectively. The L1NormDiff method is called on the components of these vectors which correspond to the pages in L, and it estimates the value of EL(y + (˜uT j f)z) 1 for each page j. The estimation works as follows. First, the values of ˜uT j f are discretized uniformly into c values {a1, ..., ac}. The quantity EL(y + aiz) 1 is then computed for each discretized value of ai and stored in a table. To evaluate EL (y + az) 1 for some a ∈ [a1, ac], the closest discretized value ai is determined, and the corresponding entry in the table is used. The total running time for this method is linear in and the discretization parameter c (which we take to be a constant). We note that if exact values are desired, we have also developed an algorithm that runs in O( log ) time that is not described here. In the main loop, we compute the vector x, as defined in equation (16). The nested loops iterate over the set of pages in F that are siblings of page j. Typically, the size of this set is bounded by a constant. Finally, for each page j, the scores vector is updated over the set of non-zero components k of the vector x with k ∈ L. This set has size equal to the number of local siblings of page j, and is a subset of the total number of siblings of page j. Thus, each iteration of the main loop takes constant time, and the total running time of the main loop is O( ). Since we have assumed that the size of F will not grow larger than O(n), the total running time for the algorithm is O(n). Algorithm 3: Node Selection via Stochastic Complementation. SC-Select(F , Fout, f, k) Input: F : zero-one adjacency matrix of size corresponding to the current local subgraph, Fout: zero-one outlink matrix from F to global subgraph, f: PageRank of F , k: number of pages to return Output: pages: set of k pages to crawl next {Compute outlink sums for local subgraph} foreach (page j ∈ F ) o[j] ← k:(j,k)∈F F[j, k] end {Compute scalar ˜uT j f for each global node j } foreach (page j ∈ Fout) g[j] ← (1 − α) 1 +1 foreach (page k : (k, j) ∈ Fout) g[j] ← g[j] + α f[k] o[k]+1 end end {Compute vectors y and z as in (17) and (18) } y ← −(1 − α) e ( +1) z ← (1 − w)−1 ˜s {Approximate y + g[j] ∗ z 1 for all values g[j]} norm diffs ←L1NormDiffs(g, ELy, ELz) foreach (page j ∈ Fout) {Compute sparse vector x as in (19)} x ← 0 foreach (page k : (k, j) ∈ Fout) foreach (page k : (k, k ) ∈ F )) x[k ] ← x[k ] − f[k] o[k](o[k]+1) end end x ← αx scores[j] ← norm diffs[j] foreach (k : x[k] > 0 and page k ∈ L) scores[j] ← scores[j] − |y[k] + g[j] ∗ z[k]| +|x[k]+y[k]+g[j]∗z[k])| end end Return k pages with highest scores 5.2.2 PageRank Flows We now present an intuitive analysis of the stochastic complementation method by decomposing the change in PageRank in terms of ‘leaks" and ‘flows". This analysis is motivated by the decomposition given in (15). PageRank ‘flow" is the increase in the local PageRanks originating from global page j. The flows are represented by the non-negative vector (˜uT j f)z (equations (15) and (18)). The scalar ˜uT j f can be thought of as the total amount of PageRank flow that page j has available to distribute. The vector z dictates how the flow is allocated to the local domain; the flow that local page k receives is proportional to (within a constant factor due to the random surfer vector) the expected number of its inlinks. The PageRank ‘leaks" represent the decrease in PageRank resulting from the addition of page j. The leakage can be quantified in terms of the non-positive vectors x and y (equations (16) and (17)). For vector x, we can see from equation (19) that the amount of PageRank leaked by a local page is proportional to the weighted sum of the Page121 Research Track Paper Ranks of its siblings. Thus, pages that have siblings with higher PageRanks (and low outlink counts) will experience more leakage. The leakage caused by y is an artifact of the random surfer vector. We will next show that if only the ‘flow" term, (˜uT j f)z, is considered, then the resulting method is very similar to a heuristic proposed by Cho et al. [6] that has been widely used for the Crawling Through URL Ordering problem. This heuristic is computationally cheaper, but as we will see later, not as effective as the Stochastic Complementation method. Our node selection strategy chooses global nodes that have the largest influence (equation (7)). If this influence is approximated using only ‘flows", the optimal node j∗ is: j∗ = argmaxj EL ˜uT j fz 1 = argmaxj ˜uT j f EL z 1 = argmaxj ˜uT j f = argmaxj α(DF + diag(uj))−1 uj + (1 − α) e + 1 , f = argmaxjfT (DF + diag(uj))−1 uj. The resulting page selection score can be expressed as a sum of the PageRanks of each local page k that links to j, where each PageRank value is normalized by o[k]+1. Interestingly, the normalization that arises in our method differs from the heuristic given in [6], which normalizes by o[k]. The algorithm PF-Select, which is omitted due to lack of space, first computes the quantity fT (DF +diag(uj))−1 uj for each global page j, and then returns the pages with the k largest scores. To see that the running time for this algorithm is O(n), note that the computation involved in this method is a subset of that needed for the SC-Select method (Algorithm 3), which was shown to have a running time of O(n). 6. EXPERIMENTS In this section, we provide experimental evidence to verify the effectiveness of our algorithms. We first outline our experimental methodology and then provide results across a variety of local domains. 6.1 Methodology Given the limited resources available at an academic institution, crawling a section of the web that is of the same magnitude as that indexed by Google or Yahoo! is clearly infeasible. Thus, for a given local domain, we approximate the global graph by crawling a local neighborhood around the domain that is several orders of magnitude larger than the local subgraph. Even though such a graph is still orders of magnitude smaller than the ‘true" global graph, we contend that, even if there exist some highly influential pages that are very far away from our local domain, it is unrealistic for any local node selection algorithm to find them. Such pages also tend to be highly unrelated to pages within the local domain. When explaining our node selection strategies in section 5, we made the simplifying assumption that our local graph contained no dangling nodes. This assumption was only made to ease our analysis. Our implementation efficiently handles dangling links by replacing each zero column of our adjacency matrix with the uniform vector. We evaluate the algorithm using the two node selection strategies given in Section 5.2, and also against the following baseline methods: • Random: Nodes are chosen uniformly at random among the known global nodes. • OutlinkCount: Global nodes with the highest number of outlinks from the local domain are chosen. At each iteration of the FindGlobalPR algorithm, we evaluate performance by computing the difference between the current PageRank estimate of the local domain, ELf ELf 1 , and the global PageRank of the local domain ELg ELg 1 . All PageRank calculations were performed using the uniform random surfer vector. Across all experiments, we set the random surfer parameter α, to be .85, and used a convergence threshold of 10−6 . We evaluate the difference between the local and global PageRank vectors using three different metrics: the L1 and L∞ norms, and Kendall"s tau. The L1 norm measures the sum of the absolute value of the differences between the two vectors, and the L∞ norm measures the absolute value of the largest difference. Kendall"s tau metric is a popular rank correlation measure used to compare PageRanks [2, 11]. This metric can be computed by counting the number of pairs of pairs that agree in ranking, and subtracting from that the number of pairs of pairs that disagree in ranking. The final value is then normalized by the total number of n 2 such pairs, resulting in a [−1, 1] range, where a negative score signifies anti-correlation among rankings, and values near one correspond to strong rank correlation. 6.2 Results Our experiments are based on two large web crawls and were downloaded using the web crawler that is part of the Nutch open source search engine project [18]. All crawls were restricted to only ‘http" pages, and to limit the number of dynamically generated pages that we crawl, we ignored all pages with urls containing any of the characters ‘?", ‘*", ‘@", or ‘=". The first crawl, which we will refer to as the ‘edu" dataset, was seeded by homepages of the top 100 graduate computer science departments in the USA, as rated by the US News and World Report [16], and also by the home pages of their respective institutions. A crawl of depth 5 was performed, restricted to pages within the ‘.edu" domain, resulting in a graph with approximately 4.7 million pages and 22.9 million links. The second crawl was seeded by the set of pages under the ‘politics" hierarchy in the dmoz open directory project[17]. We crawled all pages up to four links away, which yielded a graph with 4.4 million pages and 17.3 million links. Within the ‘edu" crawl, we identified the five site-specific domains corresponding to the websites of the top five graduate computer science departments, as ranked by the US News and World Report. This yielded local domains of various sizes, from 10,626 (UIUC) to 59,895 (Berkeley). For each of these site-specific domains with size n, we performed 50 iterations of the FindGlobalPR algorithm to crawl a total of 2n additional nodes. Figure 2(a) gives the (L1) difference from the PageRank estimate at each iteration to the global PageRank, for the Berkeley local domain. The performance of this dataset was representative of the typical performance across the five computer science sitespecific local domains. Initially, the L1 difference between the global and local PageRanks ranged from .0469 (Stanford) to .149 (MIT). For the first several iterations, the 122 Research Track Paper 0 5 10 15 20 25 30 35 40 45 50 0.015 0.02 0.025 0.03 0.035 0.04 0.045 0.05 0.055 Number of Iterations GlobalandLocalPageRankDifference(L1) Stochastic Complement PageRank Flow Outlink Count Random 0 10 20 30 40 50 0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 Number of Iterations GlobalandLocalPageRankDifference(L1) Stochastic Complement PageRank Flow Outlink Count Random 0 5 10 15 20 25 0.16 0.18 0.2 0.22 0.24 0.26 0.28 0.3 0.32 0.34 Number of Iterations GlobalandLocalPageRankDifference(L1) Stochastic Complement PageRank Flow Outlink Count Random (a) www.cs.berkeley.edu (b) www.enterstageright.com (c) Politics Figure 2: L1 difference between the estimated and true global PageRanks for (a) Berkeley"s computer science website, (b) the site-specific domain, www.enterstageright.com, and (c) the ‘politics" topic-specific domain. The stochastic complement method outperforms all other methods across various domains. three link-based methods all outperform the random selection heuristic. After these initial iterations, the random heuristic tended to be more competitive with (or even outperform, as in the Berkeley local domain) the outlink count and PageRank flow heuristics. In all tests, the stochastic complementation method either outperformed, or was competitive with, the other methods. Table 1 gives the average difference between the final estimated global PageRanks and the true global PageRanks for various distance measures. Algorithm L1 L∞ Kendall Stoch. Comp. .0384 .00154 .9257 PR Flow .0470 .00272 .8946 Outlink .0419 .00196 .9053 Random .0407 .00204 .9086 Table 1: Average final performance of various node selection strategies for the five site-specific computer science local domains. Note that Kendall"s Tau measures similarity, while the other metrics are dissimilarity measures. Stochastic Complementation clearly outperforms the other methods in all metrics. Within the ‘politics" dataset, we also performed two sitespecific tests for the largest websites in the crawl: www.adamsmith.org, the website for the London based Adam Smith Institute, and www.enterstageright.com, an online conservative journal. As with the ‘edu" local domains, we ran our algorithm for 50 iterations, crawling a total of 2n nodes. Figure 2 (b) plots the results for the www.enterstageright.com domain. In contrast to the ‘edu" local domains, the Random and OutlinkCount methods were not competitive with either the SC-Select or the PF-Select methods. Among all datasets and all node selection methods, the stochastic complementation method was most impressive in this dataset, realizing a final estimate that differed only .0279 from the global PageRank, a ten-fold improvement over the initial local PageRank difference of .299. For the Adam Smith local domain, the initial difference between the local and global PageRanks was .148, and the final estimates given by the SC-Select, PF-Select, OutlinkCount, and Random methods were .0208, .0193, .0222, and .0356, respectively. Within the ‘politics" dataset, we constructed four topicspecific local domains. The first domain consisted of all pages in the dmoz politics category, and also all pages within each of these sites up to two links away. This yielded a local domain of 90,811 pages, and the results are given in figure 2 (c). Because of the larger size of the topic-specific domains, we ran our algorithm for only 25 iterations to crawl a total of n nodes. We also created topic-specific domains from three political sub-topics: liberalism, conservatism, and socialism. The pages in these domains were identified by their corresponding dmoz categories. For each sub-topic, we set the local domain to be all pages within three links from the corresponding dmoz category pages. Table 2 summarizes the performance of these three topic-specific domains, and also the larger political domain. To quantify a global page j"s effect on the global PageRank values of pages in the local domain, we define page j"s impact to be its PageRank value, g[j], normalized by the fraction of its outlinks pointing to the local domain: impact(j) = oL[j] o[j] · g[j], where, oL[j] is the number of outlinks from page j to pages in the local domain L, and o[j] is the total number of j"s outlinks. In terms of the random surfer model, the impact of page j is the probability that the random surfer (1) is currently at global page j in her random walk and (2) takes an outlink to a local page, given that she has already decided not to jump to a random page. For the politics local domain, we found that many of the pages with high impact were in fact political pages that should have been included in the dmoz politics topic, but were not. For example, the two most influential global pages were the political search engine www.askhenry.com, and the home page of the online political magazine, www.policyreview.com. Among non-political pages, the home page of the journal Education Next was most influential. The journal is freely available online and contains articles regarding various aspect of K-12 education in America. To provide some anecdotal evidence for the effectiveness of our page selection methods, we note that the SC-Select method chose 11 pages within the www.educationnext.org domain, the PF-Select method discovered 7 such pages, while the OutlinkCount and Random methods found only 6 pages each. For the conservative political local domain, the socialist website www.ornery.org had a very high impact score. This 123 Research Track Paper All Politics: Algorithm L1 L2 Kendall Stoch. Comp. .1253 .000700 .8671 PR Flow .1446 .000710 .8518 Outlink .1470 .00225 .8642 Random .2055 .00203 .8271 Conservativism: Algorithm L1 L2 Kendall Stoch. Comp. .0496 .000990 .9158 PR Flow .0554 .000939 .9028 Outlink .0602 .00527 .9144 Random .1197 .00102 .8843 Liberalism: Algorithm L1 L2 Kendall Stoch. Comp. .0622 .001360 .8848 PR Flow .0799 .001378 .8669 Outlink .0763 .001379 .8844 Random .1127 .001899 .8372 Socialism: Algorithm L1 L∞ Kendall Stoch. Comp. .04318 .00439 .9604 PR Flow .0450 .004251 .9559 Outlink .04282 .00344 .9591 Random .0631 .005123 .9350 Table 2: Final performance among node selection strategies for the four political topic-specific crawls. Note that Kendall"s Tau measures similarity, while the other metrics are dissimilarity measures. was largely due to a link from the front page of this site to an article regarding global warming published by the National Center for Public Policy Research, a conservative research group in Washington, DC. Not surprisingly, the global PageRank of this article (which happens to be on the home page of the NCCPR, www.nationalresearch.com), was approximately .002, whereas the local PageRank of this page was only .00158. The SC-Select method yielded a global PageRank estimate of approximately .00182, the PFSelect method estimated a value of .00167, and the Random and OutlinkCount methods yielded values of .01522 and .00171, respectively. 7. RELATED WORK The node selection framework we have proposed is similar to the url ordering for crawling problem proposed by Cho et al. in [6]. Whereas our framework seeks to minimize the difference between the global and local PageRank, the objective used in [6] is to crawl the most highly (globally) ranked pages first. They propose several node selection algorithms, including the outlink count heuristic, as well as a variant of our PF-Select algorithm which they refer to as the ‘PageRank ordering metric". They found this method to be most effective in optimizing their objective, as did a recent survey of these methods by Baeza-Yates et al. [1]. Boldi et al. also experiment within a similar crawling framework in [2], but quantify their results by comparing Kendall"s rank correlation between the PageRanks of the current set of crawled pages and those of the entire global graph. They found that node selection strategies that crawled pages with the highest global PageRank first actually performed worse (with respect to Kendall"s Tau correlation between the local and global PageRanks) than basic depth first or breadth first strategies. However, their experiments differ from our work in that our node selection algorithms do not use (or have access to) global PageRank values. Many algorithmic improvements for computing exact PageRank values have been proposed [9, 10, 14]. If such algorithms are used to compute the global PageRanks of our local domain, they would all require O(N) computation, storage, and bandwidth, where N is the size of the global domain. This is in contrast to our method, which approximates the global PageRank and scales linearly with the size of the local domain. Wang and Dewitt [22] propose a system where the set of web servers that comprise the global domain communicate with each other to compute their respective global PageRanks. For a given web server hosting n pages, the computational, bandwidth, and storage requirements are also linear in n. One drawback of this system is that the number of distinct web servers that comprise the global domain can be very large. For example, our ‘edu" dataset contains websites from over 3,200 different universities; coordinating such a system among a large number of sites can be very difficult. Gan, Chen, and Suel propose a method for estimating the PageRank of a single page [5] which uses only constant bandwidth, computation, and space. Their approach relies on the availability of a remote connectivity server that can supply the set of inlinks to a given page, an assumption not used in our framework. They experimentally show that a reasonable estimate of the node"s PageRank can be obtained by visiting at most a few hundred nodes. Using their algorithm for our problem would require that either the entire global domain first be downloaded or a connectivity server be used, both of which would lead to very large web graphs. 8. CONCLUSIONS AND FUTURE WORK The internet is growing exponentially, and in order to navigate such a large repository as the web, global search engines have established themselves as a necessity. Along with the ubiquity of these large-scale search engines comes an increase in search users" expectations. By providing complete and isolated coverage of a particular web domain, localized search engines are an effective outlet to quickly locate content that could otherwise be difficult to find. In this work, we contend that the use of global PageRank in a localized search engine can improve performance. To estimate the global PageRank, we have proposed an iterative node selection framework where we select which pages from the global frontier to crawl next. Our primary contribution is our stochastic complementation page selection algorithm. This method crawls nodes that will most significantly impact the local domain and has running time linear in the number of nodes in the local domain. Experimentally, we validate these methods across a diverse set of local domains, including seven site-specific domains and four topic-specific domains. We conclude that by crawling an additional n or 2n pages, our methods find an estimate of the global PageRanks that is up to ten times better than just using the local PageRanks. Furthermore, we demonstrate that our algorithm consistently outperforms other existing heuristics. 124 Research Track Paper Often times, topic-specific domains are discovered using a focused web crawler which considers a page"s content in conjunction with link anchor text to decide which pages to crawl next [4]. Although such crawlers have proven to be quite effective in discovering topic-related content, many irrelevant pages are also crawled in the process. Typically, these pages are deleted and not indexed by the localized search engine. These pages can of course provide valuable information regarding the global PageRank of the local domain. One way to integrate these pages into our framework is to start the FindGlobalPR algorithm with the current subgraph F equal to the set of pages that were crawled by the focused crawler. The global PageRank estimation framework, along with the node selection algorithms presented, all require O(n) computation per iteration and bandwidth proportional to the number of pages crawled, Tk. If the number of iterations T is relatively small compared to the number of pages crawled per iteration, k, then the bottleneck of the algorithm will be the crawling phase. However, as the number of iterations increases (relative to k), the bottleneck will reside in the node selection computation. In this case, our algorithms would benefit from constant factor optimizations. Recall that the FindGlobalPR algorithm (Algorithm 2) requires that the PageRanks of the current expanded local domain be recomputed in each iteration. Recent work by Langville and Meyer [12] gives an algorithm to quickly recompute PageRanks of a given webgraph if a small number of nodes are added. This algorithm was shown to give speedup of five to ten times on some datasets. We plan to investigate this and other such optimizations as future work. In this paper, we have objectively evaluated our methods by measuring how close our global PageRank estimates are to the actual global PageRanks. To determine the benefit of using global PageRanks in a localized search engine, we suggest a user study in which users are asked to rate the quality of search results for various search queries. For some queries, only the local PageRanks are used in ranking, and for the remaining queries, local PageRanks and the approximate global PageRanks, as computed by our algorithms, are used. The results of such a study can then be analyzed to determine the added benefit of using the global PageRanks computed by our methods, over just using the local PageRanks. Acknowledgements. This research was supported by NSF grant CCF-0431257, NSF Career Award ACI-0093404, and a grant from Sabre, Inc. 9. REFERENCES [1] R. Baeza-Yates, M. Marin, C. Castillo, and A. Rodriguez. Crawling a country: better strategies than breadth-first for web page ordering. World-Wide Web Conference, 2005. [2] P. Boldi, M. Santini, and S. Vigna. Do your worst to make the best: paradoxical effects in pagerank incremental computations. Workshop on Web Graphs, 3243:168-180, 2004. [3] S. Brin and L. Page. The anatomy of a large-scale hypertextual web search engine. Computer Networks and ISDN Systems, 33(1-7):107-117, 1998. [4] S. Chakrabarti, M. van den Berg, and B. Dom. Focused crawling: a new approach to topic-specific web resource discovery. World-Wide Web Conference, 1999. [5] Y. Chen, Q. Gan, and T. Suel. Local methods for estimating pagerank values. Conference on Information and Knowledge Management, 2004. [6] J. Cho, H. Garcia-Molina, and L. Page. Efficient crawling through url ordering. World-Wide Web Conference, 1998. [7] T. H. Haveliwala and S. D. Kamvar. The second eigenvalue of the Google matrix. Technical report, Stanford University, 2003. [8] T. Joachims, F. Radlinski, L. Granka, A. Cheng, C. Tillekeratne, and A. Patel. Learning retrieval functions from implicit feedback. http://www.cs.cornell.edu/People/tj/career. [9] S. D. Kamvar, T. H. Haveliwala, C. D. Manning, and G. H. Golub. Exploiting the block structure of the web for computing pagerank. World-Wide Web Conference, 2003. [10] S. D. Kamvar, T. H. Haveliwala, C. D. Manning, and G. H. Golub. Extrapolation methods for accelerating pagerank computation. World-Wide Web Conference, 2003. [11] A. N. Langville and C. D. Meyer. Deeper inside pagerank. Internet Mathematics, 2004. [12] A. N. Langville and C. D. Meyer. Updating the stationary vector of an irreducible markov chain with an eye on Google"s pagerank. SIAM Journal on Matrix Analysis, 2005. [13] P. Lyman, H. R. Varian, K. Swearingen, P. Charles, N. Good, L. L. Jordan, and J. Pal. How much information 2003? School of Information Management and System, University of California at Berkely, 2003. [14] F. McSherry. A uniform approach to accelerated pagerank computation. World-Wide Web Conference, 2005. [15] C. D. Meyer. Stochastic complementation, uncoupling markov chains, and the theory of nearly reducible systems. SIAM Review, 31:240-272, 1989. [16] US News and World Report. http://www.usnews.com. [17] Dmoz open directory project. http://www.dmoz.org. [18] Nutch open source search engine. http://www.nutch.org. [19] F. Radlinski and T. Joachims. Query chains: learning to rank from implicit feedback. ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, 2005. [20] S. Raghavan and H. Garcia-Molina. Crawling the hidden web. In Proceedings of the Twenty-seventh International Conference on Very Large Databases, 2001. [21] T. Tin Tang, D. Hawking, N. Craswell, and K. Griffiths. Focused crawling for both topical relevance and quality of medical information. Conference on Information and Knowledge Management, 2005. [22] Y. Wang and D. J. DeWitt. Computing pagerank in a distributed internet search system. Proceedings of the 30th VLDB Conference, 2004. 125 Research Track Paper

algorithm;localized search engine;web community;experimentation;subgraph;link-based ranking;topic-specific domain;large-scale search engine;local domain;global pagerank;crawling problem;global graph

train_H-84

Event Threading within News Topics

With the overwhelming volume of online news available today, there is an increasing need for automatic techniques to analyze and present news to the user in a meaningful and efficient manner. Previous research focused only on organizing news stories by their topics into a flat hierarchy. We believe viewing a news topic as a flat collection of stories is too restrictive and inefficient for a user to understand the topic quickly. In this work, we attempt to capture the rich structure of events and their dependencies in a news topic through our event models. We call the process of recognizing events and their dependencies event threading. We believe our perspective of modeling the structure of a topic is more effective in capturing its semantics than a flat list of on-topic stories. We formally define the novel problem, suggest evaluation metrics and present a few techniques for solving the problem. Besides the standard word based features, our approaches take into account novel features such as temporal locality of stories for event recognition and time-ordering for capturing dependencies. Our experiments on a manually labeled data sets show that our models effectively identify the events and capture dependencies among them.

1. INTRODUCTION News forms a major portion of information disseminated in the world everyday. Common people and news analysts alike are very interested in keeping abreast of new things that happen in the news, but it is becoming very difficult to cope with the huge volumes of information that arrives each day. Hence there is an increasing need for automatic techniques to organize news stories in a way that helps users interpret and analyze them quickly. This problem is addressed by a research program called Topic Detection and Tracking (TDT) [3] that runs an open annual competition on standardized tasks of news organization. One of the shortcomings of current TDT evaluation is its view of news topics as flat collection of stories. For example, the detection task of TDT is to arrange a collection of news stories into clusters of topics. However, a topic in news is more than a mere collection of stories: it is characterized by a definite structure of inter-related events. This is indeed recognized by TDT which defines a topic as ‘a set of news stories that are strongly related by some seminal realworld event" where an event is defined as ‘something that happens at a specific time and location" [3]. For example, when a bomb explodes in a building, that is the seminal event that triggers the topic. Other events in the topic may include the rescue attempts, the search for perpetrators, arrests and trials and so on. We see that there is a pattern of dependencies between pairs of events in the topic. In the above example, the event of rescue attempts is ‘influenced" by the event of bombing and so is the event of search for perpetrators. In this work we investigate methods for modeling the structure of a topic in terms of its events. By structure, we mean not only identifying the events that make up a topic, but also establishing dependencies-generally causal-among them. We call the process of recognizing events and identifying dependencies among them event threading, an analogy to email threading that shows connections between related email messages. We refer to the resulting interconnected structure of events as the event model of the topic. Although this paper focuses on threading events within an existing news topic, we expect that such event based dependency structure more accurately reflects the structure of news than strictly bounded topics do. From a user"s perspective, we believe that our view of a news topic as a set of interconnected events helps him/her get a quick overview of the topic and also allows him/her navigate through the topic faster. The rest of the paper is organized as follows. In section 2, we discuss related work. In section 3, we define the problem and use an example to illustrate threading of events within a news topic. In section 4, we describe how we built the corpus for our problem. Section 5 presents our evaluation techniques while section 6 describes the techniques we use for modeling event structure. In section 7 we present our experiments and results. Section 8 concludes the paper with a few observations on our results and comments on future work. 446 2. RELATED WORK The process of threading events together is related to threading of electronic mail only by name for the most part. Email usually incorporates a strong structure of referenced messages and consistently formatted subject headings-though information retrieval techniques are useful when the structure breaks down [7]. Email threading captures reference dependencies between messages and does not attempt to reflect any underlying real-world structure of the matter under discussion. Another area of research that looks at the structure within a topic is hierarchical text classification of topics [9, 6]. The hierarchy within a topic does impose a structure on the topic, but we do not know of an effort to explore the extent to which that structure reflects the underlying event relationships. Barzilay and Lee [5] proposed a content structure modeling technique where topics within text are learnt using unsupervised methods, and a linear order of these topics is modeled using hidden Markov models. Our work differs from theirs in that we do not constrain the dependency to be linear. Also their algorithms are tuned to work on specific genres of topics such as earthquakes, accidents, etc., while we expect our algorithms to generalize over any topic. In TDT, researchers have traditionally considered topics as flatclusters [1]. However, in TDT-2003, a hierarchical structure of topic detection has been proposed and [2] made useful attempts to adopt the new structure. However this structure still did not explicitly model any dependencies between events. In a work closest to ours, Makkonen [8] suggested modeling news topics in terms of its evolving events. However, the paper stopped short of proposing any models to the problem. Other related work that dealt with analysis within a news topic includes temporal summarization of news topics [4]. 3. PROBLEM DEFINITION AND NOTATION In this work, we have adhered to the definition of event and topic as defined in TDT. We present some definitions (in italics) and our interpretations (regular-faced) below for clarity. 1. Story: A story is a news article delivering some information to users. In TDT, a story is assumed to refer to only a single topic. In this work, we also assume that each story discusses a single event. In other words, a story is the smallest atomic unit in the hierarchy (topic event story). Clearly, both the assumptions are not necessarily true in reality, but we accept them for simplicity in modeling. 2. Event: An event is something that happens at some specific time and place [10]. In our work, we represent an event by a set of stories that discuss it. Following the assumption of atomicity of a story, this means that any set of distinct events can be represented by a set of non-overlapping clusters of news stories. 3. Topic: A set of news stories strongly connected by a seminal event. We expand on this definition and interpret a topic as a series of related events. Thus a topic can be represented by clusters of stories each representing an event and a set of (directed or undirected) edges between pairs of these clusters representing the dependencies between these events. We will describe this representation of a topic in more detail in the next section. 4. Topic detection and tracking (TDT) :Topic detection detects clusters of stories that discuss the same topic; Topic tracking detects stories that discuss a previously known topic [3]. Thus TDT concerns itself mainly with clustering stories into topics that discuss them. 5. Event threading: Event threading detects events within in a topic, and also captures the dependencies among the events. Thus the main difference between event threading and TDT is that we focus our modeling effort on microscopic events rather than larger topics. Additionally event threading models the relatedness or dependencies between pairs of events in a topic while TDT models topics as unrelated clusters of stories. We first define our problem and representation of our model formally and then illustrate with the help of an example. We are given a set of Ò news stories Ë ×½ ¡ ¡ ¡ ×Ò on a given topic Ì and their time of publication. We define a set of events ½ ¡ ¡ ¡ Ñ with the following constraints: ¾ ¾ Ë (1) × Ø (2) × ¾ × Ø × ¾ (3) While the first constraint says that each event is an element in the power set of S, the second constraint ensures that each story can belong to at most one event. The last constraint tells us that every story belongs to one of the events in . In fact this allows us to define a mapping function from stories to events as follows: ´× µ iff × ¾ (4) Further, we also define a set of directed edges ´ µ which denote dependencies between events. It is important to explain what we mean by this directional dependency: While the existence of an edge itself represents relatedness of two events, the direction could imply causality or temporal-ordering. By causal dependency we mean that the occurrence of event B is related to and is a consequence of the occurrence of event A. By temporal ordering, we mean that event B happened after event A and is related to A but is not necessarily a consequence of A. For example, consider the following two events: ‘plane crash" (event A) and ‘subsequent investigations" (event B) in a topic on a plane crash incident. Clearly, the investigations are a result of the crash. Hence an arrow from A to B falls under the category of causal dependency. Now consider the pair of events ‘Pope arrives in Cuba"(event A) and ‘Pope meets Castro"(event B) in a topic that discusses Pope"s visit to Cuba. Now events A and B are closely related through their association with the Pope and Cuba but event B is not necessarily a consequence of the occurrence of event A. An arrow in such scenario captures what we call time ordering. In this work, we do not make an attempt to distinguish between these two kinds of dependencies and our models treats them as identical. A simpler (and hence less controversial) choice would be to ignore direction in the dependencies altogether and consider only undirected edges. This choice definitely makes sense as a first step but we chose the former since we believe directional edges make more sense to the user as they provide a more illustrative flow-chart perspective to the topic. To make the idea of event threading more concrete, consider the example of TDT3 topic 30005, titled ‘Osama bin Laden"s Indictment" (in the 1998 news). This topic has 23 stories which form 5 events. An event model of this topic can be represented as in figure 1. Each box in the figure indicates an event in the topic of Osama"s indictment. The occurrence of event 2, namely ‘Trial and Indictment of Osama" is dependent on the event of ‘evidence gathered by CIA", i.e., event 1. Similarly, event 2 influences the occurrences of events 3, 4 and 5, namely ‘Threats from Militants", ‘Reactions 447 from Muslim World" and ‘announcement of reward". Thus all the dependencies in the example are causal. Extending our notation further, we call an event A a parent of B and B the child of A, if ´ µ ¾ . We define an event model Å ´ µ to be a tuple of the set of events and set of dependencies. Trial and (5) (3) (4) CIA announces reward Muslim world Reactions from Islamic militants Threats from (2) (1) Osama Indictment of CIA gathered by Evidence Figure 1: An event model of TDT topic ‘Osama bin Laden"s indictment". Event threading is strongly related to topic detection and tracking, but also different from it significantly. It goes beyond topics, and models the relationships between events. Thus, event threading can be considered as a further extension of topic detection and tracking and is more challenging due to at least the following difficulties. 1. The number of events is unknown. 2. The granularity of events is hard to define. 3. The dependencies among events are hard to model. 4. Since it is a brand new research area, no standard evaluation metrics and benchmark data is available. In the next few sections, we will describe our attempts to tackle these problems. 4. LABELED DATA We picked 28 topics from the TDT2 corpus and 25 topics from the TDT3 corpus. The criterion we used for selecting a topic is that it should contain at least 15 on-topic stories from CNN headline news. If the topic contained more than 30 CNN stories, we picked only the first 30 stories to keep the topic short enough for annotators. The reason for choosing only CNN as the source is that the stories from this source tend to be short and precise and do not tend to digress or drift too far away from the central theme. We believe modeling such stories would be a useful first step before dealing with more complex data sets. We hired an annotator to create truth data. Annotation includes defining the event membership for each story and also the dependencies. We supervised the annotator on a set of three topics that we did our own annotations on and then asked her to annotate the 28 topics from TDT2 and 25 topics from TDT3. In identifying events in a topic, the annotator was asked to broadly follow the TDT definition of an event, i.e., ‘something that happens at a specific time and location". The annotator was encouraged to merge two events A and B into a single event C if any of the stories discusses both A and B. This is to satisfy our assumption that each story corresponds to a unique event. The annotator was also encouraged to avoid singleton events, events that contain a single news story, if possible. We realized from our own experience that people differ in their perception of an event especially when the number of stories in that event is small. As part of the guidelines, we instructed the annotator to assign titles to all the events in each topic. We believe that this would help make her understanding of the events more concrete. We however, do not use or model these titles in our algorithms. In defining dependencies between events, we imposed no restrictions on the graph structure. Each event could have single, multiple or no parents. Further, the graph could have cycles or orphannodes. The annotator was however instructed to assign a dependency from event A to event B if and only if the occurrence of B is ‘either causally influenced by A or is closely related to A and follows A in time". From the annotated topics, we created a training set of 26 topics and a test set of 27 topics by merging the 28 topics from TDT2 and 25 from TDT3 and splitting them randomly. Table 1 shows that the training and test sets have fairly similar statistics. Feature Training set Test set Num. topics 26 27 Avg. Num. Stories/Topic 28.69 26.74 Avg. Doc. Len. 64.60 64.04 Avg. Num. Stories/Event 5.65 6.22 Avg. Num. Events/Topic 5.07 4.29 Avg. Num. Dependencies/Topic 3.07 2.92 Avg. Num. Dependencies/Event 0.61 0.68 Avg. Num. Days/Topic 30.65 34.48 Table 1: Statistics of annotated data 5. EVALUATION A system can generate some event model ż ´ ¼ ¼µ using certain algorithms, which is usually different from the truth model Å ´ µ (we assume the annotator did not make any mistake). Comparing a system event model ż with the true model Å requires comparing the entire event models including their dependency structure. And different event granularities may bring huge discrepancy between ż and Å. This is certainly non-trivial as even testing whether two graphs are isomorphic has no known polynomial time solution. Hence instead of comparing the actual structure we examine a pair of stories at a time and verify if the system and true labels agree on their event-memberships and dependencies. Specifically, we compare two kinds of story pairs: ¯ Cluster pairs ( ´Åµ): These are the complete set of unordered pairs ´× × µ of stories × and × that fall within the same event given a model Å. Formally, ´Åµ ´× × µ × × ¾ Ë ´× µ ´× µ (5) where is the function in Å that maps stories to events as defined in equation 4. ¯ Dependency pairs ( ´Åµ): These are the set of all ordered pairs of stories ´× × µ such that there is a dependency from the event of × to the event of × in the model Å. ´Åµ ´× × µ ´ ´× µ ´× µµ ¾ (6) Note the story pair is ordered here, so ´× × µ is not equivalent to ´× × µ. In our evaluation, a correct pair with wrong 448 (B->D) Cluster pairs (A,C) Dependency pairs (A->B) (C->B) (B->D) D,E D,E (D,E) (D,E) (A->C) (A->E) (B->C) (B->E) (B->E) Cluster precision: 1/2 Cluster Recall: 1/2 Dependency Recall: 2/6 Dependency Precision: 2/4 (A->D) True event model System event model A,B C A,C B Cluster pairs (A,B) Dependency pairs Figure 2: Evaluation measures direction will be considered a mistake. As we mentioned earlier in section 3, ignoring the direction may make the problem simpler, but we will lose the expressiveness of our representation. Given these two sets of story pairs corresponding to the true event model Å and the system event model ż, we define recall and precision for each category as follows. ¯ Cluster Precision (CP): It is the probability that two randomly selected stories × and × are in the same true-event given that they are in the same system event. È È´ ´× µ ´× µ ¼´× µ ¼´× µµ ´Åµ ´Å¼µ ´Å¼µ (7) where ¼ is the story-event mapping function corresponding to the model ż. ¯ Cluster Recall(CR): It is the probability that two randomly selected stories × and × are in the same system-event given that they are in the same true event. Ê È´ ¼´× µ ¼´× µ ´× µ ´× µµ ´Åµ ´Å¼µ ´Åµ (8) ¯ Dependency Precision(DP): It is the probability that there is a dependency between the events of two randomly selected stories × and × in the true model Å given that they have a dependency in the system model ż. Note that the direction of dependency is important in comparison. È È´´ ´× µ ´× µµ ¾ ´ ¼´× µ ¼´× µµ ¾ ¼µ ´Åµ ´Å¼µ ´Å¼µ (9) ¯ Dependency Recall(DR): It is the probability that there is a dependency between the events of two randomly selected stories × and × in the system model ż given that they have a dependency in the true model Å. Again, the direction of dependency is taken into consideration. Ê È´´ ¼´× µ ¼´× µµ ¾ ¼ ´ ´× µ ´× µµ ¾ µ ´Åµ ´Å¼µ ´Åµ (10) The measures are illustrated by an example in figure 2. We also combine these measures using the well known F1-measure commonly used in text classification and other research areas as shown below. ¾ ¢ È ¢ Ê È · Ê ¾ ¢ È ¢ Ê È · Ê Â ¾ ¢ ¢ · (11) where and are the cluster and dependency F1-measures respectively and  is the Joint F1-measure ( ) that we use to measure the overall performance. 6. TECHNIQUES The task of event modeling can be split into two parts: clustering the stories into unique events in the topic and constructing dependencies among them. In the following subsections, we describe techniques we developed for each of these sub-tasks. 6.1 Clustering Each topic is composed of multiple events, so stories must be clustered into events before we can model the dependencies among them. For simplicity, all stories in the same topic are assumed to be available at one time, rather than coming in a text stream. This task is similar to traditional clustering but features other than word distributions may also be critical in our application. In many text clustering systems, the similarity between two stories is the inner product of their tf-idf vectors, hence we use it as one of our features. Stories in the same event tend to follow temporal locality, so the time stamp of each story can be a useful feature. Additionally, named-entities such as person and location names are another obvious feature when forming events. Stories in the same event tend to be related to the same person(s) and locations(s). In this subsection, we present an agglomerative clustering algorithm that combines all these features. In our experiments, however, we study the effect of each feature on the performance separately using modified versions of this algorithm. 6.1.1 Agglomerative clustering with time decay (ACDT) We initialize our events to singleton events (clusters), i.e., each cluster contains exactly one story. So the similarity between two events, to start with, is exactly the similarity between the corresponding stories. The similarity Û×ÙÑ´×½ ×¾µ between two stories ×½ and ×¾ is given by the following formula: Û×ÙÑ´×½ ×¾µ ½ Ó×´×½ ×¾µ · ¾ÄÓ ´×½ ×¾µ · ¿È Ö´×½ ×¾µ (12) Here ½, ¾, ¿ are the weights on different features. In this work, we determined them empirically, but in the future, one can consider more sophisticated learning techniques to determine them. Ó×´×½ ×¾µ is the cosine similarity of the term vectors. ÄÓ ´×½ ×¾µ is 1 if there is some location that appears in both stories, otherwise it is 0. È Ö´×½ ×¾µ is similarly defined for person name. We use time decay when calculating similarity of story pairs, i.e., the larger time difference between two stories, the smaller their similarities. The time period of each topic differs a lot, from a few days to a few months. So we normalize the time difference using the whole duration of that topic. The time decay adjusted similarity 449 × Ñ´×½ ×¾µ is given by × Ñ´×½ ×¾µ Û×ÙÑ´×½ ×¾µ « ؽ Ø¾ Ì (13) where ؽ and ؾ are the time stamps for story 1 and 2 respectively. T is the time difference between the earliest and the latest story in the given topic. « is the time decay factor. In each iteration, we find the most similar event pair and merge them. We have three different ways to compute the similarity between two events Ù and Ú: ¯ Average link: In this case the similarity is the average of the similarities of all pairs of stories between Ù and Ú as shown below: × Ñ´ Ù Ú µ È×Ù¾ Ù È×Ú¾ Ú × Ñ´×Ù ×Ú µ Ù Ú (14) ¯ Complete link: The similarity between two events is given by the smallest of the pair-wise similarities. × Ñ´ Ù Ú µ Ñ Ò ×Ù¾ Ù ×Ú¾ Ú × Ñ´×Ù ×Ú µ (15) ¯ Single link: Here the similarity is given by the best similarity between all pairs of stories. × Ñ´ Ù Ú µ Ñ Ü ×Ù¾ Ù ×Ú¾ Ú × Ñ´×Ù ×Ú µ (16) This process continues until the maximum similarity falls below the threshold or the number of clusters is smaller than a given number. 6.2 Dependency modeling Capturing dependencies is an extremely hard problem because it may require a ‘deeper understanding" of the events in question. A human annotator decides on dependencies not just based on the information in the events but also based on his/her vast repertoire of domain-knowledge and general understanding of how things operate in the world. For example, in Figure 1 a human knows ‘Trial and indictment of Osama" is influenced by ‘Evidence gathered by CIA" because he/she understands the process of law in general. We believe a robust model should incorporate such domain knowledge in capturing dependencies, but in this work, as a first step, we will rely on surface-features such as time-ordering of news stories and word distributions to model them. Our experiments in later sections demonstrate that such features are indeed useful in capturing dependencies to a large extent. In this subsection, we describe the models we considered for capturing dependencies. In the rest of the discussion in this subsection, we assume that we are already given the mapping ¼ Ë and we focus only on modeling the edges ¼. First we define a couple of features that the following models will employ. First we define a 1-1 time-ordering function Ø Ë ½ ¡ ¡ ¡ Ò that sorts stories in ascending order by their time of publication. Now, the event-time-ordering function Ø is defined as follows. Ø ½ ¡ ¡ ¡ Ñ × Ø Ù Ú ¾ Ø ´ Ùµ Ø ´ Úµ ´µ Ñ Ò ×Ù¾ Ù Ø´×Ùµ Ñ Ò ×Ú¾ Ú Ø´×Úµ (17) In other words, Ø time-orders events based on the time-ordering of their respective first stories. We will also use average cosine similarity between two events as a feature and it is defined as follows. Ú Ë Ñ´ Ù Ú µ È×Ù¾ Ù È×Ú¾ Ú Ó×´×Ù ×Ú µ Ù Ú (18) 6.2.1 Complete-Link model In this model, we assume that there are dependencies between all pairs of events. The direction of dependency is determined by the time-ordering of the first stories in the respective events. Formally, the system edges are defined as follows. ¼ ´ Ù Ú µ Ø ´ Ùµ Ø ´ Ú µ (19) where Ø is the event-time-ordering function. In other words, the dependency edge is directed from event Ù to event Ú , if the first story in event Ù is earlier than the first story in event Ú . We point out that this is not to be confused with the complete-link algorithm in clustering. Although we use the same names, it will be clear from the context which one we refer to. 6.2.2 Simple Thresholding This model is an extension of the complete link model with an additional constraint that there is a dependency between any two events Ù and Ú only if the average cosine similarity between event Ù and event Ú is greater than a threshold Ì. Formally, ¼ ´ Ù Úµ Ú Ë Ñ´ Ù Ú µ Ì Ø ´ Ùµ Ø ´ Ú µ (20) 6.2.3 Nearest Parent Model In this model, we assume that each event can have at most one parent. We define the set of dependencies as follows. ¼ ´ Ù Úµ Ú Ë Ñ´ Ù Ú µ Ì Ø ´ Úµ Ø ´ Ùµ · ½ (21) Thus, for each event Ú , the nearest parent model considers only the event preceding it as defined by Ø as a potential candidate. The candidate is assigned as the parent only if the average similarity exceeds a pre-defined threshold Ì. 6.2.4 Best Similarity Model This model also assumes that each event can have at most one parent. An event Ú is assigned a parent Ù if and only if Ù is the most similar earlier event to Ú and the similarity exceeds a threshold Ì. Mathematically, this can be expressed as: ¼ ´ Ù Ú µ Ú Ë Ñ´ Ù Úµ Ì Ù Ö Ñ Ü Û Ø ´ Ûµ Ø ´ Úµ Ú Ë Ñ´ Û Ú µ (22) 6.2.5 Maximum Spanning Tree model In this model, we first build a maximum spanning tree (MST) using a greedy algorithm on the following fully connected weighted, undirected graph whose vertices are the events and whose edges are defined as follows: ´ Ù Ú µ Û´ Ù Ú µ Ú Ë Ñ´ Ù Úµ (23) Let ÅËÌ´ µ be the set of edges in the maximum spanning tree of ¼. Now our directed dependency edges are defined as follows. ¼ ´ Ù Ú µ ´ Ù Ú µ ¾ ÅËÌ´ µ Ø ´ Ùµ Ø ´ Úµ Ú Ë Ñ´ Ù Ú µ Ì (24) 450 Thus in this model, we assign dependencies between the most similar events in the topic. 7. EXPERIMENTS Our experiments consists of three parts. First we modeled only the event clustering part (defining the mapping function ¼) using clustering algorithms described in section 6.1. Then we modeled only the dependencies by providing to the system the true clusters and running only the dependency algorithms of section 6.2. Finally, we experimented with combinations of clustering and dependency algorithms to produce the complete event model. This way of experimentation allows us to compare the performance of our algorithms in isolation and in association with other components. The following subsections present the three parts of our experimentation. 7.1 Clustering We have tried several variations of the Ì algorithm to study the effects of various features on the clustering performance. All the parameters are learned by tuning on the training set. We also tested the algorithms on the test set with parameters fixed at their optimal values learned from training. We used agglomerative clusModel best T CP CR CF P-value cos+1-lnk 0.15 0.41 0.56 0.43cos+all-lnk 0.00 0.40 0.62 0.45cos+Loc+avg-lnk 0.07 0.37 0.74 0.45cos+Per+avg-lnk 0.07 0.39 0.70 0.46cos+TD+avg-lnk 0.04 0.45 0.70 0.53 2.9e-4* cos+N(T)+avg-lnk - 0.41 0.62 0.48 7.5e-2 cos+N(T)+T+avg-lnk 0.03 0.42 0.62 0.49 2.4e-2* cos+TD+N(T)+avg-lnk - 0.44 0.66 0.52 7.0e-3* cos+TD+N(T)+T+avg-lnk 0.03 0.47 0.64 0.53 1.1e-3* Baseline(cos+avg-lnk) 0.05 0.39 0.67 0.46Table 2: Comparison of agglomerative clustering algorithms (training set) tering based on only cosine similarity as our clustering baseline. The results on the training and test sets are in Table 2 and 3 respectively. We use the Cluster F1-measure (CF) averaged over all topics as our evaluation criterion. Model CP CR CF P-value cos+1-lnk 0.43 0.49 0.39cos+all-lnk 0.43 0.62 0.47cos+Loc+avg-lnk 0.37 0.73 0.45cos+Per+avg-lnk 0.44 0.62 0.45cos+TD+avg-lnk 0.48 0.70 0.54 0.014* cos+N(T)+avg-lnk 0.41 0.71 0.51 0.31 cos+N(T)+T+avg-lnk 0.43 0.69* 0.52 0.14 cos+TD+N(T)+avg-lnk 0.43 0.76 0.54 0.025* cos+TD+N(T)+T+avg-lnk 0.47 0.69 0.54 0.0095* Baseline(cos+avg-lnk) 0.44 0.67 0.50Table 3: Comparison of agglomerative clustering algorithms (test set) P-value marked with a £ means that it is a statistically significant improvement over the baseline (95% confidence level, one tailed T-test). The methods shown in table 2 and 3 are: ¯ Baseline: tf-idf vector weight, cosine similarity, average link in clustering. In equation 12, ½ ½, ¾ ¿ ¼. And « ¼ in equation 13. This F-value is the maximum obtained by tuning the threshold. ¯ cos+1-lnk: Single link comparison (see equation 16) is used where similarity of two clusters is the maximum of all story pairs, other configurations are the same as the baseline run. ¯ cos+all-lnk: Complete link algorithm of equation 15 is used. Similar to single link but it takes the minimum similarity of all story pairs. ¯ cos+Loc+avg-lnk: Location names are used when calculating similarity. ¾ ¼ ¼ in equation 12. All algorithms starting from this one use average link (equation 14), since single link and complete link do not show any improvement of performance. ¯ cos+Per+avg-lnk: ¿ ¼ ¼ in equation 12, i.e., we put some weight on person names in the similarity. ¯ cos+TD+avg-lnk: Time Decay coefficient « ½ in equation 13, which means the similarity between two stories will be decayed to ½ if they are at different ends of the topic. ¯ cos+N(T)+avg-lnk: Use the number of true events to control the agglomerative clustering algorithm. When the number of clusters is fewer than that of truth events, stop merging clusters. ¯ cos+N(T)+T+avg-lnk: similar to N(T) but also stop agglomeration if the maximal similarity is below the threshold Ì. ¯ cos+TD:+N(T)+avg-lnk: similar to N(T) but the similarities are decayed, « ½ in equation 13. ¯ cos+TD+N(T)+T+avg-lnk: similar to TD+N(Truth) but calculation halts when the maximal similarity is smaller than the threshold Ì. Our experiments demonstrate that single link and complete link similarities perform worse than average link, which is reasonable since average link is less sensitive to one or two story pairs. We had expected locations and person names to improve the result, but it is not the case. Analysis of topics shows that many on-topic stories share the same locations or persons irrespective of the event they belong to, so these features may be more useful in identifying topics rather than events. Time decay is successful because events are temporally localized, i.e., stories discussing the same event tend to be adjacent to each other in terms of time. Also we noticed that providing the number of true events improves the performance since it guides the clustering algorithm to get correct granularity. However, for most applications, it is not available. We used it only as a cheat experiment for comparison with other algorithms. On the whole, time decay proved to the most powerful feature besides cosine similarity on both training and test sets. 7.2 Dependencies In this subsection, our goal is to model only dependencies. We use the true mapping function and by implication the true events Î . We build our dependency structure ¼ using all the five models described in section 6.2. We first train our models on the 26 training topics. Training involves learning the best threshold Ì for each of the models. We then test the performances of all the trained models on the 27 test topics. We evaluate our performance 451 using the average values of Dependency Precision (DP), Dependency Recall (DR) and Dependency F-measure (DF). We consider the complete-link model to be our baseline since for each event, it trivially considers all earlier events to be parents. Table 4 lists the results on the training set. We see that while all the algorithms except MST outperform the baseline complete-link algorithm , the nearest Parent algorithm is statistically significant from the baseline in terms of its DF-value using a one-tailed paired T-test at 95% confidence level. Model best Ì DP DR DF P-value Nearest Parent 0.025 0.55 0.62 0.56 0.04* Best Similarity 0.02 0.51 0.62 0.53 0.24 MST 0.0 0.46 0.58 0.48Simple Thresh. 0.045 0.45 0.76 0.52 0.14 Complete-link - 0.36 0.93 0.48Table 4: Results on the training set: Best Ì is the optimal value of the threshold Ì. * indicates the corresponding model is statistically significant compared to the baseline using a one-tailed, paired T-test at 95% confidence level. In table 5 we present the comparison of the models on the test set. Here, we do not use any tuning but set the threshold to the corresponding optimal values learned from the training set. The results throw some surprises: The nearest parent model, which was significantly better than the baseline on training set, turns out to be worse than the baseline on the test set. However all the other models are better than the baseline including the best similarity which is statistically significant. Notice that all the models that perform better than the baseline in terms of DF, actually sacrifice their recall performance compared to the baseline, but improve on their precision substantially thereby improving their performance on the DF-measure. We notice that both simple-thresholding and best similarity are better than the baseline on both training and test sets although the improvement is not significant. On the whole, we observe that the surface-level features we used capture the dependencies to a reasonable level achieving a best value of 0.72 DF on the test set. Although there is a lot of room for improvement, we believe this is a good first step. Model DP DR DF P-value Nearest Parent 0.61 0.60 0.60Best Similarity 0.71 0.74 0.72 0.04* MST 0.70 0.68 0.69 0.22 Simple Thresh. 0.57 0.75 0.64 0.24 Baseline (Complete-link) 0.50 0.94 0.63Table 5: Results on the test set 7.3 Combining Clustering and Dependencies Now that we have studied the clustering and dependency algorithms in isolation, we combine the best performing algorithms and build the entire event model. Since none of the dependency algorithms has been shown to be consistently and significantly better than the others, we use all of them in our experimentation. From the clustering techniques, we choose the best performing Cos+TD. As a baseline, we use a combination of the baselines in each components, i.e., cos for clustering and complete-link for dependencies. Note that we need to retrain all the algorithms on the training set because our objective function to optimize is now JF, the joint F-measure. For each algorithm, we need to optimize both the clustering threshold and the dependency threshold. We did this empirically on the training set and the optimal values are listed in table 6. The results on the training set, also presented in table 6, indicate that cos+TD+Simple-Thresholding is significantly better than the baseline in terms of the joint F-value JF, using a one-tailed paired Ttest at 95% confidence level. On the whole, we notice that while the clustering performance is comparable to the experiments in section 7.1, the overall performance is undermined by the low dependency performance. Unlike our experiments in section 7.2 where we had provided the true clusters to the system, in this case, the system has to deal with deterioration in the cluster quality. Hence the performance of the dependency algorithms has suffered substantially thereby lowering the overall performance. The results on the test set present a very similar story as shown in table 7. We also notice a fair amount of consistency in the performance of the combination algorithms. cos+TD+Simple-Thresholding outperforms the baseline significantly. The test set results also point to the fact that the clustering component remains a bottleneck in achieving an overall good performance. 8. DISCUSSION AND CONCLUSIONS In this paper, we have presented a new perspective of modeling news topics. Contrary to the TDT view of topics as flat collection of news stories, we view a news topic as a relational structure of events interconnected by dependencies. In this paper, we also proposed a few approaches for both clustering stories into events and constructing dependencies among them. We developed a timedecay based clustering approach that takes advantage of temporallocalization of news stories on the same event and showed that it performs significantly better than the baseline approach based on cosine similarity. Our experiments also show that we can do fairly well on dependencies using only surface-features such as cosinesimilarity and time-stamps of news stories as long as true events are provided to the system. However, the performance deteriorates rapidly if the system has to discover the events by itself. Despite that discouraging result, we have shown that our combined algorithms perform significantly better than the baselines. Our results indicate modeling dependencies can be a very hard problem especially when the clustering performance is below ideal level. Errors in clustering have a magnifying effect on errors in dependencies as we have seen in our experiments. Hence, we should focus not only on improving dependencies but also on clustering at the same time. As part of our future work, we plan to investigate further into the data and discover new features that influence clustering as well as dependencies. And for modeling dependencies, a probabilistic framework should be a better choice since there is no definite answer of yes/no for the causal relations among some events. We also hope to devise an iterative algorithm which can improve clustering and dependency performance alternately as suggested by one of the reviewers. We also hope to expand our labeled corpus further to include more diverse news sources and larger and more complex event structures. Acknowledgments We would like to thank the three anonymous reviewers for their valuable comments. This work was supported in part by the Center 452 Model Cluster T Dep. T CP CR CF DP DR DF JF P-value cos+TD+Nearest-Parent 0.055 0.02 0.51 0.53 0.49 0.21 0.19 0.19 0.27cos+TD+Best-Similarity 0.04 0.02 0.45 0.70 0.53 0.21 0.33 0.23 0.32cos+TD+MST 0.04 0.00 0.45 0.70 0.53 0.22 0.35 0.25 0.33cos+TD+Simple-Thresholding 0.065 0.02 0.56 0.47 0.48 0.23 0.61 0.32 0.38 0.0004* Baseline (cos+Complete-link) 0.10 - 0.58 0.31 0.38 0.20 0.67 0.30 0.33Table 6: Combined results on the training set Model CP CR CF DP DR DF JF P-value cos+TD+Nearest Parent 0.57 0.50 0.50 0.27 0.19 0.21 0.30cos+TD+Best Similarity 0.48 0.70 0.54 0.31 0.27 0.26 0.35cos+TD+MST 0.48 0.70 0.54 0.31 0.30 0.28 0.37cos+TD+Simple Thresholding 0.60 0.39 0.44 0.32 0.66 0.42 0.43 0.0081* Baseline (cos+Complete-link) 0.66 0.27 0.36 0.30 0.72 0.43 0.39Table 7: Combined results on the test set for Intelligent Information Retrieval and in part by SPAWARSYSCENSD grant number N66001-02-1-8903. Any opinions, findings and conclusions or recommendations expressed in this material are the authors" and do not necessarily reflect those of the sponsor. 9. REFERENCES [1] J. Allan, J. Carbonell, G. Doddington, J. Yamron, and Y. Yang. Topic detection and tracking pilot study: Final report. In Proceedings of the DARPA Broadcast News Transcription and Understanding Workshop, pages 194-218, 1998. [2] J. Allan, A. Feng, and A. Bolivar. Flexible intrinsic evaluation of hierarchical clustering for tdt. volume In the Proc. of the ACM Twelfth International Conference on Information and Knowledge Management, pages 263-270, Nov 2003. [3] James Allan, editor. Topic Detection and Tracking:Event based Information Organization. Kluwer Academic Publishers, 2000. [4] James Allan, Rahul Gupta, and Vikas Khandelwal. Temporal summaries of new topics. In Proceedings of the 24th annual international ACM SIGIR conference on Research and development in information retrieval, pages 10-18. ACM Press, 2001. [5] Regina Barzilay and Lillian Lee. Catching the drift: Probabilistic content models, with applications to generation and summarization. In Proceedings of Human Language Technology Conference and North American Chapter of the Association for Computational Linguistics(HLT-NAACL), pages 113-120, 2004. [6] D. Lawrie and W. B. Croft. Discovering and comparing topic hierarchies. In Proceedings of RIAO 2000 Conference, pages 314-330, 1999. [7] David D. Lewis and Kimberly A. Knowles. Threading electronic mail: a preliminary study. Inf. Process. Manage., 33(2):209-217, 1997. [8] Juha Makkonen. Investigations on event evolution in tdt. In Proceedings of HLT-NAACL 2003 Student Workshop, pages 43-48, 2004. [9] Aixin Sun and Ee-Peng Lim. Hierarchical text classification and evaluation. In Proceedings of the 2001 IEEE International Conference on Data Mining, pages 521-528. IEEE Computer Society, 2001. [10] Yiming Yang, Jaime Carbonell, Ralf Brown, Thomas Pierce, Brian T. Archibald, and Xin Liu. Learning approaches for detecting and tracking news events. In IEEE Intelligent Systems Special Issue on Applications of Intelligent Information Retrieval, volume 14 (4), pages 32-43, 1999. 453

dependency recall;topic detection;correct granularity;flatcluster;thread;simple thresholding;mapping function;temporal locality;event threading;seminal event;time ordering;cluster;topic cluster;inter-related event;quick overview;directed edge;dependency precision;hidden markov model;cluster of topic;dependency;cosine similarity;novel feature;event;atomicity;flat hierarchy;microscopic event;automatic technique;agglomerative clustering;dependency f-measure;maximum spanning tree;event recognition;temporallocalization;event model;term vector;timedecay;news organization;time-ordering

train_H-85

Learning User Interaction Models for Predicting Web Search Result Preferences

Evaluating user preferences of web search results is crucial for search engine development, deployment, and maintenance. We present a real-world study of modeling the behavior of web search users to predict web search result preferences. Accurate modeling and interpretation of user behavior has important applications to ranking, click spam detection, web search personalization, and other tasks. Our key insight to improving robustness of interpreting implicit feedback is to model query-dependent deviations from the expected noisy user behavior. We show that our model of clickthrough interpretation improves prediction accuracy over state-of-the-art clickthrough methods. We generalize our approach to model user behavior beyond clickthrough, which results in higher preference prediction accuracy than models based on clickthrough information alone. We report results of a large-scale experimental evaluation that show substantial improvements over published implicit feedback interpretation methods.

1. INTRODUCTION Relevance measurement is crucial to web search and to information retrieval in general. Traditionally, search relevance is measured by using human assessors to judge the relevance of query-document pairs. However, explicit human ratings are expensive and difficult to obtain. At the same time, millions of people interact daily with web search engines, providing valuable implicit feedback through their interactions with the search results. If we could turn these interactions into relevance judgments, we could obtain large amounts of data for evaluating, maintaining, and improving information retrieval systems. Recently, automatic or implicit relevance feedback has developed into an active area of research in the information retrieval community, at least in part due to an increase in available resources and to the rising popularity of web search. However, most traditional IR work was performed over controlled test collections and carefully-selected query sets and tasks. Therefore, it is not clear whether these techniques will work for general real-world web search. A significant distinction is that web search is not controlled. Individual users may behave irrationally or maliciously, or may not even be real users; all of this affects the data that can be gathered. But the amount of the user interaction data is orders of magnitude larger than anything available in a non-web-search setting. By using the aggregated behavior of large numbers of users (and not treating each user as an individual expert) we can correct for the noise inherent in individual interactions, and generate relevance judgments that are more accurate than techniques not specifically designed for the web search setting. Furthermore, observations and insights obtained in laboratory settings do not necessarily translate to real world usage. Hence, it is preferable to automatically induce feedback interpretation strategies from large amounts of user interactions. Automatically learning to interpret user behavior would allow systems to adapt to changing conditions, changing user behavior patterns, and different search settings. We present techniques to automatically interpret the collective behavior of users interacting with a web search engine to predict user preferences for search results. Our contributions include: • A distributional model of user behavior, robust to noise within individual user sessions, that can recover relevance preferences from user interactions (Section 3). • Extensions of existing clickthrough strategies to include richer browsing and interaction features (Section 4). • A thorough evaluation of our user behavior models, as well as of previously published state-of-the-art techniques, over a large set of web search sessions (Sections 5 and 6). We discuss our results and outline future directions and various applications of this work in Section 7, which concludes the paper. 2. BACKGROUND AND RELATED WORK Ranking search results is a fundamental problem in information retrieval. The most common approaches in the context of the web use both the similarity of the query to the page content, and the overall quality of a page [3, 20]. A state-ofthe-art search engine may use hundreds of features to describe a candidate page, employing sophisticated algorithms to rank pages based on these features. Current search engines are commonly tuned on human relevance judgments. Human annotators rate a set of pages for a query according to perceived relevance, creating the gold standard against which different ranking algorithms can be evaluated. Reducing the dependence on explicit human judgments by using implicit relevance feedback has been an active topic of research. Several research groups have evaluated the relationship between implicit measures and user interest. In these studies, both reading time and explicit ratings of interest are collected. Morita and Shinoda [14] studied the amount of time that users spent reading Usenet news articles and found that reading time could predict a user"s interest levels. Konstan et al. [13] showed that reading time was a strong predictor of user interest in their GroupLens system. Oard and Kim [15] studied whether implicit feedback could substitute for explicit ratings in recommender systems. More recently, Oard and Kim [16] presented a framework for characterizing observable user behaviors using two dimensions-the underlying purpose of the observed behavior and the scope of the item being acted upon. Goecks and Shavlik [8] approximated human labels by collecting a set of page activity measures while users browsed the World Wide Web. The authors hypothesized correlations between a high degree of page activity and a user"s interest. While the results were promising, the sample size was small and the implicit measures were not tested against explicit judgments of user interest. Claypool et al. [6] studied how several implicit measures related to the interests of the user. They developed a custom browser called the Curious Browser to gather data, in a computer lab, about implicit interest indicators and to probe for explicit judgments of Web pages visited. Claypool et al. found that the time spent on a page, the amount of scrolling on a page, and the combination of time and scrolling have a strong positive relationship with explicit interest, while individual scrolling methods and mouse-clicks were not correlated with explicit interest. Fox et al. [7] explored the relationship between implicit and explicit measures in Web search. They built an instrumented browser to collect data and then developed Bayesian models to relate implicit measures and explicit relevance judgments for both individual queries and search sessions. They found that clickthrough was the most important individual variable but that predictive accuracy could be improved by using additional variables, notably dwell time on a page. Joachims [9] developed valuable insights into the collection of implicit measures, introducing a technique based entirely on clickthrough data to learn ranking functions. More recently, Joachims et al. [10] presented an empirical evaluation of interpreting clickthrough evidence. By performing eye tracking studies and correlating predictions of their strategies with explicit ratings, the authors showed that it is possible to accurately interpret clickthrough events in a controlled, laboratory setting. A more comprehensive overview of studies of implicit measures is described in Kelly and Teevan [12]. Unfortunately, the extent to which existing research applies to real-world web search is unclear. In this paper, we build on previous research to develop robust user behavior interpretation models for the real web search setting. 3. LEARNING USER BEHAVIOR MODELS As we noted earlier, real web search user behavior can be noisy in the sense that user behaviors are only probabilistically related to explicit relevance judgments and preferences. Hence, instead of treating each user as a reliable expert, we aggregate information from many unreliable user search session traces. Our main approach is to model user web search behavior as if it were generated by two components: a relevance component - queryspecific behavior influenced by the apparent result relevance, and a background component - users clicking indiscriminately. Our general idea is to model the deviations from the expected user behavior. Hence, in addition to basic features, which we will describe in detail in Section 3.2, we compute derived features that measure the deviation of the observed feature value for a given search result from the expected values for a result, with no query-dependent information. We motivate our intuitions with a particularly important behavior feature, result clickthrough, analyzed next, and then introduce our general model of user behavior that incorporates other user actions (Section 3.2). 3.1 A Case Study in Click Distributions As we discussed, we aggregate statistics across many user sessions. A click on a result may mean that some user found the result summary promising; it could also be caused by people clicking indiscriminately. In general, individual user behavior, clickthrough and otherwise, is noisy, and cannot be relied upon for accurate relevance judgments. The data set is described in more detail in Section 5.2. For the present it suffices to note that we focus on a random sample of 3,500 queries that were randomly sampled from query logs. For these queries we aggregate click data over more than 120,000 searches performed over a three week period. We also have explicit relevance judgments for the top 10 results for each query. Figure 3.1 shows the relative clickthrough frequency as a function of result position. The aggregated click frequency at result position p is calculated by first computing the frequency of a click at p for each query (i.e., approximating the probability that a randomly chosen click for that query would land on position p). These frequencies are then averaged across queries and normalized so that relative frequency of a click at the top position is 1. The resulting distribution agrees with previous observations that users click more often on top-ranked results. This reflects the fact that search engines do a reasonable job of ranking results as well as biases to click top results and noisewe attempt to separate these components in the analysis that follows. 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 1 3 5 7 9 11 13 15 17 19 21 23 25 27 29 result position RelativeClickFrequency Figure 3.1: Relative click frequency for top 30 result positions over 3,500 queries and 120,000 searches. First we consider the distribution of clicks for the relevant documents for these queries. Figure 3.2 reports the aggregated click distribution for queries with varying Position of Top Relevant document (PTR). While there are many clicks above the first relevant document for each distribution, there are clearly peaks in click frequency for the first relevant result. For example, for queries with top relevant result in position 2, the relative click frequency at that position (second bar) is higher than the click frequency at other positions for these queries. Nevertheless, many users still click on the non-relevant results in position 1 for such queries. This shows a stronger property of the bias in the click distribution towards top results - users click more often on results that are ranked higher, even when they are not relevant. 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 1 2 3 5 10 result position relativeclickfrequency PTR=1 PTR=2 PTR=3 PTR=5 PTR=10 Background Figure 3.2: Relative click frequency for queries with varying PTR (Position of Top Relevant document). -0.06 -0.04 -0.02 0 0.02 0.04 0.06 0.08 0.1 0.12 0.14 0.16 1 2 3 5 10 result position correctedrelativeclickfrequency PTR=1 PTR=2 PTR=3 PTR=5 PTR=10 Figure 3.3: Relative corrected click frequency for relevant documents with varying PTR (Position of Top Relevant). If we subtract the background distribution of Figure 3.1 from the mixed distribution of Figure 3.2, we obtain the distribution in Figure 3.3, where the remaining click frequency distribution can be interpreted as the relevance component of the results. Note that the corrected click distribution correlates closely with actual result relevance as explicitly rated by human judges. 3.2 Robust User Behavior Model Clicks on search results comprise only a small fraction of the post-search activities typically performed by users. We now introduce our techniques for going beyond the clickthrough statistics and explicitly modeling post-search user behavior. Although clickthrough distributions are heavily biased towards top results, we have just shown how the ‘relevance-driven" click distribution can be recovered by correcting for the prior, background distribution. We conjecture that other aspects of user behavior (e.g., page dwell time) are similarly distorted. Our general model includes two feature types for describing user behavior: direct and deviational where the former is the directly measured values, and latter is deviation from the expected values estimated from the overall (query-independent) distributions for the corresponding directly observed features. More formally, we postulate that the observed value o of a feature f for a query q and result r can be expressed as a mixture of two components: ),,()(),,( frqrelfCfrqo += (1) where )( fC is the prior background distribution for values of f aggregated across all queries, and rel(q,r,f) is the component of the behavior influenced by the relevance of the result r. As illustrated above with the clickthrough feature, if we subtract the background distribution (i.e., the expected clickthrough for a result at a given position) from the observed clickthrough frequency at a given position, we can approximate the relevance component of the clickthrough value1 . In order to reduce the effect of individual user variations in behavior, we average observed feature values across all users and search sessions for each query-URL pair. This aggregation gives additional robustness of not relying on individual noisy user interactions. In summary, the user behavior for a query-URL pair is represented by a feature vector that includes both the directly observed features and the derived, corrected feature values. We now describe the actual features we use to represent user behavior. 3.3 Features for Representing User Behavior Our goal is to devise a sufficiently rich set of features that allow us to characterize when a user will be satisfied with a web search result. Once the user has submitted a query, they perform many different actions (reading snippets, clicking results, navigating, refining their query) which we capture and summarize. This information was obtained via opt-in client-side instrumentation from users of a major web search engine. This rich representation of user behavior is similar in many respects to the recent work by Fox et al. [7]. An important difference is that many of our features are (by design) query specific whereas theirs was (by design) a general, queryindependent model of user behavior. Furthermore, we include derived, distributional features computed as described above. The features we use to represent user search interactions are summarized in Table 3.1. For clarity, we organize the features into the groups Query-text, Clickthrough, and Browsing. Query-text features: Users decide which results to examine in more detail by looking at the result title, URL, and summary - in some cases, looking at the original document is not even necessary. To model this aspect of user experience we defined features to characterize the nature of the query and its relation to the snippet text. These include features such as overlap between the words in title and in query (TitleOverlap), the fraction of words shared by the query and the result summary (SummaryOverlap), etc. Browsing features: Simple aspects of the user web page interactions can be captured and quantified. These features are used to characterize interactions with pages beyond the results page. For example, we compute how long users dwell on a page (TimeOnPage) or domain (TimeOnDomain), and the deviation of dwell time from expected page dwell time for a query. These features allows us to model intra-query diversity of page browsing behavior (e.g., navigational queries, on average, are likely to have shorter page dwell time than transactional or informational queries). We include both the direct features and the derived features described above. Clickthrough features: Clicks are a special case of user interaction with the search engine. We include all the features necessary to learn the clickthrough-based strategies described in Sections 4.1 and 4.4. For example, for a query-URL pair we provide the number of clicks for the result (ClickFrequency), as 1 Of course, this is just a rough estimate, as the observed background distribution also includes the relevance component. well as whether there was a click on result below or above the current URL (IsClickBelow, IsClickAbove). The derived feature values such as ClickRelativeFrequency and ClickDeviation are computed as described in Equation 1. Query-text features TitleOverlap Fraction of shared words between query and title SummaryOverlap Fraction of shared words between query and summary QueryURLOverlap Fraction of shared words between query and URL QueryDomainOverlap Fraction of shared words between query and domain QueryLength Number of tokens in query QueryNextOverlap Average fraction of words shared with next query Browsing features TimeOnPage Page dwell time CumulativeTimeOnPage Cumulative time for all subsequent pages after search TimeOnDomain Cumulative dwell time for this domain TimeOnShortUrl Cumulative time on URL prefix, dropping parameters IsFollowedLink 1 if followed link to result, 0 otherwise IsExactUrlMatch 0 if aggressive normalization used, 1 otherwise IsRedirected 1 if initial URL same as final URL, 0 otherwise IsPathFromSearch 1 if only followed links after query, 0 otherwise ClicksFromSearch Number of hops to reach page from query AverageDwellTime Average time on page for this query DwellTimeDeviation Deviation from overall average dwell time on page CumulativeDeviation Deviation from average cumulative time on page DomainDeviation Deviation from average time on domain ShortURLDeviation Deviation from average time on short URL Clickthrough features Position Position of the URL in Current ranking ClickFrequency Number of clicks for this query, URL pair ClickRelativeFrequency Relative frequency of a click for this query and URL ClickDeviation Deviation from expected click frequency IsNextClicked 1 if there is a click on next position, 0 otherwise IsPreviousClicked 1 if there is a click on previous position, 0 otherwise IsClickAbove 1 if there is a click above, 0 otherwise IsClickBelow 1 if there is click below, 0 otherwise Table 3.1: Features used to represent post-search interactions for a given query and search result URL 3.4 Learning a Predictive Behavior Model Having described our features, we now turn to the actual method of mapping the features to user preferences. We attempt to learn a general implicit feedback interpretation strategy automatically instead of relying on heuristics or insights. We consider this approach to be preferable to heuristic strategies, because we can always mine more data instead of relying (only) on our intuition and limited laboratory evidence. Our general approach is to train a classifier to induce weights for the user behavior features, and consequently derive a predictive model of user preferences. The training is done by comparing a wide range of implicit behavior measures with explicit user judgments for a set of queries. For this, we use a large random sample of queries in the search query log of a popular web search engine, the sets of results (identified by URLs) returned for each of the queries, and any explicit relevance judgments available for each query/result pair. We can then analyze the user behavior for all the instances where these queries were submitted to the search engine. To learn the mapping from features to relevance preferences, we use a scalable implementation of neural networks, RankNet [4], capable of learning to rank a set of given items. More specifically, for each judged query we check if a result link has been judged. If so, the label is assigned to the query/URL pair and to the corresponding feature vector for that search result. These vectors of feature values corresponding to URLs judged relevant or non-relevant by human annotators become our training set. RankNet has demonstrated excellent performance in learning to rank objects in a supervised setting, hence we use RankNet for our experiments. 4. PREDICTING USER PREFERENCES In our experiments, we explore several models for predicting user preferences. These models range from using no implicit user feedback to using all available implicit user feedback. Ranking search results to predict user preferences is a fundamental problem in information retrieval. Most traditional IR and web search approaches use a combination of page and link features to rank search results, and a representative state-ofthe-art ranking system will be used as our baseline ranker (Section 4.1). At the same time, user interactions with a search engine provide a wealth of information. A commonly considered type of interaction is user clicks on search results. Previous work [9], as described above, also examined which results were skipped (e.g., ‘skip above" and ‘skip next") and other related strategies to induce preference judgments from the users" skipping over results and not clicking on following results. We have also added refinements of these strategies to take into account the variability observed in realistic web scenarios.. We describe these strategies in Section 4.2. As clickthroughs are just one aspect of user interaction, we extend the relevance estimation by introducing a machine learning model that incorporates clicks as well as other aspects of user behavior, such as follow-up queries and page dwell time (Section 4.3). We conclude this section by briefly describing our baseline - a state-of-the-art ranking algorithm used by an operational web search engine. 4.1 Baseline Model A key question is whether browsing behavior can provide information absent from existing explicit judgments used to train an existing ranker. For our baseline system we use a state-of-theart page ranking system currently used by a major web search engine. Hence, we will call this system Current for the subsequent discussion. While the specific algorithms used by the search engine are beyond the scope of this paper, the algorithm ranks results based on hundreds of features such as query to document similarity, query to anchor text similarity, and intrinsic page quality. The Current web search engine rankings provide a strong system for comparison and experiments of the next two sections. 4.2 Clickthrough Model If we assume that every user click was motivated by a rational process that selected the most promising result summary, we can then interpret each click as described in Joachims et al.[10]. By studying eye tracking and comparing clicks with explicit judgments, they identified a few basic strategies. We discuss the two strategies that performed best in their experiments, Skip Above and Skip Next. Strategy SA (Skip Above): For a set of results for a query and a clicked result at position p, all unclicked results ranked above p are predicted to be less relevant than the result at p. In addition to information about results above the clicked result, we also have information about the result immediately following the clicked one. Eye tracking study performed by Joachims et al. [10] showed that users usually consider the result immediately following the clicked result in current ranking. Their Skip Next strategy uses this observation to predict that a result following the clicked result at p is less relevant than the clicked result, with accuracy comparable to the SA strategy above. For better coverage, we combine the SA strategy with this extension to derive the Skip Above + Skip Next strategy: Strategy SA+N (Skip Above + Skip Next): This strategy predicts all un-clicked results immediately following a clicked result as less relevant than the clicked result, and combines these predictions with those of the SA strategy above. We experimented with variations of these strategies, and found that SA+N outperformed both SA and the original Skip Next strategy, so we will consider the SA and SA+N strategies in the rest of the paper. These strategies are motivated and empirically tested for individual users in a laboratory setting. As we will show, these strategies do not work as well in real web search setting due to inherent inconsistency and noisiness of individual users" behavior. The general approach for using our clickthrough models directly is to filter clicks to those that reflect higher-than-chance click frequency. We then use the same SA and SA+N strategies, but only for clicks that have higher-than-expected frequency according to our model. For this, we estimate the relevance component rel(q,r,f) of the observed clickthrough feature f as the deviation from the expected (background) clickthrough distribution )( fC . Strategy CD (deviation d): For a given query, compute the observed click frequency distribution o(r, p) for all results r in positions p. The click deviation for a result r in position p, dev(r, p) is computed as: )(),(),( pCproprdev −= where C(p) is the expected clickthrough at position p. If dev(r,p)>d, retain the click as input to the SA+N strategy above, and apply SA+N strategy over the filtered set of click events. The choice of d selects the tradeoff between recall and precision. While the above strategy extends SA and SA+N, it still assumes that a (filtered) clicked result is preferred over all unclicked results presented to the user above a clicked position. However, for informational queries, multiple results may be clicked, with varying frequency. Hence, it is preferable to individually compare results for a query by considering the difference between the estimated relevance components of the click distribution of the corresponding query results. We now define a generalization of the previous clickthrough interpretation strategy: Strategy CDiff (margin m): Compute deviation dev(r,p) for each result r1...rn in position p. For each pair of results ri and rj, predict preference of ri over rj iff dev(ri,pi)-dev(ri,pj)>m. As in CD, the choice of m selects the tradeoff between recall and precision. The pairs may be preferred in the original order or in reverse of it. Given the margin, two results might be effectively indistinguishable, but only one can possibly be preferred over the other. Intuitively, CDiff generalizes the skip idea above to include cases where the user skipped (i.e., clicked less than expected) on uj and preferred (i.e., clicked more than expected) on ui. Furthermore, this strategy allows for differentiation within the set of clicked results, making it more appropriate to noisy user behavior. CDiff and CD are complimentary. CDiff is a generalization of the clickthrough frequency model of CD, but it ignores the positional information used in CD. Hence, combining the two strategies to improve coverage is a natural approach: Strategy CD+CDiff (deviation d, margin m): Union of CD and CDiff predictions. Other variations of the above strategies were considered, but these five methods cover the range of observed performance. 4.3 General User Behavior Model The strategies described in the previous section generate orderings based solely on observed clickthrough frequencies. As we discussed, clickthrough is just one, albeit important, aspect of user interactions with web search engine results. We now present our general strategy that relies on the automatically derived predictive user behavior models (Section 3). The UserBehavior Strategy: For a given query, each result is represented with the features in Table 3.1. Relative user preferences are then estimated using the learned user behavior model described in Section 3.4. Recall that to learn a predictive behavior model we used the features from Table 3.1 along with explicit relevance judgments as input to RankNet which learns an optimal weighting of features to predict preferences. This strategy models user interaction with the search engine, allowing it to benefit from the wisdom of crowds interacting with the results and the pages beyond. As our experiments in the subsequent sections demonstrate, modeling a richer set of user interactions beyond clickthroughs results in more accurate predictions of user preferences. 5. EXPERIMENTAL SETUP We now describe our experimental setup. We first describe the methodology used, including our evaluation metrics (Section 5.1). Then we describe the datasets (Section 5.2) and the methods we compared in this study (Section 5.3). 5.1 Evaluation Methodology and Metrics Our evaluation focuses on the pairwise agreement between preferences for results. This allows us to compare to previous work [9,10]. Furthermore, for many applications such as tuning ranking functions, pairwise preference can be used directly for training [1,4,9]. The evaluation is based on comparing preferences predicted by various models to the correct preferences derived from the explicit user relevance judgments. We discuss other applications of our models beyond web search ranking in Section 7. To create our set of test pairs we take each query and compute the cross-product between all search results, returning preferences for pairs according to the order of the associated relevance labels. To avoid ambiguity in evaluation, we discard all ties (i.e., pairs with equal label). In order to compute the accuracy of our preference predictions with respect to the correct preferences, we adapt the standard Recall and Precision measures [20]. While our task of computing pairwise agreement is different from the absolute relevance ranking task, the metrics are used in the similar way. Specifically, we report the average query recall and precision. For our task, Query Precision and Query Recall for a query q are defined as: • Query Precision: Fraction of predicted preferences for results for q that agree with preferences obtained from explicit human judgment. • Query Recall: Fraction of preferences obtained from explicit human judgment for q that were correctly predicted. The overall Recall and Precision are computed as the average of Query Recall and Query Precision, respectively. A drawback of this evaluation measure is that some preferences may be more valuable than others, which pairwise agreement does not capture. We discuss this issue further when we consider extensions to the current work in Section 7. 5.2 Datasets For evaluation we used 3,500 queries that were randomly sampled from query logs(for a major web search engine. For each query the top 10 returned search results were manually rated on a 6-point scale by trained judges as part of ongoing relevance improvement effort. In addition for these queries we also had user interaction data for more than 120,000 instances of these queries. The user interactions were harvested from anonymous browsing traces that immediately followed a query submitted to the web search engine. This data collection was part of voluntary opt-in feedback submitted by users from October 11 through October 31. These three weeks (21 days) of user interaction data was filtered to include only the users in the English-U.S. market. In order to better understand the effect of the amount of user interaction data available for a query on accuracy, we created subsets of our data (Q1, Q10, and Q20) that contain different amounts of interaction data: • Q1: Human-rated queries with at least 1 click on results recorded (3500 queries, 28,093 query-URL pairs) • Q10: Queries in Q1 with at least 10 clicks (1300 queries, 18,728 query-URL pairs). • Q20: Queries in Q1 with at least 20 clicks (1000 queries total, 12,922 query-URL pairs). These datasets were collected as part of normal user experience and hence have different characteristics than previously reported datasets collected in laboratory settings. Furthermore, the data size is order of magnitude larger than any study reported in the literature. 5.3 Methods Compared We considered a number of methods for comparison. We compared our UserBehavior model (Section 4.3) to previously published implicit feedback interpretation techniques and some variants of these approaches (Section 4.2), and to the current search engine ranking based on query and page features alone (Section 4.1). Specifically, we compare the following strategies: • SA: The Skip Above clickthrough strategy (Section 4.2) • SA+N: A more comprehensive extension of SA that takes better advantage of current search engine ranking. • CD: Our refinement of SA+N that takes advantage of our mixture model of clickthrough distribution to select trusted clicks for interpretation (Section 4.2). • CDiff: Our generalization of the CD strategy that explicitly uses the relevance component of clickthrough probabilities to induce preferences between search results (Section 4.2). • CD+CDiff: The strategy combining CD and CDiff as the union of predicted preferences from both (Section 4.2). • UserBehavior: We order predictions based on decreasing highest score of any page. In our preliminary experiments we observed that higher ranker scores indicate higher confidence in the predictions. This heuristic allows us to do graceful recall-precision tradeoff using the score of the highest ranked result to threshold the queries (Section 4.3) • Current: Current search engine ranking (section 4.1). Note that the Current ranker implementation was trained over a superset of the rated query/URL pairs in our datasets, but using the same truth labels as we do for our evaluation. Training/Test Split: The only strategy for which splitting the datasets into training and test was required was the UserBehavior method. To evaluate UserBehavior we train and validate on 75% of labeled queries, and test on the remaining 25%. The sampling was done per query (i.e., all results for a chosen query were included in the respective dataset, and there was no overlap in queries between training and test sets). It is worth noting that both the ad-hoc SA and SA+N, as well as the distribution-based strategies (CD, CDiff, and CD+CDiff), do not require a separate training and test set, since they are based on heuristics for detecting anomalous click frequencies for results. Hence, all strategies except for UserBehavior were tested on the full set of queries and associated relevance preferences, while UserBehavior was tested on a randomly chosen hold-out subset of the queries as described above. To make sure we are not favoring UserBehavior, we also tested all other strategies on the same hold-out test sets, resulting in the same accuracy results as testing over the complete datasets. 6. RESULTS We now turn to experimental evaluation of predicting relevance preference of web search results. Figure 6.1 shows the recall-precision results over the Q1 query set (Section 5.2). The results indicate that previous click interpretation strategies, SA and SA+N perform suboptimally in this setting, exhibiting precision 0.627 and 0.638 respectively. Furthermore, there is no mechanism to do recall-precision trade-off with SA and SA+N, as they do not provide prediction confidence. In contrast, our clickthrough distribution-based techniques CD and CD+CDiff exhibit somewhat higher precision than SA and SA+N (0.648 and 0.717 at Recall of 0.08, maximum achieved by SA or SA+N). SA+N SA 0.6 0.62 0.64 0.66 0.68 0.7 0.72 0.74 0.76 0.78 0.8 0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 Recall Precision SA SA+N CD CDiff CD+CDiff UserBehavior Current Figure 6.1: Precision vs. Recall of SA, SA+N, CD, CDiff, CD+CDiff, UserBehavior, and Current relevance prediction methods over the Q1 dataset. Interestingly, CDiff alone exhibits precision equal to SA (0.627) at the same recall at 0.08. In contrast, by combining CD and CDiff strategies (CD+CDiff method) we achieve the best performance of all clickthrough-based strategies, exhibiting precision of above 0.66 for recall values up to 0.14, and higher at lower recall levels. Clearly, aggregating and intelligently interpreting clickthroughs, results in significant gain for realistic web search, than previously described strategies. However, even the CD+CDiff clickthrough interpretation strategy can be improved upon by automatically learning to interpret the aggregated clickthrough evidence. But first, we consider the best performing strategy, UserBehavior. Incorporating post-search navigation history in addition to clickthroughs (Browsing features) results in the highest recall and precision among all methods compared. Browse exhibits precision of above 0.7 at recall of 0.16, significantly outperforming our Baseline and clickthrough-only strategies. Furthermore, Browse is able to achieve high recall (as high as 0.43) while maintaining precision (0.67) significantly higher than the baseline ranking. To further analyze the value of different dimensions of implicit feedback modeled by the UserBehavior strategy, we consider each group of features in isolation. Figure 6.2 reports Precision vs. Recall for each feature group. Interestingly, Query-text alone has low accuracy (only marginally better than Random). Furthermore, Browsing features alone have higher precision (with lower maximum recall achieved) than considering all of the features in our UserBehavior model. Applying different machine learning methods for combining classifier predictions may increase performance of using all features for all recall values. 0.5 0.55 0.6 0.65 0.7 0.75 0.8 0.01 0.05 0.09 0.13 0.17 0.21 0.25 0.29 0.33 0.37 0.41 0.45 Recall Precision All Features Clickthrough Query-text Browsing Figure 6.2: Precision vs. recall for predicting relevance with each group of features individually. 0.65 0.67 0.69 0.71 0.73 0.75 0.77 0.79 0.81 0.83 0.85 0.01 0.05 0.09 0.13 0.17 0.21 0.25 0.29 0.33 0.37 0.41 0.45 0.49 Recall Precision CD+CDiff:Q1 UserBehavior:Q1 CD+CDiff:Q10 UserBehavior:Q10 CD+CDiff:Q20 UserBehavior:Q20 Figure 6.3: Recall vs. Precision of CD+CDiff and UserBehavior for query sets Q1, Q10, and Q20 (queries with at least 1, at least 10, and at least 20 clicks respectively). Interestingly, the ranker trained over Clickthrough-only features achieves substantially higher recall and precision than human-designed clickthrough-interpretation strategies described earlier. For example, the clickthrough-trained classifier achieves 0.67 precision at 0.42 Recall vs. the maximum recall of 0.14 achieved by the CD+CDiff strategy. Our clickthrough and user behavior interpretation strategies rely on extensive user interaction data. We consider the effects of having sufficient interaction data available for a query before proposing a re-ranking of results for that query. Figure 6.3 reports recall-precision curves for the CD+CDiff and UserBehavior methods for different test query sets with at least 1 click (Q1), 10 clicks (Q10) and 20 clicks (Q20) available per query. Not surprisingly, CD+CDiff improves with more clicks. This indicates that accuracy will improve as more user interaction histories become available, and more queries from the Q1 set will have comprehensive interaction histories. Similarly, the UserBehavior strategy performs better for queries with 10 and 20 clicks, although the improvement is less dramatic than for CD+CDiff. For queries with sufficient clicks, CD+CDiff exhibits precision comparable with Browse at lower recall. 0 0.05 0.1 0.15 0.2 7 12 17 21 Days of user interaction data harvested Recall CD+CDiff UserBehavior Figure 6.4: Recall of CD+CDiff and UserBehavior strategies at fixed minimum precision 0.7 for varying amounts of user activity data (7, 12, 17, 21 days). Our techniques often do not make relevance predictions for search results (i.e., if no interaction data is available for the lower-ranked results), consequently maintaining higher precision at the expense of recall. In contrast, the current search engine always makes a prediction for every result for a given query. As a consequence, the recall of Current is high (0.627) at the expense of lower precision As another dimension of acquiring training data we consider the learning curve with respect to amount (days) of training data available. Figure 6.4 reports the Recall of CD+CDiff and UserBehavior strategies for varying amounts of training data collected over time. We fixed minimum precision for both strategies at 0.7 as a point substantially higher than the baseline (0.625). As expected, Recall of both strategies improves quickly with more days of interaction data examined. We now briefly summarize our experimental results. We showed that by intelligently aggregating user clickthroughs across queries and users, we can achieve higher accuracy on predicting user preferences. Because of the skewed distribution of user clicks our clickthrough-only strategies have high precision, but low recall (i.e., do not attempt to predict relevance of many search results). Nevertheless, our CD+CDiff clickthrough strategy outperforms most recent state-of-the-art results by a large margin (0.72 precision for CD+CDiff vs. 0.64 for SA+N) at the highest recall level of SA+N. Furthermore, by considering the comprehensive UserBehavior features that model user interactions after the search and beyond the initial click, we can achieve substantially higher precision and recall than considering clickthrough alone. Our UserBehavior strategy achieves recall of over 0.43 with precision of over 0.67 (with much higher precision at lower recall levels), substantially outperforms the current search engine preference ranking and all other implicit feedback interpretation methods. 7. CONCLUSIONS AND FUTURE WORK Our paper is the first, to our knowledge, to interpret postsearch user behavior to estimate user preferences in a real web search setting. We showed that our robust models result in higher prediction accuracy than previously published techniques. We introduced new, robust, probabilistic techniques for interpreting clickthrough evidence by aggregating across users and queries. Our methods result in clickthrough interpretation substantially more accurate than previously published results not specifically designed for web search scenarios. Our methods" predictions of relevance preferences are substantially more accurate than the current state-of-the-art search result ranking that does not consider user interactions. We also presented a general model for interpreting post-search user behavior that incorporates clickthrough, browsing, and query features. By considering the complete search experience after the initial query and click, we demonstrated prediction accuracy far exceeding that of interpreting only the limited clickthrough information. Furthermore, we showed that automatically learning to interpret user behavior results in substantially better performance than the human-designed ad-hoc clickthrough interpretation strategies. Another benefit of automatically learning to interpret user behavior is that such methods can adapt to changing conditions and changing user profiles. For example, the user behavior model on intranet search may be different from the web search behavior. Our general UserBehavior method would be able to adapt to these changes by automatically learning to map new behavior patterns to explicit relevance ratings. A natural application of our preference prediction models is to improve web search ranking [1]. In addition, our work has many potential applications including click spam detection, search abuse detection, personalization, and domain-specific ranking. For example, our automatically derived behavior models could be trained on examples of search abuse or click spam behavior instead of relevance labels. Alternatively, our models could be used directly to detect anomalies in user behavior - either due to abuse or to operational problems with the search engine. While our techniques perform well on average, our assumptions about clickthrough distributions (and learning the user behavior models) may not hold equally well for all queries. For example, queries with divergent access patterns (e.g., for ambiguous queries with multiple meanings) may result in behavior inconsistent with the model learned for all queries. Hence, clustering queries and learning different predictive models for each query type is a promising research direction. Query distributions also change over time, and it would be productive to investigate how that affects the predictive ability of these models. Furthermore, some predicted preferences may be more valuable than others, and we plan to investigate different metrics to capture the utility of the predicted preferences. As we showed in this paper, using the wisdom of crowds can give us accurate interpretation of user interactions even in the inherently noisy web search setting. Our techniques allow us to automatically predict relevance preferences for web search results with accuracy greater than the previously published methods. The predicted relevance preferences can be used for automatic relevance evaluation and tuning, for deploying search in new settings, and ultimately for improving the overall web search experience. 8. REFERENCES [1] E. Agichtein, E. Brill, and S. Dumais, Improving Web Search Ranking by Incorporating User Behavior, in Proceedings of the ACM Conference on Research and Development on Information Retrieval (SIGIR), 2006 [2] J. Allan. HARD Track Overview in TREC 2003: High Accuracy Retrieval from Documents. In Proceedings of TREC 2003, 24-37, 2004. [3] S. Brin and L. Page, The Anatomy of a Large-scale Hypertextual Web Search Engine,. In Proceedings of WWW7, 107-117, 1998. [4] C.J.C. Burges, T. Shaked, E. Renshaw, A. Lazier, M. Deeds, N. Hamilton, and G. Hullender, Learning to Rank using Gradient Descent, in Proceedings of the International Conference on Machine Learning (ICML), 2005 [5] D.M. Chickering, The WinMine Toolkit, Microsoft Technical Report MSR-TR-2002-103, 2002 [6] M. Claypool, D. Brown, P. Lee and M. Waseda. Inferring user interest, in IEEE Internet Computing. 2001 [7] S. Fox, K. Karnawat, M. Mydland, S. T. Dumais and T. White. Evaluating implicit measures to improve the search experience. In ACM Transactions on Information Systems, 2005 [8] J. Goecks and J. Shavlick. Learning users" interests by unobtrusively observing their normal behavior. In Proceedings of the IJCAI Workshop on Machine Learning for Information Filtering. 1999. [9] T. Joachims, Optimizing Search Engines Using Clickthrough Data, in Proceedings of the ACM Conference on Knowledge Discovery and Datamining (SIGKDD), 2002 [10] T. Joachims, L. Granka, B. Pang, H. Hembrooke and G. Gay, Accurately Interpreting Clickthrough Data as Implicit Feedback, in Proceedings of the ACM Conference on Research and Development on Information Retrieval (SIGIR), 2005 [11] T. Joachims, Making Large-Scale SVM Learning Practical. Advances in Kernel Methods, in Support Vector Learning, MIT Press, 1999 [12] D. Kelly and J. Teevan, Implicit feedback for inferring user preference: A bibliography. In SIGIR Forum, 2003 [13] J. Konstan, B. Miller, D. Maltz, J. Herlocker, L. Gordon and J. Riedl. GroupLens: Applying collaborative filtering to usenet news. In Communications of ACM, 1997. [14] M. Morita, and Y. Shinoda, Information filtering based on user behavior analysis and best match text retrieval. In Proceedings of the ACM Conference on Research and Development on Information Retrieval (SIGIR), 1994 [15] D. Oard and J. Kim. Implicit feedback for recommender systems. in Proceedings of AAAI Workshop on Recommender Systems. 1998 [16] D. Oard and J. Kim. Modeling information content using observable behavior. In Proceedings of the 64th Annual Meeting of the American Society for Information Science and Technology. 2001 [17] P. Pirolli, The Use of Proximal Information Scent to Forage for Distal Content on the World Wide Web. In Working with Technology in Mind: Brunswikian. Resources for Cognitive Science and Engineering, Oxford University Press, 2004 [18] F. Radlinski and T. Joachims, Query Chains: Learning to Rank from Implicit Feedback, in Proceedings of the ACM Conference on Knowledge Discovery and Data Mining (KDD), ACM, 2005 [19] F. Radlinski and T. Joachims, Evaluating the Robustness of Learning from Implicit Feedback, in the ICML Workshop on Learning in Web Search, 2005 [20] G. Salton and M. McGill. Introduction to modern information retrieval. McGraw-Hill, 1983 [21] E.M. Voorhees, D. Harman, Overview of TREC, 2001

precision measure;relevance measurement;recall measure;predictive model;induce weight;clickthrough;top relevant document position;search abuse detection;predictive behavior model;low recall;information retrieval;predict relevance preference;position of top relevant document;explicit relevance judgment;user preference;interpret implicit relevance feedback;follow-up query;web search ranking;click spam detection;domain-specific ranking;page dwell time;implicit feedback;user behavior model;personalization

train_H-87

Robustness of Adaptive Filtering Methods In a Cross-benchmark Evaluation

This paper reports a cross-benchmark evaluation of regularized logistic regression (LR) and incremental Rocchio for adaptive filtering. Using four corpora from the Topic Detection and Tracking (TDT) forum and the Text Retrieval Conferences (TREC) we evaluated these methods with non-stationary topics at various granularity levels, and measured performance with different utility settings. We found that LR performs strongly and robustly in optimizing T11SU (a TREC utility function) while Rocchio is better for optimizing Ctrk (the TDT tracking cost), a high-recall oriented objective function. Using systematic cross-corpus parameter optimization with both methods, we obtained the best results ever reported on TDT5, TREC10 and TREC11. Relevance feedback on a small portion (0.05~0.2%) of the TDT5 test documents yielded significant performance improvements, measuring up to a 54% reduction in Ctrk and a 20.9% increase in T11SU (with β=0.1), compared to the results of the top-performing system in TDT2004 without relevance feedback information.

1. INTRODUCTION Adaptive filtering (AF) has been a challenging research topic in information retrieval. The task is for the system to make an online topic membership decision (yes or no) for every document, as soon as it arrives, with respect to each pre-defined topic of interest. Starting from 1997 in the Topic Detection and Tracking (TDT) area and 1998 in the Text Retrieval Conferences (TREC), benchmark evaluations have been conducted by NIST under the following conditions[6][7][8][3][4]: • A very small number (1 to 4) of positive training examples was provided for each topic at the starting point. • Relevance feedback was available but only for the systemaccepted documents (with a yes decision) in the TREC evaluations for AF. • Relevance feedback (RF) was not allowed in the TDT evaluations for AF (or topic tracking in the TDT terminology) until 2004. • TDT2004 was the first time that TREC and TDT metrics were jointly used in evaluating AF methods on the same benchmark (the TDT5 corpus) where non-stationary topics dominate. The above conditions attempt to mimic realistic situations where an AF system would be used. That is, the user would be willing to provide a few positive examples for each topic of interest at the start, and might or might not be able to provide additional labeling on a small portion of incoming documents through relevance feedback. Furthermore, topics of interest might change over time, with new topics appearing and growing, and old topics shrinking and diminishing. These conditions make adaptive filtering a difficult task in statistical learning (online classification), for the following reasons: 1) it is difficult to learn accurate models for prediction based on extremely sparse training data; 2) it is not obvious how to correct the sampling bias (i.e., relevance feedback on system-accepted documents only) during the adaptation process; 3) it is not well understood how to effectively tune parameters in AF methods using cross-corpus validation where the validation and evaluation topics do not overlap, and the documents may be from different sources or different epochs. None of these problems is addressed in the literature of statistical learning for batch classification where all the training data are given at once. The first two problems have been studied in the adaptive filtering literature, including topic profile adaptation using incremental Rocchio, Gaussian-Exponential density models, logistic regression in a Bayesian framework, etc., and threshold optimization strategies using probabilistic calibration or local fitting techniques [1][2][9][10][11][12][13]. Although these works provide valuable insights for understanding the problems and possible solutions, it is difficult to draw conclusions regarding the effectiveness and robustness of current methods because the third problem has not been thoroughly investigated. Addressing the third issue is the main focus in this paper. We argue that robustness is an important measure for evaluating and comparing AF methods. By robust we mean consistent and strong performance across benchmark corpora with a systematic method for parameter tuning across multiple corpora. Most AF methods have pre-specified parameters that may influence the performance significantly and that must be determined before the test process starts. Available training examples, on the other hand, are often insufficient for tuning the parameters. In TDT5, for example, there is only one labeled training example per topic at the start; parameter optimization on such training data is doomed to be ineffective. This leaves only one option (assuming tuning on the test set is not an alternative), that is, choosing an external corpus as the validation set. Notice that the validation-set topics often do not overlap with the test-set topics, thus the parameter optimization is performed under the tough condition that the validation data and the test data may be quite different from each other. Now the important question is: which methods (if any) are robust under the condition of using cross-corpus validation to tune parameters? Current literature does not offer an answer because no thorough investigation on the robustness of AF methods has been reported. In this paper we address the above question by conducting a cross-benchmark evaluation with two effective approaches in AF: incremental Rocchio and regularized logistic regression (LR). Rocchio-style classifiers have been popular in AF, with good performance in benchmark evaluations (TREC and TDT) if appropriate parameters were used and if combined with an effective threshold calibration strategy [2][4][7][8][9][11][13]. Logistic regression is a classical method in statistical learning, and one of the best in batch-mode text categorization [15][14]. It was recently evaluated in adaptive filtering and was found to have relatively strong performance (Section 5.1). Furthermore, a recent paper [13] reported that the joint use of Rocchio and LR in a Bayesian framework outperformed the results of using each method alone on the TREC11 corpus. Stimulated by those findings, we decided to include Rocchio and LR in our crossbenchmark evaluation for robustness testing. Specifically, we focus on how much the performance of these methods depends on parameter tuning, what the most influential parameters are in these methods, how difficult (or how easy) to optimize these influential parameters using cross-corpus validation, how strong these methods perform on multiple benchmarks with the systematic tuning of parameters on other corpora, and how efficient these methods are in running AF on large benchmark corpora. The organization of the paper is as follows: Section 2 introduces the four benchmark corpora (TREC10 and TREC11, TDT3 and TDT5) used in this study. Section 3 analyzes the differences among the TREC and TDT metrics (utilities and tracking cost) and the potential implications of those differences. Section 4 outlines the Rocchio and LR approaches to AF, respectively. Section 5 reports the experiments and results. Section 6 concludes the main findings in this study. 2. BENCHMARK CORPORA We used four benchmark corpora in our study. Table 1 shows the statistics about these data sets. TREC10 was the evaluation benchmark for adaptive filtering in TREC 2001, consisting of roughly 806,791 Reuters news stories from August 1996 to August 1997 with 84 topic labels (subject categories)[7]. The first two weeks (August 20th to 31st , 1996) of documents is the training set, and the remaining 11 & ½ months (from September 1st , 1996 to August 19th , 1997) is the test set. TREC11 was the evaluation benchmark for adaptive filtering in TREC 2002, consisting of the same set of documents as those in TREC10 but with a slightly different splitting point for the training and test sets. The TREC11 topics (50) are quite different from those in TREC10; they are queries for retrieval with relevance judgments by NIST assessors [8]. TDT3 was the evaluation benchmark in the TDT2001 dry run1 . The tracking part of the corpus consists of 71,388 news stories from multiple sources in English and Mandarin (AP, NYT, CNN, ABC, NBC, MSNBC, Xinhua, Zaobao, Voice of America and PRI the World) in the period of October to December 1998. Machine-translated versions of the non-English stories (Xinhua, Zaobao and VOA Mandarin) are provided as well. The splitting point for training-test sets is different for each topic in TDT. TDT5 was the evaluation benchmark in TDT2004 [4]. The tracking part of the corpus consists of 407,459 news stories in the period of April to September, 2003 from 15 news agents or broadcast sources in English, Arabic and Mandarin, with machine-translated versions of the non-English stories. We only used the English versions of those documents in our experiments for this paper. The TDT topics differ from TREC topics both conceptually and statistically. Instead of generic, ever-lasting subject categories (as those in TREC), TDT topics are defined at a finer level of granularity, for events that happen at certain times and locations, and that are born and die, typically associated with a bursty distribution over chronologically ordered news stories. The average size of TDT topics (events) is two orders of magnitude smaller than that of the TREC10 topics. Figure 1 compares the document densities of a TREC topic (Civil Wars) and two TDT topics (Gunshot and APEC Summit Meeting, respectively) over a 3-month time period, where the area under each curve is normalized to one. The granularity differences among topics and the corresponding non-stationary distributions make the cross-benchmark evaluation interesting. For example, algorithms favoring large and stable topics may not work well for short-lasting and nonstationary topics, and vice versa. Cross-benchmark evaluations allow us to test this hypothesis and possibly identify the weaknesses in current approaches to adaptive filtering in tracking the drifting trends of topics. 1 http://www.ldc.upenn.edu/Projects/TDT2001/topics.html Table 1: Statistics of benchmark corpora for adaptive filtering evaluations N(tr) is the number of the initial training documents; N(ts) is the number of the test documents; n+ is the number of positive examples of a predefined topic; * is an average over all the topics. 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 Week P(topic|week) Gunshot (TDT5) APEC Summit Meeting (TDT3) Civil War(TREC10) Figure 1: The temporal nature of topics 3. METRICS To make our results comparable to the literature, we decided to use both TREC-conventional and TDT-conventional metrics in our evaluation. 3.1 TREC11 metrics Let A, B, C and D be, respectively, the numbers of true positives, false alarms, misses and true negatives for a specific topic, and DCBAN +++= be the total number of test documents. The TREC-conventional metrics are defined as: Precision )/( BAA += , Recall )/( CAA += )(2 )21( CABA A F +++ + = β β β ( ) η ηηβ ηβ − −+− = 1 ),/()(max 11 , CABA SUT where parameters β and η were set to 0.5 and -0.5 respectively in TREC10 (2001) and TREC11 (2002). For evaluating the performance of a system, the performance scores are computed for individual topics first and then averaged over topics (macroaveraging). 3.2 TDT metrics The TDT-conventional metric for topic tracking is defined as: famisstrk PTPwPTPwTC ))(1()()( 21 −+= where P(T) is the percentage of documents on topic T, missP is the miss rate by the system on that topic, faP is the false alarm rate, and 1w and 2w are the costs (pre-specified constants) for a miss and a false alarm, respectively. The TDT benchmark evaluations (since 1997) have used the settings of 11 =w , 1.02 =w and 02.0)( =TP for all topics. For evaluating the performance of a system, Ctrk is computed for each topic first and then the resulting scores are averaged for a single measure (the topic-weighted Ctrk). To make the intuition behind this measure transparent, we substitute the terms in the definition of Ctrk as follows: N CA TP + =)( , N DB TP + =− )(1 , CA C Pmiss + = , DB B Pfa + = , )( 1 )( 21 21 BwCw N DB B N DB w CA C N CA wTCtrk +⋅= + ⋅ + ⋅+ + ⋅ + ⋅= Clearly, trkC is the average cost per error on topic T, with 1w and 2w controlling the penalty ratio for misses vs. false alarms. In addition to trkC , TDT2004 also employed 1.011 =βSUT as a utility metric. To distinguish this from the 5.011 =βSUT in TREC11, we call former TDT5SU in the rest of this paper. Corpus #Topics N(tr) N(ts) Avg n+ (tr) Avg n+ (ts) Max n+ (ts) Min n+ (ts) #Topics per doc (ts) TREC10 84 20,307 783,484 2 9795.3 39,448 38 1.57 TREC11 50 80.664 726,419 3 378.0 597 198 1.12 TDT3 53 18,738* 37,770* 4 79.3 520 1 1.06 TDT5 111 199,419* 207,991* 1 71.3 710 1 1.01 3.3 The correlations and the differences From an optimization point of view, TDT5SU and T11SU are both utility functions while Ctrk is a cost function. Our objective is to maximize the former or to minimize the latter on test documents. The differences and correlations among these objective functions can be analyzed through the shared counts of A, B, C and D in their definitions. For example, both TDT5SU and T11SU are positively correlated to the values of A and D, and negatively correlated to the values of B and C; the only difference between them is in their penalty ratios for misses vs. false alarms, i.e., 10:1 in TDT5SU and 2:1 in T11SU. The Ctrk function, on the other hand, is positively correlated to the values of C and B, and negatively correlated to the values of A and D; hence, it is negatively correlated to T11SU and TDT5SU. More importantly, there is a subtle and major difference between Ctrk and the utility functions: T11SU and TDT5SU. That is, Ctrk has a very different penalty ratio for misses vs. false alarms: it favors recall-oriented systems to an extreme. At first glance, one would think that the penalty ratio in Ctrk is 10:1 since 11 =w and 1.02 =w . However, this is not true if 02.0)( =TP is an inaccurate estimate of the on-topic documents on average for the test corpus. Using TDT3 as an example, the true percentage is: 002.0 37770 3.79 )( ≈= + = N n TP where N is the average size of the test sets in TDT3, and n+ is the average number of positive examples per topic in the test sets. Using 02.0)(ˆ =TP as an (inaccurate) estimate of 0.002 enlarges the intended penalty ratio of 10:1 to 100:1, roughly speaking. To wit: )1.010( 1 1.010 )3.7937770(37770 3.79 1011.0101 1011.0101 ))(1(2)(1 )02.01(202.01)( BC NN B N C B N C DB B N CA N C faPTPwmissPTPw faPwmissPwTtrkC ×+×=×+×≈ − ×−×+××= + + ×−×+××= ×−⋅+××= −×+××= ⎟ ⎠ ⎞ ⎜ ⎝ ⎛ ⎟ ⎠ ⎞ ⎜ ⎝ ⎛ ρρ where 10 002.0 02.0 )( )(ˆ === TP TP ρ is the factor of enlargement in the estimation of P(T) compared to the truth. Comparing the above result to formula 2, we can see the actual penalty ratio for misses vs. false alarms was 100:1 in the evaluations on TDT3 using Ctrk. Similarly, we can compute the enlargement factor for TDT5 using the statistics in Table 1 as follows: 3.58 991,207/3.71 02.0 )( )(ˆ === TP TP ρ which means the actual penalty ratio for misses vs. false alarms in the evaluation on TDT5 using Ctrk was approximately 583:1. The implications of the above analysis are rather significant: • Ctrk defined in the same formula does not necessarily mean the same objective function in evaluation; instead, the optimization criterion depends on the test corpus. • Systems optimized for Ctrk would not optimize TDT5SU (and T11SU) because the former favors high-recall oriented to an extreme while the latter does not. • Parameters tuned on one corpus (e.g., TDT3) might not work for an evaluation on another corpus (say, TDT5) unless we account for the previously-unknown subtle dependency of Ctrk on data. • Results in Ctrk in the past years of TDT evaluations may not be directly comparable to each other because the evaluation collections changed most years and hence the penalty ratio in Ctrk varied. Although these problems with Ctrk were not originally anticipated, it offered an opportunity to examine the ability of systems in trading off precision for extreme recall. This was a challenging part of the TDT2004 evaluation for AF. Comparing the metrics in TDT and TREC from a utility or cost optimization point of view is important for understanding the evaluation results of adaptive filtering methods. This is the first time this issue is explicitly analyzed, to our knowledge. 4. METHODS 4.1 Incremental Rocchio for AF We employed a common version of Rocchio-style classifiers which computes a prototype vector per topic (T) as follows: |)(| ' |)(| )()( )(')( TD d TD d TqTp TDdTDd − ∈ + ∈ ∑∑ −+ −+= rr rr rr γβα The first term on the RHS is the weighted vector representation of topic description whose elements are terms weights. The second term is the weighted centroid of the set )(TD+ of positive training examples, each of which is a vector of withindocument term weights. The third term is the weighted centroid of the set )(TD− of negative training examples which are the nearest neighbors of the positive centroid. The three terms are given pre-specified weights of βα, and γ , controlling the relative influence of these components in the prototype. The prototype of a topic is updated each time the system makes a yes decision on a new document for that topic. If relevance feedback is available (as is the case in TREC adaptive filtering), the new document is added to the pool of either )(TD+ or )(TD− , and the prototype is recomputed accordingly; if relevance feedback is not available (as is the case in TDT event tracking), the system"s prediction (yes) is treated as the truth, and the new document is added to )(TD+ for updating the prototype. Both cases are part of our experiments in this paper (and part of the TDT 2004 evaluations for AF). To distinguish the two, we call the first case simply Rocchio and the second case PRF Rocchio where PRF stands for pseudorelevance feedback. The predictions on a new document are made by computing the cosine similarity between each topic prototype and the document vector, and then comparing the resulting scores against a threshold: ⎩ ⎨ ⎧ − + =− )( )( ))),((cos( no yes dTpsign new θ rr Threshold calibration in incremental Rocchio is a challenging research topic. Multiple approaches have been developed. The simplest is to use a universal threshold for all topics, tuned on a validation set and fixed during the testing phase. More elaborate methods include probabilistic threshold calibration which converts the non-probabilistic similarity scores to probabilities (i.e., )|( dTP r ) for utility optimization [9][13], and margin-based local regression for risk reduction [11]. It is beyond the scope of this paper to compare all the different ways to adapt Rocchio-style methods for AF. Instead, our focus here is to investigate the robustness of Rocchio-style methods in terms of how much their performance depends on elaborate system tuning, and how difficult (or how easy) it is to get good performance through cross-corpus parameter optimization. Hence, we decided to use a relatively simple version of Rocchio as the baseline, i.e., with a universal threshold tuned on a validation corpus and fixed for all topics in the testing phase. This simple version of Rocchio has been commonly used in the past TDT benchmark evaluations for topic tracking, and had strong performance in the TDT2004 evaluations for adaptive filtering with and without relevance feedback (Section 5.1). Results of more complex variants of Rocchio are also discussed when relevant. 4.2 Logistic Regression for AF Logistic regression (LR) estimates the posterior probability of a topic given a document using a sigmoid function )1/(1),|1( xw ewxyP rrrr ⋅− +== where x r is the document vector whose elements are term weights, w r is the vector of regression coefficients, and }1,1{ −+∈y is the output variable corresponding to yes or no with respect to a particular topic. Given a training set of labeled documents { }),(,),,( 11 nn yxyxD r L r = , the standard regression problem is defined as to find the maximum likelihood estimates of the regression coefficients (the model parameters): { } { } { }))exp(1(1logminarg )|(logmaxarg)|(maxarg ii xwyn i w wDP w wDP w mlw rr r r r r r r ⋅−+∑ == == This is a convex optimization problem which can be solved using a standard conjugate gradient algorithm in O(INF) time for training per topic, where I is the average number of iterations needed for convergence, and N and F are the number of training documents and number of features respectively [14]. Once the regression coefficients are optimized on the training data, the filtering prediction on each incoming document is made as: ( ) ⎩ ⎨ ⎧ − + =− )( )( ),|( no yes wxyPsign optnew θ rr Note that w r is constantly updated whenever a new relevance judgment is available in the testing phase of AF, while the optimal threshold optθ is constant, depending only on the predefined utility (or cost) function for evaluation. If T11SU is the metric, for example, with the penalty ratio of 2:1 for misses and false alarms (Section 3.1), the optimal threshold for LR is 33.0)12/(1 =+ for all topics. We modified the standard (above) version of LR to allow more flexible optimization criteria as follows: ⎭ ⎬ ⎫ ⎩ ⎨ ⎧ −++= ∑= ⋅− 2 1 )1log()(minarg μλ rrr rr r weysw n i xwy i w map ii where )( iys is taken to be α , β and γ for query, positive and negative documents respectively, which are similar to those in Rocchio, giving different weights to the three kinds of training examples: topic descriptions (queries), on-topic documents and off-topic documents. The second term in the objective function is for regularization, equivalent to adding a Gaussian prior to the regression coefficients with mean μ r and covariance variance matrix Ι⋅λ2/1 , where Ι is the identity matrix. Tuning λ (≥0) is theoretically justified for reducing model complexity (the effective degree of freedom) and avoiding over-fitting on training data [5]. How to find an effective μ r is an open issue for research, depending on the user"s belief about the parameter space and the optimal range. The solution of the modified objective function is called the Maximum A Posteriori (MAP) estimate, which reduces to the maximum likelihood solution for standard LR if 0=λ . 5. EVALUATIONS We report our empirical findings in four parts: the TDT2004 official evaluation results, the cross-corpus parameter optimization results, and the results corresponding to the amounts of relevance feedback. 5.1 TDT2004 benchmark results The TDT2004 evaluations for adaptive filtering were conducted by NIST in November 2004. Multiple research teams participated and multiple runs from each team were allowed. Ctrk and TDT5SU were used as the metrics. Figure 2 and Figure 3 show the results; the best run from each team was selected with respect to Ctrk or TDT5SU, respectively. Our Rocchio (with adaptive profiles but fixed universal threshold for all topics) had the best result in Ctrk, and our logistic regression had the best result in TDT5SU. All the parameters of our runs were tuned on the TDT3 corpus. Results for other sites are also listed anonymously for comparison. Ctrk Ours 0.0324 Site2 0.0467 Site3 0.1366 Site4 0.2438 Metric = Ctrk (the lower the better) 0.0324 0.0467 0.1366 0.2438 0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 Ours Site2 Site3 Site4 Figure 2: TDT2004 results in Ctrk of systems using true relevance feedback. (Ours is the Rocchio method.) We also put the 1st and 3rd quartiles as sticks for each site.2 T11SU Ours 0.7328 Site3 0.7281 Site2 0.6672 Site4 0.382 Metric = TDT5SU (the higher the better) 0.7328 0.7281 0.6672 0.382 0 0.2 0.4 0.6 0.8 1 Ours Site3 Site2 Site4 Figure 3:TDT2004 results in TDT5SU of systems using true relevance feedback. (Ours is LR with 0=μ r and 005.0=λ ). CTrk Ours 0.0707 Site2 0.1545 Site5 0.5669 Site4 0.6507 Site6 0.8973 Primary Topic Traking Results in TDT2004 0.0707 0.8973 0.6507 0.1545 0.5669 0 0.2 0.4 0.6 0.8 1 1.2 Ours Site2 Site5 Site4 Site6 Ctrk Figure 4:TDT2004 results in Ctrk of systems without using true relevance feedback. (Ours is PRF Rocchio.) Adaptive filtering without using true relevance feedback was also a part of the evaluations. In this case, systems had only one labeled training example per topic during the entire training and testing processes, although unlabeled test documents could be used as soon as predictions on them were made. Such a setting has been conventional for the Topic Tracking task in TDT until 2004. Figure 4 shows the summarized official submissions from each team. Our PRF Rocchio (with a fixed threshold for all the topics) had the best performance. 2 We use quartiles rather than standard deviations since the former is more resistant to outliers. 5.2 Cross-corpus parameter optimization How much the strong performance of our systems depends on parameter tuning is an important question. Both Rocchio and LR have parameters that must be prespecified before the AF process. The shared parameters include the sample weightsα , β and γ , the sample size of the negative training documents (i.e., )(TD− ), the term-weighting scheme, and the maximal number of non-zero elements in each document vector. The method-specific parameters include the decision threshold in Rocchio, and μ r , λ and MI (the maximum number of iterations in training) in LR. Given that we only have one labeled example per topic in the TDT5 training sets, it is impossible to effectively optimize these parameters on the training data, and we had to choose an external corpus for validation. Among the choices of TREC10, TREC11 and TDT3, we chose TDT3 (c.f. Section 2) because it is most similar to TDT5 in terms of the nature of the topics (Section 2). We optimized the parameters of our systems on TDT3, and fixed those parameters in the runs on TDT5 for our submissions to TDT2004. We also tested our methods on TREC10 and TREC11 for further analysis. Since exhaustive testing of all possible parameter settings is computationally intractable, we followed a step-wise forward chaining procedure instead: we pre-specified an order of the parameters in a method (Rocchio or LR), and then tuned one parameter at the time while fixing the settings of the remaining parameters. We repeated this procedure for several passes as time allowed. 0.05 0.26 0.67 0.69 0.00 0.10 0.20 0.30 0.40 0.50 0.60 0.70 0.80 0.90 0.02 0.03 0.04 0.06 0.08 0.1 0.15 0.2 0.25 0.3 Threshold TDT5SU TDT3 TDT5 TREC10 TREC11 Figure 5: Performance curves of adaptive Rocchio Figure 5 compares the performance curves in TDT5SU for Rocchio on TDT3, TDT5, TREC10 and TREC11 when the decision threshold varied. These curves peak at different locations: the TDT3-optimal is closest to the TDT5-optimal while the TREC10-optimal and TREC1-optimal are quite far away from the TDT5-optimal. If we were using TREC10 or TREC11 instead of TDT3 as the validation corpus for TDT5, or if the TDT3 corpus were not available, we would have difficulty in obtaining strong performance for Rocchio in TDT2004. The difficulty comes from the ad-hoc (non-probabilistic) scores generated by the Rocchio method: the distribution of the scores depends on the corpus, making cross-corpus threshold optimization a tricky problem. Logistic regression has less difficulty with respect to threshold tuning because it produces probabilistic scores of )|1Pr( xy = upon which the optimal threshold can be directly computed if probability estimation is accurate. Given the penalty ratio for misses vs. false alarms as 2:1 in T11SU, 10:1 in TDT5SU and 583:1 in Ctrk (Section 3.3), the corresponding optimal thresholds (t) are 0.33, 0.091 and 0.0017 respectively. Although the theoretical threshold could be inaccurate, it still suggests the range of near-optimal settings. With these threshold settings in our experiments for LR, we focused on the crosscorpus validation of the Bayesian prior parameters, that is, μ r and λ. Table 2 summarizes the results 3 . We measured the performance of the runs on TREC10 and TREC11 using T11SU, and the performance of the runs on TDT3 and TDT5 using TDT5SU. For comparison we also include the best results of Rocchio-based methods on these corpora, which are our own results of Rocchio on TDT3 and TDT5, and the best results reported by NIST for TREC10 and TREC11. From this set of results, we see that LR significantly outperformed Rocchio on all the corpora, even in the runs of standard LR without any tuning, i.e. λ=0. This empirical finding is consistent with a previous report [13] for LR on TREC11 although our results of LR (0.585~0.608 in T11SU) are stronger than the results (0.49 for standard LR and 0.54 for LR using Rocchio prototype as the prior) in that report. More importantly, our cross-benchmark evaluation gives strong evidence for the robustness of LR. The robustness, we believe, comes from the probabilistic nature of the system-generated scores. That is, compared to the ad-hoc scores in Rocchio, the normalized posterior probabilities make the threshold optimization in LR a much easier problem. Moreover, logistic regression is known to converge towards the Bayes classifier asymptotically while Rocchio classifiers" parameters do not. Another interesting observation in these results is that the performance of LR did not improve when using a Rocchio prototype as the mean in the prior; instead, the performance decreased in some cases. This observation does not support the previous report by [13], but we are not surprised because we are not convinced that Rocchio prototypes are more accurate than LR models for topics in the early stage of the AF process, and we believe that using a Rocchio prototype as the mean in the Gaussian prior would introduce undesirable bias to LR. We also believe that variance reduction (in the testing phase) should be controlled by the choice of λ (but not μ r ), for which we conducted the experiments as shown in Figure 6. Table 2: Results of LR with different Bayesian priors Corpus TDT3 TDT5 TREC10 TREC11 LR(μ=0,λ=0) 0.7562 0.7737 0.585 0.5715 LR(μ=0,λ=0.01) 0.8384 0.7812 0.6077 0.5747 LR(μ=roc*,λ=0.01) 0.8138 0.7811 0.5803 0.5698 Best Rocchio 0.6628 0.6917 0.4964 0.475 3 The LR results (0.77~0.78) on TDT5 in this table are better than our TDT2004 official result (0.73) because parameter optimization has been improved afterwards. 4 The TREC10-best result (0.496 by Oracle) is only available in T10U which is not directly comparable to the scores in T11SU, just indicative. *: μ r was set to the Rocchio prototype 0 0.2 0.4 0.6 0.8 0.000 0.001 0.005 0.050 0.500 Lambda Performance Ctrk on TDT3 TDT5SU on TDT3 TDT5SU on TDT5 T11SU on TREC11 Figure 6: LR with varying lambda. The performance of LR is summarized with respect to λ tuning on the corpora of TREC10, TREC11 and TDT3. The performance on each corpus was measured using the corresponding metrics, that is, T11SU for the runs on TREC10 and TREC11, and TDT5SU and Ctrk for the runs on TDT3,. In the case of maximizing the utilities, the safe interval for λ is between 0 and 0.01, meaning that the performance of regularized LR is stable, the same as or improved slightly over the performance of standard LR. In the case of minimizing Ctrk, the safe range for λ is between 0 and 0.1, and setting λ between 0.005 and 0.05 yielded relatively large improvements over the performance of standard LR because training a model for extremely high recall is statistically more tricky, and hence more regularization is needed. In either case, tuning λ is relatively safe, and easy to do successfully by cross-corpus tuning. Another influential choice in our experiment settings is term weighting: we examined the choices of binary, TF and TF-IDF (the ltc version) schemes. We found TF-IDF most effective for both Rocchio and LR, and used this setting in all our experiments. 5.3 Percentages of labeled data How much relevance feedback (RF) would be needed during the AF process is a meaningful question in real-world applications. To answer it, we evaluated Rocchio and LR on TDT with the following settings: • Basic Rocchio, no adaptation at all • PRF Rocchio, updating topic profiles without using true relevance feedback; • Adaptive Rocchio, updating topic profiles using relevance feedback on system-accepted documents plus 10 documents randomly sampled from the pool of systemrejected documents; • LR with 0 rr =μ , 01.0=λ and threshold = 0.004; • All the parameters in Rocchio tuned on TDT3. Table 3 summarizes the results in Ctrk: Adaptive Rocchio with relevance feedback on 0.6% of the test documents reduced the tracking cost by 54% over the result of the PRF Rocchio, the best system in the TDT2004 evaluation for topic tracking without relevance feedback information. Incremental LR, on the other hand, was weaker but still impressive. Recall that Ctrk is an extremely high-recall oriented metric, causing frequent updating of profiles and hence an efficiency problem in LR. For this reason we set a higher threshold (0.004) instead of the theoretically optimal threshold (0.0017) in LR to avoid an untolerable computation cost. The computation time in machine-hours was 0.33 for the run of adaptive Rocchio and 14 for the run of LR on TDT5 when optimizing Ctrk. Table 4 summarizes the results in TDT5SU; adaptive LR was the winner in this case, with relevance feedback on 0.05% of the test documents improving the utility by 20.9% over the results of PRF Rocchio. Table 3: AF methods on TDT5 (Performance in Ctrk) Base Roc PRF Roc Adp Roc LR % of RF 0% 0% 0.6% 0.2% Ctrk 0.076 0.0707 0.0324 0.0382 ±% +7% (baseline) -54% -46% Table 4: AF methods on TDT5 (Performance in TDT5SU) Base Roc PRF Roc Adp Roc LR(λ=.01) % of RF 0% 0% 0.04% 0.05% TDT5SU 0.57 0.6452 0.69 0.78 ±% -11.7% (baseline) +6.9% +20.9% Evidently, both Rocchio and LR are highly effective in adaptive filtering, in terms of using of a small amount of labeled data to significantly improve the model accuracy in statistical learning, which is the main goal of AF. 5.4 Summary of Adaptation Process After we decided the parameter settings using validation, we perform the adaptive filtering in the following steps for each topic: 1) Train the LR/Rocchio model using the provided positive training examples and 30 randomly sampled negative examples; 2) For each document in the test corpus: we first make a prediction about relevance, and then get relevance feedback for those (predicted) positive documents. 3) Model and IDF statistics will be incrementally updated if we obtain its true relevance feedback. 6. CONCLUDING REMARKS We presented a cross-benchmark evaluation of incremental Rocchio and incremental LR in adaptive filtering, focusing on their robustness in terms of performance consistency with respect to cross-corpus parameter optimization. Our main conclusions from this study are the following: • Parameter optimization in AF is an open challenge but has not been thoroughly studied in the past. • Robustness in cross-corpus parameter tuning is important for evaluation and method comparison. • We found LR more robust than Rocchio; it had the best results (in T11SU) ever reported on TDT5, TREC10 and TREC11 without extensive tuning. • We found Rocchio performs strongly when a good validation corpus is available, and a preferred choice when optimizing Ctrk is the objective, favoring recall over precision to an extreme. For future research we want to study explicit modeling of the temporal trends in topic distributions and content drifting. Acknowledgments This material is based upon work supported in parts by the National Science Foundation (NSF) under grant IIS-0434035, by the DoD under award 114008-N66001992891808 and by the Defense Advanced Research Project Agency (DARPA) under Contract No. NBCHD030010. Any opinions, findings and conclusions or recommendations expressed in this material are those of the author(s) and do not necessarily reflect the views of the sponsors. 7. REFERENCES [1] J. Allan. Incremental relevance feedback for information filtering. In SIGIR-96, 1996. [2] J. Callan. Learning while filtering documents. In SIGIR-98, 224-231, 1998. [3] J. Fiscus and G. Duddington. Topic detection and tracking overview. In Topic detection and tracking: event-based information organization, 17-31, 2002. [4] J. Fiscus and B. Wheatley. Overview of the TDT 2004 Evaluation and Results. In TDT-04, 2004. [5] T. Hastie, R. Tibshirani and J. Friedman. Elements of Statistical Learning. Springer, 2001. [6] S. Robertson and D. Hull. The TREC-9 filtering track final report. In TREC-9, 2000. [7] S. Robertson and I. Soboroff. The TREC-10 filtering track final report. In TREC-10, 2001. [8] S. Robertson and I. Soboroff. The TREC 2002 filtering track report. In TREC-11, 2002. [9] S. Robertson and S. Walker. Microsoft Cambridge at TREC-9. In TREC-9, 2000. [10] R. Schapire, Y. Singer and A. Singhal. Boosting and Rocchio applied to text filtering. In SIGIR-98, 215-223, 1998. [11] Y. Yang and B. Kisiel. Margin-based local regression for adaptive filtering. In CIKM-03, 2003. [12] Y. Zhang and J. Callan. Maximum likelihood estimation for filtering thresholds. In SIGIR-01, 2001. [13] Y. Zhang. Using Bayesian priors to combine classifiers for adaptive filtering. In SIGIR-04, 2004. [14] J. Zhang and Y. Yang. Robustness of regularized linear classification methods in text categorization. In SIGIR-03: 190-197, 2003. [15] T. Zhang, F. J. Oles. Text Categorization Based on Regularized Linear Classification Methods. Inf. Retr. 4(1): 5-31 (2001).

logistic regression;adaptive filter;gaussian;topic track;validation set;cost function;topic detection;statistical learning;posterior probability;sigmoid function;information retrieval;relevance feedback;optimization criterion;cross-benchmark evaluation;topic tracking;bias;systematic method for parameter tuning across multiple corpora;rocchio-style method;rocchio;cross-corpus parameter optimization;penalty ratio;robustness;granularity difference;adaptive filtering;external corpus;utility function;probabilistic threshold calibration;regularization;lr

train_H-88

Controlling Overlap in Content-Oriented XML Retrieval

The direct application of standard ranking techniques to retrieve individual elements from a collection of XML documents often produces a result set in which the top ranks are dominated by a large number of elements taken from a small number of highly relevant documents. This paper presents and evaluates an algorithm that re-ranks this result set, with the aim of minimizing redundant content while preserving the benefits of element retrieval, including the benefit of identifying topic-focused components contained within relevant documents. The test collection developed by the INitiative for the Evaluation of XML Retrieval (INEX) forms the basis for the evaluation.

1. INTRODUCTION The representation of documents in XML provides an opportunity for information retrieval systems to take advantage of document structure, returning individual document components when appropriate, rather than complete documents in all circumstances. In response to a user query, an XML information retrieval system might return a mixture of paragraphs, sections, articles, bibliographic entries and other components. This facility is of particular benefit when a collection contains very long documents, such as product manuals or books, where the user should be directed to the most relevant portions of these documents. <article> <fm> <atl>Text Compression for Dynamic Document Databases</atl> <au>Alistair Moffat</au> <au>Justin Zobel</au> <au>Neil Sharman</au> <abs><p><b>Abstract</b> For ...</p></abs> </fm> <bdy> <sec><st>INTRODUCTION</st> <ip1>Modern document databases...</ip1> <p>There are good reasons to compress...</p> </sec> <sec><st>REDUCING MEMORY REQUIREMENTS</st>... <ss1><st>2.1 Method A</st>... </sec> ... </bdy> </article> Figure 1: A journal article encoded in XML. Figure 1 provides an example of a journal article encoded in XML, illustrating many of the important characteristics of XML documents. Tags indicate the beginning and end of each element, with elements varying widely in size, from one word to thousands of words. Some elements, such as paragraphs and sections, may be reasonably presented to the user as retrieval results, but others are not appropriate. Elements overlap each other - articles contain sections, sections contain subsections, and subsections contain paragraphs. Each of these characteristics affects the design of an XML IR system, and each leads to fundamental problems that must be solved in an successful system. Most of these fundamental problems can be solved through the careful adaptation of standard IR techniques, but the problems caused by overlap are unique to this area [4,11] and form the primary focus of this paper. The article of figure 1 may be viewed as an XML tree, as illustrated in figure 2. Formally, a collection of XML documents may be represented as a forest of ordered, rooted trees, consisting of a set of nodes N and a set of directed edges E connecting these nodes. For each node x ∈ N , the notation x.parent refers to the parent node of x, if one exists, and the notation x.children refers to the set of child nodes sec bdyfm atl au au au abs p b st ip1 sec st ss1 st article p Figure 2: Example XML tree. of x. Since an element may be represented by the node at its root, the output of an XML IR system may be viewed as a ranked list of the top-m nodes. The direct application of a standard relevance ranking technique to a set of XML elements can produce a result in which the top ranks are dominated by many structurally related elements. A high scoring section is likely to contain several high scoring paragraphs and to be contained in an high scoring article. For example, many of the elements in figure 2 would receive a high score on the keyword query text index compression algorithms. If each of these elements are presented to a user as an individual and separate result, she may waste considerable time reviewing and rejecting redundant content. One possible solution is to report only the highest scoring element along a given path in the tree, and to remove from the lower ranks any element containing it, or contained within it. Unfortunately, this approach destroys some of the possible benefits of XML IR. For example, an outer element may contain a substantial amount of information that does not appear in an inner element, but the inner element may be heavily focused on the query topic and provide a short overview of the key concepts. In such cases, it is reasonable to report elements which contain, or are contained in, higher ranking elements. Even when an entire book is relevant, a user may still wish to have the most important paragraphs highlighted, to guide her reading and to save time [6]. This paper presents a method for controlling overlap. Starting with an initial element ranking, a re-ranking algorithm adjusts the scores of lower ranking elements that contain, or are contained within, higher ranking elements, reflecting the fact that this information may now be redundant. For example, once an element representing a section appears in the ranking, the scores for the paragraphs it contains and the article that contains it are reduced. The inspiration for this strategy comes partially from recent work on structured documents retrieval, where terms appearing in different fields, such as the title and body, are given different weights [20]. Extending that approach, the re-ranking algorithm varies weights dynamically as elements are processed. The remainder of the paper is organized as follows: After a discussion of background work and evaluation methodology, a baseline retrieval method is presented in section 4. This baseline method represents a reasonable adaptation of standard IR technology to XML. Section 5 then outlines a strategy for controlling overlap, using the baseline method as a starting point. A re-ranking algorithm implementing this strategy is presented in section 6 and evaluated in section 7. Section 8 discusses an extended version of the algorithm. 2. BACKGROUND This section provides a general overview of XML information retrieval and discusses related work, with an emphasis on the fundamental problems mentioned in the introduction. Much research in the area of XML retrieval views it from a traditional database perspective, being concerned with such problems as the implementation of structured query languages [5] and the processing of joins [1]. Here, we take a content oriented IR perceptive, focusing on XML documents that primarily contain natural language data and queries that are primarily expressed in natural language. We assume that these queries indicate only the nature of desired content, not its structure, and that the role of the IR system is to determine which elements best satisfy the underlying information need. Other IR research has considered mixed queries, in which both content and structural requirements are specified [2,6,14,17,23]. 2.1 Term and Document Statistics In traditional information retrieval applications the standard unit of retrieval is taken to be the document. Depending on the application, this term might be interpreted to encompass many different objects, including web pages, newspaper articles and email messages. When applying standard relevance ranking techniques in the context of XML IR, a natural approach is to treat each element as a separate document, with term statistics available for each [16]. In addition, most ranking techniques require global statistics (e.g. inverse document frequency) computed over the collection as a whole. If we consider this collection to include all elements that might be returned by the system, a specific occurrence of a term may appear in several different documents, perhaps in elements representing a paragraph, a subsection, a section and an article. It is not appropriate to compute inverse document frequency under the assumption that the term is contained in all of these elements, since the number of elements that contain a term depends entirely on the structural arrangement of the documents [13,23]. 2.2 Retrievable Elements While an XML IR system might potentially retrieve any element, many elements may not be appropriate as retrieval results. This is usually the case when elements contain very little text [10]. For example, a section title containing only the query terms may receive a high score from a ranking algorithm, but alone it would be of limited value to a user, who might prefer the actual section itself. Other elements may reflect the document"s physical, rather than logical, structure, which may have little or no meaning to a user. An effective XML IR system must return only those elements that have sufficient content to be usable and are able to stand alone as independent objects [15,18]. Standard document components such as paragraphs, sections, subsections, and abstracts usually meet these requirements; titles, italicized phrases, and individual metadata fields often do not. 2.3 Evaluation Methodology Over the past three years, the INitiative for the Evaluation of XML Retrieval (INEX) has encouraged research into XML information retrieval technology [7,8]. INEX is an experimental conference series, similar to TREC, with groups from different institutions completing one or more experimental tasks using their own tools and systems, and comparing their results at the conference itself. Over 50 groups participated in INEX 2004, and the conference has become as influential in the area of XML IR as TREC is in other IR areas. The research described in this paper, as well as much of the related work it cites, depends on the test collections developed by INEX. Overlap causes considerable problems with retrieval evaluation, and the INEX organizers and participants have wrestled with these problems since the beginning. While substantial progress has been made, these problem are still not completely solved. Kazai et al. [11] provide a detailed exposition of the overlap problem in the context of INEX retrieval evaluation and discuss both current and proposed evaluation metrics. Many of these metrics are applied to evaluate the experiments reported in this paper, and they are briefly outlined in the next section. 3. INEX 2004 Space limitations prevent the inclusion of more than a brief summary of INEX 2004 tasks and evaluation methodology. For detailed information, the proceedings of the conference itself should be consulted [8]. 3.1 Tasks For the main experimental tasks, INEX 2004 participants were provided with a collection of 12,107 articles taken from the IEEE Computer Societies magazines and journals between 1995 and 2002. Each document is encoded in XML using a common DTD, with the document of figures 1 and 2 providing one example. At INEX 2004, the two main experimental tasks were both adhoc retrieval tasks, investigating the performance of systems searching a static collection using previously unseen topics. The two tasks differed in the types of topics they used. For one task, the content-only or CO task, the topics consist of short natural language statements with no direct reference to the structure of the documents in the collection. For this task, the IR system is required to select the elements to be returned. For the other task, the contentand-structure or CAS task, the topics are written in an XML query language [22] and contain explicit references to document structure, which the IR system must attempt to satisfy. Since the work described in this paper is directed at the content-only task, where the IR system receives no guidance regarding the elements to return, the CAS task is ignored in the remainder of our description. In 2004, 40 new CO topics were selected by the conference organizers from contributions provided by the conference participants. Each topic includes a short keyword query, which is executed over the collection by each participating group on their own XML IR system. Each group could submit up to three experimental runs consisting of the top m = 1500 elements for each topic. 3.2 Relevance Assessment Since XML IR is concerned with locating those elements that provide complete coverage of a topic while containing as little extraneous information as possible, simple relevant vs. not relevant judgments are not sufficient. Instead, the INEX organizers adopted two dimensions for relevance assessment: The exhaustivity dimension reflects the degree to which an element covers the topic, and the specificity dimension reflects the degree to which an element is focused on the topic. A four-point scale is used in both dimensions. Thus, a (3,3) element is highly exhaustive and highly specific, a (1,3) element is marginally exhaustive and highly specific, and a (0,0) element is not relevant. Additional information on the assessment methodology may be found in Piwowarski and Lalmas [19], who provide a detailed rationale. 3.3 Evaluation Metrics The principle evaluation metric used at INEX 2004 is a version of mean average precision (MAP), adjusted by various quantization functions to give different weights to different elements, depending on their exhaustivity and specificity values. One variant, the strict quantization function gives a weight of 1 to (3,3) elements and a weight of 0 to all others. This variant is essentially the familiar MAP value, with (3,3) elements treated as relevant and all other elements treated as not relevant. Other quantization functions are designed to give partial credit to elements which are near misses, due to a lack or exhaustivity and/or specificity. Both the generalized quantization function and the specificity-oriented generalization (sog) function credit elements according to their degree of relevance [11], with the second function placing greater emphasis on specificity. This paper reports results of this metric using all three of these quantization functions. Since this metric was first introduced at INEX 2002, it is generally referred as the inex-2002 metric. The inex-2002 metric does not penalize overlap. In particular, both the generalized and sog quantization functions give partial credit to a near miss even when a (3,3) element overlapping it is reported at a higher rank. To address this problem, Kazai et al. [11] propose an XML cumulated gain metric, which compares the cumulated gain [9] of a ranked list to an ideal gain vector. This ideal gain vector is constructed from the relevance judgments by eliminating overlap and retaining only best element along a given path. Thus, the XCG metric rewards retrieval runs that avoid overlap. While XCG was not used officially at INEX 2004, a version of it is likely to be used in the future. At INEX 2003, yet another metric was introduced to ameliorate the perceived limitations of the inex-2002 metric. This inex-2003 metric extends the definitions of precision and recall to consider both the size of reported components and the overlap between them. Two versions were created, one that considered only component size and another that considered both size and overlap. While the inex-2003 metric exhibits undesirable anomalies [11], and was not used in 2004, values are reported in the evaluation section to provide an additional instrument for investigating overlap. 4. BASELINE RETRIEVAL METHOD This section provides an overview of baseline XML information retrieval method currently used in the MultiText IR system, developed by the Information Retrieval Group at the University of Waterloo [3]. This retrieval method results from the adaptation and tuning of the Okapi BM25 measure [21] to the XML information retrieval task. The MultiText system performed respectably at INEX 2004, placing in the top ten under all of the quantization functions, and placing first when the quantization function emphasized exhaustivity. To support retrieval from XML and other structured document types, the system provides generalized queries of the form: rank X by Y where X is a sub-query specifying a set of document elements to be ranked and Y is a vector of sub-queries specifying individual retrieval terms. For our INEX 2004 runs, the sub-query X specified a list of retrievable elements as those with tag names as follows: abs app article bb bdy bm fig fm ip1 li p sec ss1 ss2 vt This list includes bibliographic entries (bb) and figure captions (fig) as well as paragraphs, sections and subsections. Prior to INEX 2004, the INEX collection and the INEX 2003 relevance judgments were manually analyzed to select these tag names. Tag names were selected on the basis of their frequency in the collection, the average size of their associated elements, and the relative number of positive relevance judgments they received. Automating this selection process is planned as future work. For INEX 2004, the term vector Y was derived from the topic by splitting phrases into individual words, eliminating stopwords and negative terms (those starting with -), and applying a stemmer. For example, keyword field of topic 166 +"tree edit distance" + XML -image became the four-term query "$tree" "$edit" "$distance" "$xml" where the $ operator within a quoted string stems the term that follows it. Our implementation of Okapi BM25 is derived from the formula of Robertson et al. [21] by setting parameters k2 = 0 and k3 = ∞. Given a term set Q, an element x is assigned the score t∈Q w(1) qt (k1 + 1)xt K + xt (1) where w(1) = log ¡ D − Dt + 0.5 Dt + 0.5 ¢ D = number of documents in the corpus Dt = number of documents containing t qt = frequency that t occurs in the topic xt = frequency that t occurs in x K = k1((1 − b) + b · lx/lavg) lx = length of x lavg = average document length 0.07 0.08 0.09 0.10 0.11 0.12 0.13 0.14 0.15 0 2 4 6 8 10 12 14 16 MeanAveragePrecision(inex-2002) k1 strict generalized sog Figure 3: Impact of k1 on inex-2002 mean average precision with b = 0.75 (INEX 2003 CO topics). Prior to INEX 2004, the INEX 2003 topics and judgments were used to tune the b and k1 parameters, and the impact of this tuning is discussed later in this section. For the purposes of computing document-level statistics (D, Dt and lavg) a document is defined to be an article. These statistics are used for ranking all element types. Following the suggestion of Kamps et al. [10], the retrieval results are filtered to eliminate very short elements, those less than 25 words in length. The use of article statistics for all element types might be questioned. This approach may be justified by viewing the collection as a set of articles to be searched using standard document-oriented techniques, where only articles may be returned. The score computed for an element is essentially the score it would receive if it were added to the collection as a new document, ignoring the minor adjustments needed to the document-level statistics. Nonetheless, we plan to examine this issue again in the future. In our experience, the performance of BM25 typically benefits from tuning the b and k1 parameters to the collection, whenever training queries are available for this purpose. Prior to INEX 2004, we trained the MultiText system using the INEX 2003 queries. As a starting point we used the values b = 0.75 and k1 = 1.2, which perform well on TREC adhoc collections and are used as default values in our system. The results were surprising. Figure 3 shows the result of varying k1 with b = 0.75 on the MAP values under three quantization functions. In our experience, optimal values for k1 are typically in the range 0.0 to 2.0. In this case, large values are required for good performance. Between k1 = 1.0 and k1 = 6.0 MAP increases by over 15% under the strict quantization. Similar improvements are seen under the generalized and sog quantizations. In contrast, our default value of b = 0.75 works well under all quantization functions (figure 4). After tuning over a wide range of values under several quantization functions, we selected values of k = 10.0 and b = 0.80 for our INEX 2004 experiments, and these values are used for the experiments reported in section 7. 0.07 0.08 0.09 0.10 0.11 0.12 0.13 0.14 0.15 0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0 MeanAveragePrecision(inex-2002) b strict generalized sog Figure 4: Impact of b on inex-2002 mean average precision with k1 = 10 (INEX 2003 CO topics). 5. CONTROLLING OVERLAP Starting with an element ranking generated by the baseline method described in the previous section, elements are re-ranked to control overlap by iteratively adjusting the scores of those elements containing or contained in higher ranking elements. At a conceptual level, re-ranking proceeds as follows: 1. Report the highest ranking element. 2. Adjust the scores of the unreported elements. 3. Repeat steps 1 and 2 until m elements are reported. One approach to adjusting the scores of unreported elements in step 2 might be based on the Okapi BM25 scores of the involved elements. For example, assume a paragraph with score p is reported in step 1. In step 2, the section containing the paragraph might then have its score s lowered by an amount α · p to reflect the reduced contribution the paragraph should make to the section"s score. In a related context, Robertson et al. [20] argue strongly against the linear combination of Okapi scores in this fashion. That work considers the problem of assigning different weights to different document fields, such as the title and body associated with Web pages. A common approach to this problem scores the title and body separately and generates a final score as a linear combination of the two. Robertson et al. discuss the theoretical flaws in this approach and demonstrate experimentally that it can actually harm retrieval effectiveness. Instead, they apply the weights at the term frequency level, with an occurrence of a query term t in the title making a greater contribution to the score than an occurrence in the body. In equation 1, xt becomes α0 · yt + α1 · zt, where yt is the number of times t occurs in the title and zt is the number of times t occurs in the body. Translating this approach to our context, the contribution of terms appearing in elements is dynamically reduced as they are reported. The next section presents and analysis a simple re-ranking algorithm that follows this strategy. The algorithm is evaluated experimentally in section 7. One limitation of the algorithm is that the contribution of terms appearing in reported elements is reduced by the same factor regardless of the number of reported elements in which it appears. In section 8 the algorithm is extended to apply increasing weights, lowering the score, when a term appears in more than one reported element. 6. RE-RANKING ALGORITHM The re-ranking algorithm operates over XML trees, such as the one appearing in figure 2. Input to the algorithm is a list of n elements ranked according to their initial BM25 scores. During the initial ranking the XML tree is dynamically re-constructed to include only those nodes with nonzero BM25 scores, so n may be considerably less than |N |. Output from the algorithm is a list of the top m elements, ranked according to their adjusted scores. An element is represented by the node x ∈ N at its root. Associated with this node are fields storing the length of element, term frequencies, and other information required by the re-ranking algorithm, as follows: x.f - term frequency vector x.g - term frequency adjustments x.l - element length x.score - current Okapi BM25 score x.reported - boolean flag, initially false x.children - set of child nodes x.parent - parent node, if one exists These fields are populated during the initial ranking process, and updated as the algorithm progresses. The vector x.f contains term frequency information corresponding to each term in the query. The vector x.g is initially zero and is updated by the algorithm as elements are reported. The score field contains the current BM25 score for the element, which will change as the values in x.g change. The score is computed using equation 1, with the xt value for each term determined by a combination of the values in x.f and x.g. Given a term t ∈ Q, let ft be the component of x.f corresponding to t, and let gt be the component of x.g corresponding to t, then: xt = ft − α · gt (2) For processing by the re-ranking algorithm, nodes are stored in priority queues, ordered by decreasing score. Each priority queue PQ supports three operations: PQ.front() - returns the node with greatest score PQ.add (x) - adds node x to the queue PQ.remove(x) - removes node x from the queue When implemented using standard data structures, the front operation requires O(1) time, and the other operations require O(log n) time, where n is the size of the queue. The core of the re-ranking algorithm is presented in figure 5. The algorithm takes as input the priority queue S containing the initial ranking, and produces the top-m reranked nodes in the priority queue F. After initializing F to be empty on line 1, the algorithm loops m times over lines 215, transferring at least one node from S to F during each iteration. At the start of each iteration, the unreported node at the front of S has the greatest adjusted score, and it is removed and added to F. The algorithm then traverses the 1 F ← ∅ 2 for i ← 1 to m do 3 x ← S.front() 4 S.remove(x) 5 x.reported ← true 6 F.add(x) 7 8 foreach y ∈ x.children do 9 Down (y) 10 end do 11 12 if x is not a root node then 13 Up (x, x.parent) 14 end if 15 end do Figure 5: Re-Ranking Algorithm - As input, the algorithm takes a priority queue S, containing XML nodes ranked by their initial scores, and returns its results in priority queue F, ranked by adjusted scores. 1 Up(x, y) ≡ 2 S.remove(y) 3 y.g ← y.g + x.f − x.g 4 recompute y.score 5 S.add(y) 6 if y is not a root node then 7 Up (x, y.parent) 8 end if 9 10 Down(x) ≡ 11 if not x.reported then 12 S.remove(x) 14 x.g ← x.f 15 recompute x.score 16 if x.score > 0 then 17 F.add(x) 18 end if 19 x.reported ← true 20 foreach y ∈ x.children do 21 Down (y) 22 end do 23 end if Figure 6: Tree traversal routines called by the reranking algorithm. 0.0 0.02 0.04 0.06 0.08 0.10 0.12 0.14 0.16 0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0 0.25 0.26 0.27 0.28 0.29 0.30 0.31 0.32 0.33 0.34 0.35 0.36 MeanAveragePrecision(inex-2002) XMLCumulatedGain(XCG) alpha MAP (strict) MAP (generalized) MAP (sog) XCG (sog2) Figure 7: Impact of α on XCG and inex-2002 MAP (INEX 2004 CO topics; assessment set I). node"s ancestors (lines 8-10) and descendants (lines 12-14) adjusting the scores of these nodes. The tree traversal routines, Up and Down are given in figure 6. The Up routine removes each ancestor node from S, adjusts its term frequency values, recomputes its score, and adds it back into S. The adjustment of the term frequency values (line 3) adds to y.g only the previously unreported term occurrences in x. Re-computation of the score on line 4 uses equations 1 and 2. The Down routine performs a similar operation on each descendant. However, since the contents of each descendant are entirely contained in a reported element its final score may be computed, and it is removed from S and added to F. In order to determine the time complexity of the algorithm, first note that a node may be an argument to Down at most once. Thereafter, the reported flag of its parent is true. During each call to Down a node may be moved from S to F, requiring O(log n) time. Thus, the total time for all calls to Down is O(n log n), and we may temporarily ignore lines 8-10 of figure 5 when considering the time complexity of the loop over lines 2-15. During each iteration of this loop, a node and each of its ancestors are removed from a priority queue and then added back into a priority queue. Since a node may have at most h ancestors, where h is the maximum height of any tree in the collection, each of the m iterations requires O(h log n) time. Combining these observations produces an overall time complexity of O((n + mh) log n). In practice, re-ranking an INEX result set requires less than 200ms on a three-year-old desktop PC. 7. EVALUATION None of the metrics described in section 3.3 is a close fit with the view of overlap advocated by this paper. Nonetheless, when taken together they provide insight into the behaviour of the re-ranking algorithm. The INEX evaluation packages (inex_eval and inex_eval_ng) were used to compute values for the inex-2002 and inex-2003 metrics. Values for the XCG metrics were computed using software supplied by its inventors [11]. Figure 7 plots the three variants of inex-2002 MAP metric together with the XCG metric. Values for these metrics 0.0 0.02 0.04 0.06 0.08 0.10 0.12 0.14 0.16 0.18 0.20 0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0 MeanAveragePrecision(inex-2003) alpha strict, overlap not considered strict, overlap considered generalized, overlap not considered generalized, overlap considered Figure 8: Impact of α on inex-2003 MAP (INEX 2004 CO topics; assessment set I). are plotted for values of α between 0.0 and 1.0. Recalling that the XCG metric is designed to penalize overlap, while the inex-2002 metric ignores overlap, the conflict between the metrics is obvious. The MAP values at one extreme (α = 0.0) and the XCG value at the other extreme (α = 1.0) represent retrieval performance comparable to the best systems at INEX 2004 [8,12]. Figure 8 plots values of the inex-2003 MAP metric for two quantizations, with and without consideration of overlap. Once again, conflict is apparent, with the influence of α substantially lessened when overlap is considered. 8. EXTENDED ALGORITHM One limitation of the re-ranking algorithm is that a single weight α is used to adjust the scores of both the ancestors and descendants of reported elements. An obvious extension is to use different weights in these two cases. Furthermore, the same weight is used regardless of the number of times an element is contained in a reported element. For example, a paragraph may form part of a reported section and then form part of a reported article. Since the user may now have seen this paragraph twice, its score should be further lowered by increasing the value of the weight. Motivated by these observations, the re-ranking algorithm may be extended with a series of weights 1 = β0 ≥ β1 ≥ β2 ≥ ... ≥ βM ≥ 0. where βj is the weight applied to a node that has been a descendant of a reported node j times. Note that an upper bound on M is h, the maximum height of any XML tree in the collection. However, in practice M is likely to be relatively small (perhaps 3 or 4). Figure 9 presents replacements for the Up and Down routines of figure 6, incorporating this series of weights. One extra field is required in each node, as follows: x.j - down count The value of x.j is initially set to zero in all nodes and is incremented each time Down is called with x as its argument. When computing the score of node, the value of x.j selects 1 Up(x, y) ≡ 2 if not y.reported then 3 S.remove(y) 4 y.g ← y.g + x.f − x.g 5 recompute y.score 6 S.add(y) 8 if y is not a root node then 9 Up (x, y.parent) 10 end if 11 end if 12 13 Down(x) ≡ 14 if x.j < M then 15 x.j ← x.j + 1 16 if not x.reported then 17 S.remove(x) 18 recompute x.score 19 S.add(x) 20 end if 21 foreach y ∈ x.children do 22 Down (y) 23 end do 24 end if Figure 9: Extended tree traversal routines. the weight to be applied to the node by adjusting the value of xt in equation 1, as follows: xt = βx.j · (ft − α · gt) (3) where ft and gt are the components of x.f and x.g corresponding to term t. A few additional changes are required to extend Up and Down. The Up routine returns immediately (line 2) if its argument has already been reported, since term frequencies have already been adjusted in its ancestors. The Down routine does not report its argument, but instead recomputes its score and adds it back into S. A node cannot be an argument to Down more than M +1 times, which in turn implies an overall time complexity of O((nM + mh) log n). Since M ≤ h and m ≤ n, the time complexity is also O(nh log n). 9. CONCLUDING DISCUSSION When generating retrieval results over an XML collection, some overlap in the results should be tolerated, and may be beneficial. For example, when a highly exhaustive and fairly specific (3,2) element contains a much smaller (2,3) element, both should be reported to the user, and retrieval algorithms and evaluation metrics should respect this relationship. The algorithm presented in this paper controls overlap by weighting the terms occurring in reported elements to reflect their reduced importance. Other approaches may also help to control overlap. For example, when XML retrieval results are presented to users it may be desirable to cluster structurally related elements together, visually illustrating the relationships between them. While this style of user interface may help a user cope with overlap, the strategy presented in this paper continues to be applicable, by determining the best elements to include in each cluster. At Waterloo, we continue to develop and test our ideas for INEX 2005. In particular, we are investigating methods for learning the α and βj weights. We are also re-evaluating our approach to document statistics and examining appropriate adjustments to the k1 parameter as term weights change [20]. 10. ACKNOWLEDGMENTS Thanks to Gabriella Kazai and Arjen de Vries for providing an early version of their software for computing the XCG metric, and thanks to Phil Tilker and Stefan B¨uttcher for their help with the experimental evaluation. In part, funding for this project was provided by IBM Canada through the National Institute for Software Research. 11. REFERENCES [1] N. Bruno, N. Koudas, and D. Srivastava. Holistic twig joins: Optimal XML pattern matching. In Proceedings of the 2002 ACM SIGMOD International Conference on the Management of Data, pages 310-321, Madison, Wisconsin, June 2002. [2] D. Carmel, Y. S. Maarek, M. Mandelbrod, Y. Mass, and A. Soffer. Searching XML documents via XML fragments. In Proceedings of the 26th Annual International ACM SIGIR Conference on Research and Development in Information Retrieval, pages 151-158, Toronto, Canada, 2003. [3] C. L. A. Clarke and P. L. Tilker. MultiText experiments for INEX 2004. In INEX 2004 Workshop Proceedings, 2004. Published in LNCS 3493 [8]. [4] A. P. de Vries, G. Kazai, and M. Lalmas. Tolerance to irrelevance: A user-effort oriented evaluation of retrieval systems without predefined retrieval unit. In RIAO 2004 Conference Proceedings, pages 463-473, Avignon, France, April 2004. [5] D. DeHaan, D. Toman, M. P. Consens, and M. T. ¨Ozsu. A comprehensive XQuery to SQL translation using dynamic interval encoding. In Proceedings of the 2003 ACM SIGMOD International Conference on the Management of Data, San Diego, June 2003. [6] N. Fuhr and K. Großjohann. XIRQL: A query language for information retrieval in XML documents. In Proceedings of the 24th Annual International ACM SIGIR Conference on Research and Development in Information Retrieval, pages 172-180, New Orleans, September 2001. [7] N. Fuhr, M. Lalmas, and S. Malik, editors. Initiative for the Evaluation of XML Retrieval. Proceedings of the Second Workshop (INEX 2003), Dagstuhl, Germany, December 2003. [8] N. Fuhr, M. Lalmas, S. Malik, and Zolt´an Szl´avik, editors. Initiative for the Evaluation of XML Retrieval. Proceedings of the Third Workshop (INEX 2004), Dagstuhl, Germany, December 2004. Published as Advances in XML Information Retrieval, Lecture Notes in Computer Science, volume 3493, Springer, 2005. [9] K. J¨avelin and J. Kek¨al¨ainen. Cumulated gain-based evaluation of IR techniques. ACM Transactions on Information Systems, 20(4):422-446, 2002. [10] J. Kamps, M. de Rijke, and B. Sigurbj¨ornsson. Length normalization in XML retrieval. In Proceedings of the 27th Annual International ACM SIGIR Conference on Research and Development in Information Retrieval, pages 80-87, Sheffield, UK, July 2004. [11] G. Kazai, M. Lalmas, and A. P. de Vries. The overlap problem in content-oriented XML retrieval evaluation. In Proceedings of the 27th Annual International ACM SIGIR Conference on Research and Development in Information Retrieval, pages 72-79, Sheffield, UK, July 2004. [12] G. Kazai, M. Lalmas, and A. P. de Vries. Reliability tests for the XCG and inex-2002 metrics. In INEX 2004 Workshop Proceedings, 2004. Published in LNCS 3493 [8]. [13] J. Kek¨al¨ainen, M. Junkkari, P. Arvola, and T. Aalto. TRIX 2004 - Struggling with the overlap. In INEX 2004 Workshop Proceedings, 2004. Published in LNCS 3493 [8]. [14] S. Liu, Q. Zou, and W. W. Chu. Configurable indexing and ranking for XML information retrieval. In Proceedings of the 27th Annual International ACM SIGIR Conference on Research and Development in Information Retrieval, pages 88-95, Sheffield, UK, July 2004. [15] Y. Mass and M. Mandelbrod. Retrieving the most relevant XML components. In INEX 2003 Workshop Proceedings, Dagstuhl, Germany, December 2003. [16] Y. Mass and M. Mandelbrod. Component ranking and automatic query refinement for XML retrieval. In INEX 2004 Workshop Proceedings, 2004. Published in LNCS 3493 [8]. [17] P. Ogilvie and J. Callan. Hierarchical language models for XML component retrieval. In INEX 2004 Workshop Proceedings, 2004. Published in LNCS 3493 [8]. [18] J. Pehcevski, J. A. Thom, and A. Vercoustre. Hybrid XML retrieval re-visited. In INEX 2004 Workshop Proceedings, 2004. Published in LNCS 3493 [8]. [19] B. Piwowarski and M. Lalmas. Providing consistent and exhaustive relevance assessments for XML retrieval evaluation. In Proceedings of the 13th ACM Conference on Information and Knowledge Management, pages 361-370, Washington, DC, November 2004. [20] S. Robertson, H. Zaragoza, and M. Taylor. Simple BM25 extension to multiple weighted fields. In Proceedings of the 13th ACM Conference on Information and Knowledge Management, pages 42-50, Washington, DC, November 2004. [21] S. E. Robertson, S. Walker, and M. Beaulieu. Okapi at TREC-7: Automatic ad-hoc, filtering, VLC and interactive track. In Proceedings of the Seventh Text REtrieval Conference, Gaithersburg, MD, November 1998. [22] A. Trotman and B. Sigurbj¨ornsson. NEXI, now and next. In INEX 2004 Workshop Proceedings, 2004. Published in LNCS 3493 [8]. [23] J. Vittaut, B. Piwowarski, and P. Gallinari. An algebra for structured queries in bayesian networks. In INEX 2004 Workshop Proceedings, 2004. Published in LNCS 3493 [8].

baseline retrieval;inex;extended tree traversal routine;ideal gain vector;xml ir;xml cumulated gain metric;priority queue;cumulated gain;multitext system;term frequency vector;rank;sog quantization;xml;information retrieval;time complexity;xcg metric reward retrieval;re-ranking algorithm

train_H-90

Context-Sensitive Information Retrieval Using Implicit Feedback

A major limitation of most existing retrieval models and systems is that the retrieval decision is made based solely on the query and document collection; information about the actual user and search context is largely ignored. In this paper, we study how to exploit implicit feedback information, including previous queries and clickthrough information, to improve retrieval accuracy in an interactive information retrieval setting. We propose several contextsensitive retrieval algorithms based on statistical language models to combine the preceding queries and clicked document summaries with the current query for better ranking of documents. We use the TREC AP data to create a test collection with search context information, and quantitatively evaluate our models using this test set. Experiment results show that using implicit feedback, especially the clicked document summaries, can improve retrieval performance substantially.

1. INTRODUCTION In most existing information retrieval models, the retrieval problem is treated as involving one single query and a set of documents. From a single query, however, the retrieval system can only have very limited clue about the user"s information need. An optimal retrieval system thus should try to exploit as much additional context information as possible to improve retrieval accuracy, whenever it is available. Indeed, context-sensitive retrieval has been identified as a major challenge in information retrieval research[2]. There are many kinds of context that we can exploit. Relevance feedback [14] can be considered as a way for a user to provide more context of search and is known to be effective for improving retrieval accuracy. However, relevance feedback requires that a user explicitly provides feedback information, such as specifying the category of the information need or marking a subset of retrieved documents as relevant documents. Since it forces the user to engage additional activities while the benefits are not always obvious to the user, a user is often reluctant to provide such feedback information. Thus the effectiveness of relevance feedback may be limited in real applications. For this reason, implicit feedback has attracted much attention recently [11, 13, 18, 17, 12]. In general, the retrieval results using the user"s initial query may not be satisfactory; often, the user would need to revise the query to improve the retrieval/ranking accuracy [8]. For a complex or difficult information need, the user may need to modify his/her query and view ranked documents with many iterations before the information need is completely satisfied. In such an interactive retrieval scenario, the information naturally available to the retrieval system is more than just the current user query and the document collection - in general, all the interaction history can be available to the retrieval system, including past queries, information about which documents the user has chosen to view, and even how a user has read a document (e.g., which part of a document the user spends a lot of time in reading). We define implicit feedback broadly as exploiting all such naturally available interaction history to improve retrieval results. A major advantage of implicit feedback is that we can improve the retrieval accuracy without requiring any user effort. For example, if the current query is java, without knowing any extra information, it would be impossible to know whether it is intended to mean the Java programming language or the Java island in Indonesia. As a result, the retrieved documents will likely have both kinds of documents - some may be about the programming language and some may be about the island. However, any particular user is unlikely searching for both types of documents. Such an ambiguity can be resolved by exploiting history information. For example, if we know that the previous query from the user is cgi programming, it would strongly suggest that it is the programming language that the user is searching for. Implicit feedback was studied in several previous works. In [11], Joachims explored how to capture and exploit the clickthrough information and demonstrated that such implicit feedback information can indeed improve the search accuracy for a group of people. In [18], a simulation study of the effectiveness of different implicit feedback algorithms was conducted, and several retrieval models designed for exploiting clickthrough information were proposed and evaluated. In [17], some existing retrieval algorithms are adapted to improve search results based on the browsing history of a user. Other related work on using context includes personalized search [1, 3, 4, 7, 10], query log analysis [5], context factors [12], and implicit queries [6]. While the previous work has mostly focused on using clickthrough information, in this paper, we use both clickthrough information and preceding queries, and focus on developing new context-sensitive language models for retrieval. Specifically, we develop models for using implicit feedback information such as query and clickthrough history of the current search session to improve retrieval accuracy. We use the KL-divergence retrieval model [19] as the basis and propose to treat context-sensitive retrieval as estimating a query language model based on the current query and any search context information. We propose several statistical language models to incorporate query and clickthrough history into the KL-divergence model. One challenge in studying implicit feedback models is that there does not exist any suitable test collection for evaluation. We thus use the TREC AP data to create a test collection with implicit feedback information, which can be used to quantitatively evaluate implicit feedback models. To the best of our knowledge, this is the first test set for implicit feedback. We evaluate the proposed models using this data set. The experimental results show that using implicit feedback information, especially the clickthrough data, can substantially improve retrieval performance without requiring additional effort from the user. The remaining sections are organized as follows. In Section 2, we attempt to define the problem of implicit feedback and introduce some terms that we will use later. In Section 3, we propose several implicit feedback models based on statistical language models. In Section 4, we describe how we create the data set for implicit feedback experiments. In Section 5, we evaluate different implicit feedback models on the created data set. Section 6 is our conclusions and future work. 2. PROBLEM DEFINITION There are two kinds of context information we can use for implicit feedback. One is short-term context, which is the immediate surrounding information which throws light on a user"s current information need in a single session. A session can be considered as a period consisting of all interactions for the same information need. The category of a user"s information need (e.g., kids or sports), previous queries, and recently viewed documents are all examples of short-term context. Such information is most directly related to the current information need of the user and thus can be expected to be most useful for improving the current search. In general, short-term context is most useful for improving search in the current session, but may not be so helpful for search activities in a different session. The other kind of context is long-term context, which refers to information such as a user"s education level and general interest, accumulated user query history and past user clickthrough information; such information is generally stable for a long time and is often accumulated over time. Long-term context can be applicable to all sessions, but may not be as effective as the short-term context in improving search accuracy for a particular session. In this paper, we focus on the short-term context, though some of our methods can also be used to naturally incorporate some long-term context. In a single search session, a user may interact with the search system several times. During interactions, the user would continuously modify the query. Therefore for the current query Qk (except for the first query of a search session) , there is a query history, HQ = (Q1, ..., Qk−1) associated with it, which consists of the preceding queries given by the same user in the current session. Note that we assume that the session boundaries are known in this paper. In practice, we need techniques to automatically discover session boundaries, which have been studied in [9, 16]. Traditionally, the retrieval system only uses the current query Qk to do retrieval. But the short-term query history clearly may provide useful clues about the user"s current information need as seen in the java example given in the previous section. Indeed, our previous work [15] has shown that the short-term query history is useful for improving retrieval accuracy. In addition to the query history, there may be other short-term context information available. For example, a user would presumably frequently click some documents to view. We refer to data associated with these actions as clickthrough history. The clickthrough data may include the title, summary, and perhaps also the content and location (e.g., the URL) of the clicked document. Although it is not clear whether a viewed document is actually relevant to the user"s information need, we may safely assume that the displayed summary/title information about the document is attractive to the user, thus conveys information about the user"s information need. Suppose we concatenate all the displayed text information about a document (usually title and summary) together, we will also have a clicked summary Ci in each round of retrieval. In general, we may have a history of clicked summaries C1, ..., Ck−1. We will also exploit such clickthrough history HC = (C1, ..., Ck−1) to improve our search accuracy for the current query Qk. Previous work has also shown positive results using similar clickthrough information [11, 17]. Both query history and clickthrough history are implicit feedback information, which naturally exists in interactive information retrieval, thus no additional user effort is needed to collect them. In this paper, we study how to exploit such information (HQ and HC ), develop models to incorporate the query history and clickthrough history into a retrieval ranking function, and quantitatively evaluate these models. 3. LANGUAGE MODELS FOR CONTEXTSENSITIVEINFORMATIONRETRIEVAL Intuitively, the query history HQ and clickthrough history HC are both useful for improving search accuracy for the current query Qk. An important research question is how we can exploit such information effectively. We propose to use statistical language models to model a user"s information need and develop four specific context-sensitive language models to incorporate context information into a basic retrieval model. 3.1 Basic retrieval model We use the Kullback-Leibler (KL) divergence method [19] as our basic retrieval method. According to this model, the retrieval task involves computing a query language model θQ for a given query and a document language model θD for a document and then computing their KL divergence D(θQ||θD), which serves as the score of the document. One advantage of this approach is that we can naturally incorporate the search context as additional evidence to improve our estimate of the query language model. Formally, let HQ = (Q1, ..., Qk−1) be the query history and the current query be Qk. Let HC = (C1, ..., Ck−1) be the clickthrough history. Note that Ci is the concatenation of all clicked documents" summaries in the i-th round of retrieval since we may reasonably treat all these summaries equally. Our task is to estimate a context query model, which we denote by p(w|θk), based on the current query Qk, as well as the query history HQ and clickthrough history HC . We now describe several different language models for exploiting HQ and HC to estimate p(w|θk). We will use c(w, X) to denote the count of word w in text X, which could be either a query or a clicked document"s summary or any other text. We will use |X| to denote the length of text X or the total number of words in X. 3.2 Fixed Coefficient Interpolation (FixInt) Our first idea is to summarize the query history HQ with a unigram language model p(w|HQ) and the clickthrough history HC with another unigram language model p(w|HC ). Then we linearly interpolate these two history models to obtain the history model p(w|H). Finally, we interpolate the history model p(w|H) with the current query model p(w|Qk). These models are defined as follows. p(w|Qi) = c(w, Qi) |Qi| p(w|HQ) = 1 k − 1 i=k−1 i=1 p(w|Qi) p(w|Ci) = c(w, Ci) |Ci| p(w|HC ) = 1 k − 1 i=k−1 i=1 p(w|Ci) p(w|H) = βp(w|HC ) + (1 − β)p(w|HQ) p(w|θk) = αp(w|Qk) + (1 − α)p(w|H) where β ∈ [0, 1] is a parameter to control the weight on each history model, and where α ∈ [0, 1] is a parameter to control the weight on the current query and the history information. If we combine these equations, we see that p(w|θk) = αp(w|Qk) + (1 − α)[βp(w|HC ) + (1 − β)p(w|HQ)] That is, the estimated context query model is just a fixed coefficient interpolation of three models p(w|Qk), p(w|HQ), and p(w|HC ). 3.3 Bayesian Interpolation (BayesInt) One possible problem with the FixInt approach is that the coefficients, especially α, are fixed across all the queries. But intuitively, if our current query Qk is very long, we should trust the current query more, whereas if Qk has just one word, it may be beneficial to put more weight on the history. To capture this intuition, we treat p(w|HQ) and p(w|HC ) as Dirichlet priors and Qk as the observed data to estimate a context query model using Bayesian estimator. The estimated model is given by p(w|θk) = c(w, Qk) + µp(w|HQ) + νp(w|HC ) |Qk| + µ + ν = |Qk| |Qk| + µ + ν p(w|Qk)+ µ + ν |Qk| + µ + ν [ µ µ + ν p(w|HQ)+ ν µ + ν p(w|HC )] where µ is the prior sample size for p(w|HQ) and ν is the prior sample size for p(w|HC ). We see that the only difference between BayesInt and FixInt is the interpolation coefficients are now adaptive to the query length. Indeed, when viewing BayesInt as FixInt, we see that α = |Qk| |Qk|+µ+ν , β = ν ν+µ , thus with fixed µ and ν, we will have a query-dependent α. Later we will show that such an adaptive α empirically performs better than a fixed α. 3.4 Online Bayesian Updating (OnlineUp) Both FixInt and BayesInt summarize the history information by averaging the unigram language models estimated based on previous queries or clicked summaries. This means that all previous queries are treated equally and so are all clicked summaries. However, as the user interacts with the system and acquires more knowledge about the information in the collection, presumably, the reformulated queries will become better and better. Thus assigning decaying weights to the previous queries so as to trust a recent query more than an earlier query appears to be reasonable. Interestingly, if we incrementally update our belief about the user"s information need after seeing each query, we could naturally obtain decaying weights on the previous queries. Since such an incremental online updating strategy can be used to exploit any evidence in an interactive retrieval system, we present it in a more general way. In a typical retrieval system, the retrieval system responds to every new query entered by the user by presenting a ranked list of documents. In order to rank documents, the system must have some model for the user"s information need. In the KL divergence retrieval model, this means that the system must compute a query model whenever a user enters a (new) query. A principled way of updating the query model is to use Bayesian estimation, which we discuss below. 3.4.1 Bayesian updating We first discuss how we apply Bayesian estimation to update a query model in general. Let p(w|φ) be our current query model and T be a new piece of text evidence observed (e.g., T can be a query or a clicked summary). To update the query model based on T, we use φ to define a Dirichlet prior parameterized as Dir(µT p(w1|φ), ..., µT p(wN |φ)) where µT is the equivalent sample size of the prior. We use Dirichlet prior because it is a conjugate prior for multinomial distributions. With such a conjugate prior, the predictive distribution of φ (or equivalently, the mean of the posterior distribution of φ is given by p(w|φ) = c(w, T) + µT p(w|φ) |T| + µT (1) where c(w, T) is the count of w in T and |T| is the length of T. Parameter µT indicates our confidence in the prior expressed in terms of an equivalent text sample comparable with T. For example, µT = 1 indicates that the influence of the prior is equivalent to adding one extra word to T. 3.4.2 Sequential query model updating We now discuss how we can update our query model over time during an interactive retrieval process using Bayesian estimation. In general, we assume that the retrieval system maintains a current query model φi at any moment. As soon as we obtain some implicit feedback evidence in the form of a piece of text Ti, we will update the query model. Initially, before we see any user query, we may already have some information about the user. For example, we may have some information about what documents the user has viewed in the past. We use such information to define a prior on the query model, which is denoted by φ0. After we observe the first query Q1, we can update the query model based on the new observed data Q1. The updated query model φ1 can then be used for ranking documents in response to Q1. As the user views some documents, the displayed summary text for such documents C1 (i.e., clicked summaries) can serve as some new data for us to further update the query model to obtain φ1. As we obtain the second query Q2 from the user, we can update φ1 to obtain a new model φ2. In general, we may repeat such an updating process to iteratively update the query model. Clearly, we see two types of updating: (1) updating based on a new query Qi; (2) updating based on a new clicked summary Ci. In both cases, we can treat the current model as a prior of the context query model and treat the new observed query or clicked summary as observed data. Thus we have the following updating equations: p(w|φi) = c(w, Qi) + µip(w|φi−1) |Qi| + µi p(w|φi) = c(w, Ci) + νip(w|φi) |Ci| + νi where µi is the equivalent sample size for the prior when updating the model based on a query, while νi is the equivalent sample size for the prior when updating the model based on a clicked summary. If we set µi = 0 (or νi = 0) we essentially ignore the prior model, thus would start a completely new query model based on the query Qi (or the clicked summary Ci). On the other hand, if we set µi = +∞ (or νi = +∞) we essentially ignore the observed query (or the clicked summary) and do not update our model. Thus the model remains the same as if we do not observe any new text evidence. In general, the parameters µi and νi may have different values for different i. For example, at the very beginning, we may have very sparse query history, thus we could use a smaller µi, but later as the query history is richer, we can consider using a larger µi. But in our experiments, unless otherwise stated, we set them to the same constants, i.e., ∀i, j, µi = µj, νi = νj. Note that we can take either p(w|φi) or p(w|φi) as our context query model for ranking documents. This suggests that we do not have to wait until a user enters a new query to initiate a new round of retrieval; instead, as soon as we collect clicked summary Ci, we can update the query model and use p(w|φi) to immediately rerank any documents that a user has not yet seen. To score documents after seeing query Qk, we use p(w|φk), i.e., p(w|θk) = p(w|φk) 3.5 Batch Bayesian updating (BatchUp) If we set the equivalent sample size parameters to fixed constant, the OnlineUp algorithm would introduce a decaying factor - repeated interpolation would cause the early data to have a low weight. This may be appropriate for the query history as it is reasonable to believe that the user becomes better and better at query formulation as time goes on, but it is not necessarily appropriate for the clickthrough information, especially because we use the displayed summary, rather than the actual content of a clicked document. One way to avoid applying a decaying interpolation to the clickthrough data is to do OnlineUp only for the query history Q = (Q1, ..., Qi−1), but not for the clickthrough data C. We first buffer all the clickthrough data together and use the whole chunk of clickthrough data to update the model generated through running OnlineUp on previous queries. The updating equations are as follows. p(w|φi) = c(w, Qi) + µip(w|φi−1) |Qi| + µi p(w|ψi) = i−1 j=1 c(w, Cj) + νip(w|φi) i−1 j=1 |Cj| + νi where µi has the same interpretation as in OnlineUp, but νi now indicates to what extent we want to trust the clicked summaries. As in OnlineUp, we set all µi"s and νi"s to the same value. And to rank documents after seeing the current query Qk, we use p(w|θk) = p(w|ψk) 4. DATA COLLECTION In order to quantitatively evaluate our models, we need a data set which includes not only a text database and testing topics, but also query history and clickthrough history for each topic. Since there is no such data set available to us, we have to create one. There are two choices. One is to extract topics and any associated query history and clickthrough history for each topic from the log of a retrieval system (e.g., search engine). But the problem is that we have no relevance judgments on such data. The other choice is to use a TREC data set, which has a text database, topic description and relevance judgment file. Unfortunately, there are no query history and clickthrough history data. We decide to augment a TREC data set by collecting query history and clickthrough history data. We select TREC AP88, AP89 and AP90 data as our text database, because AP data has been used in several TREC tasks and has relatively complete judgments. There are altogether 242918 news articles and the average document length is 416 words. Most articles have titles. If not, we select the first sentence of the text as the title. For the preprocessing, we only do case folding and do not do stopword removal or stemming. We select 30 relatively difficult topics from TREC topics 1-150. These 30 topics have the worst average precision performance among TREC topics 1-150 according to some baseline experiments using the KL-Divergence model with Bayesian prior smoothing [20]. The reason why we select difficult topics is that the user then would have to have several interactions with the retrieval system in order to get satisfactory results so that we can expect to collect a relatively richer query history and clickthrough history data from the user. In real applications, we may also expect our models to be most useful for such difficult topics, so our data collection strategy reflects the real world applications well. We index the TREC AP data set and set up a search engine and web interface for TREC AP news articles. We use 3 subjects to do experiments to collect query history and clickthrough history data. Each subject is assigned 10 topics and given the topic descriptions provided by TREC. For each topic, the first query is the title of the topic given in the original TREC topic description. After the subject submits the query, the search engine will do retrieval and return a ranked list of search results to the subject. The subject will browse the results and maybe click one or more results to browse the full text of article(s). The subject may also modify the query to do another search. For each topic, the subject composes at least 4 queries. In our experiment, only the first 4 queries for each topic are used. The user needs to select the topic number from a selection menu before submitting the query to the search engine so that we can easily detect the session boundary, which is not the focus of our study. We use a relational database to store user interactions, including the submitted queries and clicked documents. For each query, we store the query terms and the associated result pages. And for each clicked document, we store the summary as shown on the search result page. The summary of the article is query dependent and is computed online using fixed-length passage retrieval (KL divergence model with Bayesian prior smoothing). Among 120 (4 for each of 30 topics) queries which we study in the experiment, the average query length is 3.71 words. Altogether there are 91 documents clicked to view. So on average, there are around 3 clicks per topic. The average length of clicked summary FixInt BayesInt OnlineUp BatchUp Query (α = 0.1, β = 1.0) (µ = 0.2, ν = 5.0) (µ = 5.0, ν = 15.0) (µ = 2.0, ν = 15.0) MAP pr@20docs MAP pr@20docs MAP pr@20docs MAP pr@20docs q1 0.0095 0.0317 0.0095 0.0317 0.0095 0.0317 0.0095 0.0317 q2 0.0312 0.1150 0.0312 0.1150 0.0312 0.1150 0.0312 0.1150 q2 + HQ + HC 0.0324 0.1117 0.0345 0.1117 0.0215 0.0733 0.0342 0.1100 Improve. 3.8% -2.9% 10.6% -2.9% -31.1% -36.3% 9.6% -4.3% q3 0.0421 0.1483 0.0421 0.1483 0.0421 0.1483 0.0421 0.1483 q3 + HQ + HC 0.0726 0.1967 0.0816 0.2067 0.0706 0.1783 0.0810 0.2067 Improve 72.4% 32.6% 93.8% 39.4% 67.7% 20.2% 92.4% 39.4% q4 0.0536 0.1933 0.0536 0.1933 0.0536 0.1933 0.0536 0.1933 q4 + HQ + HC 0.0891 0.2233 0.0955 0.2317 0.0792 0.2067 0.0950 0.2250 Improve 66.2% 15.5% 78.2% 19.9% 47.8% 6.9% 77.2% 16.4% Table 1: Effect of using query history and clickthrough data for document ranking. is 34.4 words. Among 91 clicked documents, 29 documents are judged relevant according to TREC judgment file. This data set is publicly available 1 . 5. EXPERIMENTS 5.1 Experiment design Our major hypothesis is that using search context (i.e., query history and clickthrough information) can help improve search accuracy. In particular, the search context can provide extra information to help us estimate a better query model than using just the current query. So most of our experiments involve comparing the retrieval performance using the current query only (thus ignoring any context) with that using the current query as well as the search context. Since we collected four versions of queries for each topic, we make such comparisons for each version of queries. We use two performance measures: (1) Mean Average Precision (MAP): This is the standard non-interpolated average precision and serves as a good measure of the overall ranking accuracy. (2) Precision at 20 documents (pr@20docs): This measure does not average well, but it is more meaningful than MAP and reflects the utility for users who only read the top 20 documents. In all cases, the reported figure is the average over all of the 30 topics. We evaluate the four models for exploiting search context (i.e., FixInt, BayesInt, OnlineUp, and BatchUp). Each model has precisely two parameters (α and β for FixInt; µ and ν for others). Note that µ and ν may need to be interpreted differently for different methods. We vary these parameters and identify the optimal performance for each method. We also vary the parameters to study the sensitivity of our algorithms to the setting of the parameters. 5.2 Result analysis 5.2.1 Overall effect of search context We compare the optimal performances of four models with those using the current query only in Table 1. A row labeled with qi is the baseline performance and a row labeled with qi + HQ + HC is the performance of using search context. We can make several observations from this table: 1. Comparing the baseline performances indicates that on average reformulated queries are better than the previous queries with the performance of q4 being the best. Users generally formulate better and better queries. 2. Using search context generally has positive effect, especially when the context is rich. This can be seen from the fact that the 1 http://sifaka.cs.uiuc.edu/ir/ucair/QCHistory.zip improvement for q4 and q3 is generally more substantial compared with q2. Actually, in many cases with q2, using the context may hurt the performance, probably because the history at that point is sparse. When the search context is rich, the performance improvement can be quite substantial. For example, BatchUp achieves 92.4% improvement in the mean average precision over q3 and 77.2% improvement over q4. (The generally low precisions also make the relative improvement deceptively high, though.) 3. Among the four models using search context, the performances of FixInt and OnlineUp are clearly worse than those of BayesInt and BatchUp. Since BayesInt performs better than FixInt and the main difference between BayesInt and FixInt is that the former uses an adaptive coefficient for interpolation, the results suggest that using adaptive coefficient is quite beneficial and a Bayesian style interpolation makes sense. The main difference between OnlineUp and BatchUp is that OnlineUp uses decaying coefficients to combine the multiple clicked summaries, while BatchUp simply concatenates all clicked summaries. Therefore the fact that BatchUp is consistently better than OnlineUp indicates that the weights for combining the clicked summaries indeed should not be decaying. While OnlineUp is theoretically appealing, its performance is inferior to BayesInt and BatchUp, likely because of the decaying coefficient. Overall, BatchUp appears to be the best method when we vary the parameter settings. We have two different kinds of search context - query history and clickthrough data. We now look into the contribution of each kind of context. 5.2.2 Using query history only In each of four models, we can turn off the clickthrough history data by setting parameters appropriately. This allows us to evaluate the effect of using query history alone. We use the same parameter setting for query history as in Table 1. The results are shown in Table 2. Here we see that in general, the benefit of using query history is very limited with mixed results. This is different from what is reported in a previous study [15], where using query history is consistently helpful. Another observation is that the context runs perform poorly at q2, but generally perform (slightly) better than the baselines for q3 and q4. This is again likely because at the beginning the initial query, which is the title in the original TREC topic description, may not be a good query; indeed, on average, performances of these first-generation queries are clearly poorer than those of all other user-formulated queries in the later generations. Yet another observation is that when using query history only, the BayesInt model appears to be better than other models. Since the clickthrough data is ignored, OnlineUp and BatchUp FixInt BayesInt OnlineUp BatchUp Query (α = 0.1, β = 0) (µ = 0.2,ν = 0) (µ = 5.0,ν = +∞) (µ = 2.0, ν = +∞) MAP pr@20docs MAP pr@20docs MAP pr@20docs MAP pr@20docs q2 0.0312 0.1150 0.0312 0.1150 0.0312 0.1150 0.0312 0.1150 q2 + HQ 0.0097 0.0317 0.0311 0.1200 0.0213 0.0783 0.0287 0.0967 Improve. -68.9% -72.4% -0.3% 4.3% -31.7% -31.9% -8.0% -15.9% q3 0.0421 0.1483 0.0421 0.1483 0.0421 0.1483 0.0421 0.1483 q3 + HQ 0.0261 0.0917 0.0451 0.1517 0.0444 0.1333 0.0455 0.1450 Improve -38.2% -38.2% 7.1% 2.3% 5.5% -10.1% 8.1% -2.2% q4 0.0536 0.1933 0.0536 0.1933 0.0536 0.1933 0.0536 0.1933 q4 + HQ 0.0428 0.1467 0.0537 0.1917 0.0550 0.1733 0.0552 0.1917 Improve -20.1% -24.1% 0.2% -0.8% 3.0% -10.3% 3.0% -0.8% Table 2: Effect of using query history only for document ranking. µ 0 0.5 1 2 3 4 5 6 7 8 9 q2 + HQ MAP 0.0312 0.0313 0.0308 0.0287 0.0257 0.0231 0.0213 0.0194 0.0183 0.0182 0.0164 q3 + HQ MAP 0.0421 0.0442 0.0441 0.0455 0.0457 0.0458 0.0444 0.0439 0.0430 0.0390 0.0335 q4 + HQ MAP 0.0536 0.0546 0.0547 0.0552 0.0544 0.0548 0.0550 0.0541 0.0534 0.0525 0.0513 Table 3: Average Precision of BatchUp using query history only are essentially the same algorithm. The displayed results thus reflect the variation caused by parameter µ. A smaller setting of 2.0 is seen better than a larger value of 5.0. A more complete picture of the influence of the setting of µ can be seen from Table 3, where we show the performance figures for a wider range of values of µ. The value of µ can be interpreted as how many words we regard the query history is worth. A larger value thus puts more weight on the history and is seen to hurt the performance more when the history information is not rich. Thus while for q4 the best performance tends to be achieved for µ ∈ [2, 5], only when µ = 0.5 we see some small benefit for q2. As we would expect, an excessively large µ would hurt the performance in general, but q2 is hurt most and q4 is barely hurt, indicating that as we accumulate more and more query history information, we can put more and more weight on the history information. This also suggests that a better strategy should probably dynamically adjust parameters according to how much history information we have. The mixed query history results suggest that the positive effect of using implicit feedback information may have largely come from the use of clickthrough history, which is indeed true as we discuss in the next subsection. 5.2.3 Using clickthrough history only We now turn off the query history and only use the clicked summaries plus the current query. The results are shown in Table 4. We see that the benefit of using clickthrough information is much more significant than that of using query history. We see an overall positive effect, often with significant improvement over the baseline. It is also clear that the richer the context data is, the more improvement using clicked summaries can achieve. Other than some occasional degradation of precision at 20 documents, the improvement is fairly consistent and often quite substantial. These results show that the clicked summary text is in general quite useful for inferring a user"s information need. Intuitively, using the summary text, rather than the actual content of the document, makes more sense, as it is quite possible that the document behind a seemingly relevant summary is actually non-relevant. 29 out of the 91 clicked documents are relevant. Updating the query model based on such summaries would bring up the ranks of these relevant documents, causing performance improvement. However, such improvement is really not beneficial for the user as the user has already seen these relevant documents. To see how much improvement we have achieved on improving the ranks of the unseen relevant documents, we exclude these 29 relevant documents from our judgment file and recompute the performance of BayesInt and the baseline using the new judgment file. The results are shown in Table 5. Note that the performance of the baseline method is lower due to the removal of the 29 relevant documents, which would have been generally ranked high in the results. From Table 5, we see clearly that using clicked summaries also helps improve the ranks of unseen relevant documents significantly. Query BayesInt(µ = 0, ν = 5.0) MAP pr@20docs q2 0.0263 0.100 q2 + HC 0.0314 0.100 Improve. 19.4% 0% q3 0.0331 0.125 q3 + HC 0.0661 0.178 Improve 99.7% 42.4% q4 0.0442 0.165 q4 + HC 0.0739 0.188 Improve 67.2% 13.9% Table 5: BayesInt evaluated on unseen relevant documents One remaining question is whether the clickthrough data is still helpful if none of the clicked documents is relevant. To answer this question, we took out the 29 relevant summaries from our clickthrough history data HC to obtain a smaller set of clicked summaries HC , and re-evaluated the performance of the BayesInt method using HC with the same setting of parameters as in Table 4. The results are shown in Table 6. We see that although the improvement is not as substantial as in Table 4, the average precision is improved across all generations of queries. These results should be interpreted as very encouraging as they are based on only 62 non-relevant clickthroughs. In reality, a user would more likely click some relevant summaries, which would help bring up more relevant documents as we have seen in Table 4 and Table 5. FixInt BayesInt OnlineUp BatchUp Query (α = 0.1, β = 1) (µ = 0, ν = 5.0) (µk = 5.0, ν = 15, ∀i < k, µi = +∞) (µ = 0, ν = 15) MAP pr@20docs MAP pr@20docs MAP pr@20docs MAP pr@20docs q2 0.0312 0.1150 0.0312 0.1150 0.0312 0.1150 0.0312 0.1150 q2 + HC 0.0324 0.1117 0.0338 0.1133 0.0358 0.1300 0.0344 0.1167 Improve. 3.8% -2.9% 8.3% -1.5% 14.7% 13.0% 10.3% 1.5% q3 0.0421 0.1483 0.0421 0.1483 0.04210 0.1483 0.0420 0.1483 q3 + HC 0.0726 0.1967 0.0766 0.2033 0.0622 0.1767 0.0513 0.1650 Improve 72.4% 32.6% 81.9% 37.1% 47.7% 19.2% 21.9% 11.3% q4 0.0536 0.1930 0.0536 0.1930 0.0536 0.1930 0.0536 0.1930 q4 + HC 0.0891 0.2233 0.0925 0.2283 0.0772 0.2217 0.0623 0.2050 Improve 66.2% 15.5% 72.6% 18.1% 44.0% 14.7% 16.2% 6.1% Table 4: Effect of using clickthrough data only for document ranking. Query BayesInt(µ = 0, ν = 5.0) MAP pr@20docs q2 0.0312 0.1150 q2 + HC 0.0313 0.0950 Improve. 0.3% -17.4% q3 0.0421 0.1483 q3 + HC 0.0521 0.1820 Improve 23.8% 23.0% q4 0.0536 0.1930 q4 + HC 0.0620 0.1850 Improve 15.7% -4.1% Table 6: Effect of using only non-relevant clickthrough data 5.2.4 Additive effect of context information By comparing the results across Table 1, Table 2 and Table 4, we can see that the benefit of the query history information and that of clickthrough information are mostly additive, i.e., combining them can achieve better performance than using each alone, but most improvement has clearly come from the clickthrough information. In Table 7, we show this effect for the BatchUp method. 5.2.5 Parameter sensitivity All four models have two parameters to control the relative weights of HQ, HC , and Qk, though the parameterization is different from model to model. In this subsection, we study the parameter sensitivity for BatchUp, which appears to perform relatively better than others. BatchUp has two parameters µ and ν. We first look at µ. When µ is set to 0, the query history is not used at all, and we essentially just use the clickthrough data combined with the current query. If we increase µ, we will gradually incorporate more information from the previous queries. In Table 8, we show how the average precision of BatchUp changes as we vary µ with ν fixed to 15.0, where the best performance of BatchUp is achieved. We see that the performance is mostly insensitive to the change of µ for q3 and q4, but is decreasing as µ increases for q2. The pattern is also similar when we set ν to other values. In addition to the fact that q1 is generally worse than q2, q3, and q4, another possible reason why the sensitivity is lower for q3 and q4 may be that we generally have more clickthrough data available for q3 and q4 than for q2, and the dominating influence of the clickthrough data has made the small differences caused by µ less visible for q3 and q4. The best performance is generally achieved when µ is around 2.0, which means that the past query information is as useful as about 2 words in the current query. Except for q2, there is clearly some tradeoff between the current query and the previous queries Query MAP pr@20docs q2 0.0312 0.1150 q2 + HQ 0.0287 0.0967 Improve. -8.0% -15.9% q2 + HC 0.0344 0.1167 Improve. 10.3% 1.5% q2 + HQ + HC 0.0342 0.1100 Improve. 9.6% -4.3% q3 0.0421 0.1483 q3 + HQ 0.0455 0.1450 Improve 8.1% -2.2% q3 + HC 0.0513 0.1650 Improve 21.9% 11.3% q3 + HQ + HC 0.0810 0.2067 Improve 92.4% 39.4% q4 0.0536 0.1930 q4 + HQ 0.0552 0.1917 Improve 3.0% -0.8% q4 + HC 0.0623 0.2050 Improve 16.2% 6.1% q4 + HQ + HC 0.0950 0.2250 Improve 77.2% 16.4% Table 7: Additive benefit of context information and using a balanced combination of them achieves better performance than using each of them alone. We now turn to the other parameter ν. When ν is set to 0, we only use the clickthrough data; When ν is set to +∞, we only use the query history and the current query. With µ set to 2.0, where the best performance of BatchUp is achieved, we vary ν and show the results in Table 9. We see that the performance is also not very sensitive when ν ≤ 30, with the best performance often achieved at ν = 15. This means that the combined information of query history and the current query is as useful as about 15 words in the clickthrough data, indicating that the clickthrough information is highly valuable. Overall, these sensitivity results show that BatchUp not only performs better than other methods, but also is quite robust. 6. CONCLUSIONS AND FUTURE WORK In this paper, we have explored how to exploit implicit feedback information, including query history and clickthrough history within the same search session, to improve information retrieval performance. Using the KL-divergence retrieval model as the basis, we proposed and studied four statistical language models for context-sensitive information retrieval, i.e., FixInt, BayesInt, OnlineUp and BatchUp. We use TREC AP Data to create a test set µ 0 1 2 3 4 5 6 7 8 9 10 MAP 0.0386 0.0366 0.0342 0.0315 0.0290 0.0267 0.0250 0.0236 0.0229 0.0223 0.0219 q2 + HQ + HC pr@20 0.1333 0.1233 0.1100 0.1033 0.1017 0.0933 0.0833 0.0767 0.0783 0.0767 0.0750 MAP 0.0805 0.0807 0.0811 0.0814 0.0813 0.0808 0.0804 0.0799 0.0795 0.0790 0.0788 q3 + HQ + HC pr@20 0.210 0.2150 0.2067 0.205 0.2067 0.205 0.2067 0.2067 0.2050 0.2017 0.2000 MAP 0.0929 0.0947 0.0950 0.0940 0.0941 0.0940 0.0942 0.0937 0.0936 0.0932 0.0929 q4 + HQ + HC pr@20 0.2183 0.2217 0.2250 0.2217 0.2233 0.2267 0.2283 0.2333 0.2333 0.2350 0.2333 Table 8: Sensitivity of µ in BatchUp ν 0 1 2 5 10 15 30 100 300 500 MAP 0.0278 0.0287 0.0296 0.0315 0.0334 0.0342 0.0328 0.0311 0.0296 0.0290 q2 + HQ + HC pr@20 0.0933 0.0950 0.0950 0.1000 0.1050 0.1100 0.1150 0.0983 0.0967 0.0967 MAP 0.0728 0.0739 0.0751 0.0786 0.0809 0.0811 0.0770 0.0634 0.0511 0.0491 q3 + HQ + HC pr@20 0.1917 0.1933 0.1950 0.2100 0.2000 0.2067 0.2017 0.1783 0.1600 0.1550 MAP 0.0895 0.0903 0.0914 0.0932 0.0944 0.0950 0.0919 0.0761 0.0664 0.0625 q4 + HQ + HC pr@20 0.2267 0.2233 0.2283 0.2317 0.2233 0.2250 0.2283 0.2200 0.2067 0.2033 Table 9: Sensitivity of ν in BatchUp for evaluating implicit feedback models. Experiment results show that using implicit feedback, especially clickthrough history, can substantially improve retrieval performance without requiring any additional user effort. The current work can be extended in several ways: First, we have only explored some very simple language models for incorporating implicit feedback information. It would be interesting to develop more sophisticated models to better exploit query history and clickthrough history. For example, we may treat a clicked summary differently depending on whether the current query is a generalization or refinement of the previous query. Second, the proposed models can be implemented in any practical systems. We are currently developing a client-side personalized search agent, which will incorporate some of the proposed algorithms. We will also do a user study to evaluate effectiveness of these models in the real web search. Finally, we should further study a general retrieval framework for sequential decision making in interactive information retrieval and study how to optimize some of the parameters in the context-sensitive retrieval models. 7. ACKNOWLEDGMENTS This material is based in part upon work supported by the National Science Foundation under award numbers IIS-0347933 and IIS-0428472. We thank the anonymous reviewers for their useful comments. 8. REFERENCES [1] E. Adar and D. Karger. Haystack: Per-user information environments. In Proceedings of CIKM 1999, 1999. [2] J. Allan and et al. Challenges in information retrieval and language modeling. Workshop at University of Amherst, 2002. [3] K. Bharat. Searchpad: Explicit capture of search context to support web search. In Proceeding of WWW 2000, 2000. [4] W. B. Croft, S. Cronen-Townsend, and V. Larvrenko. Relevance feedback and personalization: A language modeling perspective. In Proeedings of Second DELOS Workshop: Personalisation and Recommender Systems in Digital Libraries, 2001. [5] H. Cui, J.-R. Wen, J.-Y. Nie, and W.-Y. Ma. Probabilistic query expansion using query logs. In Proceedings of WWW 2002, 2002. [6] S. T. Dumais, E. Cutrell, R. Sarin, and E. Horvitz. Implicit queries (IQ) for contextualized search (demo description). In Proceedings of SIGIR 2004, page 594, 2004. [7] L. Finkelstein, E. Gabrilovich, Y. Matias, E. Rivlin, Z. Solan, G. Wolfman, and E. Ruppin. Placing search in context: The concept revisited. In Proceedings of WWW 2002, 2001. [8] C. Huang, L. Chien, and Y. Oyang. Query session based term suggestion for interactive web search. In Proceedings of WWW 2001, 2001. [9] X. Huang, F. Peng, A. An, and D. Schuurmans. Dynamic web log session identification with statistical language models. Journal of the American Society for Information Science and Technology, 55(14):1290-1303, 2004. [10] G. Jeh and J. Widom. Scaling personalized web search. In Proceeding of WWW 2003, 2003. [11] T. Joachims. Optimizing search engines using clickthrough data. In Proceedings of SIGKDD 2002, 2002. [12] D. Kelly and N. J. Belkin. Display time as implicit feedback: Understanding task effects. In Proceedings of SIGIR 2004, 2004. [13] D. Kelly and J. Teevan. Implicit feedback for inferring user preference. SIGIR Forum, 32(2), 2003. [14] J. Rocchio. Relevance feedback information retrieval. In The Smart Retrieval System-Experiments in Automatic Document Processing, pages 313-323, Kansas City, MO, 1971. Prentice-Hall. [15] X. Shen and C. Zhai. Exploiting query history for document ranking in interactive information retrieval (poster). In Proceedings of SIGIR 2003, 2003. [16] S. Sriram, X. Shen, and C. Zhai. A session-based search engine (poster). In Proceedings of SIGIR 2004, 2004. [17] K. Sugiyama, K. Hatano, and M. Yoshikawa. Adaptive web search based on user profile constructed without any effort from users. In Proceedings of WWW 2004, 2004. [18] R. W. White, J. M. Jose, C. J. van Rijsbergen, and I. Ruthven. A simulated study of implicit feedback models. In Proceedings of ECIR 2004, pages 311-326, 2004. [19] C. Zhai and J. Lafferty. Model-based feedback in the KL-divergence retrieval model. In Proceedings of CIKM 2001, 2001. [20] C. Zhai and J. Lafferty. A study of smoothing methods for language models applied to ad-hoc information retrieval. In Proceedings of SIGIR 2001, 2001.

kl-divergence retrieval model;context-sensitive language;long-term context;interactive retrieval;trec datum set;query expansion;mean average precision;fixed coefficient interpolation;query history information;implicit feedback information;current query;query history;context;retrieval accuracy;relevance feedback;clickthrough information;short-term context;bayesian estimation

train_H-92

Improving Web Search Ranking by Incorporating User Behavior Information

We show that incorporating user behavior data can significantly improve ordering of top results in real web search setting. We examine alternatives for incorporating feedback into the ranking process and explore the contributions of user feedback compared to other common web search features. We report results of a large scale evaluation over 3,000 queries and 12 million user interactions with a popular web search engine. We show that incorporating implicit feedback can augment other features, improving the accuracy of a competitive web search ranking algorithms by as much as 31% relative to the original performance.

1. INTRODUCTION Millions of users interact with search engines daily. They issue queries, follow some of the links in the results, click on ads, spend time on pages, reformulate their queries, and perform other actions. These interactions can serve as a valuable source of information for tuning and improving web search result ranking and can compliment more costly explicit judgments. Implicit relevance feedback for ranking and personalization has become an active area of research. Recent work by Joachims and others exploring implicit feedback in controlled environments have shown the value of incorporating implicit feedback into the ranking process. Our motivation for this work is to understand how implicit feedback can be used in a large-scale operational environment to improve retrieval. How does it compare to and compliment evidence from page content, anchor text, or link-based features such as inlinks or PageRank? While it is intuitive that user interactions with the web search engine should reveal at least some information that could be used for ranking, estimating user preferences in real web search settings is a challenging problem, since real user interactions tend to be more noisy than commonly assumed in the controlled settings of previous studies. Our paper explores whether implicit feedback can be helpful in realistic environments, where user feedback can be noisy (or adversarial) and a web search engine already uses hundreds of features and is heavily tuned. To this end, we explore different approaches for ranking web search results using real user behavior obtained as part of normal interactions with the web search engine. The specific contributions of this paper include: • Analysis of alternatives for incorporating user behavior into web search ranking (Section 3). • An application of a robust implicit feedback model derived from mining millions of user interactions with a major web search engine (Section 4). • A large scale evaluation over real user queries and search results, showing significant improvements derived from incorporating user feedback (Section 6). We summarize our findings and discuss extensions to the current work in Section 7, which concludes the paper. 2. BACKGROUND AND RELATED WORK Ranking search results is a fundamental problem in information retrieval. Most common approaches primarily focus on similarity of query and a page, as well as the overall page quality [3,4,24]. However, with increasing popularity of search engines, implicit feedback (i.e., the actions users take when interacting with the search engine) can be used to improve the rankings. Implicit relevance measures have been studied by several research groups. An overview of implicit measures is compiled in Kelly and Teevan [14]. This research, while developing valuable insights into implicit relevance measures, was not applied to improve the ranking of web search results in realistic settings. Closely related to our work, Joachims [11] collected implicit measures in place of explicit measures, introducing a technique based entirely on clickthrough data to learn ranking functions. Fox et al. [8] explored the relationship between implicit and explicit measures in Web search, and developed Bayesian models to correlate implicit measures and explicit relevance judgments for both individual queries and search sessions. This work considered a wide range of user behaviors (e.g., dwell time, scroll time, reformulation patterns) in addition to the popular clickthrough behavior. However, the modeling effort was aimed at predicting explicit relevance judgments from implicit user actions and not specifically at learning ranking functions. Other studies of user behavior in web search include Pharo and Järvelin [19], but were not directly applied to improve ranking. More recently, Joachims et al. [12] presented an empirical evaluation of interpreting clickthrough evidence. By performing eye tracking studies and correlating predictions of their strategies with explicit ratings, the authors showed that it is possible to accurately interpret clickthroughs in a controlled, laboratory setting. Unfortunately, the extent to which previous research applies to real-world web search is unclear. At the same time, while recent work (e.g., [26]) on using clickthrough information for improving web search ranking is promising, it captures only one aspect of the user interactions with web search engines. We build on existing research to develop robust user behavior interpretation techniques for the real web search setting. Instead of treating each user as a reliable expert, we aggregate information from multiple, unreliable, user search session traces, as we describe in the next two sections. 3. INCORPORATING IMPLICIT FEEDBACK We consider two complementary approaches to ranking with implicit feedback: (1) treating implicit feedback as independent evidence for ranking results, and (2) integrating implicit feedback features directly into the ranking algorithm. We describe the two general ranking approaches next. The specific implicit feedback features are described in Section 4, and the algorithms for interpreting and incorporating implicit feedback are described in Section 5. 3.1 Implicit Feedback as Independent Evidence The general approach is to re-rank the results obtained by a web search engine according to observed clickthrough and other user interactions for the query in previous search sessions. Each result is assigned a score according to expected relevance/user satisfaction based on previous interactions, resulting in some preference ordering based on user interactions alone. While there has been significant work on merging multiple rankings, we adapt a simple and robust approach of ignoring the original rankers" scores, and instead simply merge the rank orders. The main reason for ignoring the original scores is that since the feature spaces and learning algorithms are different, the scores are not directly comparable, and re-normalization tends to remove the benefit of incorporating classifier scores. We experimented with a variety of merging functions on the development set of queries (and using a set of interactions from a different time period from final evaluation sets). We found that a simple rank merging heuristic combination works well, and is robust to variations in score values from original rankers. For a given query q, the implicit score ISd is computed for each result d from available user interaction features, resulting in the implicit rank Id for each result. We compute a merged score SM(d) for d by combining the ranks obtained from implicit feedback, Id with the original rank of d, Od: ¡ ¢ £ + + + + = otherwise O dforexistsfeedbackimplicitif OI w wOIdS d dd I IddM 1 1 1 1 1 1 ),,,( where the weight wI is a heuristically tuned scaling factor representing the relative importance of the implicit feedback. The query results are ordered in by decreasing values of SM to produce the final ranking. One special case of this model arises when setting wI to a very large value, effectively forcing clicked results to be ranked higher than un-clicked results - an intuitive and effective heuristic that we will use as a baseline. Applying more sophisticated classifier and ranker combination algorithms may result in additional improvements, and is a promising direction for future work. The approach above assumes that there are no interactions between the underlying features producing the original web search ranking and the implicit feedback features. We now relax this assumption by integrating implicit feedback features directly into the ranking process. 3.2 Ranking with Implicit Feedback Features Modern web search engines rank results based on a large number of features, including content-based features (i.e., how closely a query matches the text or title or anchor text of the document), and query-independent page quality features (e.g., PageRank of the document or the domain). In most cases, automatic (or semiautomatic) methods are developed for tuning the specific ranking function that combines these feature values. Hence, a natural approach is to incorporate implicit feedback features directly as features for the ranking algorithm. During training or tuning, the ranker can be tuned as before but with additional features. At runtime, the search engine would fetch the implicit feedback features associated with each query-result URL pair. This model requires a ranking algorithm to be robust to missing values: more than 50% of queries to web search engines are unique, with no previous implicit feedback available. We now describe such a ranker that we used to learn over the combined feature sets including implicit feedback. 3.3 Learning to Rank Web Search Results A key aspect of our approach is exploiting recent advances in machine learning, namely trainable ranking algorithms for web search and information retrieval (e.g., [5, 11] and classical results reviewed in [3]). In our setting, explicit human relevance judgments (labels) are available for a set of web search queries and results. Hence, an attractive choice to use is a supervised machine learning technique to learn a ranking function that best predicts relevance judgments. RankNet is one such algorithm. It is a neural net tuning algorithm that optimizes feature weights to best match explicitly provided pairwise user preferences. While the specific training algorithms used by RankNet are beyond the scope of this paper, it is described in detail in [5] and includes extensive evaluation and comparison with other ranking methods. An attractive feature of RankNet is both train- and run-time efficiency - runtime ranking can be quickly computed and can scale to the web, and training can be done over thousands of queries and associated judged results. We use a 2-layer implementation of RankNet in order to model non-linear relationships between features. Furthermore, RankNet can learn with many (differentiable) cost functions, and hence can automatically learn a ranking function from human-provided labels, an attractive alternative to heuristic feature combination techniques. Hence, we will also use RankNet as a generic ranker to explore the contribution of implicit feedback for different ranking alternatives. 4. IMPLICIT USER FEEDBACK MODEL Our goal is to accurately interpret noisy user feedback obtained as by tracing user interactions with the search engine. Interpreting implicit feedback in real web search setting is not an easy task. We characterize this problem in detail in [1], where we motivate and evaluate a wide variety of models of implicit user activities. The general approach is to represent user actions for each search result as a vector of features, and then train a ranker on these features to discover feature values indicative of relevant (and nonrelevant) search results. We first briefly summarize our features and model, and the learning approach (Section 4.2) in order to provide sufficient information to replicate our ranking methods and the subsequent experiments. 4.1 Representing User Actions as Features We model observed web search behaviors as a combination of a ``background"" component (i.e., query- and relevance-independent noise in user behavior, including positional biases with result interactions), and a ``relevance"" component (i.e., query-specific behavior indicative of relevance of a result to a query). We design our features to take advantage of aggregated user behavior. The feature set is comprised of directly observed features (computed directly from observations for each query), as well as queryspecific derived features, computed as the deviation from the overall query-independent distribution of values for the corresponding directly observed feature values. The features used to represent user interactions with web search results are summarized in Table 4.1. This information was obtained via opt-in client-side instrumentation from users of a major web search engine. We include the traditional implicit feedback features such as clickthrough counts for the results, as well as our novel derived features such as the deviation of the observed clickthrough number for a given query-URL pair from the expected number of clicks on a result in the given position. We also model the browsing behavior after a result was clicked - e.g., the average page dwell time for a given query-URL pair, as well as its deviation from the expected (average) dwell time. Furthermore, the feature set was designed to provide essential information about the user experience to make feedback interpretation robust. For example, web search users can often determine whether a result is relevant by looking at the result title, URL, and summary - in many cases, looking at the original document is not necessary. To model this aspect of user experience we include features such as overlap in words in title and words in query (TitleOverlap) and the fraction of words shared by the query and the result summary. Clickthrough features Position Position of the URL in Current ranking ClickFrequency Number of clicks for this query, URL pair ClickProbability Probability of a click for this query and URL ClickDeviation Deviation from expected click probability IsNextClicked 1 if clicked on next position, 0 otherwise IsPreviousClicked 1 if clicked on previous position, 0 otherwise IsClickAbove 1 if there is a click above, 0 otherwise IsClickBelow 1 if there is click below, 0 otherwise Browsing features TimeOnPage Page dwell time CumulativeTimeOnPage Cumulative time for all subsequent pages after search TimeOnDomain Cumulative dwell time for this domain TimeOnShortUrl Cumulative time on URL prefix, no parameters IsFollowedLink 1 if followed link to result, 0 otherwise IsExactUrlMatch 0 if aggressive normalization used, 1 otherwise IsRedirected 1 if initial URL same as final URL, 0 otherwise IsPathFromSearch 1 if only followed links after query, 0 otherwise ClicksFromSearch Number of hops to reach page from query AverageDwellTime Average time on page for this query DwellTimeDeviation Deviation from average dwell time on page CumulativeDeviation Deviation from average cumulative dwell time DomainDeviation Deviation from average dwell time on domain Query-text features TitleOverlap Words shared between query and title SummaryOverlap Words shared between query and snippet QueryURLOverlap Words shared between query and URL QueryDomainOverlap Words shared between query and URL domain QueryLength Number of tokens in query QueryNextOverlap Fraction of words shared with next query Table 4.1: Some features used to represent post-search navigation history for a given query and search result URL. Having described our feature set, we briefly review our general method for deriving a user behavior model. 4.2 Deriving a User Feedback Model To learn to interpret the observed user behavior, we correlate user actions (i.e., the features in Table 4.1 representing the actions) with the explicit user judgments for a set of training queries. We find all the instances in our session logs where these queries were submitted to the search engine, and aggregate the user behavior features for all search sessions involving these queries. Each observed query-URL pair is represented by the features in Table 4.1, with values averaged over all search sessions, and assigned one of six possible relevance labels, ranging from Perfect to Bad, as assigned by explicit relevance judgments. These labeled feature vectors are used as input to the RankNet training algorithm (Section 3.3) which produces a trained user behavior model. This approach is particularly attractive as it does not require heuristics beyond feature engineering. The resulting user behavior model is used to help rank web search resultseither directly or in combination with other features, as described below. 5. EXPERIMENTAL SETUP The ultimate goal of incorporating implicit feedback into ranking is to improve the relevance of the returned web search results. Hence, we compare the ranking methods over a large set of judged queries with explicit relevance labels provided by human judges. In order for the evaluation to be realistic we obtained a random sample of queries from web search logs of a major search engine, with associated results and traces for user actions. We describe this dataset in detail next. Our metrics are described in Section 5.2 that we use to evaluate the ranking alternatives, listed in Section 5.3 in the experiments of Section 6. 5.1 Datasets We compared our ranking methods over a random sample of 3,000 queries from the search engine query logs. The queries were drawn from the logs uniformly at random by token without replacement, resulting in a query sample representative of the overall query distribution. On average, 30 results were explicitly labeled by human judges using a six point scale ranging from Perfect down to Bad. Overall, there were over 83,000 results with explicit relevance judgments. In order to compute various statistics, documents with label Good or better will be considered relevant, and with lower labels to be non-relevant. Note that the experiments were performed over the results already highly ranked by a web search engine, which corresponds to a typical user experience which is limited to the small number of the highly ranked results for a typical web search query. The user interactions were collected over a period of 8 weeks using voluntary opt-in information. In total, over 1.2 million unique queries were instrumented, resulting in over 12 million individual interactions with the search engine. The data consisted of user interactions with the web search engine (e.g., clicking on a result link, going back to search results, etc.) performed after a query was submitted. These actions were aggregated across users and search sessions and converted to features in Table 4.1. To create the training, validation, and test query sets, we created three different random splits of 1,500 training, 500 validation, and 1000 test queries. The splits were done randomly by query, so that there was no overlap in training, validation, and test queries. 5.2 Evaluation Metrics We evaluate the ranking algorithms over a range of accepted information retrieval metrics, namely Precision at K (P(K)), Normalized Discounted Cumulative Gain (NDCG), and Mean Average Precision (MAP). Each metric focuses on a deferent aspect of system performance, as we describe below. • Precision at K: As the most intuitive metric, P(K) reports the fraction of documents ranked in the top K results that are labeled as relevant. In our setting, we require a relevant document to be labeled Good or higher. The position of relevant documents within the top K is irrelevant, and hence this metric measure overall user satisfaction with the top K results. • NDCG at K: NDCG is a retrieval measure devised specifically for web search evaluation [10]. For a given query q, the ranked results are examined from the top ranked down, and the NDCG computed as: = +−= K j jr qq jMN 1 )( )1log(/)12( Where Mq is a normalization constant calculated so that a perfect ordering would obtain NDCG of 1; and each r(j) is an integer relevance label (0=Bad and 5=Perfect) of result returned at position j. Note that unlabeled and Bad documents do not contribute to the sum, but will reduce NDCG for the query pushing down the relevant labeled documents, reducing their contributions. NDCG is well suited to web search evaluation, as it rewards relevant documents in the top ranked results more heavily than those ranked lower. • MAP: Average precision for each query is defined as the mean of the precision at K values computed after each relevant document was retrieved. The final MAP value is defined as the mean of average precisions of all queries in the test set. This metric is the most commonly used single-value summary of a run over a set of queries. 5.3 Ranking Methods Compared Recall that our goal is to quantify the effectiveness of implicit behavior for real web search. One dimension is to compare the utility of implicit feedback with other information available to a web search engine. Specifically, we compare effectiveness of implicit user behaviors with content-based matching, static page quality features, and combinations of all features. • BM25F: As a strong web search baseline we used the BM25F scoring, which was used in one of the best performing systems in the TREC 2004 Web track [23,27]. BM25F and its variants have been extensively described and evaluated in IR literature, and hence serve as a strong, reproducible baseline. The BM25F variant we used for our experiments computes separate match scores for each field for a result document (e.g., body text, title, and anchor text), and incorporates query-independent linkbased information (e.g., PageRank, ClickDistance, and URL depth). The scoring function and field-specific tuning is described in detail in [23]. Note that BM25F does not directly consider explicit or implicit feedback for tuning. • RN: The ranking produced by a neural net ranker (RankNet, described in Section 3.3) that learns to rank web search results by incorporating BM25F and a large number of additional static and dynamic features describing each search result. This system automatically learns weights for all features (including the BM25F score for a document) based on explicit human labels for a large set of queries. A system incorporating an implementation of RankNet is currently in use by a major search engine and can be considered representative of the state of the art in web search. • BM25F-RerankCT: The ranking produced by incorporating clickthrough statistics to reorder web search results ranked by BM25F above. Clickthrough is a particularly important special case of implicit feedback, and has been shown to correlate with result relevance. This is a special case of the ranking method in Section 3.1, with the weight wI set to 1000 and the ranking Id is simply the number of clicks on the result corresponding to d. In effect, this ranking brings to the top all returned web search results with at least one click (and orders them in decreasing order by number of clicks). The relative ranking of the remainder of results is unchanged and they are inserted below all clicked results. This method serves as our baseline implicit feedback reranking method. BM25F-RerankAll The ranking produced by reordering the BM25F results using all user behavior features (Section 4). This method learns a model of user preferences by correlating feature values with explicit relevance labels using the RankNet neural net algorithm (Section 4.2). At runtime, for a given query the implicit score Ir is computed for each result r with available user interaction features, and the implicit ranking is produced. The merged ranking is computed as described in Section 3.1. Based on the experiments over the development set we fix the value of wI to 3 (the effect of the wI parameter for this ranker turned out to be negligible). • BM25F+All: Ranking derived by training the RankNet (Section 3.3) learner over the features set of the BM25F score as well as all implicit feedback features (Section 3.2). We used the 2-layer implementation of RankNet [5] trained on the queries and labels in the training and validation sets. • RN+All: Ranking derived by training the 2-layer RankNet ranking algorithm (Section 3.3) over the union of all content, dynamic, and implicit feedback features (i.e., all of the features described above as well as all of the new implicit feedback features we introduced). The ranking methods above span the range of the information used for ranking, from not using the implicit or explicit feedback at all (i.e., BM25F) to a modern web search engine using hundreds of features and tuned on explicit judgments (RN). As we will show next, incorporating user behavior into these ranking systems dramatically improves the relevance of the returned documents. 6. EXPERIMENTAL RESULTS Implicit feedback for web search ranking can be exploited in a number of ways. We compare alternative methods of exploiting implicit feedback, both by re-ranking the top results (i.e., the BM25F-RerankCT and BM25F-RerankAll methods that reorder BM25F results), as well as by integrating the implicit features directly into the ranking process (i.e., the RN+ALL and BM25F+All methods which learn to rank results over the implicit feedback and other features). We compare our methods over strong baselines (BM25F and RN) over the NDCG, Precision at K, and MAP measures defined in Section 5.2. The results were averaged over three random splits of the overall dataset. Each split contained 1500 training, 500 validation, and 1000 test queries, all query sets disjoint. We first present the results over all 1000 test queries (i.e., including queries for which there are no implicit measures so we use the original web rankings). We then drill down to examine the effects on reranking for the attempted queries in more detail, analyzing where implicit feedback proved most beneficial. We first experimented with different methods of re-ranking the output of the BM25F search results. Figures 6.1 and 6.2 report NDCG and Precision for BM25F, as well as for the strategies reranking results with user feedback (Section 3.1). Incorporating all user feedback (either in reranking framework or as features to the learner directly) results in significant improvements (using two-tailed t-test with p=0.01) over both the original BM25F ranking as well as over reranking with clickthrough alone. The improvement is consistent across the top 10 results and largest for the top result: NDCG at 1 for BM25F+All is 0.622 compared to 0.518 of the original results, and precision at 1 similarly increases from 0.5 to 0.63. Based on these results we will use the direct feature combination (i.e., BM25F+All) ranker for subsequent comparisons involving implicit feedback. 0.5 0.52 0.54 0.56 0.58 0.6 0.62 0.64 0.66 0.68 1 2 3 4 5 6 7 8 9 10K NDCG BM25 BM25-Rerank-CT BM25-Rerank-All BM25+All Figure 6.1: NDCG at K for BM25F, BM25F-RerankCT, BM25F-Rerank-All, and BM25F+All for varying K 0.35 0.4 0.45 0.5 0.55 0.6 0.65 1 3 5 10 K Precision BM25 BM25-Rerank-CT BM25-Rerank-All BM25+All Figure 6.2: Precision at K for BM25F, BM25F-RerankCT, BM25F-Rerank-All, and BM25F+All for varying K Interestingly, using clickthrough alone, while giving significant benefit over the original BM25F ranking, is not as effective as considering the full set of features in Table 4.1. While we analyze user behavior (and most effective component features) in a separate paper [1], it is worthwhile to give a concrete example of the kind of noise inherent in real user feedback in web search setting. 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 1 2 3 5 Result position Relativeclickfrequency PTR=2 PTR=3 PTR=5 Figure 6.3: Relative clickthrough frequency for queries with varying Position of Top Relevant result (PTR). If users considered only the relevance of a result to their query, they would click on the topmost relevant results. Unfortunately, as Joachims and others have shown, presentation also influences which results users click on quite dramatically. Users often click on results above the relevant one presumably because the short summaries do not provide enough information to make accurate relevance assessments and they have learned that on average topranked items are relevant. Figure 6.3 shows relative clickthrough frequencies for queries with known relevant items at positions other than the first position; the position of the top relevant result (PTR) ranges from 2-10 in the figure. For example, for queries with first relevant result at position 5 (PTR=5), there are more clicks on the non-relevant results in higher ranked positions than on the first relevant result at position 5. As we will see, learning over a richer behavior feature set, results in substantial accuracy improvement over clickthrough alone. We now consider incorporating user behavior into a much richer feature set, RN (Section 5.3) used by a major web search engine. RN incorporates BM25F, link-based features, and hundreds of other features. Figure 6.4 reports NDCG at K and Figure 6.5 reports Precision at K. Interestingly, while the original RN rankings are significantly more accurate than BM25F alone, incorporating implicit feedback features (BM25F+All) results in ranking that significantly outperforms the original RN rankings. In other words, implicit feedback incorporates sufficient information to replace the hundreds of other features available to the RankNet learner trained on the RN feature set. 0.5 0.52 0.54 0.56 0.58 0.6 0.62 0.64 0.66 0.68 0.7 1 2 3 4 5 6 7 8 9 10K NDCG RN RN+All BM25 BM25+All Figure 6.4: NDCG at K for BM25F, BM25F+All, RN, and RN+All for varying K Furthermore, enriching the RN features with implicit feedback set exhibits significant gain on all measures, allowing RN+All to outperform all other methods. This demonstrates the complementary nature of implicit feedback with other features available to a state of the art web search engine. 0.4 0.45 0.5 0.55 0.6 0.65 1 3 5 10 K Precision RN RN+All BM25 BM25+All Figure 6.5: Precision at K for BM25F, BM25F+All, RN, and RN+All for varying K We summarize the performance of the different ranking methods in Table 6.1. We report the Mean Average Precision (MAP) score for each system. While not intuitive to interpret, MAP allows quantitative comparison on a single metric. The gains marked with * are significant at p=0.01 level using two tailed t-test. MAP Gain P(1) Gain BM25F 0.184 - 0.503BM25F-Rerank-CT 0.215 0.031* 0.577 0.073* BM25F-RerankImplicit 0.218 0.003 0.605 0.028* BM25F+Implicit 0.222 0.004 0.620 0.015* RN 0.215 - 0.597RN+All 0.248 0.033* 0.629 0.032* Table 6.1: Mean Average Precision (MAP) for all strategies. So far we reported results averaged across all queries in the test set. Unfortunately, less than half had sufficient interactions to attempt reranking. Out of the 1000 queries in test, between 46% and 49%, depending on the train-test split, had sufficient interaction information to make predictions (i.e., there was at least 1 search session in which at least 1 result URL was clicked on by the user). This is not surprising: web search is heavy-tailed, and there are many unique queries. We now consider the performance on the queries for which user interactions were available. Figure 6.6 reports NDCG for the subset of the test queries with the implicit feedback features. The gains at top 1 are dramatic. The NDCG at 1 of BM25F+All increases from 0.6 to 0.75 (a 31% relative gain), achieving performance comparable to RN+All operating over a much richer feature set. 0.6 0.65 0.7 0.75 0.8 1 3 5 10K NDCG RN RN+All BM25 BM25+All Figure 6.6: NDCG at K for BM25F, BM25F+All, RN, and RN+All on test queries with user interactions Similarly, gains on precision at top 1 are substantial (Figure 6.7), and are likely to be apparent to web search users. When implicit feedback is available, the BM25F+All system returns relevant document at top 1 almost 70% of the time, compared 53% of the time when implicit feedback is not considered by the original BM25F system. 0.45 0.5 0.55 0.6 0.65 0.7 1 3 5 10K Precision RN RN+All BM25 BM25+All Figure 6.7: Precision at K NDCG at K for BM25F, BM25F+All, RN, and RN+All on test queries with user interactions We summarize the results on the MAP measure for attempted queries in Table 6.2. MAP improvements are both substantial and significant, with improvements over the BM25F ranker most pronounced. Method MAP Gain P(1) Gain RN 0.269 0.632 RN+All 0.321 0.051 (19%) 0.693 0.061(10%) BM25F 0.236 0.525 BM25F+All 0.292 0.056 (24%) 0.687 0.162 (31%) Table 6.2: Mean Average Precision (MAP) on attempted queries for best performing methods We now analyze the cases where implicit feedback was shown most helpful. Figure 6.8 reports the MAP improvements over the baseline BM25F run for each query with MAP under 0.6. Note that most of the improvement is for poorly performing queries (i.e., MAP < 0.1). Interestingly, incorporating user behavior information degrades accuracy for queries with high original MAP score. One possible explanation is that these easy queries tend to be navigational (i.e., having a single, highly-ranked most appropriate answer), and user interactions with lower-ranked results may indicate divergent information needs that are better served by the less popular results (with correspondingly poor overall relevance ratings). 0 50 100 150 200 250 300 350 0.1 0.2 0.3 0.4 0.5 0.6 -0.4 -0.35 -0.3 -0.25 -0.2 -0.15 -0.1 -0.05 0 0.05 0.1 0.15 0.2 Frequency Average Gain Figure 6.8: Gain of BM25F+All over original BM25F ranking To summarize our experimental results, incorporating implicit feedback in real web search setting resulted in significant improvements over the original rankings, using both BM25F and RN baselines. Our rich set of implicit features, such as time on page and deviations from the average behavior, provides advantages over using clickthrough alone as an indicator of interest. Furthermore, incorporating implicit feedback features directly into the learned ranking function is more effective than using implicit feedback for reranking. The improvements observed over large test sets of queries (1,000 total, between 466 and 495 with implicit feedback available) are both substantial and statistically significant. 7. CONCLUSIONS AND FUTURE WORK In this paper we explored the utility of incorporating noisy implicit feedback obtained in a real web search setting to improve web search ranking. We performed a large-scale evaluation over 3,000 queries and more than 12 million user interactions with a major search engine, establishing the utility of incorporating noisy implicit feedback to improve web search relevance. We compared two alternatives of incorporating implicit feedback into the search process, namely reranking with implicit feedback and incorporating implicit feedback features directly into the trained ranking function. Our experiments showed significant improvement over methods that do not consider implicit feedback. The gains are particularly dramatic for the top K=1 result in the final ranking, with precision improvements as high as 31%, and the gains are substantial for all values of K. Our experiments showed that implicit user feedback can further improve web search performance, when incorporated directly with popular content- and link-based features. Interestingly, implicit feedback is particularly valuable for queries with poor original ranking of results (e.g., MAP lower than 0.1). One promising direction for future work is to apply recent research on automatically predicting query difficulty, and only attempt to incorporate implicit feedback for the difficult queries. As another research direction we are exploring methods for extending our predictions to the previously unseen queries (e.g., query clustering), which should further improve the web search experience of users. ACKNOWLEDGMENTS We thank Chris Burges and Matt Richardson for an implementation of RankNet for our experiments. We also thank Robert Ragno for his valuable suggestions and many discussions. 8. REFERENCES [1] E. Agichtein, E. Brill, S. Dumais, and R.Ragno, Learning User Interaction Models for Predicting Web Search Result Preferences. In Proceedings of the ACM Conference on Research and Development on Information Retrieval (SIGIR), 2006 [2] J. Allan, HARD Track Overview in TREC 2003, High Accuracy Retrieval from Documents, 2003 [3] R. Baeza-Yates and B. Ribeiro-Neto, Modern Information Retrieval, Addison-Wesley, 1999. [4] S. Brin and L. Page, The Anatomy of a Large-scale Hypertextual Web Search Engine, in Proceedings of WWW, 1997 [5] C.J.C. Burges, T. Shaked, E. Renshaw, A. Lazier, M. Deeds, N. Hamilton, G. Hullender, Learning to Rank using Gradient Descent, in Proceedings of the International Conference on Machine Learning, 2005 [6] D.M. Chickering, The WinMine Toolkit, Microsoft Technical Report MSR-TR-2002-103, 2002 [7] M. Claypool, D. Brown, P. Lee and M. Waseda. Inferring user interest. IEEE Internet Computing. 2001 [8] S. Fox, K. Karnawat, M. Mydland, S. T. Dumais and T. White. Evaluating implicit measures to improve the search experience. In ACM Transactions on Information Systems, 2005 [9] J. Goecks and J. Shavlick. Learning users" interests by unobtrusively observing their normal behavior. In Proceedings of the IJCAI Workshop on Machine Learning for Information Filtering. 1999. [10] K Jarvelin and J. Kekalainen. IR evaluation methods for retrieving highly relevant documents. In Proceedings of the ACM Conference on Research and Development on Information Retrieval (SIGIR), 2000 [11] T. Joachims, Optimizing Search Engines Using Clickthrough Data. In Proceedings of the ACM Conference on Knowledge Discovery and Datamining (SIGKDD), 2002 [12] T. Joachims, L. Granka, B. Pang, H. Hembrooke, and G. Gay, Accurately Interpreting Clickthrough Data as Implicit Feedback, Proceedings of the ACM Conference on Research and Development on Information Retrieval (SIGIR), 2005 [13] T. Joachims, Making Large-Scale SVM Learning Practical. Advances in Kernel Methods, in Support Vector Learning, MIT Press, 1999 [14] D. Kelly and J. Teevan, Implicit feedback for inferring user preference: A bibliography. In SIGIR Forum, 2003 [15] J. Konstan, B. Miller, D. Maltz, J. Herlocker, L. Gordon, and J. Riedl. GroupLens: Applying collaborative filtering to usenet news. In Communications of ACM, 1997. [16] M. Morita, and Y. Shinoda, Information filtering based on user behavior analysis and best match text retrieval. Proceedings of the ACM Conference on Research and Development on Information Retrieval (SIGIR), 1994 [17] D. Oard and J. Kim. Implicit feedback for recommender systems. In Proceedings of the AAAI Workshop on Recommender Systems. 1998 [18] D. Oard and J. Kim. Modeling information content using observable behavior. In Proceedings of the 64th Annual Meeting of the American Society for Information Science and Technology. 2001 [19] N. Pharo, N. and K. Järvelin. The SST method: a tool for analyzing web information search processes. In Information Processing & Management, 2004 [20] P. Pirolli, The Use of Proximal Information Scent to Forage for Distal Content on the World Wide Web. In Working with Technology in Mind: Brunswikian. Resources for Cognitive Science and Engineering, Oxford University Press, 2004 [21] F. Radlinski and T. Joachims, Query Chains: Learning to Rank from Implicit Feedback. In Proceedings of the ACM Conference on Knowledge Discovery and Data Mining (SIGKDD), 2005. [22] F. Radlinski and T. Joachims, Evaluating the Robustness of Learning from Implicit Feedback, in Proceedings of the ICML Workshop on Learning in Web Search, 2005 [23] S. E. Robertson, H. Zaragoza, and M. Taylor, Simple BM25 extension to multiple weighted fields, in Proceedings of the Conference on Information and Knowledge Management (CIKM), 2004 [24] G. Salton & M. McGill. Introduction to modern information retrieval. McGraw-Hill, 1983 [25] E.M. Voorhees, D. Harman, Overview of TREC, 2001 [26] G.R. Xue, H.J. Zeng, Z. Chen, Y. Yu, W.Y. Ma, W.S. Xi, and W.G. Fan, Optimizing web search using web clickthrough data, in Proceedings of the Conference on Information and Knowledge Management (CIKM), 2004 [27] H. Zaragoza, N. Craswell, M. Taylor, S. Saria, and S. Robertson. Microsoft Cambridge at TREC 13: Web and Hard Tracks. In Proceedings of TREC 2004

feedback;implicit relevance feedback;document;web search rank;web search;ranking;score;user interaction;information;information retrieval;user behavior;relevance feedback;result

train_H-95

Handling Locations in Search Engine Queries

This paper proposes simple techniques for handling place references in search engine queries, an important aspect of geographical information retrieval. We address not only the detection, but also the disambiguation of place references, by matching them explicitly with concepts at an ontology. Moreover, when a query does not reference any locations, we propose to use information from documents matching the query, exploiting geographic scopes previously assigned to these documents. Evaluation experiments, using topics from CLEF campaigns and logs from real search engine queries, show the effectiveness of the proposed approaches.

1. INTRODUCTION Search engine queries are often associated with geographical locations, either explicitly (i.e. a location reference is given as part of the query) or implicitly (i.e. the location reference is not present in the query string, but the query clearly has a local intent [17]). One of the concerns of geographical information retrieval (GIR) lies in appropriately handling such queries, bringing better targeted search results and improving user satisfaction. Nowadays, GIR is getting increasing attention. Systems that access resources on the basis of geographic context are starting to appear, both in the academic and commercial domains [4, 7]. Accurately and effectively detecting location references in search engine queries is a crucial aspect of these systems, as they are generally based on interpreting geographical terms differently from the others. Detecting locations in queries is also important for generalpropose search engines, as this information can be used to improve ranking algorithms. Queries with a local intent are best answered with localized pages, while queries without any geographical references are best answered with broad pages [5]. Text mining methods have been successfully used in GIR to detect and disambiguate geographical references in text [9], or even to infer geographic scopes for documents [1, 13]. However, this body of research has been focused on processing Web pages and full-text documents. Search engine queries are more difficult to handle, in the sense that they are very short and with implicit and subjective user intents. Moreover, the data is also noisier and more versatile in form, and we have to deal with misspellings, multilingualism and acronyms. How to automatically understand what the user intended, given a search query, without putting the burden in the user himself, remains an open text mining problem. Key challenges in handling locations over search engine queries include their detection and disambiguation, the ranking of possible candidates, the detection of false positives (i.e not all contained location names refer to geographical locations), and the detection of implied locations by the context of the query (i.e. when the query does not explicitly contain a place reference but it is nonetheless geographical). Simple named entity recognition (NER) algorithms, based on dictionary look-ups for geographical names, may introduce high false positives for queries whose location names do not constitute place references. For example the query Denzel Washington contains the place name Washington, but the query is not geographical. Queries can also be geographic without containing any explicit reference to locations at the dictionary. In these cases, place name extraction and disambiguation does not give any results, and we need to access other sources of information. This paper proposes simple and yet effective techniques for handling place references over queries. Each query is split into a triple < what,relation,where >, where what specifies the non-geographic aspect of the information need, where specifies the geographic areas of interest, and relation specifies a spatial relationship connecting what and where. When this is not possible, i.e. the query does not contain any place references, we try using information from documents matching the query, exploiting geographic scopes previously assigned to these documents. Disambiguating place references is one of the most important aspects. We use a search procedure that combines textual patterns with geographical names defined at an ontology, and we use heuristics to disambiguate the discovered references (e.g. more important places are preferred). Disambiguation results in having the where term, from the triple above, associated with the most likely corresponding concepts from the ontology. When we cannot detect any locations, we attempt to use geographical scopes previously inferred for the documents at the top search results. By doing this, we assume that the most frequent geographical scope in the results should correspond to the geographical context implicit in the query. Experiments with CLEF topics [4] and sample queries from a Web search engine show that the proposed methods are accurate, and may have applications in improving search results. The rest of this paper is organized as follows. We first formalize the problem and describe related work to our research. Next, we describe our approach for handling place names in queries, starting with the general approach for disambiguating place references over textual strings, then presenting the method for splitting a query into a < what,relation,where > triple, and finally discussing the technique for exploiting geographic scopes previously assigned to documents in the result set. Section 4 presents evaluation results. Finally, we give some conclusions and directions for future research. 2. CONCEPTS AND RELATED WORK Search engine performance depends on the ability to capture the most likely meaning of a query as intended by the user [16]. Previous studies showed that a significant portion of the queries submitted to search engines are geographic [8, 14]. A recent enhancement to search engine technology is the addition of geographic reasoning, combining geographic information systems and information retrieval in order to build search engines that find information associated with given locations. The ability to recognize and reason about the geographical terminology, given in the text documents and user queries, is a crucial aspect of these geographical information retrieval (GIR) systems [4, 7]. Extracting and distinguishing different types of entities in text is usually referred to as Named Entity Recognition (NER). For at least a decade, this has been an important text mining task, and a key feature of the Message Understanding Conferences (MUC) [3]. NER has been successfully automated with near-human performance, but the specific problem of recognizing geographical references presents additional challenges [9]. When handling named entities with a high level of detail, ambiguity problems arise more frequently. Ambiguity in geographical references is bi-directional [15]. The same name can be used for more than one location (referent ambiguity), and the same location can have more than one name (reference ambiguity). The former has another twist, since the same name can be used for locations as well as for other class of entities, such as persons or company names (referent class ambiguity). Besides the recognition of geographical expressions, GIR also requires that the recognized expressions be classified and grounded to unique identifiers [11]. Grounding the recognized expressions (e.g. associating them to coordinates or concepts at an ontology) assures that they can be used in more advanced GIR tasks. Previous works have addressed the tagging and grounding of locations in Web pages, as well as the assignment of geographic scopes to these documents [1, 7, 13]. This is a complementary aspect to the techniques described in this paper, since if we have the Web pages tagged with location information, a search engine can conveniently return pages with a geographical scope related to the scope of the query. The task of handling geographical references over documents is however considerably different from that of handling geographical references over queries. In our case, queries are usually short and often do not constitute proper sentences. Text mining techniques that make use of context information are difficult to apply for high accuracy. Previous studies have also addressed the use of text mining and automated classification techniques over search engine queries [16, 10]. However, most of these works did not consider place references or geographical categories. Again, these previously proposed methods are difficult to apply to the geographic domain. Gravano et. al. studied the classification of Web queries into two types, namely local and global [5]. They defined a query as local if its best matches on a Web search engine are likely to be local pages, such as houses for sale. A number of classification algorithms have been evaluated using search engine queries. However, their experimental results showed that only a rather low precision and recall could be achieved. The problem addressed in this paper is also slightly different, since we are trying not only to detect local queries but also to disambiguate the local of interest. Wang et. al. proposed to go further than detecting local queries, by also disambiguating the implicit local of interest [17]. The proposed approach works for both queries containing place references and queries not containing them, by looking for dominant geographic references over query logs and text from search results. In comparison, we propose simpler techniques based on matching names from a geographic ontology. Our approach looks for spatial relationships at the query string, and it also associates the place references to ontology concepts. In the case of queries not containing explicit place references, we use geographical scopes previously assigned to the documents, whereas Wang et. al. proposed to extract locations from the text of the top search results. There are nowadays many geocoding, reverse-geocoding, and mapping services on the Web that can be easily integrated with other applications. Geocoding is the process of locating points on the surface of the Earth from alphanumeric addressing data. Taking a string with an address, a geocoder queries a geographical information system and returns interpolated coordinate values for the given location. Instead of computing coordinates for a given place reference, the technique described in this paper aims at assigning references to the corresponding ontology concepts. However, if each concept at the ontology contains associated coordinate information, the approach described here could also be used to build a geocoding service. Most of such existing services are commercial in nature, and there are no technical publications describing them. A number of commercial search services have also started to support location-based searches. Google Local, for instance, initially required the user to specify a location qualifier separately from the search query. More recently, it added location look-up capabilities that extract locations from query strings. For example, in a search for Pizza Seattle, Google Local returns local results for pizza near Seattle, WA. However, the intrinsics of their solution are not published, and their approach also does not handle locationimplicit queries. Moreover, Google Local does not take spatial relations into account. In sum, there are already some studies on tagging geographical references, but Web queries pose additional challenges which have not been addressed. In this paper, we explain the proposed solutions for the identified problems. 3. HANDLINGQUERIESIN GIR SYSTEMS Most GIR queries can be parsed to < what,relation,where > triple, where the what term is used to specify the general nongeographical aspect of the information need, the where term is used to specify the geographical areas of interest, and the relation term is used to specify a spatial relationship connecting what and where. While the what term can assume any form, in order to reflect any information need, the relation and where terms should be part of a controlled vocabulary. In particular, the relation term should refer to a well-known geographical relation that the underlying GIR system can interpret (e.g. near or contained at), and the where term should be disambiguated into a set of unique identifiers, corresponding to concepts at the ontology. Different systems can use alternative schemes to take input queries from the users. Three general strategies can be identified, and GIR systems often support more than one of the following schemes: Figure 1: Strategies for processing queries in Geographical Information Retrieval systems. 1. Input to the system is a textual query string. This is the hardest case, since we need to separate the query into the three different components, and then we need to disambiguate the where term into a set of unique identifiers. 2. Input to the system is provided in two separate strings, one concerning the what term, and the other concerning the where. The relation term can be either fixed (e.g. always assume the near relation), specified together with the where string, or provided separately from the users from a set of possible choices. Although there is no need for separating query string into the different components, we still need to disambiguate the where term into a set of unique identifiers. 3. Input to the system is provided through a query string together with an unambiguous description of the geographical area of interest (e.g. a sketch in a map, spatial coordinates or a selection from a set of possible choices). No disambiguation is required, and therefore the techniques described in this paper do not have to be applied. The first two schemes depend on place name disambiguation. Figure 1 illustrates how we propose to handle geographic queries in these first two schemes. A common component is the algorithm for disambiguating place references into corresponding ontology concepts, which is described next. 3.1 From Place Names to Ontology Concepts A required task in handling GIR queries consists of associating a string containing a geographical reference with the set of corresponding concepts at the geographic ontology. We propose to do this according to the pseudo-code listed at Algorithm 1. The algorithm considers the cases where a second (or even more than one) location is given to qualify a first (e.g. Paris, France). It makes recursive calls to match each location, and relies on hierarchical part-of relations to detect if two locations share a common hierarchy path. One of the provided locations should be more general and the other more specific, in the sense that there must exist a part-of relationship among the associated concepts at the ontology (either direct or transitive). The most specific location is a sub-region of the most general, and the algorithm returns the most specific one (i.e. for Paris, France the algorithm returns the ontology concept associated with Paris, the capital city of France). We also consider the cases where a geographical type expression is used to qualify a given name (e.g. city of Lisbon or state of New York). For instance the name Lisbon can correspond to many different concepts at a geographical ontology, and type Algorithm 1 Matching a place name with ontology concepts Require: O = A geographic ontology Require: GN = A string with the geographic name to be matched 1: L = An empty list 2: INDEX = The position in GN for the first occurrence of a comma, semi-colon or bracket character 3: if INDEX is defined then 4: GN1 = The substring of GN from position 0 to INDEX 5: GN2 = The substring of GN from INDEX +1 to length(GN) 6: L1 = Algorithm1(O,GN1) 7: L2 = Algorithm1(O,GN2) 8: for each C1 in L1 do 9: for each C2 in L2 do 10: if C1 is an ancestor of C2 at O then 11: L = The list L after adding element C2 12: else if C1 is a descendant of C2 at O then 13: L = The list L after adding element C1 14: end if 15: end for 16: end for 17: else 18: GN = The string GN after removing case and diacritics 19: if GN contains a geographic type qualifier then 20: T = The substring of GN containing the type qualifier 21: GN = The substring of GN with the type qualifier removed 22: L = The list of concepts from O with name GN and type T 23: else 24: L = The list of concepts from O with name GN 25: end if 26: end if 27: return The list L qualifiers can provide useful information for disambiguation. The considered type qualifiers should also described at the ontologies (e.g. each geographic concept should be associated to a type that is also defined at the ontology, such as country, district or city). Ideally, the geographical reference provided by the user should be disambiguated into a single ontology concept. However, this is not always possible, since the user may not provide all the required information (i.e. a type expression or a second qualifying location). The output is therefore a list with the possible concepts being referred to by the user. In a final step, we propose to sort this list, so that if a single concept is required as output, we can use the one that is ranked higher. The sorting procedure reflects the likelihood of each concept being indeed the one referred to. We propose to rank concepts according to the following heuristics: 1. The geographical type expression associated with the ontology concept. For the same name, a country is more likely to be referenced than a city, and in turn a city more likely to be referenced than a street. 2. Number of ancestors at the ontology. Top places at the ontology tend to be more general, and are therefore more likely to be referenced in search engine queries. 3. Population count. Highly populated places are better known, and therefore more likely to be referenced in queries. 4. Population counts from direct ancestors at the ontology. Subregions of highly populated places are better known, and also more likely to be referenced in search engine queries. 5. Occurrence frequency over Web documents (e.g. Google counts) for the geographical names. Places names that occur more frequently over Web documents are also more likely to be referenced in search engine queries. 6. Number of descendants at the ontology. Places that have more sub-regions tend to be more general, and are therefore more likely to be mentioned in search engine queries. 7. String size for the geographical names. Short names are more likely to be mentioned in search engine queries. Algorithm 1, plus the ranking procedure, can already handle GIR queries where the where term is given separately from the what and relation terms. However, if the query is given in a single string, we require the identification of the associated < what,relation,where > triple, before disambiguating the where term into the corresponding ontology concepts. This is described in the following Section. 3.2 Handling Single Query Strings Algorithm 2 provides the mechanism for separating a query string into a < what,relation,where > triple. It uses Algorithm 1 to find the where term, disambiguating it into a set of ontology concepts. The algorithm starts by tokenizing the query string into individual words, also taking care of removing case and diacritics. We have a simple tokenizer that uses the space character as a word delimiter, but we could also have a tokenization approach similar to the proposal of Wang et. al. which relies on Web occurrence statistics to avoid breaking collocations [17]. In the future, we plan on testing if this different tokenization scheme can improve results. Next, the algorithm tests different possible splits of the query, building the what, relation and where terms through concatenations of the individual tokens. The relation term is matched against a list of possible values (e.g. near, at, around, or south of), corresponding to the operators that are supported by the GIR system. Note that is also the responsibility of the underlying GIR system to interpret the actual meaning of the different spatial relations. Algorithm 1 is used to check whether a where term constitutes a geographical reference or not. We also check if the last word in the what term belongs to a list of exceptions, containing for instance first names of people in different languages. This ensures that a query like Denzel Washington is appropriately handled. If the algorithm succeeds in finding valid relation and where terms, then the corresponding triple is returned. Otherwise, we return a triple with the what term equaling the query string, and the relation and where terms set as empty. If the entire query string constitutes a geographical reference, we return a triple with the what term set to empty, the where term equaling the query string, and the relation term set the DEFINITION (i.e. these queries should be answered with information about the given place references). The algorithm also handles query strings where more than one geographical reference is provided, using and or an equivalent preposition, together with a recursive call to Algorithm 2. A query like Diamond trade in Angola and South Africa is Algorithm 2 Get < what,relation,where > from a query string Require: O = A geographical ontology Require: Q = A non-empty string with the query 1: Q = The string Q after removing case and diacritics 2: TOKENS[0..N] = An array of strings with the individual words of Q 3: N = The size of the TOKENS array 4: for INDEX = 0 to N do 5: if INDEX = 0 then 6: WHAT = Concatenation of TOKENS[0..INDEX −1] 7: LASTWHAT = TOKENS[INDEX −1] 8: else 9: WHAT = An empty string 10: LASTWHAT = An empty string 11: end if 12: WHERE = Concatenation of TOKENS[INDEX..N] 13: RELATION = An empty string 14: for INDEX2 = INDEX to N −1 do 15: RELATION2 = Concatenation of TOKENS[INDEX..INDEX2] 16: if RELATION2 is a valid geographical relation then 17: WHERE = Concatenation of S[INDEX2 +1..N] 18: RELATION = RELATION2; 19: end if 20: end for 21: if RELATION = empty AND LASTWHAT in an exception then 22: TESTGEO = FALSE 23: else 24: TESTGEO = TRUE 25: end if 26: if TESTGEO AND Algorithm1(WHERE) <> EMPTY then 27: if WHERE ends with AND SURROUNDINGS then 28: RELATION = The string NEAR; 29: WHERE = The substring of WHERE with AND SURROUNDINGS removed 30: end if 31: if WHAT ends with AND or similar) then 32: < WHAT,RELATION,WHERE2 >= Algorithm2(WHAT) 33: WHERE = Concatenation of WHERE with WHERE2 34: end if 35: if RELATION = An empty string then 36: if WHAT = An empty string then 37: RELATION = The string DEFINITION 38: else 39: RELATION = The string CONTAINED-AT 40: end if 41: end if 42: else 43: WHAT = The string Q 44: WHERE = An empty string 45: RELATION = An empty string 46: end if 47: end for 48: return < WHAT,RELATION,WHERE > therefore appropriately handled. Finally, if the geographical reference in the query is complemented with an expression similar to and its surroundings, the spatial relation (which is assumed to be CONTAINED-AT if none is provided) is changed to NEAR. 3.3 From Search Results to Query Locality The procedures given so far are appropriate for handling queries where a place reference is explicitly mentioned. However, the fact that a query can be associated with a geographical context may not be directly observable in the query itself, but rather from the results returned. For instance, queries like recommended hotels for SIGIR 2006 or SeaFair 2006 lodging can be seen to refer to the city of Seattle. Although they do not contain an explicit place reference, we expect results to be about hotels in Seattle. In the cases where a query does not contain place references, we start by assuming that the top results from a search engine represent the most popular and correct context and usage for the query. We Topic What Relation Where TGN concepts ML concepts Vegetable Exporters of Europe Vegetable Exporters CONTAINED-AT Europe 1 1 Trade Unions in Europe Trade Unions CONTAINED-AT Europe 1 1 Roman cities in the UK and Germany Roman cities CONTAINED-AT UK and Germany 6 2 Cathedrals in Europe Cathedrals CONTAINED-AT Europe 1 1 Car bombings near Madrid Car bombings NEAR Madrid 14 2 Volcanos around Quito Volcanos NEAR Quito 4 1 Cities within 100km of Frankfurt Cities NEAR Frankfurt 3 1 Russian troops in south(ern) Caucasus Russian troops in south(ern) CONTAINED-AT Caucasus 2 1 Cities near active volcanoes (This topic could not be appropriately handled - the relation and where terms are returned empty) Japanese rice imports (This topic could not be appropriately handled - the relation and where terms are returned empty) Table 1: Example topics from the GeoCLEF evaluation campaigns and the corresponding < what,relation,where > triples. then propose to use the distributional characteristics of geographical scopes previously assigned to the documents corresponding to these top results. In a previous work, we presented a text mining approach for assigning documents with corresponding geographical scopes, defined at an ontology, that worked as an offline preprocessing stage in a GIR system [13]. This pre-processing step is a fundamental stage of GIR, and it is reasonable to assume that this kind of information would be available on any system. Similarly to Wang et. al., we could also attempt to process the results on-line, in order to detect place references in the documents [17]. However, a GIR system already requires the offline stage. For the top N documents given at the results, we check the geographic scopes that were assigned to them. If a significant portion of the results are assigned to the same scope, than the query can be seen to be related to the corresponding geographic concept. This assumption could even be relaxed, for instance by checking if the documents belong to scopes that are hierarchically related. 4. EVALUATION EXPERIMENTS We used three different ontologies in evaluation experiments, namely the Getty thesaurus of geographic names (TGN) [6] and two specific resources developed at our group, here referred to as the PT and ML ontologies [2]. TGN and ML include global geographical information in multiple languages (although TGN is considerably larger), while the PT ontology focuses on the Portuguese territory with a high detail. Place types are also different across ontologies, as for instance PT includes street names and postal addresses, whereas ML only goes to the level of cities. The reader should refer to [2, 6] for a complete description of these resources. Our initial experiments used Portuguese and English topics from the GeoCLEF 2005 and 2006 evaluation campaigns. Topics in GeoCLEF correspond to query strings that can be used as input to a GIR system [4]. ImageCLEF 2006 also included topics specifying place references, and participants were encouraged to run their GIR systems on them. Our experiments also considered this dataset. For each topic, we measured if Algorithm 2 was able to find the corresponding < what,relation,where > triple. The ontologies used in this experiment were the TGN and ML, as topics were given in multiple languages and covered the whole globe. Dataset Number of Correct Triples Time per Query Queries ML TGN ML TGN GeoCLEF05 EN 25 19 20 GeoCLEF05 PT 25 20 18 288.1 334.5 GeoCLEF06 EN 32 28 19 msec msec GeoCLEF06 PT 25 23 11 ImgCLEF06 EN 24 16 18 Table 2: Summary of results over CLEF topics. Table 1 illustrates some of the topics, and Table 2 summarizes the obtained results. The tables show that the proposed technique adequately handles most of these queries. A manual inspection of the ontology concepts that were returned for each case also revealed that the where term was being correctly disambiguated. Note that the TGN ontology indeed added some ambiguity, as for instance names like Madrid can correspond to many different places around the globe. It should also be noted that some of the considered topics are very hard for an automated system to handle. Some of them were ambiguous (e.g. in Japanese rice imports, the query can be said to refer either rice imports in Japan or imports of Japanese rice), and others contained no direct geographical references (e.g. cities near active volcanoes). Besides these very hard cases, we also missed some topics due to their usage of place adjectives and specific regions that are not defined at the ontologies (e.g. environmental concerns around the Scottish Trossachs). In a second experiment, we used a sample of around 100,000 real search engine queries. The objective was to see if a significant number of these queries were geographical in nature, also checking if the algorithm did not produce many mistakes by classifying a query as geographical when that was not the case. The Portuguese ontology was used in this experiment, and queries were taken from the logs of a Portuguese Web search engine available at www.tumba.pt. Table 3 summarizes the obtained results. Many queries were indeed geographical (around 3.4%, although previous studies reported values above 14% [8]). A manual inspection showed that the algorithm did not produce many false positives, and the geographical queries were indeed correctly split into correct < what,relation,where > triple. The few mistakes we encountered were related to place names that were more frequently used in other contexts (e.g. in Teófilo Braga we have the problem that Braga is a Portuguese district, and Teófilo Braga was a well known Portuguese writer and politician). The addition of more names to the exception list can provide a workaround for most of these cases. Value Num. Queries 110,916 Num. Queries without Geographical References 107,159 (96.6%) Num. Queries with Geographical References 3,757 (3.4%) Table 3: Results from an experiment with search engine logs. We also tested the procedure for detecting queries that are implicitly geographical with a small sample of queries from the logs. For instance, for the query Estádio do Dragão (e.g. home stadium for a soccer team from Porto), the correct geographical context can be discovered from the analysis of the results (more than 75% of the top 20 results are assigned with the scope Porto). For future work, we plan on using a larger collection of queries to evaluate this aspect. Besides queries from the search engine logs, we also plan on using the names of well-known buildings, monuments and other landmarks, as they have a strong geographical connotation. Finally, we also made a comparative experiment with 2 popular geocoders, Maporama and Microsoft"s Mappoint. The objective was to compare Algorithm 1 with other approaches, in terms of being able to correctly disambiguate a string with a place reference. Civil Parishes from Lisbon Maporama Mappoint Ours Coded refs. (out of 53) 9 (16.9%) 30 (56,6%) 15 (28.3%) Avg. Time per ref. (msec) 506.23 1235.87 143.43 Civil Parishes from Porto Maporama Mappoint Ours Coded refs. (out of 15) 0 (0%) 2 (13,3%) 5 (33.3%) Avg. Time per ref. (msec) 514.45 991.88 132.14 Table 4: Results from a comparison with geocoding services. The Portuguese ontology was used in this experiment, taking as input the names of civil parishes from the Portuguese municipalities of Lisbon and Porto, and checking if the systems were able to disambiguate the full name (e.g. Campo Grande, Lisboa or Foz do Douro, Porto) into the correct geocode. We specifically measured whether our approach was better at unambiguously returning geocodes given the place reference (i.e. return the single correct code), and providing results rapidly. Table 4 shows the obtained results, and the accuracy of our method seems comparable to the commercial geocoders. Note that for Maporama and Mappoint, the times given at Table 4 include fetching results from the Web, but we have no direct way of accessing the geocoding algorithms (in both cases, fetching static content from the Web servers takes around 125 milliseconds). Although our approach cannot unambiguously return the correct geocode in most cases (only 20 out of a total of 68 cases), it nonetheless returns results that a human user can disambiguate (e.g. for Madalena, Lisboa we return both a street and a civil parish), as opposed to the other systems that often did not produce results. Moreover, if we consider the top geocode according to the ranking procedure described in Section 3.1, or if we use a type qualifier in the name (e.g. civil parish of Campo Grande, Lisboa), our algorithm always returns the correct geocode. 5. CONCLUSIONS This paper presented simple approaches for handling place references in search engine queries. This is a hard text mining problem, as queries are often ambiguous or underspecify information needs. However, our initial experiments indicate that for many queries, the referenced places can be determined effectively. Unlike the techniques proposed by Wang et. al. [17], we mainly focused on recognizing spatial relations and associating place names to ontology concepts. The proposed techniques were employed in the prototype system that we used for participating in GeoCLEF 2006. In queries where a geographical reference is not explicitly mentioned, we propose to use the results for the query, exploiting geographic scopes previously assigned to these documents. In the future, we plan on doing a careful evaluation of this last approach. Another idea that we would like to test involves the integration of a spelling correction mechanism [12] into Algorithm 1, so that incorrectly spelled place references can be matched to ontology concepts. The proposed techniques for handling geographic queries can have many applications in improving GIR systems or even general purpose search engines. After place references are appropriately disambiguated into ontology concepts, a GIR system can use them to retrieve relevant results, through the use of appropriate index structures (e.g. indexing the spatial coordinates associated with ontology concepts) and provided that the documents are also assigned to scopes corresponding to ontology concepts. A different GIR strategy can involve query expansion, by taking the where terms from the query and using the ontology to add names from neighboring locations. In a general purpose search engine, and if a local query is detected, we can forward users to a GIR system, which should be better suited for properly handling the query. The regular Google search interface already does this, by presenting a link to Google Local when it detects a geographical query. 6. REFERENCES [1] E. Amitay, N. Har"El, R. Sivan, and A. Soffer. Web-a-Where: Geotagging Web content. In Proceedings of SIGIR-04, the 27th Conference on research and development in information retrieval, 2004. [2] M. Chaves, M. J. Silva, and B. Martins. A Geographic Knowledge Base for Semantic Web Applications. In Proceedings of SBBD-05, the 20th Brazilian Symposium on Databases, 2005. [3] N. A. Chinchor. Overview of MUC-7/MET-2. In Proceedings of MUC-7, the 7th Message Understanding Conference, 1998. [4] F. Gey, R. Larson, M. Sanderson, H. Joho, and P. Clough. GeoCLEF: the CLEF 2005 cross-language geographic information retrieval track. In Working Notes for the CLEF 2005 Workshop, 2005. [5] L. Gravano, V. Hatzivassiloglou, and R. Lichtenstein. Categorizing Web queries according to geographical locality. In Proceedings of CIKM-03, the 12th Conference on Information and knowledge management, 2003. [6] P. Harpring. Proper words in proper places: The thesaurus of geographic names. MDA Information, 3, 1997. [7] C. Jones, R. Purves, A. Ruas, M. Sanderson, M. Sester, M. van Kreveld, and R. Weibel. Spatial information retrieval and geographical ontologies: An overview of the SPIRIT project. In Proceedings of SIGIR-02, the 25th Conference on Research and Development in Information Retrieval, 2002. [8] J. Kohler. Analyzing search engine queries for the use of geographic terms, 2003. (MSc Thesis). [9] A. Kornai and B. Sundheim, editors. Proceedings of the NAACL-HLT Workshop on the Analysis of Geographic References, 2003. [10] Y. Li, Z. Zheng, and H. Dai. KDD CUP-2005 report: Facing a great challenge. SIGKDD Explorations, 7, 2006. [11] D. Manov, A. Kiryakov, B. Popov, K. Bontcheva, D. Maynard, and H. Cunningham. Experiments with geographic knowledge for information extraction. In Proceedings of the NAACL-HLT Workshop on the Analysis of Geographic References, 2003. [12] B. Martins and M. J. Silva. Spelling correction for search engine queries. In Proceedings of EsTAL-04, España for Natural Language Processing, 2004. [13] B. Martins and M. J. Silva. A graph-ranking algorithm for geo-referencing documents. In Proceedings of ICDM-05, the 5th IEEE International Conference on Data Mining, 2005. [14] L. Souza, C. J. Davis, K. Borges, T. Delboni, and A. Laender. The role of gazetteers in geographic knowledge discovery on the web. In Proceedings of LA-Web-05, the 3rd Latin American Web Congress, 2005. [15] E. Tjong, K. Sang, and F. D. Meulder. Introduction to the CoNLL-2003 shared task: Language-Independent Named Entity Recognition. In Proceedings of CoNLL-2003, the 7th Conference on Natural Language Learning, 2003. [16] D. Vogel, S. Bickel, P. Haider, R. Schimpfky, P. Siemen, S. Bridges, and T. Scheffer. Classifying search engine queries using the Web as background knowledge. SIGKDD Explorations Newsletter, 7(2):117-122, 2005. [17] L. Wang, C. Wang, X. Xie, J. Forman, Y. Lu, W.-Y. Ma, and Y. Li. Detecting dominant locations from search queries. In Proceedings of SIGIR-05, the 28th Conference on Research and development in information retrieval, 2005.

tokenization scheme;geographical type expression;geographical information retrieval;search engine query;text mine;disambiguation result;search query;geographic ontology;place reference;spelling correction mechanism;named entity recognition algorithm;web search engine;geographic context;query string;textual string;geographic ir;locationimplicit query;query process;ner

train_H-96

A Study of Factors Affecting the Utility of Implicit Relevance Feedback

Implicit relevance feedback (IRF) is the process by which a search system unobtrusively gathers evidence on searcher interests from their interaction with the system. IRF is a new method of gathering information on user interest and, if IRF is to be used in operational IR systems, it is important to establish when it performs well and when it performs poorly. In this paper we investigate how the use and effectiveness of IRF is affected by three factors: search task complexity, the search experience of the user and the stage in the search. Our findings suggest that all three of these factors contribute to the utility of IRF.

1. INTRODUCTION Information Retrieval (IR) systems are designed to help searchers solve problems. In the traditional interaction metaphor employed by Web search systems such as Yahoo! and MSN Search, the system generally only supports the retrieval of potentially relevant documents from the collection. However, it is also possible to offer support to searchers for different search activities, such as selecting the terms to present to the system or choosing which search strategy to adopt [3, 8]; both of which can be problematic for searchers. As the quality of the query submitted to the system directly affects the quality of search results, the issue of how to improve search queries has been studied extensively in IR research [6]. Techniques such as Relevance Feedback (RF) [11] have been proposed as a way in which the IR system can support the iterative development of a search query by suggesting alternative terms for query modification. However, in practice RF techniques have been underutilised as they place an increased cognitive burden on searchers to directly indicate relevant results [10]. Implicit Relevance Feedback (IRF) [7] has been proposed as a way in which search queries can be improved by passively observing searchers as they interact. IRF has been implemented either through the use of surrogate measures based on interaction with documents (such as reading time, scrolling or document retention) [7] or using interaction with browse-based result interfaces [5]. IRF has been shown to display mixed effectiveness because the factors that are good indicators of user interest are often erratic and the inferences drawn from user interaction are not always valid [7]. In this paper we present a study into the use and effectiveness of IRF in an online search environment. The study aims to investigate the factors that affect IRF, in particular three research questions: (i) is the use of and perceived quality of terms generated by IRF affected by the search task? (ii) is the use of and perceived quality of terms generated by IRF affected by the level of search experience of system users? (iii) is IRF equally used and does it generate terms that are equally useful at all search stages? This study aims to establish when, and under what circumstances, IRF performs well in terms of its use and the query modification terms selected as a result of its use. The main experiment from which the data are taken was designed to test techniques for selecting query modification terms and techniques for displaying retrieval results [13]. In this paper we use data derived from that experiment to study factors affecting the utility of IRF. 2. STUDY In this section we describe the user study conducted to address our research questions. 2.1 Systems Our study used two systems both of which suggested new query terms to the user. One system suggested terms based on the user"s interaction (IRF), the other used Explicit RF (ERF) asking the user to explicitly indicate relevant material. Both systems used the same term suggestion algorithm, [15], and used a common interface. 2.1.1 Interface Overview In both systems, retrieved documents are represented at the interface by their full-text and a variety of smaller, query-relevant representations, created at retrieval time. We used the Web as the test collection in this study and Google1 as the underlying search engine. Document representations include the document title and a summary of the document; a list of top-ranking sentences (TRS) extracted from the top documents retrieved, scored in relation to the query, a sentence in the document summary, and each summary sentence in the context it occurs in the document (i.e., with the preceding and following sentence). Each summary sentence and top-ranking sentence is regarded as a representation of the document. The default display contains the list of top-ranking sentences and the list of the first ten document titles. Interacting with a representation guides searchers to a different representation from the same document, e.g., moving the mouse over a document title displays a summary of the document. This presentation of progressively more information from documents to aid relevance assessments has been shown to be effective in earlier work [14, 16]. In Appendix A we show the complete interface to the IRF system with the document representations marked and in Appendix B we show a fragment from the ERF interface with the checkboxes used by searchers to indicate relevant information. Both systems provide an interactive query expansion feature by suggesting new query terms to the user. The searcher has the responsibility for choosing which, if any, of these terms to add to the query. The searcher can also add or remove terms from the query at will. 2.1.2 Explicit RF system This version of the system implements explicit RF. Next to each document representation are checkboxes that allow searchers to mark individual representations as relevant; marking a representation is an indication that its contents are relevant. Only the representations marked relevant by the user are used for suggesting new query terms. This system was used as a baseline against which the IRF system could be compared. 2.1.3 Implicit RF system This system makes inferences about searcher interests based on the information with which they interact. As described in Section 2.1.1 interacting with a representation highlights a new representation from the same document. To the searcher this is a way they can find out more information from a potentially interesting source. To the implicit RF system each interaction with a representation is interpreted as an implicit indication of interest in that representation; interacting with a representation is assumed to be an indication that its contents are relevant. The query modification terms are selected using the same algorithm as in the Explicit RF system. Therefore the only difference between the systems is how relevance is communicated to the system. The results of the main experiment [13] indicated that these two systems were comparable in terms of effectiveness. 2.2 Tasks Search tasks were designed to encourage realistic search behaviour by our subjects. The tasks were phrased in the form of simulated work task situations [2], i.e., short search scenarios that were designed to reflect real-life search situations and allow subjects to develop personal assessments of relevance. We devised six search topics (i.e., applying to university, allergies in the workplace, art galleries in Rome, Third Generation mobile phones, Internet music piracy and petrol prices) based on pilot testing with a small representative group of subjects. These subjects were not involved in the main experiment. For each topic, three versions of each work task situation were devised, each version differing in their predicted level of task complexity. As described in [1] task complexity is a variable that affects subject perceptions of a task and their interactive behaviour, e.g., subjects perform more filtering activities with highly complex search tasks. By developing tasks of different complexity we can assess how the nature of the task affects the subjects" interactive behaviour and hence the evidence supplied to IRF algorithms. Task complexity was varied according to the methodology described in [1], specifically by varying the number of potential information sources and types of information required, to complete a task. In our pilot tests (and in a posteriori analysis of the main experiment results) we verified that subjects reporting of individual task complexity matched our estimation of the complexity of the task. Subjects attempted three search tasks: one high complexity, one moderate complexity and one low complexity2 . They were asked to read the task, place themselves in the situation it described and find the information they felt was required to complete the task. Figure 1 shows the task statements for three levels of task complexity for one of the six search topics. HC Task: High Complexity Whilst having dinner with an American colleague, they comment on the high price of petrol in the UK compared to other countries, despite large volumes coming from the same source. Unaware of any major differences, you decide to find out how and why petrol prices vary worldwide. MC Task: Moderate Complexity Whilst out for dinner one night, one of your friends" guests is complaining about the price of petrol and the factors that cause it. Throughout the night they seem to be complaining about everything they can, reducing the credibility of their earlier statements so you decide to research which factors actually are important in determining the price of petrol in the UK. LC Task: Low Complexity While out for dinner one night, your friend complains about the rising price of petrol. However, as you have not been driving for long, you are unaware of any major changes in price. You decide to find out how the price of petrol has changed in the UK in recent years. Figure 1. Varying task complexity (Petrol Prices topic). 2.3 Subjects 156 volunteers expressed an interest in participating in our study. 48 subjects were selected from this set with the aim of populating two groups, each with 24 subjects: inexperienced (infrequent/ inexperienced searchers) and experienced (frequent/ experienced searchers). Subjects were not chosen and classified into their groups until they had completed an entry questionnaire that asked them about their search experience and computer use. The average age of the subjects was 22.83 years (maximum 51, minimum 18, σ = 5.23 years) and 75% had a university diploma or a higher degree. 47.91% of subjects had, or were pursuing, a qualification in a discipline related to Computer Science. The subjects were a mixture of students, researchers, academic staff and others, with different levels of computer and search experience. The subjects were divided into the two groups depending on their search experience, how often they searched and the types of searches they performed. All were familiar with Web searching, and some with searching in other domains. 2.4 Methodology The experiment had a factorial design; with 2 levels of search experience, 3 experimental systems (although we only report on the findings from the ERF and IRF systems) and 3 levels of search task complexity. Subjects attempted one task of each complexity, 2 The main experiment from which these results are drawn had a third comparator system which had a different interface. Each subject carried out three tasks, one on each system. We only report on the results from the ERF and IRF systems as these are the only pertinent ones for this paper. switched systems after each task and used each system once. The order in which systems were used and search tasks attempted was randomised according to a Latin square experimental design. Questionnaires used Likert scales, semantic differentials and openended questions to elicit subject opinions [4]. System logging was also used to record subject interaction. A tutorial carried out prior to the experiment allowed subjects to use a non-feedback version of the system to attempt a practice task before using the first experimental system. Experiments lasted between oneand-a-half and two hours, dependent on variables such as the time spent completing questionnaires. Subjects were offered a 5 minute break after the first hour. In each experiment: i. the subject was welcomed and asked to read an introduction to the experiments and sign consent forms. This set of instructions was written to ensure that each subject received precisely the same information. ii. the subject was asked to complete an introductory questionnaire. This contained questions about the subject"s education, general search experience, computer experience and Web search experience. iii. the subject was given a tutorial on the interface, followed by a training topic on a version of the interface with no RF. iv. the subject was given three task sheets and asked to choose one task from the six topics on each sheet. No guidelines were given to subjects when choosing a task other than they could not choose a task from any topic more than once. Task complexity was rotated by the experimenter so each subject attempted one high complexity task, one moderate complexity task and one low complexity task. v. the subject was asked to perform the search and was given 15 minutes to search. The subject could terminate a search early if they were unable to find any more information they felt helped them complete the task. vi. after completion of the search, the subject was asked to complete a post-search questionnaire. vii. the remaining tasks were attempted by the subject, following steps v. and vi. viii. the subject completed a post-experiment questionnaire and participated in a post-experiment interview. Subjects were told that their interaction may be used by the IRF system to help them as they searched. They were not told which behaviours would be used or how it would be used. We now describe the findings of our analysis. 3. FINDINGS In this section we use the data derived from the experiment to answer our research questions about the effect of search task complexity, search experience and stage in search on the use and effectiveness of IRF. We present our findings per research question. Due to the ordinal nature of much of the data non-parametric statistical testing is used in this analysis and the level of significance is set to p < .05, unless otherwise stated. We use the method proposed by [12] to determine the significance of differences in multiple comparisons and that of [9] to test for interaction effects between experimental variables, the occurrence of which we report where appropriate. All Likert scales and semantic differentials were on a 5-point scale where a rating closer to 1 signifies more agreement with the attitude statement. The category labels HC, MC and LC are used to denote the high, moderate and low complexity tasks respectively. The highest, or most positive, values in each table are shown in bold. Our analysis uses data from questionnaires, post-experiment interviews and background system logging on the ERF and IRF systems. 3.1 Search Task Searchers attempted three search tasks of varying complexity, each on a different experimental system. In this section we present an analysis on the use and usefulness of IRF for search tasks of different complexities. We present our findings in terms of the RF provided by subjects and the terms recommended by the systems. 3.1.1 Feedback We use questionnaires and system logs to gather data on subject perceptions and provision of RF for different search tasks. In the postsearch questionnaire subjects were asked about how RF was conveyed using differentials to elicit their opinion on: 1. the value of the feedback technique: How you conveyed relevance to the system (i.e. ticking boxes or viewing information) was: easy / difficult, effective/ ineffective, useful"/not useful. 2. the process of providing the feedback: How you conveyed relevance to the system made you feel: comfortable/uncomfortable, in control/not in control. The average obtained differential values are shown in Table 1 for IRF and each task category. The value corresponding to the differential All represents the mean of all differentials for a particular attitude statement. This gives some overall understanding of the subjects" feelings which can be useful as the subjects may not answer individual differentials very precisely. The values for ERF are included for reference in this table and all other tables and figures in the Findings section. Since the aim of the paper is to investigate situations in which IRF might perform well, not a direct comparison between IRF and ERF, we make only limited comparisons between these two types of feedback. Table 1. Subject perceptions of RF method (lower = better). Each cell in Table 1 summarises the subject responses for 16 tasksystem pairs (16 subjects who ran a high complexity (HC) task on the ERF system, 16 subjects who ran a medium complexity (MC) task on the ERF system, etc). Kruskal-Wallis Tests were applied to each differential for each type of RF3 . Subject responses suggested that 3 Since this analysis involved many differentials, we use a Bonferroni correction to control the experiment-wise error rate and set the alpha level (α) to .0167 and .0250 for both statements 1. and 2. respectively, i.e., .05 divided by the number of differentials. This correction reduces the number of Type I errors i.e., rejecting null hypotheses that are true. Explicit RF Implicit RF Differential HC MC LC HC MC LC Easy 2.78 2.47 2.12 1.86 1.81 1.93 Effective 2.94 2.68 2.44 2.04 2.41 2.66 Useful 2.76 2.51 2.16 1.91 2.37 2.56 All (1) 2.83 2.55 2.24 1.94 2.20 2.38 Comfortable 2.27 2.28 2.35 2.11 2.15 2.16 In control 2.01 1.97 1.93 2.73 2.68 2.61 All (2) 2.14 2.13 2.14 2.42 2.42 2.39 IRF was most effective and useful for more complex search tasks4 and that the differences in all pair-wise comparisons between tasks were significant5 . Subject perceptions of IRF elicited using the other differentials did not appear to be affected by the complexity of the search task6 . To determine whether a relationship exists between the effectiveness and usefulness of the IRF process and task complexity we applied Spearman"s Rank Order Correlation Coefficient to participant responses. The results of this analysis suggest that the effectiveness of IRF and usefulness of IRF are both related to task complexity; as task complexity increases subject preference for IRF also increases7 . On the other hand, subjects felt ERF was more effective and useful for low complexity tasks8 . Their verbal reporting of ERF, where perceived utility and effectiveness increased as task complexity decreased, supports this finding. In tasks of lower complexity the subjects felt they were better able to provide feedback on whether or not documents were relevant to the task. We analyse interaction logs generated by both interfaces to investigate the amount of RF subjects provided. To do this we use a measure of search precision that is the proportion of all possible document representations that a searcher assessed, divided by the total number they could assess. In ERF this is the proportion of all possible representations that were marked relevant by the searcher, i.e., those representations explicitly marked relevant. In IRF this is the proportion of representations viewed by a searcher over all possible representations that could have been viewed by the searcher. This proportion measures the searcher"s level of interaction with a document, we take it to measure the user"s interest in the document: the more document representations viewed the more interested we assume a user is in the content of the document. There are a maximum of 14 representations per document: 4 topranking sentences, 1 title, 1 summary, 4 summary sentences and 4 summary sentences in document context. Since the interface shows document representations from the top-30 documents, there are 420 representations that a searcher can assess. Table 2 shows proportion of representations provided as RF by subjects. Table 2. Feedback and documents viewed. Explicit RF Implicit RF Measure HC MC LC HC MC LC Proportion Feedback 2.14 2.39 2.65 21.50 19.36 15.32 Documents Viewed 10.63 10.43 10.81 10.84 12.19 14.81 For IRF there is a clear pattern: as complexity increases the subjects viewed fewer documents but viewed more representations for each document. This suggests a pattern where users are investigating retrieved documents in more depth. It also means that the amount of 4 effective: χ2 (2) = 11.62, p = .003; useful: χ2 (2) = 12.43, p = .002 5 Dunn"s post-hoc tests (multiple comparison using rank sums); all Z ≥ 2.88, all p ≤ .002 6 all χ2 (2) ≤ 2.85, all p ≥ .24 (Kruskal-Wallis Tests) 7 effective: all r ≥ 0.644, p ≤ .002; useful: all r ≥ 0.541, p ≤ .009 8 effective: χ2 (2) = 7.01, p = .03; useful: χ2 (2) = 6.59, p = .037 (Kruskal-Wallis Test); all pair-wise differences significant, all Z ≥ 2.34, all p ≤ .01 (Dunn"s post-hoc tests) feedback varies based on the complexity of the search task. Since IRF is based on the interaction of the searcher, the more they interact, the more feedback they provide. This has no effect on the number of RF terms chosen, but may affect the quality of the terms selected. Correlation analysis revealed a strong negative correlation between the number of documents viewed and the amount of feedback searchers provide9 ; as the number of documents viewed increases the proportion of feedback falls (searchers view less representations of each document). This may be a natural consequence of their being less time to view documents in a time constrained task environment but as we will show as complexity changes, the nature of information searchers interact with also appears to change. In the next section we investigate the effect of task complexity on the terms chosen as a result of IRF. 3.1.2 Terms The same RF algorithm was used to select query modification terms in all systems [16]. We use subject opinions of terms recommended by the systems as a measure of the effectiveness of IRF with respect to the terms generated for different search tasks. To test this, subjects were asked to complete two semantic differentials that completed the statement: The words chosen by the system were: relevant/irrelevant and useful/not useful. Table 3 presents average responses grouped by search task. Table 3. Subject perceptions of system terms (lower = better). Explicit RF Implicit RF Differential HC MC LC HC MC LC Relevant 2.50 2.46 2.41 1.94 2.35 2.68 Useful 2.61 2.61 2.59 2.06 2.54 2.70 Kruskal-Wallis Tests were applied within each type of RF. The results indicate that the relevance and usefulness of the terms chosen by IRF is affected by the complexity of the search task; the terms chosen are more relevant and useful when the search task is more complex. 10 Relevant here, was explained as being related to their task whereas useful was for terms that were seen as being helpful in the search task. For ERF, the results indicate that the terms generated are perceived to be more relevant and useful for less complex search tasks; although differences between tasks were not significant11 . This suggests that subject perceptions of the terms chosen for query modification are affected by task complexity. Comparison between ERF and IRF shows that subject perceptions also vary for different types of RF12 . As well as using data on relevance and utility of the terms chosen, we used data on term acceptance to measure the perceived value of the terms suggested. Explicit and Implicit RF systems made recommendations about which terms could be added to the original search query. In Table 4 we show the proportion of the top six terms 9 r = −0.696, p = .001 (Pearson"s Correlation Coefficient) 10 relevant: χ2 (2) = 13.82, p = .001; useful: χ2 (2) = 11.04, p = .004; α = .025 11 all χ2 (2) ≤ 2.28, all p ≥ .32 (Kruskal-Wallis Test) 12 all T(16) ≥ 102, all p ≤ .021, (Wilcoxon Signed-Rank Test) 13 that were shown to the searcher that were added to the search query, for each type of task and each type of RF. Table 4. Term Acceptance (percentage of top six terms). Explicit RF Implicit RFProportion of terms HC MC LC HC MC LC Accepted 65.31 67.32 68.65 67.45 67.24 67.59 The average number of terms accepted from IRF is approximately the same across all search tasks and generally the same as that of ERF14 . As Table 2 shows, subjects marked fewer documents relevant for highly complex tasks . Therefore, when task complexity increases the ERF system has fewer examples of relevant documents and the expansion terms generated may be poorer. This could explain the difference in the proportion of recommended terms accepted in ERF as task complexity increases. For IRF there is little difference in how many of the recommended terms were chosen by subjects for each level of task complexity15 . Subjects may have perceived IRF terms as more useful for high complexity tasks but this was not reflected in the proportion of IRF terms accepted. Differences may reside in the nature of the terms accepted; future work will investigate this issue. 3.1.3 Summary In this section we have presented an investigation on the effect of search task complexity on the utility of IRF. From the results there appears to be a strong relation between the complexity of the task and the subject interaction: subjects preferring IRF for highly complex tasks. Task complexity did not affect the proportion of terms accepted in either RF method, despite there being a difference in how relevant and useful subjects perceived the terms to be for different complexities; complexity may affect term selection in ways other than the proportion of terms accepted. 3.2 Search Experience Experienced searchers may interact differently and give different types of evidence to RF than inexperienced searchers. As such, levels of search experience may affect searchers" use and perceptions of IRF. In our experiment subjects were divided into two groups based on their level of search experience, the frequency with which they searched and the types of searches they performed. In this section we use their perceptions and logging to address the next research question; the relationship between the usefulness and use of IRF and the search experience of experimental subjects. The data are the same as that analysed in the previous section, but here we focus on search experience rather than the search task. 3.2.1 Feedback We analyse the results from the attitude statements described at the beginning of Section 3.1.1. (i.e., How you conveyed relevance to the system was… and How you conveyed relevance to the system made you feel…). These differentials elicited opinion from experimental subjects about the RF method used. In Table 5 we show the mean average responses for inexperienced and experienced subject groups on ERF and IRF; 24 subjects per cell. 13 This was the smallest number of query modification terms that were offered in both systems. 14 all T(16) ≥ 80, all p ≤ .31, (Wilcoxon Signed-Rank Test) 15 ERF: χ2 (2) = 3.67, p = .16; IRF: χ2 (2) = 2.55, p = .28 (KruskalWallis Tests) Table 5. Subject perceptions of RF method (lower = better). The results demonstrate a strong preference in inexperienced subjects for IRF; they found it more easy and effective than experienced subjects. 16 The differences for all other IRF differentials were not statistically significant. For all differentials, apart from in control, inexperienced subjects generally preferred IRF over ERF17 . Inexperienced subjects also felt that IRF was more difficult to control than experienced subjects18 . As these subjects have less search experience they may be less able to understand RF processes and may be more comfortable with the system gathering feedback implicitly from their interaction. Experienced subjects tended to like ERF more than inexperienced subjects and felt more comfortable with this feedback method19 . It appears from these results that experienced subjects found ERF more useful and were more at ease with the ERF process. In a similar way to Section 3.1.1 we analysed the proportion of feedback that searchers provided to the experimental systems. Our analysis suggested that search experience does not affect the amount of feedback subjects provide20 . 3.2.2 Terms We used questionnaire responses to gauge subject opinion on the relevance and usefulness of the terms from the perspective of experienced and inexperienced subjects. Table 6 shows the average differential responses obtained from both subject groups. Table 6. Subject perceptions of system terms (lower = better). Explicit RF Implicit RF Differential Inexp. Exp. Inexp. Exp. Relevant 2.58 2.44 2.33 2.21 Useful 2.88 2.63 2.33 2.23 The differences between subject groups were significant21 . Experienced subjects generally reacted to the query modification terms chosen by the system more positively than inexperienced 16 easy: U(24) = 391, p = .016; effective: U(24) = 399, p = .011; α = .0167 (Mann-Whitney Tests) 17 all T(24) ≥ 231, all p ≤ .001 (Wilcoxon Signed-Rank Test) 18 U(24) = 390, p = .018; α = .0250 (Mann-Whitney Test) 19 T(24) = 222, p = .020 (Wilcoxon Signed-Rank Test) 20 ERF: all U(24) ≤ 319, p ≥ .26, IRF: all U(24) ≤ 313, p ≥ .30 (MannWhitney Tests) 21 ERF: all U(24) ≥ 388, p ≤ .020, IRF: all U(24) ≥ 384, p ≤ .024 Explicit RF Implicit RF Differential Inexp. Exp. Inexp. Exp. Easy 2.46 2.46 1.84 1.98 Effective 2.75 2.63 2.32 2.43 Useful 2.50 2.46 2.28 2.27 All (1) 2.57 2.52 2.14 2.23 Comfortable 2.46 2.14 2.05 2.24 In control 1.96 1.98 2.73 2.64 All (2) 2.21 2.06 2.39 2.44 subjects. This finding was supported by the proportion of query modification terms these subjects accepted. In the same way as in Section 3.1.2, we analysed the number of query modification terms recommended by the system that were used by experimental subjects. Table 7 shows the average number of accepted terms per subject group. Table 7. Term Acceptance (percentage of top six terms). Explicit RF Implicit RFProportion of terms Inexp. Exp. Inexp. Exp. Accepted 63.76 70.44 64.43 71.35 Our analysis of the data show that differences between subject groups for each type of RF are significant; experienced subjects accepted more expansion terms regardless of type of RF. However, the differences between the same groups for different types of RF are not significant; subjects chose roughly the same percentage of expansion terms offered irrespective of the type of RF22 . 3.2.3 Summary In this section we have analysed data gathered from two subject groups - inexperienced searchers and experienced searchers - on how they perceive and use IRF. The results indicate that inexperienced subjects found IRF more easy and effective than experienced subjects, who in turn found the terms chosen as a result of IRF more relevant and useful. We also showed that inexperienced subjects generally accepted less recommended terms than experienced subjects, perhaps because they were less comfortable with RF or generally submitted shorter search queries. Search experience appears to affect how subjects use the terms recommended as a result of the RF process. 3.3 Search Stage From our observations of experimental subjects as they searched we conjectured that RF may be used differently at different times during a search. To test this, our third research question concerned the use and usefulness of IRF during the course of a search. In this section we investigate whether the amount of RF provided by searchers or the proportion of terms accepted are affected by how far through their search they are. For the purposes of this analysis a search begins when a subject poses the first query to the system and progresses until they terminate the search or reach the maximum allowed time for a search task of 15 minutes. We do not divide tasks based on this limit as subjects often terminated their search in less than 15 minutes. In this section we use data gathered from interaction logs and subject opinions to investigate the extent to which RF was used and the extent to which it appeared to benefit our experimental subjects at different stages in their search 3.3.1 Feedback The interaction logs for all searches on the Explicit RF and Implicit RF were analysed and each search is divided up into nine equal length time slices. This number of slices gave us an equal number per stage and was a sufficient level of granularity to identify trends in the results. Slices 1 - 3 correspond to the start of the search, 4 - 6 to the middle of the search and 7 - 9 to the end. In Figure 2 we plot the measure of precision described in Section 3.1.1 (i.e., the proportion of all possible representations that were provided as RF) at each of the 22 IRF: U(24) = 403, p = .009, ERF: U(24) = 396, p = .013 nine slices, per search task, averaged across all subjects; this allows us to see how the provision of RF was distributed during a search. The total amount of feedback for a single RF method/task complexity pairing across all nine slices corresponds to the value recorded in the first row of Table 2 (e.g., the sum of the RF for IRF/HC across all nine slices of Figure 2 is 21.50%). To simplify the statistical analysis and comparison we use the grouping of start, middle and end. 0 1 2 3 4 5 6 7 8 9 0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 Slice Search"precision"(%oftotalrepsprovidedasRF) Explicit RF/HC Explicit RF/MC Explicit RF/LC Implicit RF/HC Implicit RF/MC Implicit RF/LC Figure 2. Distribution of RF provision per search task. Figure 2 appears to show the existence of a relationship between the stage in the search and the amount of relevance information provided to the different types of feedback algorithm. These are essentially differences in the way users are assessing documents. In the case of ERF subjects provide explicit relevance assessments throughout most of the search, but there is generally a steep increase in the end phase towards the completion of the search23 . When using the IRF system, the data indicates that at the start of the search subjects are providing little relevance information24 , which corresponds to interacting with few document representations. At this stage the subjects are perhaps concentrating more on reading the retrieved results. Implicit relevance information is generally offered extensively in the middle of the search as they interact with results and it then tails off towards the end of the search. This would appear to correspond to stages of initial exploration, detailed analysis of document representations and storage and presentation of findings. Figure 2 also shows the proportion of feedback for tasks of different complexity. The results appear to show a difference25 in how IRF is used that relates to the complexity of the search task. More specifically, as complexity increases it appears as though subjects take longer to reach their most interactive point. This suggests that task complexity affects how IRF is distributed during the search and that they may be spending more time initially interpreting search results for more complex tasks. 23 IRF: all Z ≥ 1.87, p ≤ .031, ERF: start vs. end Z = 2.58, p = .005 (Dunn"s post-hoc tests). 24 Although increasing toward the end of the start stage. 25 Although not statistically significant; χ2 (2) = 3.54, p = .17 (Friedman Rank Sum Test) 3.3.2 Terms The terms recommended by the system are chosen based on the frequency of their occurrence in the relevant items. That is, nonstopword, non-query terms occurring frequently in search results regarded as relevant are likely to be recommended to the searcher for query modification. Since there is a direct association between the RF and the terms selected we use the number of terms accepted by searchers at different points in the search as an indication of how effective the RF has been up until the current point in the search. In this section we analysed the average number of terms from the top six terms recommended by Explicit RF and Implicit RF over the course of a search. The average proportion of the top six recommended terms that were accepted at each stage are shown in Table 8; each cell contains data from all 48 subjects. Table 8. Term Acceptance (proportion of top six terms). Explicit RF Implicit RFProportion of terms start middle end start middle end Accepted 66.87 66.98 67.34 61.85 68.54 73.22 The results show an apparent association between the stage in the search and the number of feedback terms subjects accept. Search stage affects term acceptance in IRF but not in ERF26 . The further into a search a searcher progresses, the more likely they are to accept terms recommended via IRF (significantly more than ERF27 ). A correlation analysis between the proportion of terms accepted at each search stage and cumulative RF (i.e., the sum of all precision at each slice in Figure 2 up to and including the end of the search stage) suggests that in both types of RF the quality of system terms improves as more RF is provided28 . 3.3.3 Summary The results from this section indicate that the location in a search affects the amount of feedback given by the user to the system, and hence the amount of information that the RF mechanism has to decide which terms to offer the user. Further, trends in the data suggest that the complexity of the task affects how subjects provide IRF and the proportion of system terms accepted. 4. DISCUSSION AND IMPLICATIONS In this section we discuss the implications of the findings presented in the previous section for each research question. 4.1 Search Task The results of our study showed that ERF was preferred for less complex tasks and IRF for more complex tasks. From observations and subject comments we perceived that when using ERF systems subjects generally forgot to provide the feedback but also employed different criteria during the ERF process (i.e., they were assessing relevance rather than expressing an interest). When the search was more complex subjects rarely found results they regarded as completely relevant. Therefore they struggled to find relevant 26 ERF: χ2 (2) = 2.22, p = .33; IRF: χ2 (2) = 7.73, p = .021 (Friedman Rank Sum Tests); IRF: all pair-wise comparisons significant at Z ≥ 1.77, all p ≤ .038 (Dunn"s post-hoc tests) 27 all T(48) ≥ 786, all p ≤ .002, (Wilcoxon Signed-Rank Test) 28 IRF: r = .712, p < .001, ERF: r = .695, p = .001 (Pearson Correlation Coefficient) information and were unable to communicate RF to the search system. In these situations subjects appeared to prefer IRF as they do not need to make a relevance decision to obtain the benefits of RF, i.e., term suggestions, whereas in ERF they do. The association between RF method and task complexity has implications for the design of user studies of RF systems and the RF systems themselves. It implies that in the design of user studies involving ERF or IRF systems care should be taken to include tasks of varying complexities, to avoid task bias. Also, in the design of search systems it implies that since different types of RF may be appropriate for different task complexities then a system that could automatically detect complexity could use both ERF and IRF simultaneously to benefit the searcher. For example, on the IRF system we noticed that as task complexity falls search behaviour shifts from results interface to retrieved documents. Monitoring such interaction across a number of studies may lead to a set of criteria that could help IR systems automatically detect task complexity and tailor support to suit. 4.2 Search Experience We analysed the affect of search experience on the utility of IRF. Our analysis revealed a general preference across all subjects for IRF over ERF. That is, the average ratings assigned to IRF were generally more positive than those assigned to ERF. However, IRF was generally liked by both subject groups (perhaps because it removed the burden of providing relevance information) and ERF was generally preferred by experienced subjects more than inexperienced subjects (perhaps because it allowed them to specify which results were used by the system when generating term recommendations). All subjects felt more in control with ERF than IRF, but for inexperienced subjects this did not appear to affect their overall preferences29 . These subjects may understand the RF process less, but may be more willing to sacrifice control over feedback in favour of IRF, a process that they perceive more positively. 4.3 Search Stage We also analysed the effects of search stage on the use and usefulness of IRF. Through analysis of this nature we can build a more complete picture of how searchers used RF and how this varies based on the RF method. The results suggest that IRF is used more in the middle of the search than at the beginning or end, whereas ERF is used more towards the end. The results also show the effects of task complexity on the IRF process and how rapidly subjects reach their most interactive point. Without an analysis of this type it would not have been possible to establish the existence of such patterns of behaviour. The findings suggest that searchers interact differently for IRF and ERF. Since ERF is not traditionally used until toward the end of the search it may be possible to incorporate both IRF and ERF into the same IR system, with IRF being used to gather evidence until subjects decide to use ERF. The development of such a system represents part of our ongoing work in this area. 5. CONCLUSIONS In this paper we have presented an investigation of Implicit Relevance Feedback (IRF). We aimed to answer three research questions about factors that may affect the provision and usefulness of IRF. These factors were search task complexity, the subjects" search experience and the stage in the search. Our overall conclusion was that all factors 29 This may also be true for experienced subjects, but the data we have is insufficient to draw this conclusion. appear to have some effect on the use and effectiveness of IRF, although the interaction effects between factors are not statistically significant. Our conclusions per each research question are: (i) IRF is generally more useful for complex search tasks, where searchers want to focus on the search task and get new ideas for their search from the system, (ii) IRF is preferred to ERF overall and generally preferred by inexperienced subjects wanting to reduce the burden of providing RF, and (iii) within a single search session IRF is affected by temporal location in a search (i.e., it is used in the middle, not the beginning or end) and task complexity. Studies of this nature are important to establish the circumstances where a promising technique such as IRF are useful and those when it is not. It is only after such studies have been run and analysed in this way can we develop an understanding of IRF that allow it to be successfully implemented in operational IR systems. 6. REFERENCES [1] Bell, D.J. and Ruthven, I. (2004). Searchers' assessments of task complexity for web searching. Proceedings of the 26th European Conference on Information Retrieval, 57-71. [2] Borlund, P. (2000). Experimental components for the evaluation of interactive information retrieval systems. Journal of Documentation. 56(1): 71-90. [3] Brajnik, G., Mizzaro, S., Tasso, C., and Venuti, F. (2002). Strategic help for user interfaces for information retrieval. Journal of the American Society for Information Science and Technology. 53(5): 343-358. [4] Busha, C.H. and Harter, S.P., (1980). Research methods in librarianship: Techniques and interpretation. Library and information science series. New York: Academic Press. [5] Campbell, I. and Van Rijsbergen, C.J. (1996). The ostensive model of developing information needs. Proceedings of the 3rd International Conference on Conceptions of Library and Information Science, 251-268. [6] Harman, D., (1992). Relevance feedback and other query modification techniques. In Information retrieval: Data structures and algorithms. New York: Prentice-Hall. [7] Kelly, D. and Teevan, J. (2003). Implicit feedback for inferring user preference. SIGIR Forum. 37(2): 18-28. [8] Koenemann, J. and Belkin, N.J. (1996). A case for interaction: A study of interactive information retrieval behavior and effectiveness. Proceedings of the ACM SIGCHI Conference on Human Factors in Computing Systems, 205-212. [9] Meddis, R., (1984). Statistics using ranks: A unified approach. Oxford: Basil Blackwell, 303-308. [10] Morita, M. and Shinoda, Y. (1994). Information filtering based on user behavior analysis and best match text retrieval. Proceedings of the 17th Annual ACM SIGIR Conference on Research and Development in Information Retrieval, 272-281. [11] Salton, G. and Buckley, C. (1990). Improving retrieval performance by relevance feedback. Journal of the American Society for Information Science. 41(4): 288-297. [12] Siegel, S. and Castellan, N.J. (1988). Nonparametric statistics for the behavioural sciences. 2nd ed. Singapore: McGraw-Hill. [13] White, R.W. (2004). Implicit feedback for interactive information retrieval. Unpublished Doctoral Dissertation, University of Glasgow, Glasgow, United Kingdom. [14] White, R.W., Jose, J.M. and Ruthven, I. (2005). An implicit feedback approach for interactive information retrieval, Information Processing and Management, in press. [15] White, R.W., Jose, J.M., Ruthven, I. and Van Rijsbergen, C.J. (2004). A simulated study of implicit feedback models. Proceedings of the 26th European Conference on Information Retrieval, 311-326. [16] Zellweger, P.T., Regli, S.H., Mackinlay, J.D., and Chang, B.-W. (2000). The impact of fluid documents on reading and browsing: An observational study. Proceedings of the ACM SIGCHI Conference on Human Factors in Computing Systems, 249-256. Appendix B. Checkboxes to mark relevant document titles in the Explicit RF system. Appendix A. Interface to Implicit RF system. 1. Top-Ranking Sentence 2. Title 3. Summary 4. Summary Sentence 5. Sentence in Context 2 3 4 5 1

implicit relevance feedback;top-ranking sentence;query modification term;interactive query expansion feature;varying complexity;explicit rf system;medium complexity;high complexity whilst;proportion feedback;moderate complexity whilst;search task complexity;browse-based result interface;search precision;relevance feedback;introductory questionnaire

train_H-97

Feature Representation for Effective Action-Item Detection

E-mail users face an ever-growing challenge in managing their inboxes due to the growing centrality of email in the workplace for task assignment, action requests, and other roles beyond information dissemination. Whereas Information Retrieval and Machine Learning techniques are gaining initial acceptance in spam filtering and automated folder assignment, this paper reports on a new task: automated action-item detection, in order to flag emails that require responses, and to highlight the specific passage(s) indicating the request(s) for action. Unlike standard topic-driven text classification, action-item detection requires inferring the sender"s intent, and as such responds less well to pure bag-of-words classification. However, using enriched feature sets, such as n-grams (up to n=4) with chi-squared feature selection, and contextual cues for action-item location improve performance by up to 10% over unigrams, using in both cases state of the art classifiers such as SVMs with automated model selection via embedded cross-validation.

1. INTRODUCTION E-mail users are facing an increasingly difficult task of managing their inboxes in the face of mounting challenges that result from rising e-mail usage. This includes prioritizing e-mails over a range of sources from business partners to family members, filtering and reducing junk e-mail, and quickly managing requests that demand From: Henry Hutchins <hhutchins@innovative.company.com> To: Sara Smith; Joe Johnson; William Woolings Subject: meeting with prospective customers Sent: Fri 12/10/2005 8:08 AM Hi All, I"d like to remind all of you that the group from GRTY will be visiting us next Friday at 4:30 p.m. The current schedule looks like this: + 9:30 a.m. Informal Breakfast and Discussion in Cafeteria + 10:30 a.m. Company Overview + 11:00 a.m. Individual Meetings (Continue Over Lunch) + 2:00 p.m. Tour of Facilities + 3:00 p.m. Sales Pitch In order to have this go off smoothly, I would like to practice the presentation well in advance. As a result, I will need each of your parts by Wednesday. Keep up the good work! -Henry Figure 1: An E-mail with emphasized Action-Item, an explicit request that requires the recipient"s attention or action. the receiver"s attention or action. Automated action-item detection targets the third of these problems by attempting to detect which e-mails require an action or response with information, and within those e-mails, attempting to highlight the sentence (or other passage length) that directly indicates the action request. Such a detection system can be used as one part of an e-mail agent which would assist a user in processing important e-mails quicker than would have been possible without the agent. We view action-item detection as one necessary component of a successful e-mail agent which would perform spam detection, action-item detection, topic classification and priority ranking, among other functions. The utility of such a detector can manifest as a method of prioritizing e-mails according to task-oriented criteria other than the standard ones of topic and sender or as a means of ensuring that the email user hasn"t dropped the proverbial ball by forgetting to address an action request. Action-item detection differs from standard text classification in two important ways. First, the user is interested both in detecting whether an email contains action items and in locating exactly where these action item requests are contained within the email body. In contrast, standard text categorization merely assigns a topic label to each text, whether that label corresponds to an e-mail folder or a controlled indexing vocabulary [12, 15, 22]. Second, action-item detection attempts to recover the email sender"s intent - whether she means to elicit response or action on the part of the receiver; note that for this task, classifiers using only unigrams as features do not perform optimally, as evidenced in our results below. Instead we find that we need more information-laden features such as higher-order n-grams. Text categorization by topic, on the other hand, works very well using just individual words as features [2, 9, 13, 17]. In fact, genre-classification, which one would think may require more than a bag-of-words approach, also works quite well using just unigram features [14]. Topic detection and tracking (TDT), also works well with unigram feature sets [1, 20]. We believe that action-item detection is one of the first clear instances of an IR-related task where we must move beyond bag-of-words to achieve high performance, albeit not too far, as bag-of-n-grams seem to suffice. We first review related work for similar text classification problems such as e-mail priority ranking and speech act identification. Then we more formally define the action-item detection problem, discuss the aspects that distinguish it from more common problems like topic classification, and highlight the challenges in constructing systems that can perform well at the sentence and document level. From there, we move to a discussion of feature representation and selection techniques appropriate for this problem and how standard text classification approaches can be adapted to smoothly move from the sentence-level detection problem to the documentlevel classification problem. We then conduct an empirical analysis that helps us determine the effectiveness of our feature extraction procedures as well as establish baselines for a number of classification algorithms on this task. Finally, we summarize this paper"s contributions and consider interesting directions for future work. 2. RELATED WORK Several other researchers have considered very similar text classification tasks. Cohen et al. [5] describe an ontology of speech acts, such as Propose a Meeting, and attempt to predict when an e-mail contains one of these speech acts. We consider action-items to be an important specific type of speech act that falls within their more general classification. While they provide results for several classification methods, their methods only make use of human judgments at the document-level. In contrast, we consider whether accuracy can be increased by using finer-grained human judgments that mark the specific sentences and phrases of interest. Corston-Oliver et al. [6] consider detecting items in e-mail to Put on a To-Do List. This classification task is very similar to ours except they do not consider simple factual questions to belong to this category. We include questions, but note that not all questions are action-items - some are rhetorical or simply social convention, How are you?. From a learning perspective, while they make use of judgments at the sentence-level, they do not explicitly compare what if any benefits finer-grained judgments offer. Additionally, they do not study alternative choices or approaches to the classification task. Instead, they simply apply a standard SVM at the sentence-level and focus primarily on a linguistic analysis of how the sentence can be logically reformulated before adding it to the task list. In this study, we examine several alternative classification methods, compare document-level and sentence-level approaches and analyze the machine learning issues implicit in these problems. Interest in a variety of learning tasks related to e-mail has been rapidly growing in the recent literature. For example, in a forum dedicated to e-mail learning tasks, Culotta et al. [7] presented methods for learning social networks from e-mail. In this work, we do not focus on peer relationships; however, such methods could complement those here since peer relationships often influence word choice when requesting an action. 3. PROBLEM DEFINITION & APPROACH In contrast to previous work, we explicitly focus on the benefits that finer-grained, more costly, sentence-level human judgments offer over coarse-grained document-level judgments. Additionally, we consider multiple standard text classification approaches and analyze both the quantitative and qualitative differences that arise from taking a document-level vs. a sentence-level approach to classification. Finally, we focus on the representation necessary to achieve the most competitive performance. 3.1 Problem Definition In order to provide the most benefit to the user, a system would not only detect the document, but it would also indicate the specific sentences in the e-mail which contain the action-items. Therefore, there are three basic problems: 1. Document detection: Classify a document as to whether or not it contains an action-item. 2. Document ranking: Rank the documents such that all documents containing action-items occur as high as possible in the ranking. 3. Sentence detection: Classify each sentence in a document as to whether or not it is an action-item. As in most Information Retrieval tasks, the weight the evaluation metric should give to precision and recall depends on the nature of the application. In situations where a user will eventually read all received messages, ranking (e.g., via precision at recall of 1) may be most important since this will help encourage shorter delays in communications between users. In contrast, high-precision detection at low recall will be of increasing importance when the user is under severe time-pressure and therefore will likely not read all mail. This can be the case for crisis managers during disaster management. Finally, sentence detection plays a role in both timepressure situations and simply to alleviate the user"s required time to gist the message. 3.2 Approach As mentioned above, the labeled data can come in one of two forms: a document-labeling provides a yes/no label for each document as to whether it contains an action-item; a phrase-labeling provides only a yes label for the specific items of interest. We term the human judgments a phrase-labeling since the user"s view of the action-item may not correspond with actual sentence boundaries or predicted sentence boundaries. Obviously, it is straightforward to generate a document-labeling consistent with a phrase-labeling by labeling a document yes if and only if it contains at least one phrase labeled yes. To train classifiers for this task, we can take several viewpoints related to both the basic problems we have enumerated and the form of the labeled data. The document-level view treats each e-mail as a learning instance with an associated class-label. Then, the document can be converted to a feature-value vector and learning progresses as usual. Applying a document-level classifier to document detection and ranking is straightforward. In order to apply it to sentence detection, one must make additional steps. For example, if the classifier predicts a document contains an action-item, then areas of the document that contain a high-concentration of words which the model weights heavily in favor of action-items can be indicated. The obvious benefit of the document-level approach is that training set collection costs are lower since the user only has to specify whether or not an e-mail contains an action-item and not the specific sentences. In the sentence-level view, each e-mail is automatically segmented into sentences, and each sentence is treated as a learning instance with an associated class-label. Since the phrase-labeling provided by the user may not coincide with the automatic segmentation, we must determine what label to assign a partially overlapping sentence when converting it to a learning instance. Once trained, applying the resulting classifiers to sentence detection is now straightforward, but in order to apply the classifiers to document detection and document ranking, the individual predictions over each sentence must be aggregated in order to make a document-level prediction. This approach has the potential to benefit from morespecific labels that enable the learner to focus attention on the key sentences instead of having to learn based on data that the majority of the words in the e-mail provide no or little information about class membership. 3.2.1 Features Consider some of the phrases that might constitute part of an action item: would like to know, let me know, as soon as possible, have you. Each of these phrases consists of common words that occur in many e-mails. However, when they occur in the same sentence, they are far more indicative of an action-item. Additionally, order can be important: consider have you versus you have. Because of this, we posit that n-grams play a larger role in this problem than is typical of problems like topic classification. Therefore, we consider all n-grams up to size 4. When using n-grams, if we find an n-gram of size 4 in a segment of text, we can represent the text as just one occurrence of the ngram or as one occurrence of the n-gram and an occurrence of each smaller n-gram contained by it. We choose the second of these alternatives since this will allow the algorithm itself to smoothly back-off in terms of recall. Methods such as na¨ıve Bayes may be hurt by such a representation because of double-counting. Since sentence-ending punctuation can provide information, we retain the terminating punctuation token when it is identifiable. Additionally, we add a beginning-of-sentence and end-of-sentence token in order to capture patterns that are often indicators at the beginning or end of a sentence. Assuming proper punctuation, these extra tokens are unnecessary, but often e-mail lacks proper punctuation. In addition, for the sentence-level classifiers that use ngrams, we additionally code for each sentence a binary encoding of the position of the sentence relative to the document. This encoding has eight associated features that represent which octile (the first eighth, second eighth, etc.) contains the sentence. 3.2.2 Implementation Details In order to compare the document-level to the sentence-level approach, we compare predictions at the document-level. We do not address how to use a document-level classifier to make predictions at the sentence-level. In order to automatically segment the text of the e-mail, we use the RASP statistical parser [4]. Since the automatically segmented sentences may not correspond directly with the phrase-level boundaries, we treat any sentence that contains at least 30% of a marked action-item segment as an action-item. When evaluating sentencedetection for the sentence-level system, we use these class labels as ground truth. Since we are not evaluating multiple segmentation approaches, this does not bias any of the methods. If multiple segmentation systems were under evaluation, one would need to use a metric that matched predicted positive sentences to phrases labeled positive. The metric would need to punish overly long true predictions as well as too short predictions. Our criteria for converting to labeled instances implicitly includes both criteria. Since the segmentation is fixed, an overly long prediction would be predicting yes for many no instances since presumably the extra length corresponds to additional segmented sentences all of which do not contain 30% of action-item. Likewise, a too short prediction must correspond to a small sentence included in the action-item but not constituting all of the action-item. Therefore, in order to consider the prediction to be too short, there will be an additional preceding/following sentence that is an action-item where we incorrectly predicted no. Once a sentence-level classifier has made a prediction for each sentence, we must combine these predictions to make both a document-level prediction and a document-level score. We use the simple policy of predicting positive when any of the sentences is predicted positive. In order to produce a document score for ranking, the confidence that the document contains an action-item is: ψ(d) = 1 n(d) s∈d|π(s)=1 ψ(s) if for any s ∈ d, π(s) = 1 1 n(d) maxs∈d ψ(s) o.w. where s is a sentence in document d, π is the classifier"s 1/0 prediction, ψ is the score the classifier assigns as its confidence that π(s) = 1, and n(d) is the greater of 1 and the number of (unigram) tokens in the document. In other words, when any sentence is predicted positive, the document score is the length normalized sum of the sentence scores above threshold. When no sentence is predicted positive, the document score is the maximum sentence score normalized by length. As in other text problems, we are more likely to emit false positives for documents with more words or sentences. Thus we include a length normalization factor. 4. EXPERIMENTAL ANALYSIS 4.1 The Data Our corpus consists of e-mails obtained from volunteers at an educational institution and cover subjects such as: organizing a research workshop, arranging for job-candidate interviews, publishing proceedings, and talk announcements. The messages were anonymized by replacing the names of each individual and institution with a pseudonym.1 After attempting to identify and eliminate duplicate e-mails, the corpus contains 744 e-mail messages. After identity anonymization, the corpora has three basic versions. Quoted material refers to the text of a previous e-mail that an author often leaves in an e-mail message when responding to the e-mail. Quoted material can act as noise when learning since it may include action-items from previous messages that are no longer relevant. To isolate the effects of quoted material, we have three versions of the corpora. The raw form contains the basic messages. The auto-stripped version contains the messages after quoted material has been automatically removed. The hand-stripped version contains the messages after quoted material has been removed by a human. Additionally, the hand-stripped version has had any xml content and e-mail signatures removed - leaving only the essential content of the message. The studies reported here are performed with the hand-stripped version. This allows us to balance the cognitive load in terms of number of tokens that must be read in the user-studies we report - including quoted material would complicate the user studies since some users might skip the material while others read it. Additionally, ensuring all quoted material is removed 1 We have an even more highly anonymized version of the corpus that can be made available for some outside experimentation. Please contact the authors for more information on obtaining this data. prevents tainting the cross-validation since otherwise a test item could occur as quoted material in a training document. 4.1.1 Data Labeling Two human annotators labeled each message as to whether or not it contained an action-item. In addition, they identified each segment of the e-mail which contained an action-item. A segment is a contiguous section of text selected by the human annotators and may span several sentences or a complete phrase contained in a sentence. They were instructed that an action item is an explicit request for information that requires the recipient"s attention or a required action and told to highlight the phrases or sentences that make up the request. Annotator 1 No Yes Annotator 2 No 391 26 Yes 29 298 Table 1: Agreement of Human Annotators at Document Level Annotator One labeled 324 messages as containing action items. Annotator Two labeled 327 messages as containing action items. The agreement of the human annotators is shown in Tables 1 and 2. The annotators are said to agree at the document-level when both marked the same document as containing no action-items or both marked at least one action-item regardless of whether the text segments were the same. At the document-level, the annotators agreed 93% of the time. The kappa statistic [3, 5] is often used to evaluate inter-annotator agreement: κ = A − R 1 − R A is the empirical estimate of the probability of agreement. R is the empirical estimate of the probability of random agreement given the empirical class priors. A value close to −1 implies the annotators agree far less often than would be expected randomly, while a value close to 1 means they agree more often than randomly expected. At the document-level, the kappa statistic for inter-annotator agreement is 0.85. This value is both strong enough to expect the problem to be learnable and is comparable with results for similar tasks [5, 6]. In order to determine the sentence-level agreement, we use each judgment to create a sentence-corpus with labels as described in Section 3.2.2, then consider the agreement over these sentences. This allows us to compare agreement over no judgments. We perform this comparison over the hand-stripped corpus since that eliminates spurious no judgments that would come from including quoted material, etc. Both annotators were free to label the subject as an action-item, but since neither did, we omit the subject line of the message as well. This only reduces the number of no agreements. This leaves 6301 automatically segmented sentences. At the sentence-level, the annotators agreed 98% of the time, and the kappa statistic for inter-annotator agreement is 0.82. In order to produce one single set of judgments, the human annotators went through each annotation where there was disagreement and came to a consensus opinion. The annotators did not collect statistics during this process but anecdotally reported that the majority of disagreements were either cases of clear annotator oversight or different interpretations of conditional statements. For example, If you would like to keep your job, come to tomorrow"s meeting implies a required action where If you would like to join Annotator 1 No Yes Annotator 2 No 5810 65 Yes 74 352 Table 2: Agreement of Human Annotators at Sentence Level the football betting pool, come to tomorrow"s meeting does not. The first would be an action-item in most contexts while the second would not. Of course, many conditional statements are not so clearly interpretable. After reconciling the judgments there are 416 e-mails with no action-items and 328 e-mails containing actionitems. Of the 328 e-mails containing action-items, 259 messages have one action-item segment; 55 messages have two action-item segments; 11 messages have three action-item segments. Two messages have four action-item segments, and one message has six action-item segments. Computing the sentence-level agreement using the reconciled gold standard judgments with each of the annotators" individual judgments gives a kappa of 0.89 for Annotator One and a kappa of 0.92 for Annotator Two. In terms of message characteristics, there were on average 132 content tokens in the body after stripping. For action-item messages, there were 115. However, by examining Figure 2 we see the length distributions are nearly identical. As would be expected for e-mail, it is a long-tailed distribution with about half the messages having more than 60 tokens in the body (this paragraph has 65 tokens). 4.2 Classifiers For this experiment, we have selected a variety of standard text classification algorithms. In selecting algorithms, we have chosen algorithms that are not only known to work well but which differ along such lines as discriminative vs. generative and lazy vs. eager. We have done this in order to provide both a competitive and thorough sampling of learning methods for the task at hand. This is important since it is easy to improve a strawman classifier by introducing a new representation. By thoroughly sampling alternative classifier choices we demonstrate that representation improvements over bag-of-words are not due to using the information in the bag-of-words poorly. 4.2.1 kNN We employ a standard variant of the k-nearest neighbor algorithm used in text classification, kNN with s-cut score thresholding [19]. We use a tfidf-weighting of the terms with a distanceweighted vote of the neighbors to compute the score before thresholding it. In order to choose the value of s for thresholding, we perform leave-one-out cross-validation over the training set. The value of k is set to be 2( log2 N + 1) where N is the number of training points. This rule for choosing k is theoretically motivated by results which show such a rule converges to the optimal classifier as the number of training points increases [8]. In practice, we have also found it to be a computational convenience that frequently leads to comparable results with numerically optimizing k via a cross-validation procedure. 4.2.2 Na¨ıve Bayes We use a standard multinomial na¨ıve Bayes classifier [16]. In using this classifier, we smoothed word and class probabilities using a Bayesian estimate (with the word prior) and a Laplace m-estimate, respectively. 0 20 40 60 80 100 120 140 160 0 200 400 600 800 1000 1200 1400 NumberofMessages Number of Tokens All Messages Action-Item Messages 0 0.02 0.04 0.06 0.08 0.1 0.12 0.14 0.16 0.18 0.2 0 200 400 600 800 1000 1200 1400 PercentageofMessages Number of Tokens All Messages Action-Item Messages Figure 2: The Histogram (left) and Distribution (right) of Message Length. A bin size of 20 words was used. Only tokens in the body after hand-stripping were counted. After stripping, the majority of words left are usually actual message content. Classifiers Document Unigram Document Ngram Sentence Unigram Sentence Ngram F1 kNN 0.6670 ± 0.0288 0.7108 ± 0.0699 0.7615 ± 0.0504 0.7790 ± 0.0460 na¨ıve Bayes 0.6572 ± 0.0749 0.6484 ± 0.0513 0.7715 ± 0.0597 0.7777 ± 0.0426 SVM 0.6904 ± 0.0347 0.7428 ± 0.0422 0.7282 ± 0.0698 0.7682 ± 0.0451 Voted Perceptron 0.6288 ± 0.0395 0.6774 ± 0.0422 0.6511 ± 0.0506 0.6798 ± 0.0913 Accuracy kNN 0.7029 ± 0.0659 0.7486 ± 0.0505 0.7972 ± 0.0435 0.8092 ± 0.0352 na¨ıve Bayes 0.6074 ± 0.0651 0.5816 ± 0.1075 0.7863 ± 0.0553 0.8145 ± 0.0268 SVM 0.7595 ± 0.0309 0.7904 ± 0.0349 0.7958 ± 0.0551 0.8173 ± 0.0258 Voted Perceptron 0.6531 ± 0.0390 0.7164 ± 0.0376 0.6413 ± 0.0833 0.7082 ± 0.1032 Table 3: Average Document-Detection Performance during Cross-Validation for Each Method and the Sample Standard Deviation (Sn−1) in italics. The best performance for each classifier is shown in bold. 4.2.3 SVM We have used a linear SVM with a tfidf feature representation and L2-norm as implemented in the SVMlight package v6.01 [11]. All default settings were used. 4.2.4 Voted Perceptron Like the SVM, the Voted Perceptron is a kernel-based learning method. We use the same feature representation and kernel as we have for the SVM, a linear kernel with tfidf-weighting and an L2-norm. The voted perceptron is an online-learning method that keeps a history of past perceptrons used, as well as a weight signifying how often that perceptron was correct. With each new training example, a correct classification increases the weight on the current perceptron and an incorrect classification updates the perceptron. The output of the classifier uses the weights on the perceptra to make a final voted classification. When used in an offline-manner, multiple passes can be made through the training data. Both the voted perceptron and the SVM give a solution from the same hypothesis space - in this case, a linear classifier. Furthermore, it is well-known that the Voted Perceptron increases the margin of the solution after each pass through the training data [10]. Since Cohen et al. [5] obtain worse results using an SVM than a Voted Perceptron with one training iteration, they conclude that the best solution for detecting speech acts may not lie in an area with a large margin. Because their tasks are highly similar to ours, we employ both classifiers to ensure we are not overlooking a competitive alternative classifier to the SVM for the basic bag-of-words representation. 4.3 Performance Measures To compare the performance of the classification methods, we look at two standard performance measures, F1 and accuracy. The F1 measure [18, 21] is the harmonic mean of precision and recall where Precision = Correct Positives Predicted Positives and Recall = Correct Positives Actual Positives . 4.4 Experimental Methodology We perform standard 10-fold cross-validation on the set of documents. For the sentence-level approach, all sentences in a document are either entirely in the training set or entirely in the test set for each fold. For significance tests, we use a two-tailed t-test [21] to compare the values obtained during each cross-validation fold with a p-value of 0.05. Feature selection was performed using the chi-squared statistic. Different levels of feature selection were considered for each classifier. Each of the following number of features was tried: 10, 25, 50, 100, 250, 750, 1000, 2000, 4000. There are approximately 4700 unigram tokens without feature selection. In order to choose the number of features to use for each classifier, we perform nested cross-validation and choose the settings that yield the optimal document-level F1 for that classifier. For this study, only the body of each e-mail message was used. Feature selection is always applied to all candidate features. That is, for the n-gram representation, the n-grams and position features are also subject to removal by the feature selection method. 4.5 Results The results for document-level classification are given in Table 3. The primary hypothesis we are concerned with is that n-grams are critical for this task; if this is true, we expect to see a significant gap in performance between the document-level classifiers that use n-grams (denoted Document Ngram) and those using only unigram features (denoted Document Unigram). Examining Table 3, we observe that this is indeed the case for every classifier except na¨ıve Bayes. This difference in performance produced by the n-gram representation is statistically significant for each classifier except for na¨ıve Bayes and the accuracy metric for kNN (see Table 4). Na¨ıve Bayes poor performance with the n-gram representation is not surprising since the bag-of-n-grams causes excessive doublecounting as mentioned in Section 3.2.1; however, na¨ıve Bayes is not hurt at the sentence-level because the sparse examples provide few chances for agglomerative effects of double counting. In either case, when a language-modeling approach is desired, modeling the n-grams directly would be preferable to na¨ıve Bayes. More importantly for the n-gram hypothesis, the n-grams lead to the best document-level classifier performance as well. As would be expected, the difference between the sentence-level n-gram representation and unigram representation is small. This is because the window of text is so small that the unigram representation, when done at the sentence-level, implicitly picks up on the power of the n-grams. Further improvement would signify that the order of the words matter even when only considering a small sentence-size window. Therefore, the finer-grained sentence-level judgments allows a unigram representation to succeed but only when performed in a small window - behaving as an n-gram representation for all practical purposes. Document Winner Sentence Winner kNN Ngram Ngram na¨ıve Bayes Unigram Ngram SVM Ngram† Ngram Voted Perceptron Ngram† Ngram Table 4: Significance results for n-grams versus unigrams for document detection using document-level and sentence-level classifiers. When the F1 result is statistically significant, it is shown in bold. When the accuracy result is significant, it is shown with a † . F1 Winner Accuracy Winner kNN Sentence Sentence na¨ıve Bayes Sentence Sentence SVM Sentence Sentence Voted Perceptron Sentence Document Table 5: Significance results for sentence-level classifiers vs. document-level classifiers for the document detection problem. When the result is statistically significant, it is shown in bold. Further highlighting the improvement from finer-grained judgments and n-grams, Figure 3 graphically depicts the edge the SVM sentence-level classifier has over the standard bag-of-words approach with a precision-recall curve. In the high precision area of the graph, the consistent edge of the sentence-level classifier is rather impressive - continuing at precision 1 out to 0.1 recall. This would mean that a tenth of the user"s action-items would be placed at the top of their action-item sorted inbox. Additionally, the large separation at the top right of the curves corresponds to the area where the optimal F1 occurs for each classifier, agreeing with the large improvement from 0.6904 to 0.7682 in F1 score. Considering the relative unexplored nature of classification at the sentence-level, this gives great hope for further increases in performance. Accuracy F1 Unigram Ngram Unigram Ngram kNN 0.9519 0.9536 0.6540 0.6686 na¨ıve Bayes 0.9419 0.9550 0.6176 0.6676 SVM 0.9559 0.9579 0.6271 0.6672 Voted Perceptron 0.8895 0.9247 0.3744 0.5164 Table 6: Performance of the Sentence-Level Classifiers at Sentence Detection Although Cohen et al. [5] observed that the Voted Perceptron with a single training iteration outperformed SVM in a set of similar tasks, we see no such behavior here. This further strengthens the evidence that an alternate classifier with the bag-of-words representation could not reach the same level of performance. The Voted Perceptron classifier does improve when the number of training iterations are increased, but it is still lower than the SVM classifier. Sentence detection results are presented in Table 6. With regard to the sentence detection problem, we note that the F1 measure gives a better feel for the remaining room for improvement in this difficult problem. That is, unlike document detection where actionitem documents are fairly common, action-item sentences are very rare. Thus, as in other text problems, the accuracy numbers are deceptively high sheerly because of the default accuracy attainable by always predicting no. Although, the results here are significantly above-random, it is unclear what level of performance is necessary for sentence detection to be useful in and of itself and not simply as a means to document ranking and classification. Figure 4: Users find action-items quicker when assisted by a classification system. Finally, when considering a new type of classification task, one of the most basic questions is whether an accurate classifier built for the task can have an impact on the end-user. In order to demonstrate the impact this task can have on e-mail users, we conducted a user study using an earlier less-accurate version of the sentence classifier - where instead of using just a single sentence, a threesentence windowed-approach was used. There were three distinct sets of e-mail in which users had to find action-items. These sets were either presented in a random order (Unordered), ordered by the classifier (Ordered), or ordered by the classifier and with the 0.4 0.5 0.6 0.7 0.8 0.9 1 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 Precision Recall Action-Item Detection SVM Performance (Post Model Selection) Document Unigram Sentence Ngram Figure 3: Both n-grams and a small prediction window lead to consistent improvements over the standard approach. center sentence in the highest confidence window highlighted (Order+help). In order to perform fair comparisons between conditions, the overall number of tokens in each message set should be approximately equal; that is, the cognitive reading load should be approximately the same before the classifier"s reordering. Additionally, users typically show practice effects by improving at the overall task and thus performing better at later message sets. This is typically handled by varying the ordering of the sets across users so that the means are comparable. While omitting further detail, we note the sets were balanced for the total number of tokens and a latin square design was used to balance practice effects. Figure 4 shows that at intervals of 5, 10, and 15 minutes, users consistently found significantly more action-items when assisted by the classifier, but were most critically aided in the first five minutes. Although, the classifier consistently aids the users, we did not gain an additional end-user impact by highlighting. As mentioned above, this might be a result of the large room for improvement that still exists for sentence detection, but anecdotal evidence suggests this might also be a result of how the information is presented to the user rather than the accuracy of sentence detection. For example, highlighting the wrong sentence near an actual action-item hurts the user"s trust, but if a vague indicator (e.g., an arrow) points to the approximate area the user is not aware of the near-miss. Since the user studies used a three sentence window, we believe this played a role as well as sentence detection accuracy. 4.6 Discussion In contrast to problems where n-grams have yielded little difference, we believe their power here stems from the fact that many of the meaningful n-grams for action-items consist of common words, e.g., let me know. Therefore, the document-level unigram approach cannot gain much leverage, even when modeling their joint probability correctly, since these words will often co-occur in the document but not necessarily in a phrase. Additionally, action-item detection is distinct from many text classification tasks in that a single sentence can change the class label of the document. As a result, good classifiers cannot rely on aggregating evidence from a large number of weak indicators across the entire document. Even though we discarded the header information, examining the top-ranked features at the document-level reveals that many of the features are names or parts of e-mail addresses that occurred in the body and are highly associated with e-mails that tend to contain many or no action-items. A few examples are terms such as org, bob, and gov. We note that these features will be sensitive to the particular distribution (senders/receivers) and thus the document-level approach may produce classifiers that transfer less readily to alternate contexts and users at different institutions. This points out that part of the problem of going beyond bag-of-words may be the methodology, and investigating such properties as learning curves and how well a model transfers may highlight differences in models which appear to have similar performance when tested on the distributions they were trained on. We are currently investigating whether the sentence-level classifiers do perform better over different test corpora without retraining. 5. FUTURE WORK While applying text classifiers at the document-level is fairly well-understood, there exists the potential for significantly increasing the performance of the sentence-level classifiers. Such methods include alternate ways of combining the predictions over each sentence, weightings other than tfidf, which may not be appropriate since sentences are small, better sentence segmentation, and other types of phrasal analysis. Additionally, named entity tagging, time expressions, etc., seem likely candidates for features that can further improve this task. We are currently pursuing some of these avenues to see what additional gains these offer. Finally, it would be interesting to investigate the best methods for combining the document-level and sentence-level classifiers. Since the simple bag-of-words representation at the document-level leads to a learned model that behaves somewhat like a context-specific prior dependent on the sender/receiver and general topic, a first choice would be to treat it as such when combining probability estimates with the sentence-level classifier. Such a model might serve as a general example for other problems where bag-of-words can establish a baseline model but richer approaches are needed to achieve performance beyond that baseline. 6. SUMMARY AND CONCLUSIONS The effectiveness of sentence-level detection argues that labeling at the sentence-level provides significant value. Further experiments are needed to see how this interacts with the amount of training data available. Sentence detection that is then agglomerated to document-level detection works surprisingly better given low recall than would be expected with sentence-level items. This, in turn, indicates that improved sentence segmentation methods could yield further improvements in classification. In this work, we examined how action-items can be effectively detected in e-mails. Our empirical analysis has demonstrated that n-grams are of key importance to making the most of documentlevel judgments. When finer-grained judgments are available, then a standard bag-of-words approach using a small (sentence) window size and automatic segmentation techniques can produce results almost as good as the n-gram based approaches. Acknowledgments This material is based upon work supported by the Defense Advanced Research Projects Agency (DARPA) under Contract No. NBCHD030010. Any opinions, findings and conclusions or recommendations expressed in this material are those of the author(s) and do not necessarily reflect the views of the Defense Advanced Research Projects Agency (DARPA), or the Department of InteriorNational Business Center (DOI-NBC). We would like to extend our sincerest thanks to Jill Lehman whose efforts in data collection were essential in constructing the corpus, and both Jill and Aaron Steinfeld for their direction of the HCI experiments. We would also like to thank Django Wexler for constructing and supporting the corpus labeling tools and Curtis Huttenhower"s support of the text preprocessing package. Finally, we gratefully acknowledge Scott Fahlman for his encouragement and useful discussions on this topic. 7. REFERENCES [1] J. Allan, J. Carbonell, G. Doddington, J. Yamron, and Y. Yang. Topic detection and tracking pilot study: Final report. In Proceedings of the DARPA Broadcast News Transcription and Understanding Workshop, Washington, D.C., 1998. [2] C. Apte, F. Damerau, and S. M. Weiss. Automated learning of decision rules for text categorization. ACM Transactions on Information Systems, 12(3):233-251, July 1994. [3] J. Carletta. Assessing agreement on classification tasks: The kappa statistic. Computational Linguistics, 22(2):249-254, 1996. [4] J. Carroll. High precision extraction of grammatical relations. In Proceedings of the 19th International Conference on Computational Linguistics (COLING), pages 134-140, 2002. [5] W. W. Cohen, V. R. Carvalho, and T. M. Mitchell. Learning to classify email into speech acts. In EMNLP-2004 (Conference on Empirical Methods in Natural Language Processing), pages 309-316, 2004. [6] S. Corston-Oliver, E. Ringger, M. Gamon, and R. Campbell. Task-focused summarization of email. In Text Summarization Branches Out: Proceedings of the ACL-04 Workshop, pages 43-50, 2004. [7] A. Culotta, R. Bekkerman, and A. McCallum. Extracting social networks and contact information from email and the web. In CEAS-2004 (Conference on Email and Anti-Spam), Mountain View, CA, July 2004. [8] L. Devroye, L. Gy¨orfi, and G. Lugosi. A Probabilistic Theory of Pattern Recognition. Springer-Verlag, New York, NY, 1996. [9] S. T. Dumais, J. Platt, D. Heckerman, and M. Sahami. Inductive learning algorithms and representations for text categorization. In CIKM "98, Proceedings of the 7th ACM Conference on Information and Knowledge Management, pages 148-155, 1998. [10] Y. Freund and R. Schapire. Large margin classification using the perceptron algorithm. Machine Learning, 37(3):277-296, 1999. [11] T. Joachims. Making large-scale svm learning practical. In B. Sch¨olkopf, C. J. Burges, and A. J. Smola, editors, Advances in Kernel Methods - Support Vector Learning, pages 41-56. MIT Press, 1999. [12] L. S. Larkey. A patent search and classification system. In Proceedings of the Fourth ACM Conference on Digital Libraries, pages 179 - 187, 1999. [13] D. D. Lewis. An evaluation of phrasal and clustered representations on a text categorization task. In SIGIR "92, Proceedings of the 15th Annual International ACM Conference on Research and Development in Information Retrieval, pages 37-50, 1992. [14] Y. Liu, J. Carbonell, and R. Jin. A pairwise ensemble approach for accurate genre classification. In Proceedings of the European Conference on Machine Learning (ECML), 2003. [15] Y. Liu, R. Yan, R. Jin, and J. Carbonell. A comparison study of kernels for multi-label text classification using category association. In The Twenty-first International Conference on Machine Learning (ICML), 2004. [16] A. McCallum and K. Nigam. A comparison of event models for naive bayes text classification. In Working Notes of AAAI "98 (The 15th National Conference on Artificial Intelligence), Workshop on Learning for Text Categorization, pages 41-48, 1998. TR WS-98-05. [17] F. Sebastiani. Machine learning in automated text categorization. ACM Computing Surveys, 34(1):1-47, March 2002. [18] C. J. van Rijsbergen. Information Retrieval. Butterworths, London, 1979. [19] Y. Yang. An evaluation of statistical approaches to text categorization. Information Retrieval, 1(1/2):67-88, 1999. [20] Y. Yang, J. Carbonell, R. Brown, T. Pierce, B. T. Archibald, and X. Liu. Learning approaches to topic detection and tracking. IEEE EXPERT, Special Issue on Applications of Intelligent Information Retrieval, 1999. [21] Y. Yang and X. Liu. A re-examination of text categorization methods. In SIGIR "99, Proceedings of the 22nd Annual International ACM Conference on Research and Development in Information Retrieval, pages 42-49, 1999. [22] Y. Yang, J. Zhang, J. Carbonell, and C. Jin. Topic-conditioned novelty detection. In Proceedings of the ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, July 2002.

document detection;svm;feature selection;sentence-level classifier;n-gram;information retrieval;action-item detection;text classification;e-mail;automated model selection;chi-squared feature selection;document ranking;embedded cross-validation;sentence detection;genre-classification;topic-driven text classification;speech act;speech act identification;text categorization;e-mail priority ranking;simple factual question

train_H-98

Using Asymmetric Distributions to Improve Text Classifier Probability Estimates

Text classifiers that give probability estimates are more readily applicable in a variety of scenarios. For example, rather than choosing one set decision threshold, they can be used in a Bayesian risk model to issue a run-time decision which minimizes a userspecified cost function dynamically chosen at prediction time. However, the quality of the probability estimates is crucial. We review a variety of standard approaches to converting scores (and poor probability estimates) from text classifiers to high quality estimates and introduce new models motivated by the intuition that the empirical score distribution for the extremely irrelevant, hard to discriminate, and obviously relevant items are often significantly different. Finally, we analyze the experimental performance of these models over the outputs of two text classifiers. The analysis demonstrates that one of these models is theoretically attractive (introducing few new parameters while increasing flexibility), computationally efficient, and empirically preferable.

1. INTRODUCTION Text classifiers that give probability estimates are more flexible in practice than those that give only a simple classification or even a ranking. For example, rather than choosing one set decision threshold, they can be used in a Bayesian risk model [8] to issue a runtime decision which minimizes the expected cost of a user-specified cost function dynamically chosen at prediction time. This can be used to minimize a linear utility cost function for filtering tasks where pre-specified costs of relevant/irrelevant are not available during training but are specified at prediction time. Furthermore, the costs can be changed without retraining the model. Additionally, probability estimates are often used as the basis of deciding which document"s label to request next during active learning [17, 23]. Effective active learning can be key in many information retrieval tasks where obtaining labeled data can be costly - severely reducing the amount of labeled data needed to reach the same performance as when new labels are requested randomly [17]. Finally, they are also amenable to making other types of cost-sensitive decisions [26] and for combining decisions [3]. However, in all of these tasks, the quality of the probability estimates is crucial. Parametric models generally use assumptions that the data conform to the model to trade-off flexibility with the ability to estimate the model parameters accurately with little training data. Since many text classification tasks often have very little training data, we focus on parametric methods. However, most of the existing parametric methods that have been applied to this task have an assumption we find undesirable. While some of these methods allow the distributions of the documents relevant and irrelevant to the topic to have different variances, they typically enforce the unnecessary constraint that the documents are symmetrically distributed around their respective modes. We introduce several asymmetric parametric models that allow us to relax this assumption without significantly increasing the number of parameters and demonstrate how we can efficiently fit the models. Additionally, these models can be interpreted as assuming the scores produced by the text classifier have three basic types of empirical behavior - one corresponding to each of the extremely irrelevant, hard to discriminate, and obviously relevant items. We first review related work on improving probability estimates and score modeling in information retrieval. Then, we discuss in further detail the need for asymmetric models. After this, we describe two specific asymmetric models and, using two standard text classifiers, na¨ıve Bayes and SVMs, demonstrate how they can be efficiently used to recalibrate poor probability estimates or produce high quality probability estimates from raw scores. We then review experiments using previously proposed methods and the asymmetric methods over several text classification corpora to demonstrate the strengths and weaknesses of the various methods. Finally, we summarize our contributions and discuss future directions. 2. RELATED WORK Parametric models have been employed to obtain probability estimates in several areas of information retrieval. Lewis & Gale [17] use logistic regression to recalibrate na¨ıve Bayes though the quality of the probability estimates are not directly evaluated; it is simply performed as an intermediate step in active learning. Manmatha et. al [20] introduced models appropriate to produce probability estimates from relevance scores returned from search engines and demonstrated how the resulting probability estimates could be subsequently employed to combine the outputs of several search engines. They use a different parametric distribution for the relevant and irrelevant classes, but do not pursue two-sided asymmetric distributions for a single class as described here. They also survey the long history of modeling the relevance scores of search engines. Our work is similar in flavor to these previous attempts to model search engine scores, but we target text classifier outputs which we have found demonstrate a different type of score distribution behavior because of the role of training data. Focus on improving probability estimates has been growing lately. Zadrozny & Elkan [26] provide a corrective measure for decision trees (termed curtailment) and a non-parametric method for recalibrating na¨ıve Bayes. In more recent work [27], they investigate using a semi-parametric method that uses a monotonic piecewiseconstant fit to the data and apply the method to na¨ıve Bayes and a linear SVM. While they compared their methods to other parametric methods based on symmetry, they fail to provide significance test results. Our work provides asymmetric parametric methods which complement the non-parametric and semi-parametric methods they propose when data scarcity is an issue. In addition, their methods reduce the resolution of the scores output by the classifier (the number of distinct values output), but the methods here do not have such a weakness since they are continuous functions. There is a variety of other work that this paper extends. Platt [22] uses a logistic regression framework that models noisy class labels to produce probabilities from the raw output of an SVM. His work showed that this post-processing method not only can produce probability estimates of similar quality to SVMs directly trained to produce probabilities (regularized likelihood kernel methods), but it also tends to produce sparser kernels (which generalize better). Finally, Bennett [1] obtained moderate gains by applying Platt"s method to the recalibration of na¨ıve Bayes but found there were more problematic areas than when it was applied to SVMs. Recalibrating poorly calibrated classifiers is not a new problem. Lindley et. al [19] first proposed the idea of recalibrating classifiers, and DeGroot & Fienberg [5, 6] gave the now accepted standard formalization for the problem of assessing calibration initiated by others [4, 24]. 3. PROBLEM DEFINITION & APPROACH Our work differs from earlier approaches primarily in three points: (1) We provide asymmetric parametric models suitable for use when little training data is available; (2) We explicitly analyze the quality of probability estimates these and competing methods produce and provide significance tests for these results; (3) We target text classifier outputs where a majority of the previous literature targeted the output of search engines. 3.1 Problem Definition The general problem we are concerned with is highlighted in Figure 1. A text classifier produces a prediction about a document and gives a score s(d) indicating the strength of its decision that the document belongs to the positive class (relevant to the topic). We assume throughout there are only two classes: the positive and the negative (or irrelevant) class ("+" and "-" respectively). There are two general types of parametric approaches. The first of these tries to fit the posterior function directly, i.e. there is one p(s|+) p(s|−) Bayes" RuleP(+) P(−) Classifier P(+| s(d)) Predict class, c(d)={+,−} confidence s(d) that c(d)=+ Document, d and give unnormalized Figure 1: We are concerned with how to perform the box highlighted in grey. The internals are for one type of approach. function estimator that performs a direct mapping of the score s to the probability P(+|s(d)). The second type of approach breaks the problem down as shown in the grey box of Figure 1. An estimator for each of the class-conditional densities (i.e. p(s|+) and p(s|−)) is produced, then Bayes" rule and the class priors are used to obtain the estimate for P(+|s(d)). 3.2 Motivation for Asymmetric Distributions Most of the previous parametric approaches to this problem either directly or indirectly (when fitting only the posterior) correspond to fitting Gaussians to the class-conditional densities; they differ only in the criterion used to estimate the parameters. We can visualize this as depicted in Figure 2. Since increasing s usually indicates increased likelihood of belonging to the positive class, then the rightmost distribution usually corresponds to p(s|+). A B C 0 0.2 0.4 0.6 0.8 1 −10 −5 0 5 10 p(s|Class={+,−}) Unnormalized Confidence Score s p(s | Class = +) p(s | Class = −) Figure 2: Typical View of Discrimination based on Gaussians However, using standard Gaussians fails to capitalize on a basic characteristic commonly seen. Namely, if we have a raw output score that can be used for discrimination, then the empirical behavior between the modes (label B in Figure 2) is often very different than that outside of the modes (labels A and C in Figure 2). Intuitively, the area between the modes corresponds to the hard examples, which are difficult for this classifier to distinguish, while the areas outside the modes are the extreme examples that are usually easily distinguished. This suggests that we may want to uncouple the scale of the outside and inside segments of the distribution (as depicted by the curve denoted as A-Gaussian in Figure 3). As a result, an asymmetric distribution may be a more appropriate choice for application to the raw output score of a classifier. Ideally (i.e. perfect classification) there will exist scores θ− and θ+ such that all examples with score greater than θ+ are relevant and all examples with scores less than θ− are irrelevant. Furthermore, no examples fall between θ− and θ+. The distance | θ− − θ+ | corresponds to the margin in some classifiers, and an attempt is often made to maximize this quantity. Because text classifiers have training data to use to separate the classes, the final behavior of the score distributions is primarily a factor of the amount of training data and the consequent separation in the classes achieved. This is in contrast to search engine retrieval where the distribution of scores is more a factor of language distribution across documents, the similarity function, and the length and type of query. Perfect classification corresponds to using two very asymmetric distributions, but in this case, the probabilities are actually one and zero and many methods will work for typical purposes. Practically, some examples will fall between θ− and θ+, and it is often important to estimate the probabilities of these examples well (since they correspond to the hard examples). Justifications can be given for both why you may find more and less examples between θ− and θ+ than outside of them, but there are few empirical reasons to believe that the distributions should be symmetric. A natural first candidate for an asymmetric distribution is to generalize a common symmetric distribution, e.g. the Laplace or the Gaussian. An asymmetric Laplace distribution can be achieved by placing two exponentials around the mode in the following manner: p(x | θ, β, γ) =    βγ β+γ exp [−β (θ − x)] x ≤ θ (β, γ > 0) βγ β+γ exp [−γ (x − θ)] x > θ (1) where θ, β, and γ are the model parameters. θ is the mode of the distribution, β is the inverse scale of the exponential to the left of the mode, and γ is the inverse scale of the exponential to the right. We will use the notation Λ(X | θ, β, γ) to refer to this distribution. 0 0.002 0.004 0.006 0.008 0.01 -300 -200 -100 0 100 200 p(s|Class={+,-}) Unnormalized Confidence Score s Gaussian A-Gaussian Figure 3: Gaussians vs. Asymmetric Gaussians. A Shortcoming of Symmetric Distributions - The vertical lines show the modes as estimated nonparametrically. We can create an asymmetric Gaussian in the same manner: p(x | θ, σl, σr) =    2√ 2π(σl+σr) exp −(x−θ)2 2σ2 l x ≤ θ (σl, σr > 0) 2√ 2π(σl+σr) exp −(x−θ)2 2σ2 r x > θ (2) where θ, σl, and σr are the model parameters. To refer to this asymmetric Gaussian, we use the notation Γ(X | θ, σl, σr). While these distributions are composed of halves, the resulting function is a single continuous distribution. These distributions allow us to fit our data with much greater flexibility at the cost of only fitting six parameters. We could instead try mixture models for each component or other extensions, but most other extensions require at least as many parameters (and can often be more computationally expensive). In addition, the motivation above should provide significant cause to believe the underlying distributions actually behave in this way. Furthermore, this family of distributions can still fit a symmetric distribution, and finally, in the empirical evaluation, evidence is presented that demonstrates this asymmetric behavior (see Figure 4). To our knowledge, neither family of distributions has been previously used in machine learning or information retrieval. Both are termed generalizations of an Asymmetric Laplace in [14], but we refer to them as described above to reflect the nature of how we derived them for this task. 3.3 Estimating the Parameters of the Asymmetric Distributions This section develops the method for finding maximum likelihood estimates (MLE) of the parameters for the above asymmetric distributions. In order to find the MLEs, we have two choices: (1) use numerical estimation to estimate all three parameters at once (2) fix the value of θ, and estimate the other two (β and γ or σl and σr) given our choice of θ, then consider alternate values of θ. Because of the simplicity of analysis in the latter alternative, we choose this method. 3.3.1 Asymmetric Laplace MLEs For D = {x1, x2, . . . , xN } where the xi are i.i.d. and X ∼ Λ(X | θ, β, γ), the likelihood is N i Λ(X | θ, β, γ). Now, we fix θ and compute the maximum likelihood for that choice of θ. Then, we can simply consider all choices of θ and choose the one with the maximum likelihood over all choices of θ. The complete derivation is omitted because of space but is available in [2]. We define the following values: Nl = | {x ∈ D | x ≤ θ} | Nr = | {x ∈ D | x > θ} | Sl = x∈D|x≤θ x Sr = x∈D|x>θ x Dl = Nlθ − Sl Dr = Sr − Nrθ. Note that Dl and Dr are the sum of the absolute differences between the x belonging to the left and right halves of the distribution (respectively) and θ. Finally the MLEs for β and γ for a fixed θ are: βMLE = N Dl + √ DrDl γMLE = N Dr + √ DrDl . (3) These estimates are not wholly unexpected since we would obtain Nl Dl if we were to estimate β independently of γ. The elegance of the formulae is that the estimates will tend to be symmetric only insofar as the data dictate it (i.e. the closer Dl and Dr are to being equal, the closer the resulting inverse scales). By continuity arguments, when N = 0, we assign β = γ = 0 where 0 is a small constant that acts to disperse the distribution to a uniform. Similarly, when N = 0 and Dl = 0, we assign β = inf where inf is a very large constant that corresponds to an extremely sharp distribution (i.e. almost all mass at θ for that half). Dr = 0 is handled similarly. Assuming that θ falls in some range [φ, ψ] dependent upon only the observed documents, then this alternative is also easily computable. Given Nl, Sl, Nr, Sr, we can compute the posterior and the MLEs in constant time. In addition, if the scores are sorted, then we can perform the whole process quite efficiently. Starting with the minimum θ = φ we would like to try, we loop through the scores once and set Nl, Sl, Nr, Sr appropriately. Then we increase θ and just step past the scores that have shifted from the right side of the distribution to the left. Assuming the number of candidate θs are O(n), this process is O(n), and the overall process is dominated by sorting the scores, O(n log n) (or expected linear time). 3.3.2 Asymmetric Gaussian MLEs For D = {x1, x2, . . . , xN } where the xi are i.i.d. and X ∼ Γ(X | θ, σl, σr), the likelihood is N i Γ(X | θ, β, γ). The MLEs can be worked out similar to the above. We assume the same definitions as above (the complete derivation omitted for space is available in [2]), and in addition, let: Sl2 = x∈D|x≤θ x2 Sr2 = x∈D|x>θ x2 Dl2 = Sl2 − Slθ + θ2 Nl Dr2 = Sr2 − Srθ + θ2 Nr. The analytical solution for the MLEs for a fixed θ is: σl,MLE = Dl2 + D 2/3 l2 D 1/3 r2 N (4) σr,MLE = Dr2 + D 2/3 r2 D 1/3 l2 N . (5) By continuity arguments, when N = 0, we assign σr = σl = inf , and when N = 0 and Dl2 = 0 (resp. Dr2 = 0), we assign σl = 0 (resp. σr = 0). Again, the same computational complexity analysis applies to estimating these parameters. 4. EXPERIMENTAL ANALYSIS 4.1 Methods For each of the methods that use a class prior, we use a smoothed add-one estimate, i.e. P(c) = |c|+1 N+2 where N is the number of documents. For methods that fit the class-conditional densities, p(s|+) and p(s|−), the resulting densities are inverted using Bayes" rule as described above. All of the methods below are fit using maximum likelihood estimates. For recalibrating a classifier (i.e. correcting poor probability estimates output by the classifier), it is usual to use the log-odds of the classifier"s estimate as s(d). The log-odds are defined to be log P (+|d) P (−|d) . The normal decision threshold (minimizing error) in terms of log-odds is at zero (i.e. P(+|d) = P(−|d) = 0.5). Since it scales the outputs to a space [−∞, ∞], the log-odds make normal (and similar distributions) applicable [19]. Lewis & Gale [17] give a more motivating viewpoint that fitting the log-odds is a dampening effect for the inaccurate independence assumption and a bias correction for inaccurate estimates of the priors. In general, fitting the log-odds can serve to boost or dampen the signal from the original classifier as the data dictate. Gaussians A Gaussian is fit to each of the class-conditional densities, using the usual maximum likelihood estimates. This method is denoted in the tables below as Gauss. Asymmetric Gaussians An asymmetric Gaussian is fit to each of the class-conditional densities using the maximum likelihood estimation procedure described above. Intervals between adjacent scores are divided by 10 in testing candidate θs, i.e. 8 points between actual scores occurring in the data set are tested. This method is denoted as A. Gauss. Laplace Distributions Even though Laplace distributions are not typically applied to this task, we also tried this method to isolate why benefit is gained from the asymmetric form. The usual MLEs were used for estimating the location and scale of a classical symmetric Laplace distribution as described in [14]. We denote this method as Laplace below. Asymmetric Laplace Distributions An asymmetric Laplace is fit to each of the class-conditional densities using the maximum likelihood estimation procedure described above. As with the asymmetric Gaussian, intervals between adjacent scores are divided by 10 in testing candidate θs. This method is denoted as A. Laplace below. Logistic Regression This method is the first of two methods we evaluated that directly fit the posterior, P(+|s(d)). Both methods restrict the set of families to a two-parameter sigmoid family; they differ primarily in their model of class labels. As opposed to the above methods, one can argue that an additional boon of these methods is they completely preserve the ranking given by the classifier. When this is desired, these methods may be more appropriate. The previous methods will mostly preserve the rankings, but they can deviate if the data dictate it. Thus, they may model the data behavior better at the cost of departing from a monotonicity constraint in the output of the classifier. Lewis & Gale [17] use logistic regression to recalibrate na¨ıve Bayes for subsequent use in active learning. The model they use is: P(+|s(d)) = exp(a + b s(d)) 1 + exp(a + b s(d)) . (6) Instead of using the probabilities directly output by the classifier, they use the loglikelihood ratio of the probabilities, log P (d|+) P (d|−) , as the score s(d). Instead of using this below, we will use the logodds ratio. This does not affect the model as it simply shifts all of the scores by a constant determined by the priors. We refer to this method as LogReg below. Logistic Regression with Noisy Class Labels Platt [22] proposes a framework that extends the logistic regression model above to incorporate noisy class labels and uses it to produce probability estimates from the raw output of an SVM. This model differs from the LogReg model only in how the parameters are estimated. The parameters are still fit using maximum likelihood estimation, but a model of noisy class labels is used in addition to allow for the possibility that the class was mislabeled. The noise is modeled by assuming there is a finite probability of mislabeling a positive example and of mislabeling a negative example; these two noise estimates are determined by the number of positive examples and the number of negative examples (using Bayes" rule to infer the probability of incorrect label). Even though the performance of this model would not be expected to deviate much from LogReg, we evaluate it for completeness. We refer to this method below as LR+Noise. 4.2 Data We examined several corpora, including the MSN Web Directory, Reuters, and TREC-AP. MSN Web Directory The MSN Web Directory is a large collection of heterogeneous web pages (from a May 1999 web snapshot) that have been hierarchically classified. We used the same train/test split of 50078/10024 documents as that reported in [9]. The MSN Web hierarchy is a seven-level hierarchy; we used all 13 of the top-level categories. The class proportions in the training set vary from 1.15% to 22.29%. In the testing set, they range from 1.14% to 21.54%. The classes are general subjects such as Health & Fitness and Travel & Vacation. Human indexers assigned the documents to zero or more categories. For the experiments below, we used only the top 1000 words with highest mutual information for each class; approximately 195K words appear in at least three training documents. Reuters The Reuters 21578 corpus [16] contains Reuters news articles from 1987. For this data set, we used the ModApte standard train/ test split of 9603/3299 documents (8676 unused documents). The classes are economic subjects (e.g., acq for acquisitions, earn for earnings, etc.) that human taggers applied to the document; a document may have multiple subjects. There are actually 135 classes in this domain (only 90 of which occur in the training and testing set); however, we only examined the ten most frequent classes since small numbers of testing examples make interpreting some performance measures difficult due to high variance.1 Limiting to the ten largest classes allows us to compare our results to previously published results [10, 13, 21, 22]. The class proportions in the training set vary from 1.88% to 29.96%. In the testing set, they range from 1.7% to 32.95%. For the experiments below we used only the top 300 words with highest mutual information for each class; approximately 15K words appear in at least three training documents. TREC-AP The TREC-AP corpus is a collection of AP news stories from 1988 to 1990. We used the same train/test split of 142791/66992 documents that was used in [18]. As described in [17] (see also [15]), the categories are defined by keywords in a keyword field. The title and body fields are used in the experiments below. There are twenty categories in total. The class proportions in the training set vary from 0.06% to 2.03%. In the testing set, they range from 0.03% to 4.32%. For the experiments described below, we use only the top 1000 words with the highest mutual information for each class; approximately 123K words appear in at least 3 training documents. 4.3 Classifiers We selected two classifiers for evaluation. A linear SVM classifier which is a discriminative classifier that does not normally output probability values, and a na¨ıve Bayes classifier whose probability outputs are often poor [1, 7] but can be improved [1, 26, 27]. 1 A separate comparison of only LogReg, LR+Noise, and A. Laplace over all 90 categories of Reuters was also conducted. After accounting for the variance, that evaluation also supported the claims made here. SVM For linear SVMs, we use the Smox toolkit which is based on Platt"s Sequential Minimal Optimization algorithm. The features were represented as continuous values. We used the raw output score of the SVM as s(d) since it has been shown to be appropriate before [22]. The normal decision threshold (assuming we are seeking to minimize errors) for this classifier is at zero. Na¨ıve Bayes The na¨ıve Bayes classifier model is a multinomial model [21]. We smoothed word and class probabilities using a Bayesian estimate (with the word prior) and a Laplace m-estimate, respectively. We use the log-odds estimated by the classifier as s(d). The normal decision threshold is at zero. 4.4 Performance Measures We use log-loss [12] and squared error [4, 6] to evaluate the quality of the probability estimates. For a document d with class c(d) ∈ {+, −} (i.e. the data have known labels and not probabilities), logloss is defined as δ(c(d), +) log P(+|d) + δ(c(d), −) log P(−|d) where δ(a, b) . = 1 if a = b and 0 otherwise. The squared error is δ(c(d), +)(1 − P(+|d))2 + δ(c(d), −)(1 − P(−|d))2 . When the class of a document is correctly predicted with a probability of one, log-loss is zero and squared error is zero. When the class of a document is incorrectly predicted with a probability of one, log-loss is −∞ and squared error is one. Thus, both measures assess how close an estimate comes to correctly predicting the item"s class but vary in how harshly incorrect predictions are penalized. We report only the sum of these measures and omit the averages for space. Their averages, average log-loss and mean squared error (MSE), can be computed from these totals by dividing by the number of binary decisions in a corpus. In addition, we also compare the error of the classifiers at their default thresholds and with the probabilities. This evaluates how the probability estimates have improved with respect to the decision threshold P(+|d) = 0.5. Thus, error only indicates how the methods would perform if a false positive was penalized the same as a false negative and not the general quality of the probability estimates. It is presented simply to provide the reader with a more complete understanding of the empirical tendencies of the methods. We use a a standard paired micro sign test [25] to determine statistical significance in the difference of all measures. Only pairs that the methods disagree on are used in the sign test. This test compares pairs of scores from two systems with the null hypothesis that the number of items they disagree on are binomially distributed. We use a significance level of p = 0.01. 4.5 Experimental Methodology As the categories under consideration in the experiments are not mutually exclusive, the classification was done by training n binary classifiers, where n is the number of classes. In order to generate the scores that each method uses to fit its probability estimates, we use five-fold cross-validation on the training data. We note that even though it is computationally efficient to perform leave-one-out cross-validation for the na¨ıve Bayes classifier, this may not be desirable since the distribution of scores can be skewed as a result. Of course, as with any application of n-fold cross-validation, it is also possible to bias the results by holding n too low and underestimating the performance of the final classifier. 4.6 Results & Discussion The results for recalibrating na¨ıve Bayes are given in Table 1a. Table 1b gives results for producing probabilistic outputs for SVMs. Log-loss Error2 Errors MSN Web Gauss -60656.41 10503.30 10754 A.Gauss -57262.26 8727.47 9675 Laplace -45363.84 8617.59 10927 A.Laplace -36765.88 6407.84† 8350 LogReg -36470.99 6525.47 8540 LR+Noise -36468.18 6534.61 8563 na¨ıve Bayes -1098900.83 17117.50 17834 Reuters Gauss -5523.14 1124.17 1654 A.Gauss -4929.12 652.67 888 Laplace -5677.68 1157.33 1416 A.Laplace -3106.95‡ 554.37‡ 726 LogReg -3375.63 603.20 786 LR+Noise -3374.15 604.80 785 na¨ıve Bayes -52184.52 1969.41 2121 TREC-AP Gauss -57872.57 8431.89 9705 A.Gauss -66009.43 7826.99 8865 Laplace -61548.42 9571.29 11442 A.Laplace -48711.55 7251.87‡ 8642 LogReg -48250.81 7540.60 8797 LR+Noise -48251.51 7544.84 8801 na¨ıve Bayes -1903487.10 41770.21 43661 Log-loss Error2 Errors MSN Web Gauss -54463.32 9090.57 10555 A. Gauss -44363.70 6907.79 8375 Laplace -42429.25 7669.75 10201 A. Laplace -31133.83 5003.32 6170 LogReg -30209.36 5158.74 6480 LR+Noise -30294.01 5209.80 6551 Linear SVM N/A N/A 6602 Reuters Gauss -3955.33 589.25 735 A. Gauss -4580.46 428.21 532 Laplace -3569.36 640.19 770 A. Laplace -2599.28 412.75 505 LogReg -2575.85 407.48 509 LR+Noise -2567.68 408.82 516 Linear SVM N/A N/A 516 TREC-AP Gauss -54620.94 6525.71 7321 A. Gauss -77729.49 6062.64 6639 Laplace -54543.19 7508.37 9033 A. Laplace -48414.39 5761.25‡ 6572‡ LogReg -48285.56 5914.04 6791 LR+Noise -48214.96 5919.25 6794 Linear SVM N/A N/A 6718 Table 1: (a) Results for na¨ıve Bayes (left) and (b) SVM (right). The best entry for a corpus is in bold. Entries that are statistically significantly better than all other entries are underlined. A † denotes the method is significantly better than all other methods except for na¨ıve Bayes. A ‡ denotes the entry is significantly better than all other methods except for A. Gauss (and na¨ıve Bayes for the table on the left). The reason for this distinction in significance tests is described in the text. We start with general observations that result from examining the performance of these methods over the various corpora. The first is that A. Laplace, LR+Noise, and LogReg, quite clearly outperform the other methods. There is usually little difference between the performance of LR+Noise and LogReg (both as shown here and on a decision by decision basis), but this is unsurprising since LR+Noise just adds noisy class labels to the LogReg model. With respect to the three different measures, LR+Noise and LogReg tend to perform slightly better (but never significantly) than A. Laplace at some tasks with respect to log-loss and squared error. However, A. Laplace always produces the least number of errors for all of the tasks, though at times the degree of improvement is not significant. In order to give the reader a better sense of the behavior of these methods, Figures 4-5 show the fits produced by the most competitive of these methods versus the actual data behavior (as estimated nonparametrically by binning) for class Earn in Reuters. Figure 4 shows the class-conditional densities, and thus only A. Laplace is shown since LogReg fits the posterior directly. Figure 5 shows the estimations of the log-odds, (i.e. log P (Earn|s(d)) P (¬Earn|s(d)) ). Viewing the log-odds (rather than the posterior) usually enables errors in estimation to be detected by the eye more easily. We can break things down as the sign test does and just look at wins and losses on the items that the methods disagree on. Looked at in this way only two methods (na¨ıve Bayes and A. Gauss) ever have more pairwise wins than A. Laplace; those two sometimes have more pairwise wins on log-loss and squared error even though the total never wins (i.e. they are dragged down by heavy penalties). In addition, this comparison of pairwise wins means that for those cases where LogReg and LR+Noise have better scores than A. Laplace, it would not be deemed significant by the sign test at any level since they do not have more wins. For example, of the 130K binary decisions over the MSN Web dataset, A. Laplace had approximately 101K pairwise wins versus LogReg and LR+Noise. No method ever has more pairwise wins than A. Laplace for the error comparison nor does any method every achieve a better total. The basic observation made about na¨ıve Bayes in previous work is that it tends to produce estimates very close to zero and one [1, 17]. This means if it tends to be right enough of the time, it will produce results that do not appear significant in a sign test that ignores size of difference (as the one here). The totals of the squared error and log-loss bear out the previous observation that when it"s wrong it"s really wrong. There are several interesting points about the performance of the asymmetric distributions as well. First, A. Gauss performs poorly because (similar to na¨ıve Bayes) there are some examples where it is penalized a large amount. This behavior results from a general tendency to perform like the picture shown in Figure 3 (note the crossover at the tails). While the asymmetric Gaussian tends to place the mode much more accurately than a symmetric Gaussian, its asymmetric flexibility combined with its distance function causes it to distribute too much mass to the outside tails while failing to fit around the mode accurately enough to compensate. Figure 3 is actually a result of fitting the two distributions to real data. As a result, at the tails there can be a large discrepancy between the likelihood of belonging to each class. Thus when there are no outliers A. Gauss can perform quite competitively, but when there is an 0 0.002 0.004 0.006 0.008 0.01 0.012 -600 -400 -200 0 200 400 p(s(d)|Class={+,-}) s(d) = naive Bayes log-odds Train Test A.Laplace 0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 -15 -10 -5 0 5 10 15 p(s(d)|Class={+,-}) s(d) = linear SVM raw score Train Test A.Laplace Figure 4: The empirical distribution of classifier scores for documents in the training and the test set for class Earn in Reuters. Also shown is the fit of the asymmetric Laplace distribution to the training score distribution. The positive class (i.e. Earn) is the distribution on the right in each graph, and the negative class (i.e. ¬Earn) is that on the left in each graph. -6 -4 -2 0 2 4 6 8 -250 -200 -150 -100 -50 0 50 100 150 LogOdds=logP(+|s(d))-logP(-|s(d)) s(d) = naive Bayes log-odds Train Test A.Laplace LogReg -5 0 5 10 15 -4 -2 0 2 4 6 LogOdds=logP(+|s(d))-logP(-|s(d)) s(d) = linear SVM raw score Train Test A.Laplace LogReg Figure 5: The fit produced by various methods compared to the empirical log-odds of the training data for class Earn in Reuters. outlier A. Gauss is penalized quite heavily. There are enough such cases overall that it seems clearly inferior to the top three methods. However, the asymmetric Laplace places much more emphasis around the mode (Figure 4) because of the different distance function (think of the sharp peak of an exponential). As a result most of the mass stays centered around the mode, while the asymmetric parameters still allow more flexibility than the standard Laplace. Since the standard Laplace also corresponds to a piecewise fit in the log-odds space, this highlights that part of the power of the asymmetric methods is their sensitivity in placing the knots at the actual modes - rather than the symmetric assumption that the means correspond to the modes. Additionally, the asymmetric methods have greater flexibility in fitting the slopes of the line segments as well. Even in cases where the test distribution differs from the training distribution (Figure 4), A. Laplace still yields a solution that gives a better fit than LogReg (Figure 5), the next best competitor. Finally, we can make a few observations about the usefulness of the various performance metrics. First, log-loss only awards a finite amount of credit as the degree to which something is correct improves (i.e. there are diminishing returns as it approaches zero), but it can infinitely penalize for a wrong estimate. Thus, it is possible for one outlier to skew the totals, but misclassifying this example may not matter for any but a handful of actual utility functions used in practice. Secondly, squared error has a weakness in the other direction. That is, its penalty and reward are bounded in [0, 1], but if the number of errors is small enough, it is possible for a method to appear better when it is producing what we generally consider unhelpful probability estimates. For example, consider a method that only estimates probabilities as zero or one (which na¨ıve Bayes tends to but doesn"t quite reach if you use smoothing). This method could win according to squared error, but with just one error it would never perform better on log-loss than any method that assigns some non-zero probability to each outcome. For these reasons, we recommend that neither of these are used in isolation as they each give slightly different insights to the quality of the estimates produced. These observations are straightforward from the definitions but are underscored by the evaluation. 5. FUTURE WORK A promising extension to the work presented here is a hybrid distribution of a Gaussian (on the outside slopes) and exponentials (on the inner slopes). From the empirical evidence presented in [22], the expectation is that such a distribution might allow more emphasis of the probability mass around the modes (as with the exponential) while still providing more accurate estimates toward the tails. Just as logistic regression allows the log-odds of the posterior distribution to be fit directly with a line, we could directly fit the log-odds of the posterior with a three-piece line (a spline) instead of indirectly doing the same thing by fitting the asymmetric Laplace. This approach may provide more power since it retains the asymmetry assumption but not the assumption that the class-conditional densities are from an asymmetric Laplace. Finally, extending these methods to the outputs of other discriminative classifiers is an open area. We are currently evaluating the appropriateness of these methods for the output of a voted perceptron [11]. By analogy to the log-odds, the operative score that appears promising is log weight perceptrons voting + weight perceptrons voting − . 6. SUMMARY AND CONCLUSIONS We have reviewed a wide variety of parametric methods for producing probability estimates from the raw scores of a discriminative classifier and for recalibrating an uncalibrated probabilistic classifier. In addition, we have introduced two new families that attempt to capitalize on the asymmetric behavior that tends to arise from learning a discrimination function. We have given an efficient way to estimate the parameters of these distributions. While these distributions attempt to strike a balance between the generalization power of parametric distributions and the flexibility that the added asymmetric parameters give, the asymmetric Gaussian appears to have too great of an emphasis away from the modes. In striking contrast, the asymmetric Laplace distribution appears to be preferable over several large text domains and a variety of performance measures to the primary competing parametric methods, though comparable performance is sometimes achieved with one of two varieties of logistic regression. Given the ease of estimating the parameters of this distribution, it is a good first choice for producing quality probability estimates. Acknowledgments We are grateful to Francisco Pereira for the sign test code, Anton Likhodedov for logistic regression code, and John Platt for the code support for the linear SVM classifier toolkit Smox. Also, we sincerely thank Chris Meek and John Platt for the very useful advice provided in the early stages of this work. Thanks also to Jaime Carbonell and John Lafferty for their useful feedback on the final versions of this paper. 7. REFERENCES [1] P. N. Bennett. Assessing the calibration of naive bayes" posterior estimates. Technical Report CMU-CS-00-155, Carnegie Mellon, School of Computer Science, 2000. [2] P. N. Bennett. Using asymmetric distributions to improve classifier probabilities: A comparison of new and standard parametric methods. Technical Report CMU-CS-02-126, Carnegie Mellon, School of Computer Science, 2002. [3] H. Bourlard and N. Morgan. A continuous speech recognition system embedding mlp into hmm. In NIPS "89, 1989. [4] G. Brier. Verification of forecasts expressed in terms of probability. Monthly Weather Review, 78:1-3, 1950. [5] M. H. DeGroot and S. E. Fienberg. The comparison and evaluation of forecasters. Statistician, 32:12-22, 1983. [6] M. H. DeGroot and S. E. Fienberg. Comparing probability forecasters: Basic binary concepts and multivariate extensions. In P. Goel and A. Zellner, editors, Bayesian Inference and Decision Techniques. Elsevier Science Publishers B.V., 1986. [7] P. Domingos and M. Pazzani. Beyond independence: Conditions for the optimality of the simple bayesian classifier. In ICML "96, 1996. [8] R. Duda, P. Hart, and D. Stork. Pattern Classification. John Wiley & Sons, Inc., 2001. [9] S. T. Dumais and H. Chen. Hierarchical classification of web content. In SIGIR "00, 2000. [10] S. T. Dumais, J. Platt, D. Heckerman, and M. Sahami. Inductive learning algorithms and representations for text categorization. In CIKM "98, 1998. [11] Y. Freund and R. Schapire. Large margin classification using the perceptron algorithm. Machine Learning, 37(3):277-296, 1999. [12] I. Good. Rational decisions. Journal of the Royal Statistical Society, Series B, 1952. [13] T. Joachims. Text categorization with support vector machines: Learning with many relevant features. In ECML "98, 1998. [14] S. Kotz, T. J. Kozubowski, and K. Podgorski. The Laplace Distribution and Generalizations: A Revisit with Applications to Communications, Economics, Engineering, and Finance. Birkh¨auser, 2001. [15] D. D. Lewis. A sequential algorithm for training text classifiers: Corrigendum and additional data. SIGIR Forum, 29(2):13-19, Fall 1995. [16] D. D. Lewis. Reuters-21578, distribution 1.0. http://www.daviddlewis.com/resources/ testcollections/reuters21578, January 1997. [17] D. D. Lewis and W. A. Gale. A sequential algorithm for training text classifiers. In SIGIR "94, 1994. [18] D. D. Lewis, R. E. Schapire, J. P. Callan, and R. Papka. Training algorithms for linear text classifiers. In SIGIR "96, 1996. [19] D. Lindley, A. Tversky, and R. Brown. On the reconciliation of probability assessments. Journal of the Royal Statistical Society, 1979. [20] R. Manmatha, T. Rath, and F. Feng. Modeling score distributions for combining the outputs of search engines. In SIGIR "01, 2001. [21] A. McCallum and K. Nigam. A comparison of event models for naive bayes text classification. In AAAI "98, Workshop on Learning for Text Categorization, 1998. [22] J. C. Platt. Probabilistic outputs for support vector machines and comparisons to regularized likelihood methods. In A. J. Smola, P. Bartlett, B. Scholkopf, and D. Schuurmans, editors, Advances in Large Margin Classifiers. MIT Press, 1999. [23] M. Saar-Tsechansky and F. Provost. Active learning for class probability estimation and ranking. In IJCAI "01, 2001. [24] R. L. Winkler. Scoring rules and the evaluation of probability assessors. Journal of the American Statistical Association, 1969. [25] Y. Yang and X. Liu. A re-examination of text categorization methods. In SIGIR "99, 1999. [26] B. Zadrozny and C. Elkan. Obtaining calibrated probability estimates from decision trees and naive bayesian classifiers. In ICML "01, 2001. [27] B. Zadrozny and C. Elkan. Reducing multiclass to binary by coupling probability estimates. In KDD "02, 2002.

classifier combination;parametric model;maximum likelihood estimate;text classification;asymmetric laplace distribution;asymmetric gaussian;cost-sensitive learn;decision threshold;logistic regression framework;posterior function;active learn;class-conditional density;symmetric distribution;text classifier;information retrieval;bayesian risk model;probability estimate;empirical score distribution;search engine retrieval

train_I-37

A Framework for Agent-Based Distributed Machine Learning and Data Mining

This paper proposes a framework for agent-based distributed machine learning and data mining based on (i) the exchange of meta-level descriptions of individual learning processes among agents and (ii) online reasoning about learning success and learning progress by learning agents. We present an abstract architecture that enables agents to exchange models of their local learning processes and introduces a number of different methods for integrating these processes. This allows us to apply existing agent interaction mechanisms to distributed machine learning tasks, thus leveraging the powerful coordination methods available in agent-based computing, and enables agents to engage in meta-reasoning about their own learning decisions. We apply this architecture to a real-world distributed clustering application to illustrate how the conceptual framework can be used in practical systems in which different learners may be using different datasets, hypotheses and learning algorithms. We report on experimental results obtained using this system, review related work on the subject, and discuss potential future extensions to the framework.

1. INTRODUCTION In the areas of machine learning and data mining (cf. [14, 17] for overviews), it has long been recognised that parallelisation and distribution can be used to improve learning performance. Various techniques have been suggested in this respect, ranging from the low-level integration of independently derived learning hypotheses (e.g. combining different classifiers to make optimal classification decisions [4, 7], model averaging of Bayesian classifiers [8], or consensusbased methods for integrating different clusterings [11]), to the high-level combination of learning results obtained by heterogeneous learning agents using meta-learning (e.g. [3, 10, 21]). All of these approaches assume homogeneity of agent design (all agents apply the same learning algorithm) and/or agent objectives (all agents are trying to cooperatively solve a single, global learning problem). Therefore, the techniques they suggest are not applicable in societies of autonomous learners interacting in open systems. In such systems, learners (agents) may not be able to integrate their datasets or learning results (because of different data formats and representations, learning algorithms, or legal restrictions that prohibit such integration [11]) and cannot always be guaranteed to interact in a strictly cooperative fashion (discovered knowledge and collected data might be economic assets that should only be shared when this is deemed profitable; malicious agents might attempt to adversely influence others" learning results, etc.). Examples for applications of this kind abound. Many distributed learning domains involve the use of sensitive data and prohibit the exchange of this data (e.g. exchange of patient data in distributed brain tumour diagnosis [2]) - however, they may permit the exchange of local learning hypotheses among different learners. In other areas, training data might be commercially valuable, so that agents would only make it available to others if those agents could provide something in return (e.g. in remote ship surveillance and tracking, where the different agencies involved are commercial service providers [1]). Furthermore, agents might have a vested interest in negatively affecting other agents" learning performance. An example for this is that of fraudulent agents on eBay which may try to prevent reputationlearning agents from the construction of useful models for detecting fraud. Viewing learners as autonomous, self-directed agents is the only appropriate view one can take in modelling these distributed learning environments: the agent metaphor becomes a necessity as oppossed to preferences for scalability, dynamic data selection, interactivity [13], which can also be achieved through (non-agent) distribution and parallelisation in principle. Despite the autonomy and self-directedness of learning agents, many of these systems exhibit a sufficient overlap in terms of individual learning goals so that beneficial cooperation might be possible if a model for flexible interaction between autonomous learners was available that allowed agents to 1. exchange information about different aspects of their own learning mechanism at different levels of detail without being forced to reveal private information that should not be disclosed, 2. decide to what extent they want to share information about their own learning processes and utilise information provided by other learners, and 3. reason about how this information can best be used to improve their own learning performance. Our model is based on the simple idea that autonomous learners should maintain meta-descriptions of their own learning processes (see also [3]) in order to be able to exchange information and reason about them in a rational way (i.e. with the overall objective of improving their own learning results). Our hypothesis is a very simple one: If we can devise a sufficiently general, abstract view of describing learning processes, we will be able to utilise the whole range of methods for (i) rational reasoning and (ii) communication and coordination offered by agent technology so as to build effective autonomous learning agents. To test this hypothesis, we introduce such an abstract architecture (section 2) and implement a simple, concrete instance of it in a real-world domain (section 3). We report on empirical results obtained with this implemented system that demonstrate the viability of our approach (section 4). Finally, we review related work (section 5) and conclude with a summary, discussion of our approach and outlook to future work on the subject (section 6). 2. ABSTRACT ARCHITECTURE Our framework is based on providing formal (meta-level) descriptions of learning processes, i.e. representations of all relevant components of the learning machinery used by a learning agent, together with information about the state of the learning process. To ensure that this framework is sufficiently general, we consider the following general description of a learning problem: Given data D ⊆ D taken from an instance space D, a hypothesis space H and an (unknown) target function c ∈ H1 , derive a function h ∈ H that approximates c as well as possible according to some performance measure g : H → Q where Q is a set of possible levels of learning performance. 1 By requiring this we are ensuring that the learning problem can be solved in principle using the given hypothesis space. This very broad definition includes a number of components of a learning problem for which more concrete specifications can be provided if we want to be more precise. For the cases of classification and clustering, for example, we can further specify the above as follows: Learning data can be described in both cases as D = ×n i=1[Ai] where [Ai] is the domain of the ith attribute and the set of attributes is A = {1, . . . , n}. For the hypothesis space we obtain H ⊆ {h|h : D → {0, 1}} in the case of classification (i.e. a subset of the set of all possible classifiers, the nature of which depends on the expressivity of the learning algorithm used) and H ⊆ {h|h : D → N, h is total with range {1, . . . , k}} in the case of clustering (i.e. a subset of all sets of possible cluster assignments that map data points to a finite number of clusters numbered 1 to k). For classification, g might be defined in terms of the numbers of false negatives and false positives with respect to some validation set V ⊆ D, and clustering might use various measures of cluster validity to evaluate the quality of a current hypothesis, so that Q = R in both cases (but other sets of learning quality levels can be imagined). Next, we introduce a notion of learning step, which imposes a uniform basic structure on all learning processes that are supposed to exchange information using our framework. For this, we assume that each learner is presented with a finite set of data D = d1, . . . dk in each step (this is an ordered set to express that the order in which the samples are used for training matters) and employs a training/update function f : H × D∗ → H which updates h given a series of samples d1, . . . , dk. In other words, one learning step always consists of applying the update function to all samples in D exactly once. We define a learning step as a tuple l = D, H, f, g, h where we require that H ⊆ H and h ∈ H. The intuition behind this definition is that each learning step completely describes one learning iteration as shown in Figure 1: in step t, the learner updates the current hypothesis ht−1 with data Dt, and evaluates the resulting new hypothesis ht according to the current performance measure gt. Such a learning step is equivalent to the following steps of computation: 1. train the algorithm on all samples in D (once), i.e. calculate ft(ht−1, Dt) = ht, 2. calculate the quality gt of the resulting hypothesis gt(ht). We denote the set of all possible learning steps by L. For ease of notation, we denote the components of any l ∈ L by D(l), H(l), f(l) and g(l), respectively. The reason why such learning step specifications use a subset H of H instead of H itself is that learners often have explicit knowledge about which hypotheses are effectively ruled out by f given h in the future (if this is not the case, we can still set H = H). A learning process is a finite, non-empty sequence l = l1 → l2 → . . . → ln of learning steps such that ∀1 ≤ i < n .h(li+1) = f(li)(h(li), D(li)) The Sixth Intl. Joint Conf. on Autonomous Agents and Multi-Agent Systems (AAMAS 07) 679 training function ht performance measure solution quality qtgtft training set Dt hypothesis hypothesis ht−1 Figure 1: A generic model of a learning step i.e. the only requirement the transition relation →⊆ L × L makes is that the new hypothesis is the result of training the old hypothesis on all available sample data that belongs to the current step. We denote the set of all possible learning processes by L (ignoring, for ease of notation, the fact that this set depends on H, D and the spaces of possible training and evaluation functions f and g). The performance trace associated with a learning process l is the sequence q1, . . . , qn ∈ Qn where qi = g(li)(h(li)), i.e. the sequence of quality values calculated by the performance measures of the individual learning steps on the respective hypotheses. Such specifications allow agents to provide a selfdescription of their learning process. However, in communication among learning agents, it is often useful to provide only partial information about one"s internal learning process rather than its full details, e.g. when advertising this information in order to enter information exchange negotiations with others. For this purpose, we will assume that learners describe their internal state in terms of sets of learning processes (in the sense of disjunctive choice) which we call learning process descriptions (LPDs) rather than by giving precise descriptions about a single, concrete learning process. This allows us to describe properties of a learning process without specifying its details exhaustively. As an example, the set {l ∈ L|∀l = l[i].D(l) ≤ 100} describes all processes that have a training set of at most 100 samples (where all the other elements are arbitrary). Likewise, {l ∈ L|∀l = l[i].D(l) = {d}} is equivalent to just providing information about a single sample {d} and no other details about the process (this can be useful to model, for example, data received from the environment). Therefore, we use ℘(L), that is the set of all LPDs, as the basis for designing content languages for communication in the protocols we specify below. In practice, the actual content language chosen will of course be more restricted and allow only for a special type of subsets of L to be specified in a compact way, and its choice will be crucial for the interactions that can occur between learning agents. For our examples below, we simply assume explicit enumeration of all possible elements of the respective sets and function spaces (D, H, etc.) extended by the use of wildcard symbols ∗ (so that our second example above would become ({d}, ∗, ∗, ∗, ∗)). 2.1 Learning agents In our framework, a learning agent is essentially a metareasoning function that operates on information about learning processes and is situated in an environment co-inhabited by other learning agents. This means that it is not only capable of meta-level control on how to learn, but in doing so it can take information into account that is provided by other agents or the environment. Although purely cooperative or hybrid cases are possible, for the purposes of this paper we will assume that agents are purely self-interested, and that while there may be a potential for cooperation considering how agents can mutually improve each others" learning performance, there is no global mechanism that can enforce such cooperative behaviour.2 Formally speaking, an agent"s learning function is a function which, given a set of histories of previous learning processes (of oneself and potentially of learning processes about which other agents have provided information) and outputs a learning step which is its next learning action. In the most general sense, our learning agent"s internal learning process update can hence be viewed as a function λ : ℘(L) → L × ℘(L) which takes a set of learning histories of oneself and others as inputs and computes a new learning step to be executed while updating the set of known learning process histories (e.g. by appending the new learning action to one"s own learning process and leaving all information about others" learning processes untouched). Note that in λ({l1, . . . ln}) = (l, {l1, . . . ln }) some elements li of the input learning process set may be descriptions of new learning data received from the environment. The λ-function can essentially be freely chosen by the agent as long as one requirement is met, namely that the learning data that is being used always stems from what has been previously observed. More formally, ∀{l1, . . . ln} ∈ ℘(L).λ({l1, . . . ln}) = (l, {l1, . . . ln }) ⇒ „ D(l) ∪ [ l =li[j] D(l ) « ⊆ [ l =li[j] D(l ) i.e. whatever λ outputs as a new learning step and updated set of learning histories, it cannot invent new data; it has to work with the samples that have been made available to it earlier in the process through the environment or from other agents (and it can of course re-train on previously used data). The goal of the agent is to output an optimal learning step in each iteration given the information that it has. One possibility of specifying this is to require that ∀{l1, . . . ln} ∈ ℘(L).λ({l1, . . . ln}) = (l, {l1, . . . ln }) ⇒ l = arg max l ∈L g(l )(h(l )) but since it will usually be unrealistic to compute the optimal next learning step in every situation, it is more useful 2 Note that our outlook is not only different from common, cooperative models of distributed machine learning and data mining, but also delineates our approach from multiagent learning systems in which agents learn about other agents [25], i.e. the learning goal itself is not affected by agents" behaviour in the environment. 680 The Sixth Intl. Joint Conf. on Autonomous Agents and Multi-Agent Systems (AAMAS 07) i j Dj Hj fj gj hj Di pD→D 1 (Di, Dj) . .. pD→D kD→D (Di, Dj) . . . . . . n/a . . . Hi . .. ... n/a fi . .. ... n/a gi . .. n/a pg→h 1 (gi, hj) . .. pg→h kg→h (gi, hj) hi . .. n/a ... Table 1: Matrix of integration functions for messages sent from learner i to j to simply use g(l )(h(l )) as a running performance measure to evaluate how well the agent is performing. This is too abstract and unspecific for our purposes: While it describes what agents should do (transform the settings for the next learning step in an optimal way), it doesn"t specify how this can be achieved in practice. 2.2 Integrating learning process information To specify how an agent"s learning process can be affected by integrating information received from others, we need to flesh out the details of how the learning steps it will perform can be modified using incoming information about learning processes described by other agents (this includes the acquisition of new learning data from the environment as a special case). In the most general case, we can specify this in terms of the potential modifications to the existing information about learning histories that can be performed using new information. For ease of presentation, we will assume that agents are stationary learning processes that can only record the previously executed learning step and only exchange information about this one individual learning step (our model can be easily extended to cater for more complex settings). Let lj = Dj, Hj, fj, gj, hj be the current state of agent j when receiving a learning process description li = Di, Hi, fi, gi, hi from agent i (for the time being, we assume that this is a specific learning step and not a more vague, disjunctive description of properties of the learning step of i). Considering all possible interactions at an abstract level, we basically obtain a matrix of possibilities for modifications of j"s learning step specification as shown in Table 1. In this matrix, each entry specifies a family of integration functions pc→c 1 , . . . , pc→c kc→c where c, c ∈ {D, H, f, g, h} and which define how agent j"s component cj will be modified using the information ci provided about (the same or a different component of) i"s learning step by applying pc→c r (ci, cj) for some r ∈ {1, . . . , kc→c }. To put it more simply, the collections of p-functions an agent j uses specifies how it will modify its own learning behaviour using information obtained from i. For the diagonal of this matrix, which contains the most common ways of integrating new information in one"s own learning model, obvious ways of modifying one"s own learning process include replacing cj by ci or ignoring ci altogether. More complex/subtle forms of learning process integration include: • Modification of Dj: append Di to Dj; filter out all elements from Dj which also appear in Di; append Di to Dj discarding all elements with attributes outside ranges which affect gj, or those elements already correctly classified by hj; • Modification of Hi: use the union/intersection of Hi and Hj; alternatively, discard elements of Hj that are inconsistent with Dj in the process of intersection or union, or filter out elements that cannot be obtained using fj (unless fj is modified at the same time) • Modification of fj: modify parameters or background knowledge of fj using information about fi; assess their relevance by simulating previous learning steps on Dj using gj and discard those that do not help improve own performance • Modification of hj: combine hj with hi using (say) logical or mathematical operators; make the use of hi contingent on a pre-integration assessment of its quality using own data Dj and gj While this list does not include fully fledged, concrete integration operations for learning processes, it is indicative of the broad range of interactions between individual agents" learning processes that our framework enables. Note that the list does not include any modifications to gj. This is because we do not allow modifications to the agent"s own quality measure as this would render the model of rational (learning) action useless (if the quality measure is relative and volatile, we cannot objectively judge learning performance). Also note that some of the above examples require consulting other elements of lj than those appearing as arguments of the p-operations; we omit these for ease of notation, but emphasise that information-rich operations will involve consulting many different aspects of lj. Apart from operations along the diagonal of the matrix, more exotic integration operations are conceivable that combine information about different components. In theory we could fill most of the matrix with entries for them, but for lack of space we list only a few examples: • Modification of Dj using fi: pre-process samples in fi, e.g. to achieve intermediate representations that fj can be applied to • Modification of Dj using hi: filter out samples from Dj that are covered by hi and build hj using fj only on remaining samples • Modification of Hj using fi: filter out hypotheses from Hj that are not realisable using fi • Modification of hj using gi: if hj is composed of several sub-components, filter out those sub-components that do not perform well according to gi • . . . Finally, many messages received from others describing properties of their learning processes will contain information about several elements of a learning step, giving rise to yet more complex operations that depend on which kinds of information are available. The Sixth Intl. Joint Conf. on Autonomous Agents and Multi-Agent Systems (AAMAS 07) 681 Figure 2: Screenshot of our simulation system, displaying online vessel tracking data for the North Sea region 3. APPLICATION EXAMPLE 3.1 Domain description As an illustration of our framework, we present an agentbased data mining system for clustering-based surveillance using AIS (Automatic Identification System [1]) data. In our application domain, different commercial and governmental agencies track the journeys of ships over time using AIS data which contains structured information automatically provided by ships equipped with shipborne mobile AIS stations to shore stations, other ships and aircrafts. This data contains the ship"s identity, type, position, course, speed, navigational status and other safety-related information. Figure 2 shows a screenshot of our simulation system. It is the task of AIS agencies to detect anomalous behaviour so as to alarm police/coastguard units to further investigate unusual, potentially suspicious behaviour. Such behaviour might include things such as deviation from the standard routes between the declared origin and destination of the journey, unexpected close encounters between different vessels on sea, or unusual patterns in the choice of destination over multiple journeys, taking the type of vessel and reported freight into account. While the reasons for such unusual behaviour may range from pure coincidence or technical problems to criminal activity (such as smuggling, piracy, terrorist/military attacks) it is obviously useful to pre-process the huge amount of vessel (tracking) data that is available before engaging in further analysis by human experts. To support this automated pre-processing task, software used by these agencies applies clustering methods in order to identify outliers and flag those as potentially suspicious entities to the human user. However, many agencies active in this domain are competing enterprises and use their (partially overlapping, but distinct) datasets and learning hypotheses (models) as assets and hence cannot be expected to collaborate in a fully cooperative way to improve overall learning results. Considering that this is the reality of the domain in the real world, it is easy to see that a framework like the one we have suggested above might be useful to exploit the cooperation potential that is not exploited by current systems. 3.2 Agent-based distributed learning system design To describe a concrete design for the AIS domain, we need to specify the following elements of the overall system: 1. The datasets and clustering algorithms available to individual agents, 2. the interaction mechanism used for exchanging descriptions of learning processes, and 3. the decision mechanism agents apply to make learning decisions. Regarding 1., our agents are equipped with their own private datasets in the form of vessel descriptions. Learning samples are represented by tuples containing data about individual vessels in terms of attributes A = {1, . . . , n} including things such as width, length, etc. with real-valued domains ([Ai] = R for all i). In terms of learning algorithm, we consider clustering with a fixed number of k clusters using the k-means and k-medoids clustering algorithms [5] (fixed meaning that the learning algorithm will always output k clusters; however, we allow agents to change the value of k over different learning cycles). This means that the hypothesis space can be defined as H = { c1, . . . , ck |ci ∈ R|A| } i.e. the set of all possible sets of k cluster centroids in |A|-dimensional Euclidean space. For each hypothesis h = c1, . . . , ck and any data point d ∈ ×n i=1[Ai] given domain [Ai] for the ith attribute of each sample, the assignment to clusters is given by C( c1, . . . , ck , d) = arg min 1≤j≤k |d − cj| i.e. d is assigned to that cluster whose centroid is closest to the data point in terms of Euclidean distance. For evaluation purposes, each dataset pertaining to a particular agent i is initially split into a training set Di and a validation Vi. Then, we generate a set of fake vessels Fi such that |Fi| = |Vi|. These two sets assess the agent"s ability to detect suspicious vessels. For this, we assign a confidence value r(h, d) to every ship d: r(h, d) = 1 |d − cC(h,d)| where C(h, d) is the index of the nearest centroid. Based on this measure, we classify any vessel in Fi ∪ Vi as fake if its r-value is below the median of all the confidences r(h, d) for d ∈ Fi ∪ Vi. With this, we can compute the quality gi(h) ∈ R as the ratio between all correctly classified vessels and all vessels in Fi ∪ Vi. As concerns 2., we use a simple Contract-Net Protocol (CNP) [20] based hypothesis trading mechanism: Before each learning iteration, agents issue (publicly broadcasted) Calls-For-Proposals (CfPs), advertising their own numerical model quality. In other words, the initiator of a CNP describes its own current learning state as (∗, ∗, ∗, gi(h), ∗) where h is their current hypothesis/model. We assume that agents are sincere when advertising their model quality, but note that this quality might be of limited relevance to other agents as they may specialise on specific regions of the data space not related to the test set of the sender of the CfP. Subsequently, (some) agents may issue bids in which they advertise, in turn, the quality of their own model. If the 682 The Sixth Intl. Joint Conf. on Autonomous Agents and Multi-Agent Systems (AAMAS 07) bids (if any) are accepted by the initiator of the protocol who issued the CfP, the agents exchange their hypotheses and the next learning iteration ensues. To describe what is necessary for 3., we have to specify (i) under which conditions agents submit bids in response to a CfP, (ii) when they accept bids in the CNP negotiation process, and (iii) how they integrate the received information in their own learning process. Concerning (i) and (ii), we employ a very simple rule that is identical in both cases: let g be one"s own model quality and g that advertised by the CfP (or highest bid, respectively). If g > g we respond to the CfP (accept the bid), else respond to the CfP (accept the bid) with probability P(g /g) and ignore (reject) it else. If two agents make a deal, they exchange their learning hypotheses (models). In our experiments, g and g are calculated by an additional agent that acts as a global validation mechanism for all agents (in a more realistic setting a comparison mechanism for different g functions would have to be provided). As for (iii), each agent uses a single model merging operator taken from the following two classes of operators (hj is the receiver"s own model and hi is the provider"s model): • ph→h (hi, hj) : - m-join: The m best clusters (in terms of coverage of Dj) from hypothesis hi are appended to hj. - m-select: The set of the m best clusters (in terms of coverage of Dj) from the union hi ∪hj is chosen as a new model. (Unlike m-join this method does not prefer own clusters over others".) • ph→D (hi, Dj) : - m-filter: The m best clusters (as above) from hi are identified and appended to a new model formed by using those samples not covered by these clusters applying the own learning algorithm fj. Whenever m is large enough to encompass all clusters, we simply write join or filter for them. In section 4 we analyse the performance of each of these two classes for different choices of m. It is noteworthy that this agent-based distributed data mining system is one of the simplest conceivable instances of our abstract architecture. While we have previously applied it also to a more complex market-based architecture using Inductive Logic Programming learners in a transport logistics domain [22], we believe that the system described here is complex enough to illustrate the key design decisions involved in using our framework and provides simple example solutions for these design issues. 4. EXPERIMENTAL RESULTS Figure 3 shows results obtained from simulations with three learning agents in the above system using the k-means and k-medoids clustering methods respectively. We partition the total dataset of 300 ships into three disjoint sets of 100 samples each and assign each of these to one learning agent. The Single Agent is learning from the whole dataset. The parameter k is set to 10 as this is the optimal value for the total dataset according to the Davies-Bouldin index [9]. For m-select we assume m = k which achieves a constant Figure 3: Performance results obtained for different integration operations in homogeneous learner societies using the k-means (top) and k-medoids (bottom) methods model size. For m-join and m-filter we assume m = 3 to limit the extent to which models increase over time. During each experiment the learning agents receive ship descriptions in batches of 10 samples. Between these batches, there is enough time to exchange the models among the agents and recompute the models if necessary. Each ship is described using width, length, draught and speed attributes with the goal of learning to detect which vessels have provided fake descriptions of their own properties. The validation set contains 100 real and 100 randomly generated fake ships. To generate sufficiently realistic properties for fake ships, their individual attribute values are taken from randomly selected ships in the validation set (so that each fake sample is a combination of attribute values of several existing ships). In these experiments, we are mainly interested in investigating whether a simple form of knowledge sharing between self-interested learning agents could improve agent performance compared to a setting of isolated learners. Thereby, we distinguish between homogeneous learner societies where all agents use the same clustering algorithm and heterogeneous ones where different agents use different algorithms. As can be seen from the performance plots in Figure 3 (homogeneous case) and 4 (heterogeneous case, two agents use the same method and one agent uses the other) this is clearly the case for the (unrestricted) join and filter integration operations (m = k) in both cases. This is quite natural, as these operations amount to sharing all available model knowledge among agents (under appropriate constraints depending on how beneficial the exchange seems to the agents). We can see that the quality of these operations is very close to the Single Agent that has access to all training data. For the restricted (m < k) m-join, m-filter and m-select methods we can also observe an interesting distinction, The Sixth Intl. Joint Conf. on Autonomous Agents and Multi-Agent Systems (AAMAS 07) 683 Figure 4: Performance results obtained for different integration operations in heterogeneous societies with the majority of learners using the k-means (top) and k-medoids (bottom) methods namely that these perform similarly to the isolated learner case in homogeneous agent groups but better than isolated learners in more heterogeneous societies. This suggests that heterogeneous learners are able to benefit even from rather limited knowledge sharing (and this is what using a rather small m = 3 amounts to given that k = 10) while this is not always true for homogeneous agents. This nicely illustrates how different learning or data mining algorithms can specialise on different parts of the problem space and then integrate their local results to achieve better individual performance. Apart from these obvious performance benefits, integrating partial learning results can also have other advantages: The m-filter operation, for example, decreases the number of learning samples and thus can speed up the learning process. The relative number of filtered examples measured in our experiments is shown in the following table. k-means k-medoids filtering 30-40 % 10-20 % m-filtering 20-30 % 5-15 % The overall conclusion we can draw from these initial experiments with our architecture is that since a very simplistic application of its principles has proven capable of improving the performance of individual learning agents, it is worthwhile investigating more complex forms of information exchange about learning processes among autonomous learners. 5. RELATED WORK We have already mentioned work on distributed (nonagent) machine learning and data mining in the introductory chapter, so in this section we shall restrict ourselves to approaches that are more closely related to our outlook on distributed learning systems. Very often, approaches that are allegedly agent-based completely disregard agent autonomy and prescribe local decision-making procedures a priori. A typical example for this type of system is the one suggested by Caragea et al. [6] which is based on a distributed support-vector machine approach where agents incrementally join their datasets together according to a fixed distributed algorithm. A similar example is the work of Weiss [24], where groups of classifier agents learn to organise their activity so as to optimise global system behaviour. The difference between this kind of collaborative agentbased learning systems [16] and our own framework is that these approaches assume a joint learning goal that is pursued collaboratively by all agents. Many approaches rely heavily on a homogeneity assumption: Plaza and Ontanon [15] suggest methods for agentbased intelligent reuse of cases in case-based reasoning but is only applicable to societies of homogeneous learners (and coined towards a specific learning method). An agentbased method for integrating distributed cluster analysis processes using density estimation is presented by Klusch et al. [13] which is also specifically designed for a particular learning algorithm. The same is true of [22, 23] which both present market-based mechanisms for aggregating the output of multiple learning agents, even though these approaches consider more interesting interaction mechanisms among learners. A number of approaches for sharing learning data [18] have also been proposed: Grecu and Becker [12] suggest an exchange of learning samples among agents, and Ghosh et al. [11] is a step in the right direction in terms of revealing only partial information about one"s learning process as it deals with limited information sharing in distributed clustering. Papyrus [3] is a system that provides a markup language for meta-description of data, hypotheses and intermediate results and allows for an exchange of all this information among different nodes, however with a strictly cooperative goal of distributing the load for massively distributed data mining tasks. The MALE system [19] was a very early multiagent learning system in which agents used a blackboard approach to communicate their hypotheses. Agents were able to critique each others" hypotheses until agreement was reached. However, all agents in this system were identical and the system was strictly cooperative. The ANIMALS system [10] was used to simulate multistrategy learning by combining two or more learning techniques (represented by heterogeneous agents) in order to overcome weaknesses in the individual algorithms, yet it was also a strictly cooperative system. As these examples show and to the best of our knowledge, there have been no previous attempts to provide a framework that can accommodate both independent and heterogeneous learning agents and this can be regarded as the main contribution of our work. 6. CONCLUSION In this paper, we outlined a generic, abstract framework for distributed machine learning and data mining. This framework constitutes, to our knowledge, the first attempt 684 The Sixth Intl. Joint Conf. on Autonomous Agents and Multi-Agent Systems (AAMAS 07) to capture complex forms of interaction between heterogeneous and/or self-interested learners in an architecture that can be used as the foundation for implementing systems that use complex interaction and reasoning mechanisms to enable agents to inform and improve their learning abilities with information provided by other learners in the system, provided that all agents engage in a sufficiently similar learning activity. To illustrate that the abstract principles of our architecture can be turned into concrete, computational systems, we described a market-based distributed clustering system which was evaluated in the domain of vessel tracking for purposes of identifying deviant or suspicious behaviour. Although our experimental results only hint at the potential of using our architecture, they underline that what we are proposing is feasible in principle and can have beneficial effects even in its most simple instantiation. Yet there is a number of issues that we have not addressed in the presentation of the architecture and its empirical evaluation: Firstly, we have not considered the cost of communication and made the implicit assumption that the required communication comes for free. This is of course inadequate if we want to evaluate our method in terms of the total effort required for producing a certain quality of learning results. Secondly, we have not experimented with agents using completely different learning algorithms (e.g. symbolic and numerical). In systems composed of completely different agents the circumstances under which successful information exchange can be achieved might be very different from those described here, and much more complex communication and reasoning methods may be necessary to achieve a useful integration of different agents" learning processes. Finally, more sophisticated evaluation criteria for such distributed learning architectures have to be developed to shed some light on what the right measures of optimality for autonomously reasoning and communicating agents should be. These issues, together with a more systematic and thorough investigation of advanced interaction and communication mechanisms for distributed, collaborating and competing agents will be the subject of our future work on the subject. Acknowledgement: We gratefully acknowledge the support of the presented research by Army Research Laboratory project N62558-03-0819 and Office for Naval Research project N00014-06-1-0232. 7. REFERENCES [1] http://www.aislive.com. [2] http://www.healthagents.com. [3] S. Bailey, R. Grossman, H. Sivakumar, and A. Turinsky. Papyrus: A System for Data Mining over Local and Wide Area Clusters and Super-Clusters. In Proc. of the Conference on Supercomputing. 1999. [4] E. Bauer and R. Kohavi. An Empirical Comparison of Voting Classification Algorithms: Bagging, Boosting, and Variants. Machine Learning, 36, 1999. [5] P. Berkhin. Survey of Clustering Data Mining Techniques, Technical Report, Accrue Software, 2002. [6] D. Caragea, A. Silvescu, and V. Honavar. Agents that Learn from Distributed Dynamic Data sources. In Proc. of the Workshop on Learning Agents, 2000. [7] N. Chawla and S. E. abd L. O. Hall. Creating ensembles of classifiers. In Proceedings of ICDM 2001, pages 580-581, San Jose, CA, USA, 2001. [8] D. Dash and G. F. Cooper. Model Averaging for Prediction with Discrete Bayesian Networks. Journal of Machine Learning Research, 5:1177-1203, 2004. [9] D. L. Davies and D. W. Bouldin. A Cluster Separation Measure. IEEE Transactions on Pattern Analysis and Machine Intelligence, 4:224-227, 1979. [10] P. Edwards and W. Davies. A Heterogeneous Multi-Agent Learning System. In Proceedings of the Special Interest Group on Cooperating Knowledge Based Systems, pages 163-184, 1993. [11] J. Ghosh, A. Strehl, and S. Merugu. A Consensus Framework for Integrating Distributed Clusterings Under Limited Knowledge Sharing. In NSF Workshop on Next Generation Data Mining, 99-108, 2002. [12] D. L. Grecu and L. A. Becker. Coactive Learning for Distributed Data Mining. In Proceedings of KDD-98, pages 209-213, New York, NY, August 1998. [13] M. Klusch, S. Lodi, and G. Moro. Agent-based distributed data mining: The KDEC scheme. In AgentLink, number 2586 in LNCS. Springer, 2003. [14] T. M. Mitchell. Machine Learning, pages 29-36. McGraw-Hill, New York, 1997. [15] S. Ontanon and E. Plaza. Recycling Data for Multi-Agent Learning. In Proc. of ICML-05, 2005. [16] L. Panait and S. Luke. Cooperative multi-agent learning: The state of the art. Autonomous Agents and Multi-Agent Systems, 11(3):387-434, 2005. [17] B. Park and H. Kargupta. Distributed Data Mining: Algorithms, Systems, and Applications. In N. Ye, editor, Data Mining Handbook, pages 341-358, 2002. [18] F. J. Provost and D. N. Hennessy. Scaling up: Distributed machine learning with cooperation. In Proc. of AAAI-96, pages 74-79. AAAI Press, 1996. [19] S. Sian. Extending learning to multiple agents: Issues and a model for multi-agent machine learning (ma-ml). In Y. Kodratoff, editor, Machine LearningEWSL-91, pages 440-456. Springer-Verlag, 1991. [20] R. Smith. The contract-net protocol: High-level communication and control in a distributed problem solver. IEEE Transactions on Computers, C-29(12):1104-1113, 1980. [21] S. J. Stolfo, A. L. Prodromidis, S. Tselepis, W. Lee, D. W. Fan, and P. K. Chan. Jam: Java Agents for Meta-Learning over Distributed Databases. In Proc. of the KDD-97, pages 74-81, USA, 1997. [22] J. Toˇziˇcka, M. Jakob, and M. Pˇechouˇcek. Market-Inspired Approach to Collaborative Learning. In Cooperative Information Agents X (CIA 2006), volume 4149 of LNCS, pages 213-227. Springer, 2006. [23] Y. Z. Wei, L. Moreau, and N. R. Jennings. Recommender systems: a market-based design. In Proceedings of AAMAS-03), pages 600-607, 2003. [24] G. Weiß. A Multiagent Perspective of Parallel and Distributed Machine Learning. In Proceedings of Agents"98, pages 226-230, 1998. [25] G. Weiss and P. Dillenbourg. What is "multi" in multi-agent learning? Collaborative-learning: Cognitive and Computational Approaches, 64-80, 1999. The Sixth Intl. Joint Conf. on Autonomous Agents and Multi-Agent Systems (AAMAS 07) 685

unsupervise cluster;distributed clustering application;multiagent learn;framework and architecture;meta-reasoning;unsupervised clustering;agent;bayesian classifier;frameworks and architecture;historical information;distribute machine learn;individual learning process;machine learning;consensusbased method;autonomous learning agent;datum mining;communication and coordination

train_I-38

Bidding Algorithms for a Distributed Combinatorial Auction

Distributed allocation and multiagent coordination problems can be solved through combinatorial auctions. However, most of the existing winner determination algorithms for combinatorial auctions are centralized. The PAUSE auction is one of a few efforts to release the auctioneer from having to do all the work (it might even be possible to get rid of the auctioneer). It is an increasing price combinatorial auction that naturally distributes the problem of winner determination amongst the bidders in such a way that they have an incentive to perform the calculation. It can be used when we wish to distribute the computational load among the bidders or when the bidders do not wish to reveal their true valuations unless necessary. PAUSE establishes the rules the bidders must obey. However, it does not tell us how the bidders should calculate their bids. We have developed a couple of bidding algorithms for the bidders in a PAUSE auction. Our algorithms always return the set of bids that maximizes the bidder"s utility. Since the problem is NP-Hard, run time remains exponential on the number of items, but it is remarkably better than an exhaustive search. In this paper we present our bidding algorithms, discuss their virtues and drawbacks, and compare the solutions obtained by them to the revenue-maximizing solution found by a centralized winner determination algorithm.

1. INTRODUCTION Both the research and practice of combinatorial auctions have grown rapidly in the past ten years. In a combinatorial auction bidders can place bids on combinations of items, called packages or bidsets, rather than just individual items. Once the bidders place their bids, it is necessary to find the allocation of items to bidders that maximizes the auctioneer"s revenue. This problem, known as the winner determination problem, is a combinatorial optimization problem and is NP-Hard [10]. Nevertheless, several algorithms that have a satisfactory performance for problem sizes and structures occurring in practice have been developed. The practical applications of combinatorial auctions include: allocation of airport takeoff and landing time slots, procurement of freight transportation services, procurement of public transport services, and industrial procurement [2]. Because of their wide applicability, one cannot hope for a general-purpose winner determination algorithm that can efficiently solve every instance of the problem. Thus, several approaches and algorithms have been proposed to address the winner determination problem. However, most of the existing winner determination algorithms for combinatorial auctions are centralized, meaning that they require all agents to send their bids to a centralized auctioneer who then determines the winners. Examples of these algorithms are CASS [3], Bidtree [11] and CABOB [12]. We believe that distributed solutions to the winner determination problem should be studied as they offer a better fit for some applications as when, for example, agents do not want to reveal their valuations to the auctioneer. The PAUSE (Progressive Adaptive User Selection Environment) auction [4, 5] is one of a few efforts to distribute the problem of winner determination amongst the bidders. PAUSE establishes the rules the participants have to adhere to so that the work is distributed amongst them. However, it is not concerned with how the bidders determine what they should bid. In this paper we present two algorithms, pausebid and cachedpausebid, which enable agents in a PAUSE auction to find the bidset that maximizes their utility. Our algorithms implement a myopic utility maximizing strategy and are guaranteed to find the bidset that maximizes the agent"s utility given the outstanding best bids at a given time. pausebid performs a branch and bound search completely from scratch every time that it is called. cachedpausebid is a caching-based algorithm which explores fewer nodes, since it caches some solutions. 694 978-81-904262-7-5 (RPS) c 2007 IFAAMAS 2. THE PAUSE AUCTION A PAUSE auction for m items has m stages. Stage 1 consists of having simultaneous ascending price open-cry auctions and during this stage the bidders can only place bids on individual items. At the end of this state we will know what the highest bid for each individual item is and who placed that bid. Each successive stage k = 2, 3, . . . , m consists of an ascending price auction where the bidders must submit bidsets that cover all items but each one of the bids must be for k items or less. The bidders are allowed to use bids that other agents have placed in previous rounds when building their bidsets, thus allowing them to find better solutions. Also, any new bidset has to have a sum of bid prices which is bigger than that of the currently winning bidset. At the end of each stage k all agents know the best bid for every subset of size k or less. Also, at any point in time after stage 1 has ended there is a standing bidset whose value increases monotonically as new bidsets are submitted. Since in the final round all agents consider all possible bidsets, we know that the final winning bidset will be one such that no agent can propose a better bidset. Note, however, that this bidset is not guaranteed to be the one that maximizes revenue since we are using an ascending price auction so the winning bid for each set will be only slightly bigger than the second highest bid for the particular set of items. That is, the final prices will not be the same as the prices in a traditional combinatorial auction where all the bidders bid their true valuation. However, there remains the open question of whether the final distribution of items to bidders found in a PAUSE auction is the same as the revenue maximizing solution. Our test results provide an answer to this question. The PAUSE auction makes the job of the auctioneer very easy. All it has to do is to make sure that each new bidset has a revenue bigger than the current winning bidset, as well as make sure that every bid in an agent"s bidset that is not his does indeed correspond to some other agents" previous bid. The computational problem shifts from one of winner determination to one of bid generation. Each agent must search over the space of all bidsets which contain at least one of its bids. The search is made easier by the fact that the agent needs to consider only the current best bids and only wants bidsets where its own utility is higher than in the current winning bidset. Each agent also has a clear incentive for performing this computation, namely, its utility only increases with each bidset it proposes (of course, it might decrease with the bidsets that others propose). Finally, the PAUSE auction has been shown to be envy-free in that at the conclusion of the auction no bidder would prefer to exchange his allocation with that of any other bidder [2]. We can even envision completely eliminating the auctioneer and, instead, have every agent perform the task of the auctioneer. That is, all bids are broadcast and when an agent receives a bid from another agent it updates the set of best bids and determines if the new bid is indeed better than the current winning bid. The agents would have an incentive to perform their computation as it will increase their expected utility. Also, any lies about other agents" bids are easily found out by keeping track of the bids sent out by every agent (the set of best bids). Namely, the only one that can increase an agent"s bid value is the agent itself. Anyone claiming a higher value for some other agent is lying. The only thing missing is an algorithm that calculates the utility-maximizing bidset for each agent. 3. PROBLEM FORMULATION A bid b is composed of three elements bitems (the set of items the bid is over), bagent (the agent that placed the bid), and bvalue (the value or price of the bid). The agents maintain a set B of the current best bids, one for each set of items of size ≤ k, where k is the current stage. At any point in the auction, after the first round, there will also be a set W ⊆ B of currently winning bids. This is the set of bids that covers all the items and currently maximizes the revenue, where the revenue of W is given by r(W) = b∈W bvalue . (1) Agent i"s value function is given by vi(S) ∈ where S is a set of items. Given an agent"s value function and the current winning bidset W we can calculate the agent"s utility from W as ui(W) = b∈W | bagent=i vi(bitems ) − bvalue . (2) That is, the agent"s utility for a bidset W is the value it receives for the items it wins in W minus the price it must pay for those items. If the agent is not winning any items then its utility is zero. The goal of the bidding agents in the PAUSE auction is to maximize their utility, subject to the constraint that their next set of bids must have a total revenue that is at least bigger than the current revenue, where is the smallest increment allowed in the auction. Formally, given that W is the current winning bidset, agent i must find a g∗ i such that r(g∗ i ) ≥ r(W) + and g∗ i = arg max g⊆2B ui(g), (3) where each g is a set of bids that covers all items and ∀b∈g (b ∈ B) or (bagent = i and bvalue > B(bitems ) and size(bitems ) ≤ k), and where B(items) is the value of the bid in B for the set items (if there is no bid for those items it returns zero). That is, each bid b in g must satisfy at least one of the two following conditions. 1) b is already in B, 2) b is a bid of size ≤ k in which the agent i bids higher than the price for the same items in B. 4. BIDDING ALGORITHMS According to the PAUSE auction, during the first stage we have only several English auctions, with the bidders submitting bids on individual items. In this case, an agent"s dominant strategy is to bid higher than the current winning bid until it reaches its valuation for that particular item. Our algorithms focus on the subsequent stages: k > 1. When k > 1, agents have to find g∗ i . This can be done by performing a complete search on B. However, this approach is computationally expensive since it produces a large search tree. Our algorithms represent alternative approaches to overcome this expensive search. 4.1 The PAUSEBID Algorithm In the pausebid algorithm (shown in Figure 1) we implement some heuristics to prune the search tree. Given that bidders want to maximize their utility and that at any given point there are likely only a few bids within B which The Sixth Intl. Joint Conf. on Autonomous Agents and Multi-Agent Systems (AAMAS 07) 695 pausebid(i, k) 1 my-bids ← ∅ 2 their-bids ← ∅ 3 for b ∈ B 4 do if bagent = i or vi(bitems ) > bvalue 5 then my-bids ← my-bids +new Bid(bitems , i, vi(bitems )) 6 else their-bids ← their-bids +b 7 for S ∈ subsets of k or fewer items such that vi(S) > 0 and ¬∃b∈Bbitems = S 8 do my-bids ← my-bids +new Bid(S, i, vi(S)) 9 bids ← my-bids + their-bids 10 g∗ ← ∅ £ Global variable 11 u∗ ← ui(W)£ Global variable 12 pbsearch(bids, ∅) 13 surplus ← b∈g∗ | bagent=i bvalue − B(bitems ) 14 if surplus = 0 15 then return g∗ 16 my-payment ← vi(g∗ ) − u∗ 17 for b ∈ g∗ | bagent = i 18 do if my-payment ≤ 0 19 then bvalue ← B(bitems ) 20 else bvalue ← B(bitems ) + my-payment ·bvalue −B(bitems ) surplus 21 return g∗ Figure 1: The pausebid algorithm which implements a branch and bound search. i is the agent and k is the current stage of the auction, for k ≥ 2. the agent can dominate, we start by defining my-bids to be the list of bids for which the agent"s valuation is higher than the current best bid, as given in B. We set the value of these bids to be the agent"s true valuation (but we won"t necessarily be bidding true valuation, as we explain later). Similarly, we set their-bids to be the rest of the bids from B. Finally, the agent"s search list is simply the concatenation of my-bids and their-bids. Note that the agent"s own bids are placed first on the search list as this will enable us to do more pruning (pausebid lines 3 to 9). The agent can now perform a branch and bound search on the branch-on-bids tree produced by these bids. This branch and bound search is implemented by pbsearch (Figure 2). Our algorithm not only implements the standard bound but it also implements other pruning techniques in order to further reduce the size of the search tree. The bound we use is the maximum utility that the agent can expect to receive from a given set of bids. We call it u∗ . Initially, u∗ is set to ui(W) (pausebid line 11) since that is the utility the agent currently receives and any solution he proposes should give him more utility. If pbsearch ever comes across a partial solution where the maximum utility the agent can expect to receive is less than u∗ then that subtree is pruned (pbsearch line 21). Note that we can determine the maximum utility only after the algorithm has searched over all of the agent"s own bids (which are first on the list) because after that we know that the solution will not include any more bids where the agent is the winner thus the agent"s utility will no longer increase. For example, pbsearch(bids, g) 1 if bids = ∅ then return 2 b ← first(bids) 3 bids ← bids −b 4 g ← g + b 5 ¯Ig ← items not in g 6 if g does not contain a bid from i 7 then return 8 if g includes all items 9 then min-payment ← max(0, r(W) + − (r(g) − ri(g)), b∈g | bagent=i B(bitems )) 10 max-utility ← vi(g) − min-payment 11 if r(g) > r(W) and max-utility ≥ u∗ 12 then g∗ ← g 13 u∗ ← max-utility 14 pbsearch(bids, g − b) £ b is Out 15 else max-revenue ← r(g) + max(h(¯Ig), hi(¯Ig)) 16 if max-revenue ≤ r(W) 17 then pbsearch(bids, g − b) £ b is Out 18 elseif bagent = i 19 then min-payment ← (r(W) + ) −(r(g) − ri(g)) − h(¯Ig) 20 max-utility ← vi(g) − min-payment 21 if max-utility > u∗ 22 then pbsearch({x ∈ bids | xitems ∩ bitems = ∅}, g) £ b is In 23 pbsearch(bids, g − b) £ b is Out 24 else 25 pbsearch({x ∈ bids | xitems ∩ bitems = ∅}, g) £ b is In 26 pbsearch(bids, g − b) £ b is Out 27 return Figure 2: The pbsearch recursive procedure where bids is the set of available bids and g is the current partial solution. if an agent has only one bid in my-bids then the maximum utility he can expect is equal to his value for the items in that bid minus the minimum possible payment we can make for those items and still come up with a set of bids that has revenue greater than r(W). The calculation of the minimum payment is shown in line 19 for the partial solution case and line 9 for the case where we have a complete solution in pbsearch. Note that in order to calculate the min-payment for the partial solution case we need an upper bound on the payments that we must make for each item. This upper bound is provided by h(S) = s∈S max b∈B | s∈bitems bvalue size(bitems) . (4) This function produces a bound identical to the one used by the Bidtree algorithm-it merely assigns to each individual item in S a value equal to the maximum bid in B divided by the number of items in that bid. To prune the branches that cannot lead to a solution with revenue greater than the current W, the algorithm considers both the values of the bids in B and the valuations of the 696 The Sixth Intl. Joint Conf. on Autonomous Agents and Multi-Agent Systems (AAMAS 07) agent. Similarly to (4) we define hi(S, k) = s∈S max S | size(S )≤k and s∈S and vi(S )>0 vi(S ) size(S ) (5) which assigns to each individual item s in S the maximum value produced by the valuation of S divided by the size of S , where S is a set for which the agent has a valuation greater than zero, contains s, and its size is less or equal than k. The algorithm uses the heuristics h and hi (lines 15 and 19 of pbsearch), to prune the just mentioned branches in the same way an A∗ algorithm uses its heuristic. A final pruning technique implemented by the algorithm is ignoring any branches where the agent has no bids in the current answer g and no more of the agent"s bids are in the list (pbsearch lines 6 and 7). The resulting g∗ found by pbsearch is thus the set of bids that has revenue bigger than r(W) and maximizes agent i"s utility. However, agent i"s bids in g∗ are still set to his own valuation and not to the lowest possible price. Lines 17 to 20 in pausebid are responsible for setting the agent"s payments so that it can achieve its maximum utility u∗ . If the agent has only one bid in g∗ then it is simply a matter of reducing the payment of that bid by u∗ from the current maximum of the agent"s true valuation. However, if the agent has more than one bid then we face the problem of how to distribute the agent"s payments among these bids. There are many ways of distributing the payments and there does not appear to be a dominant strategy for performing this distribution. We have chosen to distribute the payments in proportion to the agent"s true valuation for each set of items. pausebid assumes that the set of best bids B and the current best winning bidset W remains constant during its execution, and it returns the agent"s myopic utility-maximizing bidset (if there is one) using a branch and bound search. However it repeats the whole search at every stage. We can minimize this problem by caching the result of previous searches. 4.2 The CACHEDPAUSEBID Algorithm The cachedpausebid algorithm (shown in Figure 3) is our second approach to solve the bidding problem in the PAUSE auction. It is based in a cache table called C-Table where we store some solutions to avoid doing a complete search every time. The problem is the same; the agent i has to find g∗ i . We note that g∗ i is a bidset that contains at least one bid of the agent i. Let S be a set of items for which the agent i has a valuation such that vi(S) ≥ B(S) > 0, let gS i be a bidset over S such that r(gS i ) ≥ r(W) + and gS i = arg max g⊆2B ui(g), (6) where each g is a set of bids that covers all items and ∀b∈g (b ∈ B) or (bagent = i and bvalue > B(bitems )) and (∃b∈gbitems = S and bagent = i). That is, gS i is i"s best bidset for all items which includes a bid from i for all S items. In the PAUSE auction we cannot bid for sets of items with size greater than k. So, if we have for each set of items S for which vi(S) > 0 and size(S) ≤ k its corresponding gS i then g∗ i is the gS i that maximizes the agent"s utility. That is g∗ i = arg max {S | vi(S)>0∧size(S)≤k} ui(gS i ). (7) Each agent i implements a hash table C-Table such that C-Table[S] = gS for all S which vi(S) ≥ B(S) > 0. We can cachedpausebid(i, k, k-changed) 1 for each S in C-Table 2 do if vi(S) < B(S) 3 then remove S from C-Table 4 else if k-changed and size(S) = k 5 then B ← B + new Bid(i, S, vi(S)) 6 g∗ ← ∅ 7 u∗ ← ui(W) 8 for each S with size(S) ≤ k in C-Table 9 do ¯S ← Items − S 10 gS ← C-Table[S] £ Global variable 11 min-payment ← max(r(W) + , b∈gS B(bitems )) 12 uS ← r(gS ) − min-payment £ Global variable 13 if (k-changed and size(S) = k) or (∃b∈B bitems ⊆ ¯S and bagent = i) 14 then B ← {b ∈ B |bitems ⊆ ¯S} 15 bids ← B +{b ∈ B|bitems ⊆ ¯S and b /∈ B } 16 for b ∈ bids 17 do if vi(bitems ) > bvalue 18 then bagent ← i 19 bvalue ← vi(bitems ) 20 if k-changed and size(S) = k 21 then n ← size(bids) 22 uS ← 0 23 else n ← size(B ) 24 g ← ∅ + new Bid(S, i, vi(S)) 25 cpbsearch(bids, g, n) 26 C-Table[S] ← gS 27 if uS > u∗ and r(gS ) ≥ r(W) + 28 then surplus ← b∈gS | bagent=i bvalue − B(bitems ) 29 if surplus > 0 30 then my-payment ← vi(gS ) − ui(gS ) 31 for b ∈ gS | bagent = i 32 do if my-payment ≤ 0 33 then bvalue ← B(bitems ) 34 else bvalue ← B(bitems )+ my-payment ·bvalue −B(bitems ) surplus 35 u∗ ← ui(gS ) 36 g∗ ← gS 37 else if uS ≤ 0 and vi(S) < B(S) 38 then remove S from C-Table 39 return g∗ Figure 3: The cachedpausebid algorithm that implements a caching based search to find a bidset that maximizes the utility for the agent i. k is the current stage of the auction (for k ≥ 2), and k-changed is a boolean that is true right after the auction moved to the next stage. The Sixth Intl. Joint Conf. on Autonomous Agents and Multi-Agent Systems (AAMAS 07) 697 cpbsearch(bids, g, n) 1 if bids = ∅ or n ≤ 0 then return 2 b ← first(bids) 3 bids ← bids −b 4 g ← g + b 5 ¯Ig ← items not in g 6 if g includes all items 7 then min-payment ← max(0, r(W) + − (r(g) − ri(g)), b∈g | bagent=i B(bitems )) 8 max-utility ← vi(g) − min-payment 9 if r(g) > r(W) and max-utility ≥ uS 10 then gS ← g 11 uS ← max-utility 12 cpbsearch(bids, g − b, n − 1) £ b is Out 13 else max-revenue ← r(g) + max(h(¯Ig), hi(¯Ig)) 14 if max-revenue ≤ r(W) 15 then cpbsearch(bids, g − b, n − 1) £ b is Out 16 elseif bagent = i 17 then min-payment ← (r(W) + ) −(r(g) − ri(g)) − h(¯Ig) 18 max-utility ← vi(g) − min-payment 19 if max-utility > uS 20 then cpbsearch({x ∈ bids | xitems ∩ bitems = ∅}, g, n + 1) £ b is In 21 cpbsearch(bids, g − b, n − 1) £ b is Out 22 else 23 cpbsearch({x ∈ bids | xitems ∩ bitems = ∅}, g, n + 1) £ b is In 24 cpbsearch(bids, g − b, n − 1) £ b is Out 25 return Figure 4: The cpbsearch recursive procedure where bids is the set of available bids, g is the current partial solution and n is a value that indicates how deep in the list bids the algorithm has to search. then find g∗ by searching for the gS , stored in C-Table[S], that maximizes the agent"s utility, considering only the set of items S with size(S) ≤ k. The problem remains in maintaining the C-Table updated and avoiding to search every gS every time. cachedpausebid deals with this and other details. Let B be the set of bids that contains the new best bids, that is, B contains the bids recently added to B and the bids that have changed price (always higher), bidder, or both and were already in B. Let ¯S = Items − S be the complement of S (the set of items not included in S). cachedpausebid takes three parameters: i the agent, k the current stage of the auction, and k-changed a boolean that is true right after the auction moved to the next stage. Initially C-Table has one row or entry for each set S for which vi(S) > 0. We start by eliminating the entries corresponding to each set S for which vi(S) < B(S) from C-Table (line 3). Then, in the case that k-changed is true, for each set S with size(S) = k, we add to B a bid for that set with value equal to vi(S) and bidder agent i (line 5); this a bid that the agent is now allowed to consider. We then search for g∗ amongst the gS stored in C-Table, for this we only need to consider the sets with size(S) ≤ k (line 8). But how do we know that the gS in C-Table[S] is still the best solution for S? There are only two cases when we are not sure about that and we need to do a search to update C-Table[S]. These cases are: i) When k-changed is true and size(S) ≤ k, since there was no gS stored in C-Table for this S. ii) When there exists at least one bid in B for the set of items ¯S or a subset of it submitted by an agent different than i, since it is probable that this new bid can produce a solution better than the one stored in C-Table[S]. We handle the two cases mentioned above in lines 13 to 26 of cachedpausebid. In both of these cases, since gS must contain a bid for S we need to find a bidset that cover the missing items, that is ¯S. Thus, our search space consists of all the bids on B for the set of items ¯S or for a subset of it. We build the list bids that contains only those bids. However, we put the bids from B at the beginning of bids (line 14) since they are the ones that have changed. Then, we replace the bids in bids that have a price lower than the valuation the agent i has for those same items with a bid from agent i for those items and value equal to the agent"s valuation (lines 16-19). The recursive procedure cpbsearch, called in line 25 of cachedpausebid and shown in Figure 4, is the one that finds the new gS . cpbsearch is a slightly modified version of our branch and bound search implemented in pbsearch. The first modification is that it has a third parameter n that indicates how deep on the list bids we want to search, since it stops searching when n less or equal to zero and not only when the list bids is empty (line 1). Each time that there is a recursive call of cpbsearch n is decreased by one when a bid from bids is discarded or out (lines 12, 15, 21, and 24) and n remains the same otherwise (lines 20 and 23). We set the value of n before calling cpbsearch, to be the size of the list bids (cachedpausebid line 21) in case i), since we want cpbsearch to search over all bids; and we set n to be the number of bids from B included in bids (cachedpausebid line 23) in case ii), since we know that only the those first n bids in bids changed and can affect our current gS . Another difference with pbsearch is that the bound in cpbsearch is uS which we set to be 0 (cachedpausebid line 22) when in case i) and r(gS )−min-payment (cachedpausebid line 12) when in case ii). We call cpbsearch with g already containing a bid for S. After cpbsearch is executed we are sure that we have the right gS , so we store it in the corresponding C-Table[S] (cachedpausebid line 26). When we reach line 27 in cachedpausebid, we are sure that we have the right gS . However, agent i"s bids in gS are still set to his own valuation and not to the lowest possible price. If uS is greater than the current u∗ , lines 31 to 34 in cachedpausebid are responsible for setting the agent"s payments so that it can achieve its maximum utility uS . As in pausebid, we have chosen to distribute the payments in proportion to the agent"s true valuation for each set of items. In the case that uS less than or equal to zero and the valuation that the agent i has for the set of items S is lower than the current value of the bid in B for the same set of items, we remove the corresponding C-Table[S] since we know that is not worthwhile to keep it in the cache table (cachedpausebid line 38). The cachedpausebid function is called when k > 1 and returns the agent"s myopic utility-maximizing bidset, if there is one. It assumes that W and B remains constant during its execution. 698 The Sixth Intl. Joint Conf. on Autonomous Agents and Multi-Agent Systems (AAMAS 07) generatevalues(i, items) 1 for x ∈ items 2 do vi(x) = expd(.01) 3 for n ← 1 . . . (num-bids − items) 4 do s1, s2 ←Two random sets of items with values. 5 vi(s1 ∪ s2) = vi(s1) + vi(s2) + expd(.01) Figure 5: Algorithm for the generation of random value functions. expd(x) returns a random number taken from an exponential distribution with mean 1/x. 0 20 40 60 80 100 2 3 4 5 6 7 8 9 10 Number of Items CachedPauseBid 3 3 3 3 3 3 3 3 3 3 PauseBid + + + + + + + + + + Figure 6: Average percentage of convergence (y-axis), which is the percentage of times that our algorithms converge to the revenue-maximizing solution, as function of the number of items in the auction. 5. TEST AND COMPARISON We have implemented both algorithms and performed a series of experiments in order to determine how their solution compares to the revenue-maximizing solution and how their times compare with each other. In order to do our tests we had to generate value functions for the agents1 . The algorithm we used is shown in Figure 5. The type of valuations it generates correspond to domains where a set of agents must perform a set of tasks but there are cost savings for particular agents if they can bundle together certain subsets of tasks. For example, imagine a set of robots which must pick up and deliver items to different locations. Since each robot is at a different location and has different abilities, each one will have different preferences over how to bundle. Their costs for the item bundles are subadditive, which means that their preferences are superadditive. The first experiment we performed simply ensured the proper 1 Note that we could not use CATS [6] because it generates sets of bids for an indeterminate number of agents. It is as if you were told the set of bids placed in a combinatorial auction but not who placed each bid or even how many people placed bids, and then asked to determine the value function of every participant in the auction. 0 20 40 60 80 100 2 3 4 5 6 7 8 9 10 Number of Items CachedPauseBid 3 3 3 3 3 3 3 3 3 3 PauseBid + + + + + + + + + + Figure 7: Average percentage of revenue from our algorithms relative to maximum revenue (y-axis) as function of the number of items in the auction. functioning of our algorithms. We then compared the solutions found by both of them to the revenue-maximizing solution as found by CASS when given a set of bids that corresponds to the agents" true valuation. That is, for each agent i and each set of items S for which vi(S) > 0 we generated a bid. This set of bids was fed to CASS which implements a centralized winner determination algorithm to find the solution which maximizes revenue. Note, however, that the revenue from the PAUSE auction on all the auctions is always smaller than the revenue of the revenue-maximizing solution when the agents bid their true valuations. Since PAUSE uses English auctions the final prices (roughly) represent the second-highest valuation, plus , for that set of items. We fixed the number of agents to be 5 and we experimented with different number of items, namely from 2 to 10. We ran both algorithms 100 times for each combination. When we compared the solutions of our algorithms to the revenue-maximizing solution, we realized that they do not always find the same distribution of items as the revenue-maximizing solution (as shown in Figure 6). The cases where our algorithms failed to arrive at the distribution of the revenue-maximizing solution are those where there was a large gap between the first and second valuation for a set (or sets) of items. If the revenue-maximizing solution contains the bid (or bids) using these higher valuation then it is impossible for the PAUSE auction to find this solution because that bid (those bids) is never placed. For example, if agent i has vi(1) = 1000 and the second highest valuation for (1) is only 10 then i only needs to place a bid of 11 in order to win that item. If the revenue-maximizing solution requires that 1 be sold for 1000 then that solution will never be found because that bid will never be placed. We also found that average percentage of times that our algorithms converges to the revenue-maximizing solution decreases as the number of items increases. For 2 items is almost 100% but decreases a little bit less than 1 percent as the items increase, so that this average percentage of convergence is around 90% for 10 items. In a few instances our algorithms find different solutions this is due to the different The Sixth Intl. Joint Conf. on Autonomous Agents and Multi-Agent Systems (AAMAS 07) 699 1 10 100 1000 10000 2 3 4 5 6 7 8 9 10 Number of Items CachedPauseBid 3 3 3 3 3 3 3 3 3 PauseBid + + + + + + + + + + Figure 8: Average number of expanded nodes (y-axis) as function of items in the auction. ordering of the bids in the bids list which makes them search in different order. We know that the revenue generated by the PAUSE auction is generally lower than the revenue of the revenuemaximizing solution, but how much lower? To answer this question we calculated percentage representing the proportion of the revenue given by our algorithms relative to the revenue given by CASS. We found that the percentage of revenue of our algorithms increases in average 2.7% as the number of items increases, as shown in Figure 7. However, we found that cachedpausebid generates a higher revenue than pausebid (4.3% higher in average) except for auctions with 2 items where both have about the same percentage. Again, this difference is produced by the order of the search. In the case of 2 items both algorithms produce in average a revenue proportion of 67.4%, while in the other extreme (10 items), cachedpausebid produced in average a revenue proportion of 91.5% while pausebid produced in average a revenue proportion of 87.7%. The scalability of our algorithms can be determined by counting the number of nodes expanded in the search tree. For this we count the number of times that pbsearch gets invoked for each time that pausebid is called and the number of times that fastpausebidsearch gets invoked for each time that cachedpausebid, respectively for each of our algorithms. As expected since this is an NP-Hard problem, the number of expanded nodes does grow exponentially with the number of items (as shown in Figure 8). However, we found that cachedpausebid outperforms pausebid, since it expands in average less than half the number of nodes. For example, the average number of nodes expanded when 2 items is zero for cachedpausebid while for pausebid is 2; and in the other extreme (10 items) cachedpausebid expands in average only 633 nodes while pausebid expands in average 1672 nodes, a difference of more than 1000 nodes. Although the number of nodes expanded by our algorithms increases as function of the number of items, the actual number of nodes is a much smaller than the worst-case scenario of nn where n is the number of items. For example, for 10 items we expand slightly more than 103 nodes for the case of pausebid and less than that for the case of cachedpause0.1 1 10 100 1000 2 3 4 5 6 7 8 9 10 Number of Items CachedPauseBid 3 3 3 3 3 3 3 3 3 3 PauseBid + + + + + + + + + + Figure 9: Average time in seconds that takes to finish an auction (y-axis) as function of the number of items in the auction. bid which are much smaller numbers than 1010 . Notice also that our value generation algorithm (Figure 5) generates a number of bids that is exponential on the number of items, as might be expected in many situations. As such, these results do not support the conclusion that time grows exponentially with the number of items when the number of bids is independent of the number of items. We expect that both algorithms will grow exponentially as a function the number of bids, but stay roughly constant as the number of items grows. We wanted to make sure that less expanded nodes does indeed correspond to faster execution, especially since our algorithms execute different operations. We thus ran the same experiment with all the agents in the same machine, an Intel Centrino 2.0 GHz laptop PC with 1 GB of RAM and a 7200 RMP 60 GB hard drive, and calculated the average time that takes to finish an auction for each algorithm. As shown in Figure 9, cachedpausebid is faster than pausebid, the difference in execution speed is even more clear as the number of items increases. 6. RELATED WORK A lot of research has been done on various aspects of combinatorial auctions. We recommend [2] for a good review. However, the study of distributed winner determination algorithms for combinatorial auctions is still relatively new. One approach is given by the algorithms for distributing the winner determination problem in combinatorial auctions presented in [7], but these algorithms assume the computational entities are the items being sold and thus end up with a different type of distribution. The VSA algorithm [3] is another way of performing distributed winner determination in combinatorial auction but it assumes the bids themselves perform the computation. This algorithm also fails to converge to a solution for most cases. In [9] the authors present a distributed mechanism for calculating VCG payments in a mechanism design problem. Their mechanism roughly amounts to having each agent calculate the payments for two other agents and give these to a secure 700 The Sixth Intl. Joint Conf. on Autonomous Agents and Multi-Agent Systems (AAMAS 07) central server which then checks to make sure results from all pairs agree, otherwise a re-calculation is ordered. This general idea, which they call the redundancy principle, could also be applied to our problem but it requires the existence of a secure center agent that everyone trusts. Another interesting approach is given in [8] where the bidding agents prioritize their bids, thus reducing the set of bids that the centralized winner determination algorithm must consider, making that problem easier. Finally, in the computation procuring clock auction [1] the agents are given an everincreasing percentage of the surplus achieved by their proposed solution over the current best. As such, it assumes the agents are impartial computational entities, not the set of possible buyers as assumed by the PAUSE auction. 7. CONCLUSIONS We believe that distributed solutions to the winner determination problem should be studied as they offer a better fit for some applications as when, for example, agents do not want to reveal their valuations to the auctioneer or when we wish to distribute the computational load among the bidders. The PAUSE auction is one of a few approaches to decentralize the winner determination problem in combinatorial auctions. With this auction, we can even envision completely eliminating the auctioneer and, instead, have every agent performe the task of the auctioneer. However, while PAUSE establishes the rules the bidders must obey, it does not tell us how the bidders should calculate their bids. We have presented two algorithms, pausebid and cachedpausebid, that bidder agents can use to engage in a PAUSE auction. Both algorithms implement a myopic utility maximizing strategy that is guaranteed to find the bidset that maximizes the agent"s utility given the set of outstanding best bids at any given time, without considering possible future bids. Both algorithms find, most of the time, the same distribution of items as the revenue-maximizing solution. The cases where our algorithms failed to arrive at that distribution are those where there was a large gap between the first and second valuation for a set (or sets) of items. As it is an NP-Hard problem, the running time of our algorithms remains exponential but it is significantly better than a full search. pausebid performs a branch and bound search completely from scratch each time it is invoked. cachedpausebid caches partial solutions and performs a branch and bound search only on the few portions affected by the changes on the bids between consecutive times. cachedpausebid has a better performance since it explores fewer nodes (less than half) and it is faster. As expected the revenue generated by a PAUSE auction is lower than the revenue of a revenue-maximizing solution found by a centralized winner determination algorithm, however we found that cachedpausebid generates in average 4.7% higher revenue than pausebid. We also found that the revenue generated by our algorithms increases as function of the number of items in the auction. Our algorithms have shown that it is feasible to implement the complex coordination constraints supported by combinatorial auctions without having to resort to a centralized winner determination algorithm. Moreover, because of the design of the PAUSE auction, the agents in the auction also have an incentive to perform the required computation. Our bidding algorithms can be used by any multiagent system that would use combinatorial auctions for coordination but would rather not implement a centralized auctioneer. 8. REFERENCES [1] P. J. Brewer. Decentralized computation procurement and computational robustness in a smart market. Economic Theory, 13(1):41-92, January 1999. [2] P. Cramton, Y. Shoham, and R. Steinberg, editors. Combinatorial Auctions. MIT Press, 2006. [3] Y. Fujishima, K. Leyton-Brown, and Y. Shoham. Taming the computational complexity of combinatorial auctions: Optimal and approximate approaches. In Proceedings of the Sixteenth International Joint Conference on Artificial Intelligence, pages 548-553. Morgan Kaufmann Publishers Inc., 1999. [4] F. Kelly and R. Stenberg. A combinatorial auction with multiple winners for universal service. Management Science, 46(4):586-596, 2000. [5] A. Land, S. Powell, and R. Steinberg. PAUSE: A computationally tractable combinatorial auction. In Cramton et al. [2], chapter 6, pages 139-157. [6] K. Leyton-Brown, M. Pearson, and Y. Shoham. Towards a universal test suite for combinatorial auction algorithms. In Proceedings of the 2nd ACM conference on Electronic commerce, pages 66-76. ACM Press, 2000. http://cats.stanford.edu. [7] M. V. Narumanchi and J. M. Vidal. Algorithms for distributed winner determination in combinatorial auctions. In LNAI volume of AMEC/TADA. Springer, 2006. [8] S. Park and M. H. Rothkopf. Auctions with endogenously determined allowable combinations. Technical report, Rutgets Center for Operations Research, January 2001. RRR 3-2001. [9] D. C. Parkes and J. Shneidman. Distributed implementations of vickrey-clarke-groves auctions. In Proceedings of the Third International Joint Conference on Autonomous Agents and MultiAgent Systems, pages 261-268. ACM, 2004. [10] M. H. Rothkopf, A. Pekec, and R. M. Harstad. Computationally manageable combinational auctions. Management Science, 44(8):1131-1147, 1998. [11] T. Sandholm. An algorithm for winner determination in combinatorial auctions. Artificial Intelligence, 135(1-2):1-54, February 2002. [12] T. Sandholm, S. Suri, A. Gilpin, and D. Levine. CABOB: a fast optimal algorithm for winner determination in combinatorial auctions. Management Science, 51(3):374-391, 2005. The Sixth Intl. Joint Conf. on Autonomous Agents and Multi-Agent Systems (AAMAS 07) 701

task and resource allocation;bidding algorithm;pause auction;progressive adaptive user selection environment;branch-on-bid tree;combinatorial auction;agent;coordination;search tree;branch and bound search;distributed allocation;combinatorial optimization problem;revenue-maximizing solution

train_I-42

A Complete Distributed Constraint Optimization Method For Non-Traditional Pseudotree Arrangements

Distributed Constraint Optimization (DCOP) is a general framework that can model complex problems in multi-agent systems. Several current algorithms that solve general DCOP instances, including ADOPT and DPOP, arrange agents into a traditional pseudotree structure. We introduce an extension to the DPOP algorithm that handles an extended set of pseudotree arrangements. Our algorithm correctly solves DCOP instances for pseudotrees that include edges between nodes in separate branches. The algorithm also solves instances with traditional pseudotree arrangements using the same procedure as DPOP. We compare our algorithm with DPOP using several metrics including the induced width of the pseudotrees, the maximum dimensionality of messages and computation, and the maximum sequential path cost through the algorithm. We prove that for some problem instances it is not possible to generate a traditional pseudotree using edge-traversal heuristics that will outperform a cross-edged pseudotree. We use multiple heuristics to generate pseudotrees and choose the best pseudotree in linear space-time complexity. For some problem instances we observe significant improvements in message and computation sizes compared to DPOP.

1. INTRODUCTION Many historical problems in the AI community can be transformed into Constraint Satisfaction Problems (CSP). With the advent of distributed AI, multi-agent systems became a popular way to model the complex interactions and coordination required to solve distributed problems. CSPs were originally extended to distributed agent environments in [9]. Early domains for distributed constraint satisfaction problems (DisCSP) included job shop scheduling [1] and resource allocation [2]. Many domains for agent systems, especially teamwork coordination, distributed scheduling, and sensor networks, involve overly constrained problems that are difficult or impossible to satisfy for every constraint. Recent approaches to solving problems in these domains rely on optimization techniques that map constraints into multi-valued utility functions. Instead of finding an assignment that satisfies all constraints, these approaches find an assignment that produces a high level of global utility. This extension to the original DisCSP approach has become popular in multi-agent systems, and has been labeled the Distributed Constraint Optimization Problem (DCOP) [1]. Current algorithms that solve complete DCOPs use two main approaches: search and dynamic programming. Search based algorithms that originated from DisCSP typically use some form of backtracking [10] or bounds propagation, as in ADOPT [3]. Dynamic programming based algorithms include DPOP and its extensions [5, 6, 7]. To date, both categories of algorithms arrange agents into a traditional pseudotree to solve the problem. It has been shown in [6] that any constraint graph can be mapped into a traditional pseudotree. However, it was also shown that finding the optimal pseudotree was NP-Hard. We began to investigate the performance of traditional pseudotrees generated by current edge-traversal heuristics. We found that these heuristics often produced little parallelism as the pseudotrees tended to have high depth and low branching factors. We suspected that there could be other ways to arrange the pseudotrees that would provide increased parallelism and smaller message sizes. After exploring these other arrangements we found that cross-edged pseudotrees provide shorter depths and higher branching factors than the traditional pseudotrees. Our hypothesis was that these crossedged pseudotrees would outperform traditional pseudotrees for some problem types. In this paper we introduce an extension to the DPOP algorithm that handles an extended set of pseudotree arrangements which include cross-edged pseudotrees. We begin with a definition of 741 978-81-904262-7-5 (RPS) c 2007 IFAAMAS DCOP, traditional pseudotrees, and cross-edged pseudotrees. We then provide a summary of the original DPOP algorithm and introduce our DCPOP algorithm. We discuss the complexity of our algorithm as well as the impact of pseudotree generation heuristics. We then show that our Distributed Cross-edged Pseudotree Optimization Procedure (DCPOP) performs significantly better in practice than the original DPOP algorithm for some problem instances. We conclude with a selection of ideas for future work and extensions for DCPOP. 2. PROBLEM DEFINITION DCOP has been formalized in slightly different ways in recent literature, so we will adopt the definition as presented in [6]. A Distributed Constraint Optimization Problem with n nodes and m constraints consists of the tuple < X, D, U > where: • X = {x1,..,xn} is a set of variables, each one assigned to a unique agent • D = {d1,..,dn} is a set of finite domains for each variable • U = {u1,..,um} is a set of utility functions such that each function involves a subset of variables in X and defines a utility for each combination of values among these variables An optimal solution to a DCOP instance consists of an assignment of values in D to X such that the sum of utilities in U is maximal. Problem domains that require minimum cost instead of maximum utility can map costs into negative utilities. The utility functions represent soft constraints but can also represent hard constraints by using arbitrarily large negative values. For this paper we only consider binary utility functions involving two variables. Higher order utility functions can be modeled with minor changes to the algorithm, but they also substantially increase the complexity. 2.1 Traditional Pseudotrees Pseudotrees are a common structure used in search procedures to allow parallel processing of independent branches. As defined in [6], a pseudotree is an arrangement of a graph G into a rooted tree T such that vertices in G that share an edge are in the same branch in T. A back-edge is an edge between a node X and any node which lies on the path from X to the root (excluding X"s parent). Figure 1 shows a pseudotree with four nodes, three edges (A-B, B-C, BD), and one back-edge (A-C). Also defined in [6] are four types of relationships between nodes exist in a pseudotree: • P(X) - the parent of a node X: the single node higher in the pseudotree that is connected to X directly through a tree edge • C(X) - the children of a node X: the set of nodes lower in the pseudotree that are connected to X directly through tree edges • PP(X) - the pseudo-parents of a node X: the set of nodes higher in the pseudotree that are connected to X directly through back-edges (In Figure 1, A = PP(C)) • PC(X) - the pseudo-children of a node X: the set of nodes lower in the pseudotree that are connected to X directly through back-edges (In Figure 1, C = PC(A)) Figure 1: A traditional pseudotree. Solid line edges represent parent-child relationships and the dashed line represents a pseudo-parent-pseudo-child relationship. Figure 2: A cross-edged pseudotree. Solid line edges represent parent-child relationships, the dashed line represents a pseudoparent-pseudo-child relationship, and the dotted line represents a branch-parent-branch-child relationship. The bolded node, B, is the merge point for node E. 2.2 Cross-edged Pseudotrees We define a cross-edge as an edge from node X to a node Y that is above X but not in the path from X to the root. A cross-edged pseudotree is a traditional pseudotree with the addition of cross-edges. Figure 2 shows a cross-edged pseudotree with a cross-edge (D-E). In a cross-edged pseudotree we designate certain edges as primary. The set of primary edges defines a spanning tree of the nodes. The parent, child, pseudo-parent, and pseudo-child relationships from the traditional pseudotree are now defined in the context of this primary edge spanning tree. This definition also yields two additional types of relationships that may exist between nodes: • BP(X) - the branch-parents of a node X: the set of nodes higher in the pseudotree that are connected to X but are not in the primary path from X to the root (In Figure 2, D = BP(E)) • BC(X) - the branch-children of a node X: the set of nodes lower in the pseudotree that are connected to X but are not in any primary path from X to any leaf node (In Figure 2, E = BC(D)) 2.3 Pseudotree Generation 742 The Sixth Intl. Joint Conf. on Autonomous Agents and Multi-Agent Systems (AAMAS 07) Current algorithms usually have a pre-execution phase to generate a traditional pseudotree from a general DCOP instance. Our DCPOP algorithm generates a cross-edged pseudotree in the same fashion. First, the DCOP instance < X, D, U > translates directly into a graph with X as the set of vertices and an edge for each pair of variables represented in U. Next, various heuristics are used to arrange this graph into a pseudotree. One common heuristic is to perform a guided depth-first search (DFS) as the resulting traversal is a pseudotree, and a DFS can easily be performed in a distributed fashion. We define an edge-traversal based method as any method that produces a pseudotree in which all parent/child pairs share an edge in the original graph. This includes DFS, breadth-first search, and best-first search based traversals. Our heuristics that generate cross-edged pseudotrees use a distributed best-first search traversal. 3. DPOP ALGORITHM The original DPOP algorithm operates in three main phases. The first phase generates a traditional pseudotree from the DCOP instance using a distributed algorithm. The second phase joins utility hypercubes from children and the local node and propagates them towards the root. The third phase chooses an assignment for each domain in a top down fashion beginning with the agent at the root node. The complexity of DPOP depends on the size of the largest computation and utility message during phase two. It has been shown that this size directly corresponds to the induced width of the pseudotree generated in phase one [6]. DPOP uses polynomial time heuristics to generate the pseudotree since finding the minimum induced width pseudotree is NP-hard. Several distributed edgetraversal heuristics have been developed to find low width pseudotrees [8]. At the end of the first phase, each agent knows its parent, children, pseudo-parents, and pseudo-children. 3.1 Utility Propagation Agents located at leaf nodes in the pseudotree begin the process by calculating a local utility hypercube. This hypercube at node X contains summed utilities for each combination of values in the domains for P(X) and PP(X). This hypercube has dimensional size equal to the number of pseudo-parents plus one. A message containing this hypercube is sent to P(X). Agents located at non-leaf nodes wait for all messages from children to arrive. Once the agent at node Y has all utility messages, it calculates its local utility hypercube which includes domains for P(Y), PP(Y), and Y. The local utility hypercube is then joined with all of the hypercubes from the child messages. At this point all utilities involving node Y are known, and the domain for Y may be safely eliminated from the joined hypercube. This elimination process chooses the best utility over the domain of Y for each combination of the remaining domains. A message containing this hypercube is now sent to P(Y). The dimensional size of this hypercube depends on the number of overlapping domains in received messages and the local utility hypercube. This dynamic programming based propagation phase continues until the agent at the root node of the pseudotree has received all messages from its children. 3.2 Value Propagation Value propagation begins when the agent at the root node Z has received all messages from its children. Since Z has no parents or pseudo-parents, it simply combines the utility hypercubes received from its children. The combined hypercube contains only values for the domain for Z. At this point the agent at node Z simply chooses the assignment for its domain that has the best utility. A value propagation message with this assignment is sent to each node in C(Z). Each other node then receives a value propagation message from its parent and chooses the assignment for its domain that has the best utility given the assignments received in the message. The node adds its domain assignment to the assignments it received and passes the set of assignments to its children. The algorithm is complete when all nodes have chosen an assignment for their domain. 4. DCPOP ALGORITHM Our extension to the original DPOP algorithm, shown in Algorithm 1, shares the same three phases. The first phase generates the cross-edged pseudotree for the DCOP instance. The second phase merges branches and propagates the utility hypercubes. The third phase chooses assignments for domains at branch merge points and in a top down fashion, beginning with the agent at the root node. For the first phase we generate a pseudotree using several distributed heuristics and select the one with lowest overall complexity. The complexity of the computation and utility message size in DCPOP does not directly correspond to the induced width of the cross-edged pseudotree. Instead, we use a polynomial time method for calculating the maximum computation and utility message size for a given cross-edged pseudotree. A description of this method and the pseudotree selection process appears in Section 5. At the end of the first phase, each agent knows its parent, children, pseudo-parents, pseudo-children, branch-parents, and branch-children. 4.1 Merging Branches and Utility Propagation In the original DPOP algorithm a node X only had utility functions involving its parent and its pseudo-parents. In DCPOP, a node X is allowed to have a utility function involving a branch-parent. The concept of a branch can be seen in Figure 2 with node E representing our node X. The two distinct paths from node E to node B are called branches of E. The single node where all branches of E meet is node B, which is called the merge point of E. Agents with nodes that have branch-parents begin by sending a utility propagation message to each branch-parent. This message includes a two dimensional utility hypercube with domains for the node X and the branch-parent BP(X). It also includes a branch information structure which contains the origination node of the branch, X, the total number of branches originating from X, and the number of branches originating from X that are merged into a single representation by this branch information structure (this number starts at 1). Intuitively when the number of merged branches equals the total number of originating branches, the algorithm has reached the merge point for X. In Figure 2, node E sends a utility propagation message to its branch-parent, node D. This message has dimensions for the domains of E and D, and includes branch information with an origin of E, 2 total branches, and 1 merged branch. As in the original DPOP utility propagation phase, an agent at leaf node X sends a utility propagation message to its parent. In DCPOP this message contains dimensions for the domains of P(X) and PP(X). If node X also has branch-parents, then the utility propagation message also contains a dimension for the domain of X, and will include a branch information structure. In Figure 2, node E sends a utility propagation message to its parent, node C. This message has dimensions for the domains of E and C, and includes branch information with an origin of E, 2 total branches, and 1 merged branch. When a node Y receives utility propagation messages from all of The Sixth Intl. Joint Conf. on Autonomous Agents and Multi-Agent Systems (AAMAS 07) 743 its children and branch-children, it merges any branches with the same origination node X. The merged branch information structure accumulates the number of merged branches for X. If the cumulative total number of merged branches equals the total number of branches, then Y is the merge point for X. This means that the utility hypercubes present at Y contain all information about the valuations for utility functions involving node X. In addition to the typical elimination of the domain of Y from the utility hypercubes, we can now safely eliminate the domain of X from the utility hypercubes. To illustrate this process, we will examine what happens in the second phase for node B in Figure 2. In the second phase Node B receives two utility propagation messages. The first comes from node C and includes dimensions for domains E, B, and A. It also has a branch information structure with origin of E, 2 total branches, and 1 merged branch. The second comes from node D and includes dimensions for domains E and B. It also has a branch information structure with origin of E, 2 total branches, and 1 merged branch. Node B then merges the branch information structures from both messages because they have the same origination, node E. Since the number of merged branches originating from E is now 2 and the total branches originating from E is 2, node B now eliminates the dimensions for domain E. Node B also eliminates the dimension for its own domain, leaving only information about domain A. Node B then sends a utility propagation message to node A, containing only one dimension for the domain of A. Although not possible in DPOP, this method of utility propagation and dimension elimination may produce hypercubes at node Y that do not share any domains. In DCPOP we do not join domain independent hypercubes, but instead may send multiple hypercubes in the utility propagation message sent to the parent of Y. This lazy approach to joins helps to reduce message sizes. 4.2 Value Propagation As in DPOP, value propagation begins when the agent at the root node Z has received all messages from its children. At this point the agent at node Z chooses the assignment for its domain that has the best utility. If Z is the merge point for the branches of some node X, Z will also choose the assignment for the domain of X. Thus any node that is a merge point will choose assignments for a domain other than its own. These assignments are then passed down the primary edge hierarchy. If node X in the hierarchy has branch-parents, then the value assignment message from P(X) will contain an assignment for the domain of X. Every node in the hierarchy adds any assignments it has chosen to the ones it received and passes the set of assignments to its children. The algorithm is complete when all nodes have chosen or received an assignment for their domain. 4.3 Proof of Correctness We will prove the correctness of DCPOP by first noting that DCPOP fully extends DPOP and then examining the two cases for value assignment in DCPOP. Given a traditional pseudotree as input, the DCPOP algorithm execution is identical to DPOP. Using a traditional pseudotree arrangement no nodes have branch-parents or branch-children since all edges are either back-edges or tree edges. Thus the DCPOP algorithm using a traditional pseudotree sends only utility propagation messages that contain domains belonging to the parent or pseudo-parents of a node. Since no node has any branch-parents, no branches exist, and thus no node serves as a merge point for any other node. Thus all value propagation assignments are chosen at the node of the assignment domain. For DCPOP execution with cross-edged pseudotrees, some nodes serve as merge points. We note that any node X that is not a merge point assigns its value exactly as in DPOP. The local utility hypercube at X contains domains for X, P(X), PP(X), and BC(X). As in DPOP the value assignment message received at X includes the values assigned to P(X) and PP(X). Also, since X is not a merge point, all assignments to BC(X) must have been calculated at merge points higher in the tree and are in the value assignment message from P(X). Thus after eliminating domains for which assignments are known, only the domain of X is left. The agent at node X can now correctly choose the assignment with maximum utility for its own domain. If node X is a merge point for some branch-child Y, we know that X must be a node along the path from Y to the root, and from P(Y) and all BP(Y) to the root. From the algorithm, we know that Y necessarily has all information from C(Y), PC(Y), and BC(Y) since it waits for their messages. Node X has information about all nodes below it in the tree, which would include Y, P(Y), BP(Y), and those PP(Y) that are below X in the tree. For any PP(Y) above X in the tree, X receives the assignment for the domain of PP(Y) in the value assignment message from P(X). Thus X has utility information about all of the utility functions of which Y is a part. By eliminating domains included in the value assignment message, node X is left with a local utility hypercube with domains for X and Y. The agent at node X can now correctly choose the assignments with maximum utility for the domains of X and Y. 4.4 Complexity Analysis The first phase of DCPOP sends one message to each P(X), PP(X), and BP(X). The second phase sends one value assignment message to each C(X). Thus, DCPOP produces a linear number of messages with respect to the number of edges (utility functions) in the cross-edged pseudotree and the original DCOP instance. The actual complexity of DCPOP depends on two additional measurements: message size and computation size. Message size and computation size in DCPOP depend on the number of overlapping branches as well as the number of overlapping back-edges. It was shown in [6] that the number of overlapping back-edges is equal to the induced width of the pseudotree. In a poorly constructed cross-edged pseudotree, the number of overlapping branches at node X can be as large as the total number of descendants of X. Thus, the total message size in DCPOP in a poorly constructed instance can be space-exponential in the total number of nodes in the graph. However, in practice a well constructed cross-edged pseudotree can achieve much better results. Later we address the issue of choosing well constructed crossedged pseudotrees from a set. We introduce an additional measurement of the maximum sequential path cost through the algorithm. This measurement directly relates to the maximum amount of parallelism achievable by the algorithm. To take this measurement we first store the total computation size for each node during phase two and three. This computation size represents the number of individual accesses to a value in a hypercube at each node. For example, a join between two domains of size 4 costs 4 ∗ 4 = 16. Two directed acyclic graphs (DAG) can then be drawn; one with the utility propagation messages as edges and the phase two costs at nodes, and the other with value assignment messages and the phase three costs at nodes. The maximum sequential path cost is equal to the sum of the longest path on each DAG from the root to any leaf node. 5. HEURISTICS In our assessment of complexity in DCPOP we focused on the worst case possibly produced by the algorithm. We acknowledge 744 The Sixth Intl. Joint Conf. on Autonomous Agents and Multi-Agent Systems (AAMAS 07) Algorithm 1 DCPOP Algorithm 1: DCPOP(X; D; U) Each agent Xi executes: Phase 1: pseudotree creation 2: elect leader from all Xj ∈ X 3: elected leader initiates pseudotree creation 4: afterwards, Xi knows P(Xi), PP(Xi), BP(Xi), C(Xi), BC(Xi) and PC(Xi) Phase 2: UTIL message propagation 5: if |BP(Xi)| > 0 then 6: BRANCHXi ← |BP(Xi)| + 1 7: for all Xk ∈BP(Xi) do 8: UTILXi (Xk) ←Compute utils(Xi, Xk) 9: Send message(Xk,UTILXi (Xk),BRANCHXi ) 10: if |C(Xi)| = 0(i.e. Xi is a leaf node) then 11: UTILXi (P(Xi)) ← Compute utils(P(Xi),PP(Xi)) for all PP(Xi) 12: Send message(P(Xi), UTILXi (P(Xi)),BRANCHXi ) 13: Send message(PP(Xi), empty UTIL, empty BRANCH) to all PP(Xi) 14: activate UTIL Message handler() Phase 3: VALUE message propagation 15: activate VALUE Message handler() END ALGORITHM UTIL Message handler(Xk,UTILXk (Xi), BRANCHXk ) 16: store UTILXk (Xi),BRANCHXk (Xi) 17: if UTIL messages from all children and branch children arrived then 18: for all Bj ∈BRANCH(Xi) do 19: if Bj is merged then 20: join all hypercubes where Bj ∈UTIL(Xi) 21: eliminate Bj from the joined hypercube 22: if P(Xi) == null (that means Xi is the root) then 23: v ∗ i ← Choose optimal(null) 24: Send VALUE(Xi, v ∗ i) to all C(Xi) 25: else 26: UTILXi (P(Xi)) ← Compute utils(P(Xi), PP(Xi)) 27: Send message(P(Xi),UTILXi (P(Xi)), BRANCHXi (P(Xi))) VALUE Message handler(VALUEXi ,P(Xi)) 28: add all Xk ← v ∗ k ∈VALUEXi ,P(Xi) to agent view 29: Xi ← v ∗ i =Choose optimal(agent view) 30: Send VALUEXl , Xi to all Xl ∈C(Xi) that in real world problems the generation of the pseudotree has a significant impact on the actual performance. The problem of finding the best pseudotree for a given DCOP instance is NP-Hard. Thus a heuristic is used for generation, and the performance of the algorithm depends on the pseudotree found by the heuristic. Some previous research focused on finding heuristics to generate good pseudotrees [8]. While we have developed some heuristics that generate good cross-edged pseudotrees for use with DCPOP, our focus has been to use multiple heuristics and then select the best pseudotree from the generated pseudotrees. We consider only heuristics that run in polynomial time with respect to the number of nodes in the original DCOP instance. The actual DCPOP algorithm has worst case exponential complexity, but we can calculate the maximum message size, computation size, and sequential path cost for a given cross-edged pseudotree in linear space-time complexity. To do this, we simply run the algorithm without attempting to calculate any of the local utility hypercubes or optimal value assignments. Instead, messages include dimensional and branch information but no utility hypercubes. After each heuristic completes its generation of a pseudotree, we execute the measurement procedure and propagate the measurement information up to the chosen root in that pseudotree. The root then broadcasts the total complexity for that heuristic to all nodes. After all heuristics have had a chance to complete, every node knows which heuristic produced the best pseudotree. Each node then proceeds to begin the DCPOP algorithm using its knowledge of the pseudotree generated by the best heuristic. The heuristics used to generate traditional pseudotrees perform a distributed DFS traversal. The general distributed algorithm uses a token passing mechanism and a linear number of messages. Improved DFS based heuristics use a special procedure to choose the root node, and also provide an ordering function over the neighbors of a node to determine the order of path recursion. The DFS based heuristics used in our experiments come from the work done in [4, 8]. 5.1 The best-first cross-edged pseudotree heuristic The heuristics used to generate cross-edged pseudotrees perform a best-first traversal. A general distributed best-first algorithm for node expansion is presented in Algorithm 2. An evaluation function at each node provides the values that are used to determine the next best node to expand. Note that in this algorithm each node only exchanges its best value with its neighbors. In our experiments we used several evaluation functions that took as arguments an ordered list of ancestors and a node, which contains a list of neighbors (with each neighbor"s placement depth in the tree if it was placed). From these we can calculate branchparents, branch-children, and unknown relationships for a potential node placement. The best overall function calculated the value as ancestors−(branchparents+branchchildren) with the number of unknown relationships being a tiebreak. After completion each node has knowledge of its parent and ancestors, so it can easily determine which connected nodes are pseudo-parents, branchparents, pseudo-children, and branch-children. The complexity of the best-first traversal depends on the complexity of the evaluation function. Assuming a complexity of O(V ) for the evaluation function, which is the case for our best overall function, the best-first traversal is O(V · E) which is at worst O(n3 ). For each v ∈ V we perform a place operation, and find the next node to place using the getBestNeighbor operation. The place operation is at most O(V ) because of the sent messages. Finding the next node uses recursion and traverses only already placed The Sixth Intl. Joint Conf. on Autonomous Agents and Multi-Agent Systems (AAMAS 07) 745 Algorithm 2 Distributed Best-First Search Algorithm root ← electedleader next(root, ∅) place(node, parent) node.parent ← parent node.ancestors ← parent.ancestors ∪ parent send placement message (node, node.ancestors) to all neighbors of node next(current, previous) if current is not placed then place(current, previous) next(current, ∅) else best ← getBestNeighbor(current, previous) if best = ∅ then if previous = ∅ then terminate, all nodes are placed next(previous, ∅) else next(best, current) getBestNeighbor(current, previous) best ← ∅; score ← 0 for all n ∈ current.neighbors do if n! = previous then if n is placed then nscore ← getBestNeighbor(n, current) else nscore ← evaluate(current, n) if nscore > score then score ← nscore best ← n return best, score nodes, so it has O(V ) recursions. Each recursion performs a recursive getBestNeighbor operation that traverses all placed nodes and their neighbors. This operation is O(V · E), but results can be cached using only O(V ) space at each node. Thus we have O(V ·(V +V +V ·E)) = O(V 2 ·E). If we are smart about evaluating local changes when each node receives placement messages from its neighbors and cache the results the getBestNeighbor operation is only O(E). This increases the complexity of the place operation, but for all placements the total complexity is only O(V · E). Thus we have an overall complexity of O(V ·E+V ·(V +E)) = O(V ·E). 6. COMPARISON OF COMPLEXITY IN DPOP AND DCPOP We have already shown that given the same input, DCPOP performs the same as DPOP. We also have shown that we can accurately predict performance of a given pseudotree in linear spacetime complexity. If we use a constant number of heuristics to generate the set of pseudotrees, we can choose the best pseudotree in linear space-time complexity. We will now show that there exists a DCOP instance for which a cross-edged pseudotree outperforms all possible traditional pseudotrees (based on edge-traversal heuristics). In Figure 3(a) we have a DCOP instance with six nodes. This is a bipartite graph with each partition fully connected to the other (a) (b) (c) Figure 3: (a) The DCOP instance (b) A traditional pseudotree arrangement for the DCOP instance (c) A cross-edged pseudotree arrangement for the DCOP instance partition. In Figure 3(b) we see a traditional pseudotree arrangement for this DCOP instance. It is easy to see that any edgetraversal based heuristic cannot expand two nodes from the same partition in succession. We also see that no node can have more than one child because any such arrangement would be an invalid pseudotree. Thus any traditional pseudotree arrangement for this DCOP instance must take the form of Figure 3(b). We can see that the back-edges F-B and F-A overlap node C. Node C also has a parent E, and a back-edge with D. Using the original DPOP algorithm (or DCPOP since they are identical in this case), we find that the computation at node C involves five domains: A, B, C, D, and E. In contrast, the cross-edged pseudotree arrangement in Figure 3(c) requires only a maximum of four domains in any computation during DCPOP. Since node A is the merge point for branches from both B and C, we can see that each of the nodes D, E, and F have two overlapping branches. In addition each of these nodes has node A as its parent. Using the DCPOP algorithm we find that the computation at node D (or E or F) involves four domains: A, B, C, and D (or E or F). Since no better traditional pseudotree arrangement can be created using an edge-traversal heuristic, we have shown that DCPOP can outperform DPOP even if we use the optimal pseudotree found through edge-traversal. We acknowledge that pseudotree arrangements that allow parent-child relationships without an actual constraint can solve the problem in Figure 3(a) with maximum computation size of four domains. However, current heuristics used with DPOP do not produce such pseudotrees, and such a heuristic would be difficult to distribute since each node would require information about nodes with which it has no constraint. Also, while we do not prove it here, cross-edged pseudotrees can produce smaller message sizes than such pseudotrees even if the computation size is similar. In practice, since finding the best pseudotree arrangement is NP-Hard, we find that heuristics that produce cross-edged pseudotrees often produce significantly smaller computation and message sizes. 7. EXPERIMENTAL RESULTS 746 The Sixth Intl. Joint Conf. on Autonomous Agents and Multi-Agent Systems (AAMAS 07) Existing performance metrics for DCOP algorithms include the total number of messages, synchronous clock cycles, and message size. We have already shown that the total number of messages is linear with respect to the number of constraints in the DCOP instance. We also introduced the maximum sequential path cost (PC) as a measurement of the maximum amount of parallelism achievable by the algorithm. The maximum sequential path cost is equal to the sum of the computations performed on the longest path from the root to any leaf node. We also include as metrics the maximum computation size in number of dimensions (CD) and maximum message size in number of dimensions (MD). To analyze the relative complexity of a given DCOP instance, we find the minimum induced width (IW) of any traditional pseudotree produced by a heuristic for the original DPOP. 7.1 Generic DCOP instances For our initial tests we randomly generated two sets of problems with 3000 cases in each. Each problem was generated by assigning a random number (picked from a range) of constraints to each variable. The generator then created binary constraints until each variable reached its maximum number of constraints. The first set uses 20 variables, and the best DPOP IW ranges from 1 to 16 with an average of 8.5. The second set uses 100 variables, and the best DPOP IW ranged from 2 to 68 with an average of 39.3. Since most of the problems in the second set were too complex to actually compute the solution, we took measurements of the metrics using the techniques described earlier in Section 5 without actually solving the problem. Results are shown for the first set in Table 1 and for the second set in Table 2. For the two problem sets we split the cases into low density and high density categories. Low density cases consist of those problems that have a best DPOP IW less than or equal to half of the total number of nodes (e.g. IW ≤ 10 for the 20 node problems and IW ≤ 50 for the 100 node problems). High density problems consist of the remainder of the problem sets. In both Table 1 and Table 2 we have listed performance metrics for the original DPOP algorithm, the DCPOP algorithm using only cross-edged pseudotrees (DCPOP-CE), and the DCPOP algorithm using traditional and cross-edged pseudotrees (DCPOP-All). The pseudotrees used for DPOP were generated using 5 heuristics: DFS, DFS MCN, DFS CLIQUE MCN, DFS MCN DSTB, and DFS MCN BEC. These are all versions of the guided DFS traversal discussed in Section 5. The cross-edged pseudotrees used for DCPOP-CE were generated using 5 heuristics: MCN, LCN, MCN A-B, LCN A-B, and LCSG A-B. These are all versions of the best-first traversal discussed in Section 5. For both DPOP and DCPOP-CE we chose the best pseudotree produced by their respective 5 heuristics for each problem in the set. For DCPOP-All we chose the best pseudotree produced by all 10 heuristics for each problem in the set. For the CD and MD metrics the value shown is the average number of dimensions. For the PC metric the value shown is the natural logarithm of the maximum sequential path cost (since the actual value grows exponentially with the complexity of the problem). The final row in both tables is a measurement of improvement of DCPOP-All over DPOP. For the CD and MD metrics the value shown is a reduction in number of dimensions. For the PC metric the value shown is a percentage reduction in the maximum sequential path cost (% = DP OP −DCP OP DCP OP ∗ 100). Notice that DCPOPAll outperforms DPOP on all metrics. This logically follows from our earlier assertion that given the same input, DCPOP performs exactly the same as DPOP. Thus given the choice between the pseudotrees produced by all 10 heuristics, DCPOP-All will always outLow Density High Density Algorithm CD MD PC CD MD PC DPOP 7.81 6.81 3.78 13.34 12.34 5.34 DCPOP-CE 7.94 6.73 3.74 12.83 11.43 5.07 DCPOP-All 7.62 6.49 3.66 12.72 11.36 5.05 Improvement 0.18 0.32 13% 0.62 0.98 36% Table 1: 20 node problems Low Density High Density Algorithm CD MD PC CD MD PC DPOP 33.35 32.35 14.55 58.51 57.50 19.90 DCPOP-CE 33.49 29.17 15.22 57.11 50.03 20.01 DCPOP-All 32.35 29.57 14.10 56.33 51.17 18.84 Improvement 1.00 2.78 104% 2.18 6.33 256% Table 2: 100 node problems Figure 4: Computation Dimension Size Figure 5: Message Dimension Size The Sixth Intl. Joint Conf. on Autonomous Agents and Multi-Agent Systems (AAMAS 07) 747 Figure 6: Path Cost DCPOP Improvement Ag Mtg Vars Const IW CD MD PC 10 4 12 13.5 2.25 -0.01 -0.01 5.6% 30 14 44 57.6 3.63 0.09 0.09 10.9% 50 24 76 101.3 4.17 0.08 0.09 10.7% 100 49 156 212.9 5.04 0.16 0.20 30.0% 150 74 236 321.8 5.32 0.21 0.23 35.8% 200 99 316 434.2 5.66 0.18 0.22 29.5% Table 3: Meeting Scheduling Problems perform DPOP. Another trend we notice is that the improvement is greater for high density problems than low density problems. We show this trend in greater detail in Figures 4, 5, and 6. Notice how the improvement increases as the complexity of the problem increases. 7.2 Meeting Scheduling Problem In addition to our initial generic DCOP tests, we ran a series of tests on the Meeting Scheduling Problem (MSP) as described in [6]. The problem setup includes a number of people that are grouped into departments. Each person must attend a specified number of meetings. Meetings can be held within departments or among departments, and can be assigned to one of eight time slots. The MSP maps to a DCOP instance where each variable represents the time slot that a specific person will attend a specific meeting. All variables that belong to the same person have mutual exclusion constraints placed so that the person cannot attend more than one meeting during the same time slot. All variables that belong to the same meeting have equality constraints so that all of the participants choose the same time slot. Unary constraints are placed on each variable to account for a person"s valuation of each meeting and time slot. For our tests we generated 100 sample problems for each combination of agents and meetings. Results are shown in Table 3. The values in the first five columns represent (in left to right order), the total number of agents, the total number of meetings, the total number of variables, the average total number of constraints, and the average minimum IW produced by a traditional pseudotree. The last three columns show the same metrics we used for the generic DCOP instances, except this time we only show the improvements of DCPOP-All over DPOP. Performance is better on average for all MSP instances, but again we see larger improvements for more complex problem instances. 8. CONCLUSIONS AND FUTURE WORK We presented a complete, distributed algorithm that solves general DCOP instances using cross-edged pseudotree arrangements. Our algorithm extends the DPOP algorithm by adding additional utility propagation messages, and introducing the concept of branch merging during the utility propagation phase. Our algorithm also allows value assignments to occur at higher level merge points for lower level nodes. We have shown that DCPOP fully extends DPOP by performing the same operations given the same input. We have also shown through some examples and experimental data that DCPOP can achieve greater performance for some problem instances by extending the allowable input set to include cross-edged pseudotrees. We placed particular emphasis on the role that edge-traversal heuristics play in the generation of pseudotrees. We have shown that the performance penalty is minimal to generate multiple heuristics, and that we can choose the best generated pseudotree in linear space-time complexity. Given the importance of a good pseudotree for performance, future work will include new heuristics to find better pseudotrees. Future work will also include adapting existing DPOP extensions [5, 7] that support different problem domains for use with DCPOP. 9. REFERENCES [1] J. Liu and K. P. Sycara. Exploiting problem structure for distributed constraint optimization. In V. Lesser, editor, Proceedings of the First International Conference on Multi-Agent Systems, pages 246-254, San Francisco, CA, 1995. MIT Press. [2] P. J. Modi, H. Jung, M. Tambe, W.-M. Shen, and S. Kulkarni. A dynamic distributed constraint satisfaction approach to resource allocation. Lecture Notes in Computer Science, 2239:685-700, 2001. [3] P. J. Modi, W. Shen, M. Tambe, and M. Yokoo. An asynchronous complete method for distributed constraint optimization. In AAMAS 03, 2003. [4] A. Petcu. Frodo: A framework for open/distributed constraint optimization. Technical Report No. 2006/001 2006/001, Swiss Federal Institute of Technology (EPFL), Lausanne (Switzerland), 2006. http://liawww.epfl.ch/frodo/. [5] A. Petcu and B. Faltings. A-dpop: Approximations in distributed optimization. In poster in CP 2005, pages 802-806, Sitges, Spain, October 2005. [6] A. Petcu and B. Faltings. Dpop: A scalable method for multiagent constraint optimization. In IJCAI 05, pages 266-271, Edinburgh, Scotland, Aug 2005. [7] A. Petcu, B. Faltings, and D. Parkes. M-dpop: Faithful distributed implementation of efficient social choice problems. In AAMAS 06, pages 1397-1404, Hakodate, Japan, May 2006. [8] G. Ushakov. Solving meeting scheduling problems using distributed pseudotree-optimization procedure. Master"s thesis, ´Ecole Polytechnique F´ed´erale de Lausanne, 2005. [9] M. Yokoo, E. H. Durfee, T. Ishida, and K. Kuwabara. Distributed constraint satisfaction for formalizing distributed problem solving. In International Conference on Distributed Computing Systems, pages 614-621, 1992. [10] M. Yokoo, E. H. Durfee, T. Ishida, and K. Kuwabara. The distributed constraint satisfaction problem: Formalization and algorithms. Knowledge and Data Engineering, 10(5):673-685, 1998. 748 The Sixth Intl. Joint Conf. on Autonomous Agents and Multi-Agent Systems (AAMAS 07)

cross-edged pseudotree;teamwork coordination;distributed constraint optimization;job shop scheduling;agent;pseudotree arrangement;multi-agent coordination;edge-traversal heuristic;maximum sequential path cost;multi-valued utility function;distribute constraint satisfaction and optimization;multi-agent system;resource allocation;global utility

train_I-43

Dynamics Based Control with an Application to Area-Sweeping Problems

In this paper we introduce Dynamics Based Control (DBC), an approach to planning and control of an agent in stochastic environments. Unlike existing approaches, which seek to optimize expected rewards (e.g., in Partially Observable Markov Decision Problems (POMDPs)), DBC optimizes system behavior towards specified system dynamics. We show that a recently developed planning and control approach, Extended Markov Tracking (EMT) is an instantiation of DBC. EMT employs greedy action selection to provide an efficient control algorithm in Markovian environments. We exploit this efficiency in a set of experiments that applied multitarget EMT to a class of area-sweeping problems (searching for moving targets). We show that such problems can be naturally defined and efficiently solved using the DBC framework, and its EMT instantiation.

1. INTRODUCTION Planning and control constitutes a central research area in multiagent systems and artificial intelligence. In recent years, Partially Observable Markov Decision Processes (POMDPs) [12] have become a popular formal basis for planning in stochastic environments. In this framework, the planning and control problem is often addressed by imposing a reward function, and computing a policy (of choosing actions) that is optimal, in the sense that it will result in the highest expected utility. While theoretically attractive, the complexity of optimally solving a POMDP is prohibitive [8, 7]. We take an alternative view of planning in stochastic environments. We do not use a (state-based) reward function, but instead optimize over a different criterion, a transition-based specification of the desired system dynamics. The idea here is to view planexecution as a process that compels a (stochastic) system to change, and a plan as a dynamic process that shapes that change according to desired criteria. We call this general planning framework Dynamics Based Control (DBC). In DBC, the goal of a planning (or control) process becomes to ensure that the system will change in accordance with specific (potentially stochastic) target dynamics. As actual system behavior may deviate from that which is specified by target dynamics (due to the stochastic nature of the system), planning in such environments needs to be continual [4], in a manner similar to classical closed-loop controllers [16]. Here, optimality is measured in terms of probability of deviation magnitudes. In this paper, we present the structure of Dynamics Based Control. We show that the recently developed Extended Markov Tracking (EMT) approach [13, 14, 15] is subsumed by DBC, with EMT employing greedy action selection, which is a specific parameterization among the options possible within DBC. EMT is an efficient instantiation of DBC. To evaluate DBC, we carried out a set of experiments applying multi-target EMT to the Tag Game [11]; this is a variant on the area sweeping problem, where an agent is trying to tag a moving target (quarry) whose position is not known with certainty. Experimental data demonstrates that even with a simple model of the environment and a simple design of target dynamics, high success rates can be produced both in catching the quarry, and in surprising the quarry (as expressed by the observed entropy of the controlled agent"s position). The paper is organized as follows. In Section 2 we motivate DBC using area-sweeping problems, and discuss related work. Section 3 introduces the Dynamics Based Control (DBC) structure, and its specialization to Markovian environments. This is followed by a review of the Extended Markov Tracking (EMT) approach as a DBC-structured control regimen in Section 4. That section also discusses the limitations of EMT-based control relative to the general DBC framework. Experimental settings and results are then presented in Section 5. Section 6 provides a short discussion of the overall approach, and Section 7 gives some concluding remarks and directions for future work. 790 978-81-904262-7-5 (RPS) c 2007 IFAAMAS 2. MOTIVATION AND RELATED WORK Many real-life scenarios naturally have a stochastic target dynamics specification, especially those domains where there exists no ultimate goal, but rather system behavior (with specific properties) that has to be continually supported. For example, security guards perform persistent sweeps of an area to detect any sign of intrusion. Cunning thieves will attempt to track these sweeps, and time their operation to key points of the guards" motion. It is thus advisable to make the guards" motion dynamics appear irregular and random. Recent work by Paruchuri et al. [10] has addressed such randomization in the context of single-agent and distributed POMDPs. The goal in that work was to generate policies that provide a measure of action-selection randomization, while maintaining rewards within some acceptable levels. Our focus differs from this work in that DBC does not optimize expected rewards-indeed we do not consider rewards at all-but instead maintains desired dynamics (including, but not limited to, randomization). The Game of Tag is another example of the applicability of the approach. It was introduced in the work by Pineau et al. [11]. There are two agents that can move about an area, which is divided into a grid. The grid may have blocked cells (holes) into which no agent can move. One agent (the hunter) seeks to move into a cell occupied by the other (the quarry), such that they are co-located (this is a successful tag). The quarry seeks to avoid the hunter agent, and is always aware of the hunter"s position, but does not know how the hunter will behave, which opens up the possibility for a hunter to surprise the prey. The hunter knows the quarry"s probabilistic law of motion, but does not know its current location. Tag is an instance of a family of area-sweeping (pursuit-evasion) problems. In [11], the hunter modeled the problem using a POMDP. A reward function was defined, to reflect the desire to tag the quarry, and an action policy was computed to optimize the reward collected over time. Due to the intractable complexity of determining the optimal policy, the action policy computed in that paper was essentially an approximation. In this paper, instead of formulating a reward function, we use EMT to solve the problem, by directly specifying the target dynamics. In fact, any search problem with randomized motion, the socalled class of area sweeping problems, can be described through specification of such target system dynamics. Dynamics Based Control provides a natural approach to solving these problems. 3. DYNAMICS BASED CONTROL The specification of Dynamics Based Control (DBC) can be broken into three interacting levels: Environment Design Level, User Level, and Agent Level. • Environment Design Level is concerned with the formal specification and modeling of the environment. For example, this level would specify the laws of physics within the system, and set its parameters, such as the gravitation constant. • User Level in turn relies on the environment model produced by Environment Design to specify the target system dynamics it wishes to observe. The User Level also specifies the estimation or learning procedure for system dynamics, and the measure of deviation. In the museum guard scenario above, these would correspond to a stochastic sweep schedule, and a measure of relative surprise between the specified and actual sweeping. • Agent Level in turn combines the environment model from the Environment Design level, the dynamics estimation procedure, the deviation measure and the target dynamics specification from User Level, to produce a sequence of actions that create system dynamics as close as possible to the targeted specification. As we are interested in the continual development of a stochastic system, such as happens in classical control theory [16] and continual planning [4], as well as in our example of museum sweeps, the question becomes how the Agent Level is to treat the deviation measurements over time. To this end, we use a probability threshold-that is, we would like the Agent Level to maximize the probability that the deviation measure will remain below a certain threshold. Specific action selection then depends on system formalization. One possibility would be to create a mixture of available system trends, much like that which happens in Behavior-Based Robotic architectures [1]. The other alternative would be to rely on the estimation procedure provided by the User Level-to utilize the Environment Design Level model of the environment to choose actions, so as to manipulate the dynamics estimator into believing that a certain dynamics has been achieved. Notice that this manipulation is not direct, but via the environment. Thus, for strong enough estimator algorithms, successful manipulation would mean a successful simulation of the specified target dynamics (i.e., beyond discerning via the available sensory input). DBC levels can also have a back-flow of information (see Figure 1). For instance, the Agent Level could provide data about target dynamics feasibility, allowing the User Level to modify the requirement, perhaps focusing on attainable features of system behavior. Data would also be available about the system response to different actions performed; combined with a dynamics estimator defined by the User Level, this can provide an important tool for the environment model calibration at the Environment Design Level. UserEnv. Design Agent Model Ideal Dynamics Estimator Estimator Dynamics Feasibility System Response Data Figure 1: Data flow of the DBC framework Extending upon the idea of Actor-Critic algorithms [5], DBC data flow can provide a good basis for the design of a learning algorithm. For example, the User Level can operate as an exploratory device for a learning algorithm, inferring an ideal dynamics target from the environment model at hand that would expose and verify most critical features of system behavior. In this case, feasibility and system response data from the Agent Level would provide key information for an environment model update. In fact, the combination of feasibility and response data can provide a basis for the application of strong learning algorithms such as EM [2, 9]. 3.1 DBC for Markovian Environments For a Partially Observable Markovian Environment, DBC can be specified in a more rigorous manner. Notice how DBC discards rewards, and replaces it by another optimality criterion (structural differences are summarized in Table 1): • Environment Design level is to specify a tuple < S, A, T, O, Ω, s0 >, where: - S is the set of all possible environment states; - s0 is the initial state of the environment (which can also be viewed as a probability distribution over S); The Sixth Intl. Joint Conf. on Autonomous Agents and Multi-Agent Systems (AAMAS 07) 791 - A is the set of all possible actions applicable in the environment; - T is the environment"s probabilistic transition function: T : S ×A → Π(S). That is, T(s |a, s) is the probability that the environment will move from state s to state s under action a; - O is the set of all possible observations. This is what the sensor input would look like for an outside observer; - Ω is the observation probability function: Ω : S × A × S → Π(O). That is, Ω(o|s , a, s) is the probability that one will observe o given that the environment has moved from state s to state s under action a. • User Level, in the case of Markovian environment, operates on the set of system dynamics described by a family of conditional probabilities F = {τ : S × A → Π(S)}. Thus specification of target dynamics can be expressed by q ∈ F, and the learning or tracking algorithm can be represented as a function L : O×(A×O)∗ → F; that is, it maps sequences of observations and actions performed so far into an estimate τ ∈ F of system dynamics. There are many possible variations available at the User Level to define divergence between system dynamics; several of them are: - Trace distance or L1 distance between two distributions p and q defined by D(p(·), q(·)) = 1 2 x |p(x) − q(x)| - Fidelity measure of distance F(p(·), q(·)) = x p(x)q(x) - Kullback-Leibler divergence DKL(p(·) q(·)) = x p(x) log p(x) q(x) Notice that the latter two are not actually metrics over the space of possible distributions, but nevertheless have meaningful and important interpretations. For instance, KullbackLeibler divergence is an important tool of information theory [3] that allows one to measure the price of encoding an information source governed by q, while assuming that it is governed by p. The User Level also defines the threshold of dynamics deviation probability θ. • Agent Level is then faced with a problem of selecting a control signal function a∗ to satisfy a minimization problem as follows: a∗ = arg min a Pr(d(τa, q) > θ) where d(τa, q) is a random variable describing deviation of the dynamics estimate τa, created by L under control signal a, from the ideal dynamics q. Implicit in this minimization problem is that L is manipulated via the environment, based on the environment model produced by the Environment Design Level. 3.2 DBC View of the State Space It is important to note the complementary view that DBC and POMDPs take on the state space of the environment. POMDPs regard state as a stationary snap-shot of the environment; whatever attributes of state sequencing one seeks are reached through properties of the control process, in this case reward accumulation. This can be viewed as if the sequencing of states and the attributes of that sequencing are only introduced by and for the controlling mechanism-the POMDP policy. DBC concentrates on the underlying principle of state sequencing, the system dynamics. DBC"s target dynamics specification can use the environment"s state space as a means to describe, discern, and preserve changes that occur within the system. As a result, DBC has a greater ability to express state sequencing properties, which are grounded in the environment model and its state space definition. For example, consider the task of moving through rough terrain towards a goal and reaching it as fast as possible. POMDPs would encode terrain as state space points, while speed would be ensured by negative reward for every step taken without reaching the goalaccumulating higher reward can be reached only by faster motion. Alternatively, the state space could directly include the notion of speed. For POMDPs, this would mean that the same concept is encoded twice, in some sense: directly in the state space, and indirectly within reward accumulation. Now, even if the reward function would encode more, and finer, details of the properties of motion, the POMDP solution will have to search in a much larger space of policies, while still being guided by the implicit concept of the reward accumulation procedure. On the other hand, the tactical target expression of variations and correlations between position and speed of motion is now grounded in the state space representation. In this situation, any further constraints, e.g., smoothness of motion, speed limits in different locations, or speed reductions during sharp turns, are explicitly and uniformly expressed by the tactical target, and can result in faster and more effective action selection by a DBC algorithm. 4. EMT-BASED CONTROL AS A DBC Recently, a control algorithm was introduced called EMT-based Control [13], which instantiates the DBC framework. Although it provides an approximate greedy solution in the DBC sense, initial experiments using EMT-based control have been encouraging [14, 15]. EMT-based control is based on the Markovian environment definition, as in the case with POMDPs, but its User and Agent Levels are of the Markovian DBC type of optimality. • User Level of EMT-based control defines a limited-case target system dynamics independent of action: qEMT : S → Π(S). It then utilizes the Kullback-Leibler divergence measure to compose a momentary system dynamics estimator-the Extended Markov Tracking (EMT) algorithm. The algorithm keeps a system dynamics estimate τt EMT that is capable of explaining recent change in an auxiliary Bayesian system state estimator from pt−1 to pt, and updates it conservatively using Kullback-Leibler divergence. Since τt EMT and pt−1,t are respectively the conditional and marginal probabilities over the system"s state space, explanation simply means that pt(s ) = s τt EMT (s |s)pt−1(s), and the dynamics estimate update is performed by solving a 792 The Sixth Intl. Joint Conf. on Autonomous Agents and Multi-Agent Systems (AAMAS 07) Table 1: Structure of POMDP vs. Dynamics-Based Control in Markovian Environment Level Approach MDP Markovian DBC Environment < S, A, T, O, Ω >,where S - set of states A - set of actions Design T : S × A → Π(S) - transition O - observation set Ω : S × A × S → Π(O) User r : S × A × S → R q : S × A → Π(S) F(π∗ ) → r L(o1, ..., ot) → τ r - reward function q - ideal dynamics F - reward remodeling L - dynamics estimator θ - threshold Agent π∗ = arg max π E[ γt rt] π∗ = arg min π Prob(d(τ q) > θ) minimization problem: τt EMT = H[pt, pt−1, τt−1 EMT ] = arg min τ DKL(τ × pt−1 τt−1 EMT × pt−1) s.t. pt(s ) = s (τ × pt−1)(s , s) pt−1(s) = s (τ × pt−1)(s , s) • Agent Level in EMT-based control is suboptimal with respect to DBC (though it remains within the DBC framework), performing greedy action selection based on prediction of EMT"s reaction. The prediction is based on the environment model provided by the Environment Design level, so that if we denote by Ta the environment"s transition function limited to action a, and pt−1 is the auxiliary Bayesian system state estimator, then the EMT-based control choice is described by a∗ = arg min a∈A DKL(H[Ta × pt, pt, τt EMT ] qEMT × pt−1) Note that this follows the Markovian DBC framework precisely: the rewarding optimality of POMDPs is discarded, and in its place a dynamics estimator (EMT in this case) is manipulated via action effects on the environment to produce an estimate close to the specified target system dynamics. Yet as we mentioned, naive EMTbased control is suboptimal in the DBC sense, and has several additional limitations that do not exist in the general DBC framework (discussed in Section 4.2). 4.1 Multi-Target EMT At times, there may exist several behavioral preferences. For example, in the case of museum guards, some art items are more heavily guarded, requiring that the guards stay more often in their close vicinity. On the other hand, no corner of the museum is to be left unchecked, which demands constant motion. Successful museum security would demand that the guards adhere to, and balance, both of these behaviors. For EMT-based control, this would mean facing several tactical targets {qk}K k=1, and the question becomes how to merge and balance them. A balancing mechanism can be applied to resolve this issue. Note that EMT-based control, while selecting an action, creates a preference vector over the set of actions based on their predicted performance with respect to a given target. If these preference vectors are normalized, they can be combined into a single unified preference. This requires replacement of standard EMT-based action selection by the algorithm below [15]: • Given: - a set of target dynamics {qk}K k=1, - vector of weights w(k) • Select action as follows - For each action a ∈ A predict the future state distribution ¯pa t+1 = Ta ∗ pt; - For each action, compute Da = H(¯pa t+1, pt, PDt) - For each a ∈ A and qk tactical target, denote V (a, k) = DKL (Da qk) pt . Let Vk(a) = 1 Zk V (a, k), where Zk = a∈A V (a, k) is a normalization factor. - Select a∗ = arg min a k k=1 w(k)Vk(a) The weights vector w = (w1, ..., wK ) allows the additional tuning of importance among target dynamics without the need to redesign the targets themselves. This balancing method is also seamlessly integrated into the EMT-based control flow of operation. 4.2 EMT-based Control Limitations EMT-based control is a sub-optimal (in the DBC sense) representative of the DBC structure. It limits the User by forcing EMT to be its dynamic tracking algorithm, and replaces Agent optimization by greedy action selection. This kind of combination, however, is common for on-line algorithms. Although further development of EMT-based controllers is necessary, evidence so far suggests that even the simplest form of the algorithm possesses a great deal of power, and displays trends that are optimal in the DBC sense of the word. There are two further, EMT-specific, limitations to EMT-based control that are evident at this point. Both already have partial solutions and are subjects of ongoing research. The first limitation is the problem of negative preference. In the POMDP framework for example, this is captured simply, through The Sixth Intl. Joint Conf. on Autonomous Agents and Multi-Agent Systems (AAMAS 07) 793 the appearance of values with different signs within the reward structure. For EMT-based control, however, negative preference means that one would like to avoid a certain distribution over system development sequences; EMT-based control, however, concentrates on getting as close as possible to a distribution. Avoidance is thus unnatural in native EMT-based control. The second limitation comes from the fact that standard environment modeling can create pure sensory actions-actions that do not change the state of the world, and differ only in the way observations are received and the quality of observations received. Since the world state does not change, EMT-based control would not be able to differentiate between different sensory actions. Notice that both of these limitations of EMT-based control are absent from the general DBC framework, since it may have a tracking algorithm capable of considering pure sensory actions and, unlike Kullback-Leibler divergence, a distribution deviation measure that is capable of dealing with negative preference. 5. EMT PLAYING TAG The Game of Tag was first introduced in [11]. It is a single agent problem of capturing a quarry, and belongs to the class of area sweeping problems. An example domain is shown in Figure 2. 0 51 2 3 4 6 7 8 10 12 13 161514 17 18 19 2221 23 9 11Q A 20 Figure 2: Tag domain; an agent (A) attempts to seek and capture a quarry (Q) The Game of Tag extremely limits the agent"s perception, so that the agent is able to detect the quarry only if they are co-located in the same cell of the grid world. In the classical version of the game, co-location leads to a special observation, and the ‘Tag" action can be performed. We slightly modify this setting: the moment that both agents occupy the same cell, the game ends. As a result, both the agent and its quarry have the same motion capability, which allows them to move in four directions, North, South, East, and West. These form a formal space of actions within a Markovian environment. The state space of the formal Markovian environment is described by the cross-product of the agent and quarry"s positions. For Figure 2, it would be S = {s0, ..., s23} × {s0, ..., s23}. The effects of an action taken by the agent are deterministic, but the environment in general has a stochastic response due to the motion of the quarry. With probability q0 1 it stays put, and with probability 1 − q0 it moves to an adjacent cell further away from the 1 In our experiments this was taken to be q0 = 0.2. agent. So for the instance shown in Figure 2 and q0 = 0.1: P(Q = s9|Q = s9, A = s11) = 0.1 P(Q = s2|Q = s9, A = s11) = 0.3 P(Q = s8|Q = s9, A = s11) = 0.3 P(Q = s14|Q = s9, A = s11) = 0.3 Although the evasive behavior of the quarry is known to the agent, the quarry"s position is not. The only sensory information available to the agent is its own location. We use EMT and directly specify the target dynamics. For the Game of Tag, we can easily formulate three major trends: catching the quarry, staying mobile, and stalking the quarry. This results in the following three target dynamics: Tcatch(At+1 = si|Qt = sj, At = sa) ∝ 1 si = sj 0 otherwise Tmobile(At+1 = si|Qt = so, At = sj) ∝ 0 si = sj 1 otherwise Tstalk(At+1 = si|Qt = so, At = sj) ∝ 1 dist(si,so) Note that none of the above targets are directly achievable; for instance, if Qt = s9 and At = s11, there is no action that can move the agent to At+1 = s9 as required by the Tcatch target dynamics. We ran several experiments to evaluate EMT performance in the Tag Game. Three configurations of the domain shown in Figure 3 were used, each posing a different challenge to the agent due to partial observability. In each setting, a set of 1000 runs was performed with a time limit of 100 steps. In every run, the initial position of both the agent and its quarry was selected at random; this means that as far as the agent was concerned, the quarry"s initial position was uniformly distributed over the entire domain cell space. We also used two variations of the environment observability function. In the first version, observability function mapped all joint positions of hunter and quarry into the position of the hunter as an observation. In the second, only those joint positions in which hunter and quarry occupied different locations were mapped into the hunter"s location. The second version in fact utilized and expressed the fact that once hunter and quarry occupy the same cell the game ends. The results of these experiments are shown in Table 2. Balancing [15] the catch, move, and stalk target dynamics described in the previous section by the weight vector [0.8, 0.1, 0.1], EMT produced stable performance in all three domains. Although direct comparisons are difficult to make, the EMT performance displayed notable efficiency vis-`a-vis the POMDP approach. In spite of a simple and inefficient Matlab implementation of the EMT algorithm, the decision time for any given step averaged significantly below 1 second in all experiments. For the irregular open arena domain, which proved to be the most difficult, 1000 experiment runs bounded by 100 steps each, a total of 42411 steps, were completed in slightly under 6 hours. That is, over 4 × 104 online steps took an order of magnitude less time than the offline computation of POMDP policy in [11]. The significance of this differential is made even more prominent by the fact that, should the environment model parameters change, the online nature of EMT would allow it to maintain its performance time, while the POMDP policy would need to be recomputed, requiring yet again a large overhead of computation. We also tested the behavior cell frequency entropy, empirical measures from trial data. As Figure 4 and Figure 5 show, empir794 The Sixth Intl. Joint Conf. on Autonomous Agents and Multi-Agent Systems (AAMAS 07) A Q Q A 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 A Q Figure 3: These configurations of the Tag Game space were used: a) multiple dead-end, b) irregular open arena, c) circular corridor Table 2: Performance of the EMT-based solution in three Tag Game domains and two observability models: I) omniposition quarry, II) quarry is not at hunter"s position Model Domain Capture% E(Steps) Time/Step I Dead-ends 100 14.8 72(mSec) Arena 80.2 42.4 500(mSec) Circle 91.4 34.6 187(mSec) II Dead-ends 100 13.2 91(mSec) Arena 96.8 28.67 396(mSec) Circle 94.4 31.63 204(mSec) ical entropy grows with the length of interaction. For runs where the quarry was not captured immediately, the entropy reaches between 0.85 and 0.952 for different runs and scenarios. As the agent actively seeks the quarry, the entropy never reaches its maximum. One characteristic of the entropy graph for the open arena scenario particularly caught our attention in the case of the omniposition quarry observation model. Near the maximum limit of trial length (100 steps), entropy suddenly dropped. Further analysis of the data showed that under certain circumstances, a fluctuating behavior occurs in which the agent faces equally viable versions of quarry-following behavior. Since the EMT algorithm has greedy action selection, and the state space does not encode any form of commitment (not even speed or acceleration), the agent is locked within a small portion of cells. It is essentially attempting to simultaneously follow several courses of action, all of which are consistent with the target dynamics. This behavior did not occur in our second observation model, since it significantly reduced the set of eligible courses of action-essentially contributing to tie-breaking among them. 6. DISCUSSION The design of the EMT solution for the Tag Game exposes the core difference in approach to planning and control between EMT or DBC, on the one hand, and the more familiar POMDP approach, on the other. POMDP defines a reward structure to optimize, and influences system dynamics indirectly through that optimization. EMT discards any reward scheme, and instead measures and influences system dynamics directly. 2 Entropy was calculated using log base equal to the number of possible locations within the domain; this properly scales entropy expression into the range [0, 1] for all domains. Thus for the Tag Game, we did not search for a reward function that would encode and express our preference over the agent"s behavior, but rather directly set three (heuristic) behavior preferences as the basis for target dynamics to be maintained. Experimental data shows that these targets need not be directly achievable via the agent"s actions. However, the ratio between EMT performance and achievability of target dynamics remains to be explored. The tag game experiment data also revealed the different emphasis DBC and POMDPs place on the formulation of an environment state space. POMDPs rely entirely on the mechanism of reward accumulation maximization, i.e., formation of the action selection procedure to achieve necessary state sequencing. DBC, on the other hand, has two sources of sequencing specification: through the properties of an action selection procedure, and through direct specification within the target dynamics. The importance of the second source was underlined by the Tag Game experiment data, in which the greedy EMT algorithm, applied to a POMDP-type state space specification, failed, since target description over such a state space was incapable of encoding the necessary behavior tendencies, e.g., tie-breaking and commitment to directed motion. The structural differences between DBC (and EMT in particular), and POMDPs, prohibits direct performance comparison, and places them on complementary tracks, each within a suitable niche. For instance, POMDPs could be seen as a much more natural formulation of economic sequential decision-making problems, while EMT is a better fit for continual demand for stochastic change, as happens in many robotic or embodied-agent problems. The complementary properties of POMDPs and EMT can be further exploited. There is recent interest in using POMDPs in hybrid solutions [17], in which the POMDPs can be used together with other control approaches to provide results not easily achievable with either approach by itself. DBC can be an effective partner in such a hybrid solution. For instance, POMDPs have prohibitively large off-line time requirements for policy computation, but can be readily used in simpler settings to expose beneficial behavioral trends; this can serve as a form of target dynamics that are provided to EMT in a larger domain for on-line operation. 7. CONCLUSIONS AND FUTURE WORK In this paper, we have presented a novel perspective on the process of planning and control in stochastic environments, in the form of the Dynamics Based Control (DBC) framework. DBC formulates the task of planning as support of a specified target system dynamics, which describes the necessary properties of change within the environment. Optimality of DBC plans of action are measured The Sixth Intl. Joint Conf. on Autonomous Agents and Multi-Agent Systems (AAMAS 07) 795 0 20 40 60 80 100 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 Steps Entropy Dead−ends 0 20 40 60 80 100 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 Steps Entropy Arena 0 20 40 60 80 100 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 Steps Entropy Circle Figure 4: Observation Model I: Omniposition quarry. Entropy development with length of Tag Game for the three experiment scenarios: a) multiple dead-end, b) irregular open arena, c) circular corridor. 0 10 20 30 40 50 60 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 Steps Entropy Dead−ends 0 20 40 60 80 100 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 Steps Entropy Arena 0 20 40 60 80 100 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 Steps Entropy Circle Figure 5: Observation Model II: quarry not observed at hunter"s position. Entropy development with length of Tag Game for the three experiment scenarios: a) multiple dead-end, b) irregular open arena, c) circular corridor. with respect to the deviation of actual system dynamics from the target dynamics. We show that a recently developed technique of Extended Markov Tracking (EMT) [13] is an instantiation of DBC. In fact, EMT can be seen as a specific case of DBC parameterization, which employs a greedy action selection procedure. Since EMT exhibits the key features of the general DBC framework, as well as polynomial time complexity, we used the multitarget version of EMT [15] to demonstrate that the class of area sweeping problems naturally lends itself to dynamics-based descriptions, as instantiated by our experiments in the Tag Game domain. As enumerated in Section 4.2, EMT has a number of limitations, such as difficulty in dealing with negative dynamic preference. This prevents direct application of EMT to such problems as Rendezvous-Evasion Games (e.g., [6]). However, DBC in general has no such limitations, and readily enables the formulation of evasion games. In future work, we intend to proceed with the development of dynamics-based controllers for these problems. 8. ACKNOWLEDGMENT The work of the first two authors was partially supported by Israel Science Foundation grant #898/05, and the third author was partially supported by a grant from Israel"s Ministry of Science and Technology. 9. REFERENCES [1] R. C. Arkin. Behavior-Based Robotics. MIT Press, 1998. [2] J. A. Bilmes. A gentle tutorial of the EM algorithm and its application to parameter estimation for Gaussian mixture and Hidden Markov Models. Technical Report TR-97-021, Department of Electrical Engeineering and Computer Science, University of California at Berkeley, 1998. [3] T. M. Cover and J. A. Thomas. Elements of information theory. Wiley, 1991. [4] M. E. desJardins, E. H. Durfee, C. L. Ortiz, and M. J. Wolverton. A survey of research in distributed, continual planning. AI Magazine, 4:13-22, 1999. [5] V. R. Konda and J. N. Tsitsiklis. Actor-Critic algorithms. SIAM Journal on Control and Optimization, 42(4):1143-1166, 2003. [6] W. S. Lim. A rendezvous-evasion game on discrete locations with joint randomization. Advances in Applied Probability, 29(4):1004-1017, December 1997. [7] M. L. Littman, T. L. Dean, and L. P. Kaelbling. On the complexity of solving Markov decision problems. In Proceedings of the 11th Annual Conference on Uncertainty in Artificial Intelligence (UAI-95), pages 394-402, 1995. [8] O. Madani, S. Hanks, and A. Condon. On the undecidability of probabilistic planning and related stochastic optimization problems. Artificial Intelligence Journal, 147(1-2):5-34, July 2003. [9] R. M. Neal and G. E. Hinton. A view of the EM algorithm 796 The Sixth Intl. Joint Conf. on Autonomous Agents and Multi-Agent Systems (AAMAS 07) that justifies incremental, sparse, and other variants. In M. I. Jordan, editor, Learning in Graphical Models, pages 355-368. Kluwer Academic Publishers, 1998. [10] P. Paruchuri, M. Tambe, F. Ordonez, and S. Kraus. Security in multiagent systems by policy randomization. In Proceeding of AAMAS 2006, 2006. [11] J. Pineau, G. Gordon, and S. Thrun. Point-based value iteration: An anytime algorithm for pomdps. In International Joint Conference on Artificial Intelligence (IJCAI), pages 1025-1032, August 2003. [12] M. L. Puterman. Markov Decision Processes. Wiley Series in Probability and Mathematical Statistics: Applied Probability and Statistics Section. Wiley-Interscience Publication, New York, 1994. [13] Z. Rabinovich and J. S. Rosenschein. Extended Markov Tracking with an application to control. In The Workshop on Agent Tracking: Modeling Other Agents from Observations, at the Third International Joint Conference on Autonomous Agents and Multiagent Systems, pages 95-100, New-York, July 2004. [14] Z. Rabinovich and J. S. Rosenschein. Multiagent coordination by Extended Markov Tracking. In The Fourth International Joint Conference on Autonomous Agents and Multiagent Systems, pages 431-438, Utrecht, The Netherlands, July 2005. [15] Z. Rabinovich and J. S. Rosenschein. On the response of EMT-based control to interacting targets and models. In The Fifth International Joint Conference on Autonomous Agents and Multiagent Systems, pages 465-470, Hakodate, Japan, May 2006. [16] R. F. Stengel. Optimal Control and Estimation. Dover Publications, 1994. [17] M. Tambe, E. Bowring, H. Jung, G. Kaminka, R. Maheswaran, J. Marecki, J. Modi, R. Nair, J. Pearce, P. Paruchuri, D. Pynadath, P. Scerri, N. Schurr, and P. Varakantham. Conflicts in teamwork: Hybrids to the The Sixth Intl. Joint Conf. on Autonomous Agents and Multi-Agent Systems (AAMAS 07) 797

dynamics base control;target dynamics;partially observable markov decision problem;reward function;tag game;stochastic environment;area-sweeping problem;environment design level;user level;extended markov tracking;dynamics based control;control;game of tag;multi-agent system;action-selection randomization;agent level;system dynamics;robotic

train_I-45

Implementing Commitment-Based Interactions∗

Although agent interaction plays a vital role in MAS, and messagecentric approaches to agent interaction have their drawbacks, present agent-oriented programming languages do not provide support for implementing agent interaction that is flexible and robust. Instead, messages are provided as a primitive building block. In this paper we consider one approach for modelling agent interactions: the commitment machines framework. This framework supports modelling interactions at a higher level (using social commitments), resulting in more flexible interactions. We investigate how commitmentbased interactions can be implemented in conventional agent-oriented programming languages. The contributions of this paper are: a mapping from a commitment machine to a collection of BDI-style plans; extensions to the semantics of BDI programming languages; and an examination of two issues that arise when distributing commitment machines (turn management and race conditions) and solutions to these problems.

1. INTRODUCTION Agents are social, and agent interaction plays a vital role in multiagent systems. Consequently, design and implementation of agent interaction is an important research topic. The standard approach for designing agent interactions is messagecentric: interactions are defined by interaction protocols that give the permissible sequences of messages, specified using notations such as finite state machines, Petri nets, or Agent UML. It has been argued that this message-centric approach to interaction design is not a good match for intelligent agents. Intelligent agents should exhibit the ability to persist in achieving their goals in the face of failure (robustness) by trying different approaches (flexibility). On the other hand, when following an interaction protocol, an agent has limited flexibility and robustness: the ability to persistently try alternative means to achieving the interaction"s aim is limited to those options that the protocol"s designer provided, and in practice, message-centric design processes do not tend to lead to protocols that are flexible or robust. Recognising these limitations of the traditional approach to designing agent interactions, a number of approaches have been proposed in recent years that move away from message-centric interaction protocols, and instead consider designing agent interactions using higher-level concepts such as social commitments [8, 10, 18] or interaction goals [2]. There has also been work on richer forms of interaction in specific settings, such as teams of cooperative agents [5, 11]. However, although there has been work on designing flexible and robust agent interactions, there has been virtually no work on providing programming language support for implementing such interactions. Current Agent Oriented Programming Languages (AOPLs) do not provide support for implementing flexible and robust agent interactions using higher-level concepts than messages. Indeed, modern AOPLs [1], with virtually no exceptions, provide only simple message sending as the basis for implementing agent interaction. This paper presents what, to the best of our knowledge, is the second AOPL to support high-level, flexible, and robust agent interaction implementation. The first such language, STAPLE, was proposed a few years ago [9], but is not described in detail, and is arguably impractical for use by non-specialists, due to its logical basis and heavy reliance on temporal and modal logic. This paper presents a scheme for extending BDI-like AOPLs to support direct implementation of agent interactions that are designed using Yolum & Singh"s commitment machine (CM) framework [19]. In the remainder of this paper we briefly review commitment machines and present a simple abstraction of BDI AOPLs which lies in the common subset of languages such as Jason, 3APL, and CAN. We then present a scheme for translating commitment machines to this language, and indicate how the language needs to be extended to support this. We then extend our scheme to address a range of issues concerned with distribution, including turn tracking [7], and race conditions. 2. BACKGROUND 2.1 Commitment Machines The aim of the commitment machine framework is to allow for the definition of interactions that are more flexible than traditional message-centric approaches. A Commitment Machine (CM) [19] specifies an interaction between entities (e.g. agents, services, processes) in terms of actions that change the interaction state. This interact state consists of fluents (predicates that change value over time), but also social commitments, both base-level and conditional. A base-level social commitment is an undertaking by debtor A to creditor B to bring about condition p, denoted C(A, B, p). This is sometimes abbreviated to C(p), where it is not important to specify the identities of the entities in question. For example, a commitment by customer C to merchant M to make the fluent paid true would be written as C(C, M, paid). A conditional social commitment is an undertaking by debtor A to creditor B that should condition q become true, A will then commit to bringing about condition p. This is denoted by CC(A, B, q, p), and, where the identity of the entities involved is unimportant (or obvious), is abbreviated to CC(q p) where the arrow is a reminder of the causal link between q becoming true and the creation of a commitment to make p true. For example, a commitment to make the fluent paid true once goods have been received would be written CC(goods paid). The semantics of commitments (both base-level and conditional) is defined with rules that specify how commitments change over time. For example, the commitment C(p) (or CC(q p)) is discharged when p becomes true; and the commitment CC(q p) is replaced by C(p) when q becomes true. In this paper we use the more symmetric semantics proposed by [15] and subsequently reformalised by [14]. In brief, these semantics deal with a number of more complex cases, such as where commitments are created when conditions already hold: if p holds when CC(p q) is meant to be created, then C(q) is created instead of CC(p q). An interaction is defined by specifying the entities involved, the possible contents of the interaction state (both fluents and commitments), and (most importantly) the actions that each entity can perform along with the preconditions and effects of each action, specified as add and delete lists. A commitment machine (CM) defines a range of possible interactions that each start in some state1 , and perform actions until reaching a final state. A final state is one that has no base-level commitments. One way of visualising the interactions that are possible with a given commitment machine is to generate the finite state machine corresponding to the CM. For example, figure 1 gives the FSM2 corresponding to the NetBill [18] commitment machine: a simple CM where a customer (C) and merchant (M) attempt to trade using the following actions3 : 1 Unlike standard interaction protocols, or finite state machines, there is no designated initial state for the interaction. 2 The finite state machine is software-generated: the nodes and connections were computed by an implementation of the axioms (available from http://www.winikoff.net/CM) and were then laid out by graphviz (http://www.graphviz.org/). 3 We use the notation A(X) : P ⇒ E to indicate that action A is performed by entity X, has precondition P (with : P omitted if empty) and effect E. • sendRequest(C) ⇒ request • sendQuote(M) ⇒ offer where offer ≡ promiseGoods ∧ promiseReceipt and promiseGoods ≡ CC(M, C, accept, goods) and promiseReceipt ≡ CC(M, C, pay, receipt) • sendAccept(C) ⇒ accept where accept ≡ CC(C, M, goods, pay) • sendGoods(M) ⇒ promiseReceipt ∧ goods where promiseReceipt ≡ CC(M, C, pay, receipt) • sendEPO(C) : goods ⇒ pay • sendReceipt(M) : pay ⇒ receipt. The commitment accept is the customer"s promise to pay once goods have been sent, promiseGoods is the merchant"s promise to send the goods once the customer accepts, and promiseReceipt is the merchant"s promise to send a receipt once payment has been made. As seen in figure 1, commitment machines can support a range of interaction sequences. 2.2 An Abstract Agent ProgrammingLanguage Agent programming languages in the BDI tradition (e.g. dMARS, JAM, PRS, UM-PRS, JACK, AgentSpeak(L), Jason, 3APL, CAN, Jadex) define agent behaviour in terms of event-triggered plans, where each plan specifies what it is triggered by, under what situations it can be considered to be applicable (defined using a so-called context condition), and a plan body: a sequence of steps that can include posting events which in turn triggers further plans. Given a collection of plans and an event e that has been posted the agent first collects all plans types that are triggered by that event (the relevant plans), then evaluates the context conditions of these plans to obtain a set of applicable plan instances. One of these is chosen and is executed. We now briefly define the formal syntax and semantics of a Simple Abstract (BDI) Agent Programming Language (SAAPL). This language is intended to be an abstraction that is in the common subset of such languages as Jason [1, Chapter 1], 3APL [1, Chapter 2], and CAN [16]. Thus, it is intentionally incomplete in some areas, for instance it doesn"t commit to a particular mechanism for dealing with plan failure, since different mechanisms are used by different AOPLs. An agent program (denoted by Π) consists of a collection of plan clauses of the form e : C ← P where e is an event, C is a context condition (a logical formula over the agent"s beliefs), and P is the plan body. The plan body is built up from the following constructs. We have the empty step which always succeeds and does nothing, operations to add (+b) and delete (−b) beliefs, sending a message m to agent N (↑N m), and posting an event4 (e). These can be sequenced (P; P). C ::= b | C ∧ C | C ∨ C | ¬C | ∃x.C P ::= | +b | −b | e | ↑N m | P; P Formal semantics for this language is given in figure 2. This semantics is based on the semantics for AgentSpeak given by [12], which in turn is based on the semantics for CAN [16]. The semantics is in the style of Plotkin"s Structural Operational Semantics, and assumes that operations exist that check whether a condition 4 We use ↓N m as short hand for the event corresponding to receiving message m from agent N. 874 The Sixth Intl. Joint Conf. on Autonomous Agents and Multi-Agent Systems (AAMAS 07) Figure 1: Finite State Machine for NetBill (shaded = final states) follows from a belief set, that add a belief to a belief set, and that delete a belief from a belief set. In the case of beliefs being a set of ground atoms these operations are respectively consequence checking (B |= C), and set addition (B ∪ {b}) and deletion (B \ {b}). More sophisticated belief management methods may be used, but are not considered here. We define a basic configuration S = Q, N, B, P where Q is a (global) message queue (modelled as a sequence5 where messages are added at one end and removed from the other end), N is the name of the agent, B is the beliefs of the agent and P is the plan body being executed (i.e. the intention). We also define an agent configuration, where instead of a single plan body P there is a set of plan instances, Γ. Finally, a complete MAS is a pair Q, As of a global message queue Q and a set of agent configurations (without the queue, Q). The global message queue is a sequence of triplets of the form sender:recipient:message. A transition S0 −→ S1 specifies that executing S0 a single step yields S1. We annotate the arrow with an indication of whether the configuration in question is basic, an agent configuration, or a MAS configuration. The transition relation is defined using rules of the form S −→ S or of the form S −→ Sr S −→ Sr ; the latter are conditional with the top (numerator) being the premise and the bottom (denominator) being the conclusion. Note that there is non-determinism in SAAPL, e.g. the choice of plan to execute from a set of applicable plans. This is resolved by using selection functions: SO selects one of the applicable plan instances to handle a given event, SI selects which of the plan instances that can be executed should be executed next, and SA selects which agent should execute (a step) next. 3. IMPLEMENTING COMMITMENT-BASED INTERACTIONS In this section we present a mapping from a commitment machine to a collection of SAAPL programs (one for each role). We begin by considering the simple case of two interacting agents, and 5 The + operator is used to denote sequence concatenation. assume that the agents take turns to act. In section 4 we relax these assumptions. Each action A(X) : P ⇒ E is mapped to a number of plans: there is a plan (for agent X) with context condition P that performs the action (i.e. applies the effects E to the agent"s beliefs) and sends a message to the other agent, and a plan (for the other agent) that updates its state when a message is received from X. For example, given the action sendAccept(C) ⇒ accept we have the following plans, where each plan is preceded by M: or C: to indicate which agent that plan belongs to. Note that where the identify of the sender (respectively recipient) is obvious, i.e. the other agent, we abbreviate ↑N m to ↑m (resp. ↓N m to ↓m). Turn taking is captured through the event ı (short for interact): the agent that is active has an ı event that is being handled. Handling the event involves sending a message to the other agent, and then doing nothing until a response is received. C: ı : true ← +accept; ↑sendAccept. M: ↓sendAccept : true ← +accept; ı. If the action has a non-trivial precondition then there are two plans in the recipient: one to perform the action (if possible), and another to report an error if the action"s precondition doesn"t hold (we return to this in section 4). For example, the action sendReceipt(M) : pay ⇒ receipt generates the following plans: M: ı : pay ← +receipt; ↑sendReceipt. C: ↓sendReceipt : pay ← +receipt; ı. C: ↓sendReceipt : ¬pay ← . . . report error . . . . In addition to these plans, we also need plans to start and finish the interaction. An interaction can be completed whenever there are no base-level commitments, so both agents have the following plans: ı : ¬∃p.C(p) ← ↑done. ↓done : ¬∃p.C(p) ← . ↓done : ∃p.C(p) ← . . . report error . . . . An interaction is started by setting up an agent"s initial beliefs, and then having it begin to interact. Exactly how to do this depends on the agent platform: e.g. the agent platform in question may offer a simple way to load beliefs from a file. A generic approach that is a little cumbersome, but is portable, is to send each of the agents involved in the interaction a sequence of init messages, each The Sixth Intl. Joint Conf. on Autonomous Agents and Multi-Agent Systems (AAMAS 07) 875 Q, N, B, +b Basic −→ Q, N, B ∪ {b}, Q, N, B, −b Basic −→ Q, N, B \ {b}, Δ = {Piθ|(ti : ci ← Pi) ∈ Π ∧ tiθ = e ∧ B |= ciθ} Q, N, B, e Basic −→ Q, N, B, SO(Δ) Q, N, B, P1 Basic −→ Q , N, B , P Q, N, B, P1; P2 Basic −→ Q , N, B , P ; P2 Q, N, B, ; P Basic −→ Q, N, B, P Q, N, B, ↑NB m Basic −→ Q + N:NB:m, N, B, Q = NA:N:m + Q Q, N, B, Γ Agent −→ Q , N, B, Γ ∪ {↓NA m} P = SI(Γ) Q, N, B, P Basic −→ Q , N, B , P Q, N, B, Γ Agent −→ Q , N, B , (Γ \ {P}) ∪ {P } P = SI(Γ) P = Q, N, B, Γ Agent −→ Q, N, B, (Γ \ {P}) N, B, Γ = SA(As) Q, N, B, Γ Agent −→ Q , N, B , Γ Q, As MAS −→ Q , (As ∪ { N, B , Γ }) \ { N, B, Γ } Figure 2: Operational Semantics for SAAPL containing a belief to be added; and then send one of the agents a start message which begins the interaction. Both agents thus have the following two plans: ↓init(B) : true ← +B. ↓start : true ← ı. Figure 3 gives the SAAPL programs for both merchant and customer that implement the NetBill protocol. For conciseness the error reporting plans are omitted. We now turn to refining the context conditions. There are three refinements that we consider. Firstly, we need to prevent performing actions that have no effect on the interaction state. Secondly, an agent may want to specify that certain actions that it is able to perform should not be performed unless additional conditions hold. For example, the customer may not want to agree to the merchant"s offer unless the goods have a certain price or property. Thirdly, the context conditions of the plans that terminate the interaction need to be refined in order to avoid terminating the interaction prematurely. For each plan of the form ı : P ← +E; ↑m we replace the context condition P with the enhanced condition P ∧ P ∧ ¬E where P is any additional conditions that the agent wishes to impose, and ¬E is the negation of the effects of the action. For example, the customer"s payment plan becomes (assuming no additional conditions, i.e. no P ): ı : goods ∧ ¬pay ← +pay; ↑sendEPO. For each plan of the form ↓m : P ← +E; ı we could add ¬E to the precondition, but this is redundant, since it is already checked by the performer of the action, and if the action has no effect then Customer"s plans: ı : true ← +request; ↑sendRequest. ı : true ← +accept; ↑sendAccept. ı : goods ← +pay; ↑sendEPO. ↓sendQuote : true ← +promiseGoods; +promiseReceipt; ı. ↓sendGoods : true ← +promiseReceipt; +goods; ı. ↓sendReceipt : pay ← +receipt; ı. Merchant"s plans: ı : true ← +promiseGoods; +promiseReceipt; ↑sendQuote. ı : true ← +promiseReceipt; +goods; ↑sendGoods. ı : pay ← +receipt; ↑sendReceipt. ↓sendRequest : true ← +request; ı. ↓sendAccept : true ← +accept; ı. ↓sendEPO : goods ← +pay; ı. Shared plans (i.e. plans of both agents): ı : ¬∃p.C(p) ← ↑done. ↓done : ¬∃p.C(p) ← . ↓init(B) : true ← +B. ↓start : true ← ı. Where accept ≡ CC(goods pay) promiseGoods ≡ CC(accept goods) promiseReceipt ≡ CC(pay receipt) offer ≡ promiseGoods ∧ promiseReceipt Figure 3: SAAPL Implementation of NetBill the sender won"t perform it and send the message (see also the discussion in section 4). When specifying additional conditions (P ), some care needs to be taken to avoid situations where progress cannot be made because the only action(s) possible are prevented by additional conditions. One way of indicating preference between actions (in many agent platforms) is to reorder the agent"s plans. This is clearly safe, since actions are not prevented, just considered in a different order. The third refinement of context conditions concerns the plans that terminate the interaction. In the Commitment Machine framework any state that has no base-level commitment is final, in that the interaction may end there (or it may continue). However, only some of these final states are desirable final states. Which final states are considered to be desirable depends on the domain and the desired interaction outcome. In the NetBill example, the desirable final state is one where the goods have been sent and paid for, and a receipt issued (i.e. goods ∧ pay ∧ receipt). In order to prevent an agent from terminating the interaction too early we add this as a precondition to the termination plan: ı : goods ∧ pay ∧ receipt ∧ ¬∃p.C(p) ← ↑done. Figure 4 shows the plans that are changed from figure 3. In order to support the realisation of CMs, we need to change SAAPL in a number of ways. These changes, which are discussed below, can be applied to existing BDI languages to make them commitment machine supportive. We present the three changes, explain what they involve, and for each change explain how the change was implemented using the 3APL agent oriented programming language. The three changes are: 1. extending the beliefs of the agent so that they can contain commitments; 876 The Sixth Intl. Joint Conf. on Autonomous Agents and Multi-Agent Systems (AAMAS 07) Customer"s plans: ı : ¬request ← +request; ↑sendRequest. ı : ¬accept ← +accept; ↑sendAccept. ı : goods ∧ ¬pay ← +pay; ↑sendEPO. Merchant"s plans: ı : ¬offer ← +promiseGoods; +promiseReceipt; ↑sendQuote. ı : ¬(promiseReceipt ∧ goods) ← +promiseReceipt; +goods; ↑sendGoods. ı : pay ∧ ¬receipt ← +receipt; ↑sendReceipt. Where accept ≡ CC(goods pay) promiseGoods ≡ CC(accept goods) promiseReceipt ≡ CC(pay receipt) offer ≡ promiseGoods ∧ promiseReceipt Figure 4: SAAPL Implementation of NetBill with refined context conditions (changed plans only) 2. changing the definition of |= to encompass implied commitments; and 3. whenever a belief is added, updating existing commitments, according to the rules of commitment dynamics. Extending the notion of beliefs to encompass commitments in fact requires no change in agent platforms that are prolog-like and support terms as beliefs (e.g. Jason, 3APL, CAN). However, other agent platforms do require an extension. For example, JACK, which is an extension of Java, would require changes to support commitments that can be nested. In the case of 3APL no change is needed to support this. Whenever a context condition contains commitments, determining whether the context condition is implied by the agent"s beliefs (B |= C) needs to take into account the notion of implied commitments [15]. In brief, a commitment can be considered to follow from a belief set B if the commitment is in the belief set (C ∈ B), but also under other conditions. For example, a commitment to pay C(pay) can be considered to be implied by a belief set containing pay because the commitment may have held and been discharged when pay was made true. Similar rules apply for conditional commitments. These rules, which were introduced in [15] were subsequently re-formalised in a simpler form by [14] resulting in the four inference rules in the bottom part of figure 5. The change that needs to be made to SAAPL to support commitment machine implementations is to extend the definition of |= to include these four rules. For 3APL this was realised by having each agent include the following Prolog clauses: holds(X) :- clause(X,true). holds(c(P)) :- holds(P). holds(c(P)) :- clause(cc(Q,P),true), holds(Q). holds(cc(_,Q)) :- holds(Q). holds(cc(_,Q)) :- holds(c(Q)). The first clause simply says that anything holds if it is in agent"s beliefs (clause(X,true) is true if X is a fact). The remaining four clauses correspond respectively to the inference rules C1, C2, CC1 and CC2. To use these rules we then modify context conditions in our program so that instead of writing, for example, cc(m,c, pay, receipt) we write holds(cc(m,c, pay, receipt)). B = norm(B ∪ {b}) Q, N, B, +b −→ Q, N, B , function norm(B) B ← B for each b ∈ B do if b = C(p) ∧ B |= p then B ← B \ {b} elseif b = CC(p q) then if B |= q then B ← B \ {b} elseif B |= p then B ← (B \ {b}) ∪ {C(q)} elseif B |= C(q) then B ← B \ {b} endif endif endfor return B end function B |= P B |= C(P) C1 CC(Q P) ∈ B B |= Q B |= P C2 B |= CC(P Q) B |= Q CC1 B |= C(Q) B |= CC(P Q) CC2 Figure 5: New Operational Semantics The final change is to update commitments when a belief is added. Formally, this is done by modifying the semantic rule for belief addition so that it applies an algorithm to update commitments. The modified rule and algorithm (which mirrors the definition of norm in [14]) can be found in the top part of figure 5. For 3APL this final change was achieved by manually inserting update() after updating beliefs, and defining the following rules for update(): update() <- c(P) AND holds(P) | {Deletec(P) ; update()}, update() <- cc(P,Q) AND holds(Q) | {Deletecc(P,Q) ; update()}, update() <- cc(P,Q) AND holds(P) | {Deletecc(P,Q) ; Addc(Q) ; update()}, update() <- cc(P,Q) AND holds(c(Q)) | {Deletecc(P,Q) ; update()}, update() <- true | Skip where Deletec and Deletecc delete respectively a base-level and conditional commitment, and Addc adds a base-level commitment. One aspect that doesn"t require a change is linking commitments and actions. This is because commitments don"t trigger actions directly: they may trigger actions indirectly, but in general their effect is to prevent completion of an interaction while there are outstanding (base level) commitments. Figure 6 shows the message sequences from a number of runs of a 3APL implementation of the NetBill commitment machine6 . In order to illustrate the different possible interactions the code was modified so that each agent selected randomly from the actions that it could perform, and a number of runs were made with the customer as the initiator, and then with the merchant as the initiator. There are other possible sequences of messages, not shown, 6 Source code is available from http://www.winikoff.net/CM The Sixth Intl. Joint Conf. on Autonomous Agents and Multi-Agent Systems (AAMAS 07) 877 Figure 6: Sample runs from 3APL implementation (alternating turns) including the obvious one: request, quote, accept, goods, payment, receipt, and then done. One minor difference between the 3APL implementation and SAAPL concerns the semantics of messages. In the semantics of SAAPL (and of most AOPLs), receiving a message is treated as an event. However, in 3APL, receiving a message is modelled as the addition to the agent"s beliefs of a fact indicating that the message was received [6]. Thus in the 3APL implementation we have PG rules that are triggered by these beliefs, rather than by any event. One issue with this approach is that the belief remains there, so we need to ensure that the belief in question is either deleted once handled, or that we modify preconditions of plans to avoid handling it more than once. In our implementation we delete these received beliefs when they are handled, to avoid duplicate handling of messages. 4. BEYOND TWO PARTICIPANTS Generalising to more than two interaction participants requires revisiting how turn management is done, since it is no longer possible to assume alternating turns [7]. In fact, perhaps surprisingly, even in the two participant setting, an alternating turn setup is an unreasonable assumption! For example, consider the path (in figure 1) from state 1 to 15 (sendGoods) then to state 12 (sendAccept). The result, in an alternating turn setup, is a dead-end: there is only a single possible action in state 12, namely sendEPO, but this action is done by the customer, and it is the merchant"s turn to act! Figure 7 shows the FSM for NetBill with alternating initiative. A solution to this problem that works in this example, but doesn"t generalise7 , is to weaken the alternating turn taking regime by allowing an agent to act twice in a row if its second action is driven by a commitment. A general solution is to track whose turn it is to act. This can be done by working out which agents have actions that are able to be performed in the current state. If there is only a single active agent, then it is clearly that agent"s turn to act. However, if more than one agent is active then somehow the agents need to work out who should act next. Working this out by negotiation is not a particularly good solution for two reasons. Firstly, this negotiation has to be done at every step of the interaction where more than one agent is active (in the NetBill, this applies to seven out of sixteen states), so it is highly desirable to have a light-weight mechanism for doing this. Secondly, it is not clear how the negotiation can avoid an infinite regress situation (you go first, no, you go first, . ..) without imposing some arbitrary rule. It is also possible to resolve who should act by imposing an arbitrary rule, for example, that the customer always acts in preference to the merchant, or that each agent has a numerical priority (perhaps determined by the order in which they joined the interaction?) that determines who acts. An alternative solution, which exploits the symmetrical properties of commitment machines, is to not try and manage turn taking. 7 Consider actions A1(C) ⇒ p, A2(C) ⇒ q, and A3(M) : p ∧ q ⇒ r. Figure 7: NetBill with alternating initiative Instead of tracking and controlling whose turn it is, we simply allow the agents to act freely, and rely on the properties of the interaction space to ensure that things work out, a notion that we shall make precise, and prove, in the remainder of this section. The issue with having multiple agents be active simultaneously is that instead of all agents agreeing on the current interaction state, agents can be in different states. This can be visualised as each agent having its own copy of the FSM that it navigates through where it is possible for agents to follow different paths through the FSM. The two specific issues that need to be addressed are: 1. Can agents end up in different final states? 2. Can an agent be in a position where an error occurs because it cannot perform an action corresponding to a received message? We will show that, because actions commute under certain assumptions, agents cannot end up in different final states, and furthermore, that errors cannot occur (again, under certain assumptions). By actions commute we mean that the state resulting from performing a sequence of actions A1 . . . An is the same, regardless of the order in which the actions are performed. This means that even if agents take different paths through the FSM, they still end up in the same resulting state, because once all messages have been processed, all agents will have performed the same set of actions. This addresses the issue of ending up in different final states. We return to the possibility of errors occurring shortly. Definition 1 (Monotonicity) An action is monotonic if it does not delete8 any fluents or commitments. A Commitment Machine is 8 That is directly deletes, it is fine to discharge commitments by adding fluents/commitments. 878 The Sixth Intl. Joint Conf. on Autonomous Agents and Multi-Agent Systems (AAMAS 07) monotonic if all of its actions are monotonic. (Adapted from [14, Definition 6]) Theorem 1 If A1 and A2 are monotonic actions, then performing A1 followed by A2 has the same effect on the agent"s beliefs as performing A2 followed by A1. (Adapted from [14, Theorem 2]). This assumes that both actions can be performed. However, it is possible for the performance of A1 to disable A2 from being done. For example, if A1 has the effect +p, and A2 has precondition ¬p, then although both actions may be enabled in the initial state, they cannot be performed in either order. We can prevent this by ensuring that actions" preconditions do not contain negation (or implication), since a monotonic action cannot result in a precondition that is negation-free becoming false. Note that this restriction only applies to the original action precondition, P, not to any additional preconditions imposed by the agent (P ). This is because only P is used to determine whether another agent is able to perform the action. Thus monotonic CMs with preconditions that do not contain negations have actions that commute. However, in fact, the restriction to monotonic CMs is unnecessarily strong: all that is needed is that whenever there is a choice of agent that can act, then the possible actions are monotonic. If there is only a single agent that can act, then no restriction is needed on the actions: they may or may not be monotonic. Definition 2 (Locally Monotonic) A commitment machine is locally monotonic if for any state S either (a) only a single agent has actions that can be performed; or (b) all actions that can be performed in S are monotonic. Theorem 2 In a locally monotonic CM, once all messages have been processed, all agents will be in the same state. Furthermore, no errors can occur. Proof: Once all messages have been processed we have that all agents will have performed the same action set, perhaps in a different order. The essence of the proof is to argue that as long as agents haven"t yet converged to the same state, all actions must be monotonic, and hence that these actions commute, and cannot disable any other actions. Consider the first point of divergence, where an agent performs action A and at the same time another agent (call it XB) performs action B. Clearly, this state has actions of more than one agent enabled, so, since the CM is locally monotonic, the relevant actions must be monotonic. Therefore, after doing A, the action B must still be enabled, and so the message to do B can be processed by updating the recipient agent"s beliefs with the effects of B. Furthermore, because monotonic actions commute, the result of doing A before B is the same as doing B before A: S A −−−−−→ SA ? ? yB B ? ? y SB −−−−−→ A SAB However, what happens if the next action after A is not B, but C? Because B is enabled, and C is not done by agent XB (see below), we must have that C is also monotonic, and hence (a) the result of doing A and B and C is the same regardless of the order in which the three actions are done; and (b) C doesn"t disable B, so B can still be done after C. S A −−−−−→ SA C −−−−−→ SAC ? ? yB B ? ? y B ? ? y SB −−−−−→ A SAB −−−−−→ C SABC The reason why C cannot be done by XB is that messages are processed in the order of their arrival9 . From the perspective of XB the action B was done before C, and therefore from any other agent"s perspective the message saying that B was done must be received (and processed) before a message saying that C is done. This argument can be extended to show that once agents start taking different paths through the FSM all actions taken until the point where they converge on a single state must be monotonic, and hence it is always possible to converge (because actions aren"t disabled), so the interaction is error free; and the resulting state once convergence occurs is the same (because monotonic actions commute). This theorem gives a strong theoretical guarantee that not doing turn management will not lead to disaster. This is analogous to proving that disabling all traffic lights would not lead to any accidents, and is only possible because the refined CM axioms are symmetrical. Based on this theorem the generic transformation from CM to code should allow agents to act freely, which is achieved by simply changing ı : P ∧ P ∧ ¬E ← +E; ↑A to ı : P ∧ P ∧ ¬E ← +E; ↑A; ı For example, instead of ı : ¬request ← +request; ↑sendRequest we have ı : ¬request ← +request; ↑sendRequest; ı. One consequence of the theorem is that it is not necessary to ensure that agents process messages before continuing to interact. However, in order to avoid unnecessary parallelism, which can make debugging harder, it may still be desirable to process messages before performing actions. Figure 8 shows a number of runs from the 3APL implementation that has been modified to allow free, non-alternating, interaction. 5. DISCUSSION We have presented a scheme for mapping commitment machines to BDI platforms (using SAAPL as an exemplar), identified three changes that needed to be made to SAAPL to support CM-based interaction, and shown that turn management can be avoided in CMbased interaction, provided the CM is locally monotonic. The three changes to SAAPL, and the translation scheme from commitment machine to BDI plans are both applicable to any BDI language. As we have mentioned in section 1, there has been some work on designing flexible and robust agent interaction, but virtually no work on implementing flexible and robust interactions. We have already discussed STAPLE [9, 10]. Another piece of work that is relevant is the work by Cheong and Winikoff on their Hermes methodology [2]. Although the main focus of their work is a pragmatic design methodology, they also provide guidelines for implementing Hermes designs using BDI platforms (specifically Jadex) [3]. However, since Hermes does not yield a design that is formal, it is only possible to generate skeleton code that then needs to be completed. Also, they do not address the turn taking issue: how to decide which agent acts when more than one agent is able to act. 9 We also assume that the communication medium does not deliver messages out of order, which is the case for (e.g.) TCP. The Sixth Intl. Joint Conf. on Autonomous Agents and Multi-Agent Systems (AAMAS 07) 879 Figure 8: Sample runs from 3APL implementation (non-alternating turns) The work of Kremer and Flores (e.g. [8]) also uses commitments, and deals with implementation. However, they provide infrastructure support (CASA) rather than a programming language, and do not appear to provide assistance to a programmer seeking to implement agents. Although we have implemented the NetBill interaction using 3APL, the changes to the semantics were done by modifying our NetBill 3APL program, rather than by modifying the 3APL implementation itself. Clearly, it would be desirable to modify the semantics of 3APL (or of another language) directly, by changing the implementation. Also, although we have not done so, it should be clear that the translation from a CM to its implementation could easily be automated. Another area for further work is to look at how the assumptions required to ensure that actions commute can be relaxed. Finally, there is a need to perform empirical evaluation. There has already been some work on comparing Hermes with a conventional message-centric approach to designing interaction, and this has shown that using Hermes results in designs that are significantly more flexible and robust [4]. It would be interesting to compare commitment machines with Hermes, but, since commitment machines are a framework, not a design methodology, we need to compare Hermes with a methodology for designing interactions that results in commitment machines [13, 17]. 6. REFERENCES [1] R. H. Bordini, M. Dastani, J. Dix, and A. E. F. Seghrouchni, editors. Multi-Agent Programming: Languages, Platforms and Applications. Springer, 2005. [2] C. Cheong and M. Winikoff. Hermes: Designing goal-oriented agent interactions. In Proceedings of the 6th International Workshop on Agent-Oriented Software Engineering (AOSE-2005), July 2005. [3] C. Cheong and M. Winikoff. Hermes: Implementing goal-oriented agent interactions. In Proceedings of the Third international Workshop on Programming Multi-Agent Systems (ProMAS), July 2005. [4] C. Cheong and M. Winikoff. Hermes versus prometheus: A comparative evaluation of two agent interaction design approaches. Submitted for publication, 2007. [5] P. R. Cohen and H. J. Levesque. Teamwork. Nous, 25(4):487-512, 1991. [6] M. Dastani, J. van der Ham, and F. Dignum. Communication for goal directed agents. In Proceedings of the Agent Communication Languages and Conversation Policies Workshop, 2002. [7] F. P. Dignum and G. A. Vreeswijk. Towards a testbed for multi-party dialogues. In Advances in Agent Communication, pages 212-230. Springer, LNCS 2922, 2004. [8] R. Kremer and R. Flores. Using a performative subsumption lattice to support commitment-based conversations. In F. Dignum, V. Dignum, S. Koenig, S. Kraus, M. P. Singh, and M. Wooldridge, editors, Autonomous Agents and Multi-Agent Systems (AAMAS), pages 114-121. ACM Press, 2005. [9] S. Kumar and P. R. Cohen. STAPLE: An agent programming language based on the joint intention theory. In Proceedings of the Third International Joint Conference on Autonomous Agents & Multi-Agent Systems (AAMAS 2004), pages 1390-1391. ACM Press, July 2004. [10] S. Kumar, M. J. Huber, and P. R. Cohen. Representing and executing protocols as joint actions. In Proceedings of the First International Joint Conference on Autonomous Agents and Multi-Agent Systems, pages 543 - 550, Bologna, Italy, 15 - 19 July 2002. ACM Press. [11] M. Tambe and W. Zhang. Towards flexible teamwork in persistent teams: Extended report. Journal of Autonomous Agents and Multi-agent Systems, 2000. Special issue on Best of ICMAS 98. [12] M. Winikoff. An AgentSpeak meta-interpreter and its applications. In Third International Workshop on Programming Multi-Agent Systems (ProMAS), pages 123-138. Springer, LNCS 3862 (post-proceedings, 2006), 2005. [13] M. Winikoff. Designing commitment-based agent interactions. In Proceedings of the 2006 IEEE/WIC/ACM International Conference on Intelligent Agent Technology (IAT-06), 2006. [14] M. Winikoff. Implementing flexible and robust agent interactions using distributed commitment machines. Multiagent and Grid Systems, 2(4), 2006. [15] M. Winikoff, W. Liu, and J. Harland. Enhancing commitment machines. In J. Leite, A. Omicini, P. Torroni, and P. Yolum, editors, Declarative Agent Languages and Technologies II, number 3476 in Lecture Notes in Artificial Intelligence (LNAI), pages 198-220. Springer, 2004. [16] M. Winikoff, L. Padgham, J. Harland, and J. Thangarajah. Declarative & procedural goals in intelligent agent systems. In Proceedings of the Eighth International Conference on Principles of Knowledge Representation and Reasoning (KR2002), Toulouse, France, 2002. [17] P. Yolum. Towards design tools for protocol development. In F. Dignum, V. Dignum, S. Koenig, S. Kraus, M. P. Singh, and M. Wooldridge, editors, Autonomous Agents and Multi-Agent Systems (AAMAS), pages 99-105. ACM Press, 2005. [18] P. Yolum and M. P. Singh. Flexible protocol specification and execution: Applying event calculus planning using commitments. In Proceedings of the 1st Joint Conference on Autonomous Agents and MultiAgent Systems (AAMAS), pages 527-534, 2002. [19] P. Yolum and M. P. Singh. Reasoning about commitments in the event calculus: An approach for specifying and executing protocols. Annals of Mathematics and Artificial Intelligence (AMAI), 2004. 880 The Sixth Intl. Joint Conf. on Autonomous Agents and Multi-Agent Systems (AAMAS 07)

social commitment;commitment machine;interaction goal;race condition;agent interaction;agent-oriented programming language;bdi-style plan;bdi;commitment-based interaction;messagecentric approach;agent orient program language;herme design;belief desire intention;netbill interaction;turn tracking;commitment machine framework;belief management method

train_I-46

Modular Interpreted Systems

We propose a new class of representations that can be used for modeling (and model checking) temporal, strategic and epistemic properties of agents and their teams. Our representations borrow the main ideas from interpreted systems of Halpern, Fagin et al.; however, they are also modular and compact in the way concurrent programs are. We also mention preliminary results on model checking alternating-time temporal logic for this natural class of models.

1. INTRODUCTION The logical foundations of multi-agent systems have received much attention in recent years. Logic has been used to represent and reason about, e.g., knowledge [7], time [6], cooperation and strategic ability [3]. Lately, an increasing amount of research has focused on higher level representation languages for models of such logics, motivated mainly by the need for compact representations, and for representations that correspond more closely to the actual systems which are modeled. Multi-agent systems are open systems, in the sense that agents interact with an environment only partially known in advance. Thus, we need representations of models of multi-agent systems which are modular, in the sense that a component, such as an agent, can be replaced, removed, or added, without major changes to the representation of the whole model. However, as we argue in this paper, few existing representation languages are both modular, compact and computationally grounded on the one hand, and allow for representing properties of both knowledge and strategic ability, on the other. In this paper we present a new class of representations for models of open multi-agent systems, which are modular, compact and come with an implicit methodology for modeling and designing actual systems. The structure of the paper is as follows. First, in Section 2, we present the background of our work - that is, logics that combine time, knowledge, and strategies. More precisely: modal logics that combine branching time, knowledge, and strategies under incomplete information. We start with computation tree logic CTL, then we add knowledge (CTLK), and then we discuss two variants of alternating-time temporal logic (ATL): one for the perfect, and one for the imperfect information case. The semantics of logics like the ones presented in Section 2 are usually defined over explicit models (Kripke structures) that enumerate all possible (global) states of the system. However, enumerating these states is one of the things one mostly wants to avoid, because there are too many of them even for simple systems. Thus, we usually need representations that are more compact. Another reason for using a more specialized class of models is that general Kripke structures do not always give enough help in terms of methodology, both at the stage of design, nor at implementation. This calls for a semantics which is more grounded, in the sense that the correspondence between elements of the model, and the entities that are modeled, is more immediate. In Section 3, we present an overview of representations that have been used for modeling and model checking systems in which time, action (and possibly knowledge) are important; we mention especially representations used for theoretical analysis. We point out that the compact and/or grounded representations of temporal models do not play their role in a satisfactory way when agents" strategies are considered. Finally, in Section 4, we present our framework of modular interpreted systems (MIS), and show where it fits in the picture. We conclude with a somewhat surprising hypothesis, that model checking ability under imperfect information for MIS can be computationally cheaper than model checking perfect information. Until now, almost all complexity results were distinctly in favor of perfect information strategies (and the others were indifferent). 2. LOGICS OF TIME, KNOWLEDGE, AND STRATEGIC ABILITY First, we present the logics CTL, CTLK, ATL and ATLir that are the starting point of our study. 2.1 Branching Time: CTL Computation tree logic CTL [6] includes operators for temporal properties of systems: i.e., path quantifier E (there is a path), together with temporal operators: f(in the next state), 2 (always from now on) and U (until).1 Every occurrence of a temporal operator is immediately preceded by exactly one path quantifier (this variant of the language is sometimes called vanilla CTL). Let Π be a set of atomic propositions with a typical element p. CTL formulae ϕ are defined as follows: ϕ ::= p | ¬ϕ | ϕ ∧ ϕ | E fϕ | E2ϕ | Eϕ U ϕ. The semantics of CTL is based on Kripke models M = St, R, π , which include a nonempty set of states St, a state transition relation R ⊆ St × St, and a valuation of propositions π : Π → P(St). A path λ in M refers to a possible behavior (or computation) of system M, and can be represented as an infinite sequence of states q0q1q2... such that qiRqi+1 for every i = 0, 1, 2, .... We denote the ith state in λ by λ[i]. A q-path is a path that starts in q. Interpretation of a formula in a state q in model M is defined as follows: M, q |= p iff q ∈ π(p); M, q |= ¬ϕ iff M, q |= ϕ; M, q |= ϕ ∧ ψ iff M, q |= ϕ and M, q |= ψ; M, q |= E fϕ iff there is a q-path λ such that M, λ[1] |= ϕ; M, q |= E2ϕ iff there is a q-path λ such that M, λ[i] |= ϕ for every i ≥ 0; M, q |= Eϕ U ψ iff there is a q-path λ and i ≥ 0 such that M, λ[i] |= ψ and M, λ[j] |= ϕ for every 0 ≤ j < i. 2.2 Adding Knowledge: CTLK CTLK [19] is a straightforward combination of CTL and standard epistemic logic [10, 7]. Let Agt = {1, ..., k} be a set of agents with a typical element a. Epistemic logic uses operators for representing agents" knowledge: Kaϕ is read as agent a knows that ϕ. Models of CTLK extend models of CTL with epistemic indistinguishability relations ∼a⊆ St × St (one per agent). We assume that all ∼a are equivalences. The semantics of epistemic operators is defined as follows: M, q |= Kaϕ iff M, q |= ϕ for every q such that q ∼a q . Note that, when talking about agents" knowledge, we implicitly assume that agents may have imperfect information about the actual current state of the world (otherwise the notion of knowledge would be trivial). This does not have influence on the way we model evolution of a system as a single unit, but it will become important when particular agents and their strategies come to the fore. 2.3 Agents and Their Strategies: ATL Alternating-time temporal logic ATL [3] is a logic for reasoning about temporal and strategic properties of open computational systems (multi-agent systems in particular). The language of ATL consists of the following formulae: ϕ ::= p | ¬ϕ | ϕ ∧ ϕ | A fϕ | A 2ϕ | A ϕ U ϕ. where A ⊆ Agt. Informally, A ϕ says that agents A have a collective strategy to enforce ϕ. It should be noted that the CTL path quantifiers A, E can be expressed with ∅ , Agt respectively. The semantics of ATL is defined in so called concurrent game structures (CGSs). A CGS is a tuple M = Agt, St, Act, d, o, Π, π , 1 Additional operators A (for every path) and 3 (sometime in the future) are defined in the usual way. consisting of: a set Agt = {1, . . . , k} of agents; set St of states; valuation of propositions π : Π → P(St); set Act of atomic actions. Function d : Agt × St → P(Act) indicates the actions available to agent a ∈ Agt in state q ∈ St. Finally, o is a deterministic transition function which maps a state q ∈ St and an action profile α1, . . . , αk ∈ Actk , αi ∈ d(i, q), to another state q = o(q, α1, . . . , αk). DEFINITION 1. A (memoryless) strategy of agent a is a function sa : St → Act such that sa(q) ∈ d(a, q).2 A collective strategy SA for a team A ⊆ Agt specifies an individual strategy for each agent a ∈ A. Finally, the outcome of strategy SA in state q is defined as the set of all computations that may result from executing SA from q on: out(q, SA) = {λ = q0q1q2... | q0 = q and for every i = 1, 2, ... there exists αi−1 1 , ..., αi−1 k such that αi−1 a = SA(a)(qi−1) for each a ∈ A, αi−1 a ∈ d(a, qi−1) for each a /∈ A, and o(qi−1, αi−1 1 , ..., αi−1 k ) = qi}. The semantics of cooperation modalities is as follows: M, q |= A fϕ iff there is a collective strategy SA such that, for every λ ∈ out(q, SA), we have M, λ[1] |= ϕ; M, q |= A 2ϕ iff there exists SA such that, for every λ ∈ out(q, SA), we have M, λ[i] for every i ≥ 0; M, q |= A ϕ U ψ iff there exists SA such that for every λ ∈ out(q, SA) there is a i ≥ 0, for which M, λ[i] |= ψ, and M, λ[j] |= ϕ for every 0 ≤ j < i. 2.4 Agents with Imperfect Information: ATLir As ATL does not include incomplete information in its scope, it can be seen as a logic for reasoning about agents who always have complete knowledge about the current state of the whole system. ATLir [21] includes the same formulae as ATL, except that the cooperation modalities are presented with a subscript: A ir indicates that they address agents with imperfect information and imperfect recall. Formally, the recursive definition of ATLir formulae is: ϕ ::= p | ¬ϕ | ϕ ∧ ϕ | A ir fϕ | A ir2ϕ | A irϕ U ϕ Models of ATLir, concurrent epistemic game structures (CEGS), can be defined as tuples M = Agt, St, Act, d, o, ∼1, ..., ∼k, Π, π , where Agt, St, Act, d, o, Π, π is a CGS, and ∼1, ..., ∼k are epistemic (equivalence) relations. It is required that agents have the same choices in indistinguishable states: q ∼a q implies d(a, q) = d(a, q ). ATLir restricts the strategies that can be used by agents to uniform strategies, i.e. functions sa : St → Act, such that: (1) sa(q) ∈ d(a, q), and (2) if q ∼a q then sa(q) = sa(q ). A collective strategy is uniform if it contains only uniform individual strategies. Again, the function out(q, SA) returns the set of all paths that may result from agents A executing collective strategy SA from state q. The semantics of ATLir formulae can be defined as follows: M, q |= A ir fϕ iff there is a uniform collective strategy SA such that, for every a ∈ A, q such that q ∼a q , and λ ∈ out(SA, q ), we have M, λ[1] |= ϕ; 2 This is a deviation from the original semantics of ATL [3], where strategies assign agents" choices to sequences of states, which suggests that agents can by definition recall the whole history of each game. While the choice of one or another notion of strategy affects the semantics of the full ATL ∗ , and most ATL extensions (e.g. for games with imperfect information), it should be pointed out that both types of strategies yield equivalent semantics for pure ATL (cf. [21]). 898 The Sixth Intl. Joint Conf. on Autonomous Agents and Multi-Agent Systems (AAMAS 07) M, q |= A ir2ϕ iff there exists SA such that, for every a ∈ A, q such that q ∼a q , and λ ∈ out(SA, q ), we have M, λ[i] for every i ≥ 0; M, q |= A irϕ U ψ iff there exist SA such that, for every a ∈ A, q such that q ∼a q , and λ ∈ out(SA, q ), there is i ≥ 0 for which M, λ[i] |= ψ, and M, λ[j] |= ϕ for every 0 ≤ j < i. That is, A irϕ holds iff A have a uniform collective strategy, such that for every path that can possibly result from execution of the strategy according to at least one agent from A, ϕ is the case. 3. MODELS AND MODEL CHECKING In this section, we present and discuss various (existing) representations of systems that can be used for modeling and model checking. We believe that the two most important points of reference are in this case: (1) the modeling formalism (i.e., the logic and the semantics we use), and (2) the phenomenon, or more generally, the domain we are going to model (to which we will often refer as the real world). Our aim is a representation which is reasonably close to the real world (i.e., it is sufficiently compact and grounded), and still not too far away from the formalism (so that it e.g. easily allows for theoretical analysis of computational problems). We begin with discussing the merits of explicit modelsin our case, these are transition systems, concurrent game structures and CEGSs, presented in the previous section. 3.1 Explicit Models Obviously, an advantage of explicit models is that they are very close to the semantics of our logics (simply because they are the semantics). On the other hand, they are in many ways difficult to use to describe an actual system: • Exponential size: temporal models usually have an exponential number of states with respect to any higher-level description (e.g. Boolean variables, n-ary attributes etc.). Also, their size is exponential in the number of processes (or agents) if the evolution of a system results from joint (synchronous or asynchronous) actions of several active entities [15]. For CGSs the situation is even worse: here, also the number of transitions is exponential, even if we fix the number of states.3 In practice, this means that such representations are very seldom scalable. • Explicit models include no modularity. States in a model refer to global states of the system; transitions in the model correspond to global transitions as well, i.e., they represent (in an atomic way) everything that may happen in one single step, regardless of who has done it, to whom, and in what way. • Logics like ATL are often advertised as frameworks for modeling and reasoning about open computational systems. Ideally, one would like the elements of such a system to have as little interdependencies as possible, so that they can be plugged in and out without much hassle, for instance when we want to test various designs or implementations of the active component. In the case of a multi-agent system the 3 Another class of ATL models, alternating transition systems [2] represent transitions in a more succinct way. While we still have exponentially many states in an ATS, the number of transitions is simply quadratic wrt. to states (like for CTL models). Unfortunately, ATS are even less modular and harder to design than concurrent game structures, and they cannot be easily extended to handle incomplete information (cf. [9]). need is perhaps even more obvious. We do not only need to re-plug various designs of a single agent in the overall architecture; we usually also need to change (e.g., increase) the number of agents acting in a given environment without necessarily changing the design of the whole system. Unfortunately, ATL models are anything but open in this sense. Theoretical complexity results for explicit models are as follows. Model checking CTL and CTLK is P-complete, and can be done in time O(ml), where m is the number of transitions in the model, and l is the length of the formula [5]. Alternatively, it can be done in time O(n2 l), where n is the number of states. Model checking ATL is P-complete wrt. m, l and ΔP 3 -complete wrt. n, k, l (k being the number of agents) [3, 12, 16]. Model checking ATLir is ΔP 2complete wrt. m, l and ΔP 3 -complete wrt. n, k, l [21, 13]. 3.2 Compressed Representations Explicit representation of all states and transitions is inefficient in many ways. An alternative is to represent the state/transition space in a symbolic way [17, 18]. Such models offer some hope for feasible model checking properties of open/multi-agent systems, although it is well known that they are compact only in a fraction of all cases.4 For us, however, they are insufficient for another reason: they are merely optimized representations of explicit models. Thus, they are neither more open nor better grounded: they were meant to optimize implementation rather than facilitate design or modeling methodology. 3.3 Interpreted Systems Interpreted systems [11, 7] are held by many as a prime example of computationally grounded models of distributed systems. An interpreted system can be defined as a tuple IS = St1, ..., Stk, Stenv, R, π . St1, ..., Stk are local state spaces of agents 1, ..., k, and Stenv is the set of states of the environment. The set of global states is defined as St = St1 × ... × Stk × Stenv; R ⊆ St × St is a transition relation, and π : Π → P(St). While the transition relation encapsulates the (possible) evolution of the system over time, the epistemic dimension is defined by the local components of each global state: q1, ..., qk, qenv ∼i q1, ..., qk, qenv iff qi = qi . It is easy to see that such a representation is modular and compact as far as we are concerned with states. Moreover, it gives a natural (grounded) approach to knowledge, and suggests an intuitive methodology for modeling epistemic states. Unfortunately, the way transitions are represented in interpreted systems is neither compact, nor modular, nor grounded: the temporal aspect of the system is given by a joint transition function, exactly like in explicit models. This is not without a reason: if we separate activities of the agents too much, we cannot model interaction in the framework any more, and interaction is the most interesting thing here. But the bottom line is that the temporal dimension of an interpreted system has exponential representation. And it is almost as difficult to plug components in and out of an interpreted system, as for an ordinary CTL or ATL model, since the local activity of an agent is completely merged with his interaction with the rest of the system. 3.4 Concurrent Programs The idea of concurrent programs has been long known in the literature on distributed systems. Here, we use the formulation from [15]. A concurrent program P is composed of k concurrent processes, each described by a labeled transition system Pi = Sti, Acti, Ri, Πi, πi , where Sti is the set of local states of process 4 Representation R of an explicit model M is compact if the size of R is logarithmic with respect to the size of M. The Sixth Intl. Joint Conf. on Autonomous Agents and Multi-Agent Systems (AAMAS 07) 899 i, Acti is the set of local actions, Ri ⊆ Sti ×Acti ×Sti is a transition relation, and Πi, πi are the set of local propositions and their valuation. The behavior of program P is given by the product automaton of P1, ..., Pk under the assumption that processes work asynchronously, actions are interleaved, and synchronization is obtained through common action names. Concurrent programs have several advantages. First of all, they are modular and compact. They allow for local modeling of components - much more so than interpreted systems (not only states, but also actions are local here). Moreover, they allow for representing explicit interaction between local transitions of reactive processes, like willful communication, and synchronization. On the other hand, they do not allow for representing implicit, incidental, or not entirely benevolent interaction between processes. For example, if we want to represent the act of pushing somebody, the pushed object must explicitly execute an action of being pushed, which seems somewhat ridiculous. Side effects of actions are also not easy to model. Still, this is a minor complaint in the context of CTL, because for temporal logics we are only interested in the flow of transitions, and not in the underlying actions. For temporal reasoning about k asynchronous processes with no implicit interaction, concurrent programs seem just about perfect. The situation is different when we talk about autonomous, proactive components (like agents), acting together (cooperatively or adversely) in a common environment - and we want to address their strategies and abilities. Now, particular actions are no less important than the resulting transitions. Actions may influence other agents" local states without their consent, they may have side effects on other agents" states etc. Passing messages and/or calling procedures is by no means the only way of interaction between agents. Moreover, the availability of actions (to an agent) should not depend on the actions that will be executed by other agents at the same time - these are the outcome states that may depend on these actions! Finally, we would often like to assume that agents act synchronously. In particular, all agents play simultaneously in concurrent game structures. But, assuming synchrony and autonomy of actions, synchronization can no longer be a means of coordination. To sum up, we need a representation which is very much like concurrent programs, but allows for modeling agents that play synchronously, and which enables modeling more sophisticated interaction between agents" actions. The first postulate is easy to satisfy, as we show in the following section. The second will be addressed in Section 4. We note that model checking CTL against concurrent programs is PSPACE-complete in the number of local states and the length of the formula [15]. 3.5 Synchronous CP and Simple Reactive Modules The semantics of ATL is based on synchronous models where availability of actions does not depend on the actions currently executed by the other players. A slightly different variant of concurrent programs can be defined via synchronous product of programs, so that all agents play simultaneously.5 Unfortunately, under such interpretation, no direct interaction between agents" actions can be modeled at all. DEFINITION 2. A synchronous concurrent program consists of k concurrent processes Pi = Sti, Acti, Ri, Πi, πi with the follow5 The concept is not new, of course, and has already existed in folk knowledge, although we failed to find an explicit definition in the literature. ing unfolding to a CGS: Agt = {1, ..., k}, St = Qk i=1 Sti, Act = Sk i=1 Acti, d(i, q1, ..., qk ) = {αi | qi, αi, qi ∈ Ri for some qi ∈ Sti}, o( q1, ..., qk , α1, ..., αk) = q1, ..., qk such that qi, αi, qi ∈ Ri for every i; Π = Sk i=1 Πi, and π(p) = πi(p) for p ∈ Πi. We note that the simple reactive modules (SRML) from [22] can be seen as a particular implementation of synchronous concurrent programs. DEFINITION 3. A SRML system is a tuple Σ, Π, m1, . . . , mk , where Σ = {1, . . . , k} is a set of modules (or agents), Π is a set of Boolean variables, and, for each i ∈ Σ, we have mi = ctri, initi, updatei , where ctri ⊆ Π. Sets initi and updatei consist of guarded commands of the form φ ; v1 := ψ1; . . . ; vk := ψk, where every vj ∈ ctri, and φ, ψ1, . . . , ψk are propositional formulae over Π. It is required that ctr1, . . . ctrk partitions Π. The idea is that agent i controls the variables ctri. The init guarded commands are used to initialize the controlled variables, while the update guarded commands can change their values in each round. A guarded command is enabled if the guard φ is true in the current state of the system. In each round an enabled update guarded command is executed: each ψj is evaluated against the current state of the system, and its logical value is assigned to vj. Several guarded commands being enabled at the same time model non-deterministic choice. Model checking ATL for SRML has been proved EXPTIMEcomplete in the size of the model and the length of the formula [22]. 3.6 Concurrent Epistemic Programs Concurrent programs (both asynchronous and synchronous) can be used to encode epistemic relations too - exactly in the same way as interpreted systems do [20]. That is, when unfolding a concurrent program to a model of CTLK or ATLir, we define that q1, ..., qk ∼i q1, ..., qk iff qi = qi . Model checking CTLK against concurrent epistemic programs is PSPACE-complete [20]. SRML can be also interpreted in the same way; then, we would assume that every agent can see only the variables he controls. Concurrent epistemic programs are modular and have a grounded semantics. They are usually compact (albeit not always: for example, an agent with perfect information will always blow up the size of such a program). Still, they inherit all the problems of concurrent programs with perfect information, discussed in Section 3.4: limited interaction between components, availability of local actions depending on the actual transition etc. The problems were already important for agents with perfect information, but they become even more crucial when agents have only limited knowledge of the current situation. One of the most important applications of logics that combine strategic and epistemic properties is verification of communication protocols (e.g., in the context of security). Now, we may want to, e.g., check agents" ability to pass an information between them, without letting anybody else intercept the message. The point is that the action of intercepting is by definition enabled; we just look for a protocol in which the transition of successful interception is never carried out. So, availability of actions must be independent of the actions chosen by the other agents under incomplete information. On the other hand, interaction is arguably the most interesting feature of multi-agent systems, and it is really hard to imagine models of strategic-epistemic logics, in which it is not possible to represent communication. 3.7 Reactive Modules Reactive modules [1] can be seen as a refinement of concurrent epistemic programs (primarily used by the MOCHA model checker [4]), but they are much more powerful, expressive and 900 The Sixth Intl. Joint Conf. on Autonomous Agents and Multi-Agent Systems (AAMAS 07) grounded. We have already mentioned a very limited variant of RML (i.e., SRML). The vocabulary of RML is very close to implementations (in terms of general computational systems): the modules are essentially collections of variables, states are just valuations of variables; events/actions are variable updates. However, the sets of variables controlled by different agents can overlap, they can change over time etc. Moreover, reactive modules support incomplete information (through observability of variables), although it is not the main focus of RML. Again, the relationship between sets of observable variables (and to sets of controlled variables) is mostly left up to the designer of a system. Agents can act synchronously as well as asynchronously. To sum up, RML define a powerful framework for modeling distributed systems with various kinds of synchrony and asynchrony. However, we believe that there is still a need for a simpler and slightly more abstract class of representations. First, the framework of RML is technically complicated, involving a number auxiliary concepts and their definitions. Second, it is not always convenient to represent all that is going on in a multi-agent system as reading and/or writing from/to program variables. This view of a multi-agent system is arguably close to its computer implementation, but usually rather distant from the real world domainhence the need for a more abstract, and more conceptually flexible framework. Third, the separation of the local complexity, and the complexity of interaction is not straightforward. Our new proposal, more in the spirit of interpreted systems, takes these observations as the starting point. The proposed framework is presented in Section 4. 4. MODULAR INTERPRETED SYSTEMS The idea behind distributed systems (multi-agent systems even more so) is that we deal with several loosely coupled components, where most of the processing goes on inside components (i.e., locally), and only a small fraction of the processing occurs between the components. Interaction is crucial (which makes concurrent programs an insufficient modeling tool), but it usually consumes much less of the agent"s resources than local computations (which makes the explicit transition tables of CGS, CEGS, and interpreted systems an overkill). Modular interpreted systems, proposed here, extrapolate the modeling idea behind interpreted systems in a way that allows for a tight control of the interaction complexity. DEFINITION 4. A modular interpreted system (MIS) is defined as a tuple S = Agt, env, Act, In , where Agt = {a1, ..., ak} is a set of agents, env is the environment, Act is a set of actions, and In is a set of symbols called interaction alphabet. Each agent has the following internal structure: ai = Sti, di, outi, ini, oi, Πi, πi , where: • Sti is a set of local states, • di : Sti → P(Act) defines local availability of actions; for convenience of the notation, we additionally define the set of situated actions as Di = { qi, α | qi ∈ Sti, α ∈ di(qi)}, • outi, ini are interaction functions; outi : Di → In refers to the influence that a given situated action (of agent ai) may possibly have on the external world, and ini : Sti ×Ink → In translates external manifestations of the other agents (and the environment) into the impression that they make on ai"s transition function depending on the local state of ai, • oi : Di × In → Sti is a (deterministic) local transition function, • Πi is a set of local propositions of agent ai where we require that Πi and Πj are disjunct when i = j, and • πi : Πi → P(Sti) is a valuation of these propositions. The environment env = Stenv, outenv, inenv, oenv, Πenv, πenv has the same structure as an agent except that it does not perform actions, and that thus outenv : Stenv → In and oenv : Stenv × In → Stenv. Within our framework, we assume that every action is executed by an actor, that is, an agent. As a consequence, every actor is explicitly represented in a MIS as an agent, just like in the case of CGS and CEGS. The environment, on the other hand, represents the (passive) context of agents" actions. In practice, it serves to capture the aspects of the global state that are not observable by any of the agents. The input functions ini seem to be the fragile spots here: when given explicitly as tables, they have size exponential wrt. the number of agents (and linear wrt. the size of In). However, we can use, e.g., a construction similar to the one from [16] to represent interaction functions more compactly. DEFINITION 5. Implicit input function for state q ∈ Sti is given by a sequence ϕ1, η1 , ..., ϕn, ηn , where each ηj ∈ In is an interaction symbol, and each ϕj is a boolean combination of propositions ˆηi , with η ∈ In; ˆηi stands for η is the symbol currently generated by agent i. The input function is now defined as follows: ini(q, 1, ..., k, env) = ηj iff j is the lowest index such that {ˆ1 1, ..., ˆk k, ˆenv env} |= ϕj. It is required that ϕn ≡ , so that the mapping is effective. REMARK 1. Every ini can be encoded as an implicit input function, with each ϕj being of polynomial size with respect to the number of interaction symbols (cf. [16]). Note that, for some domains, the MIS representation of a system requires exponentially many symbols in the interaction alphabet In. In such a case, the problem is inherent to the domain, and ini will have size exponential wrt the number of agents. 4.1 Representing Agent Systems with MIS Let Stg = ( Qk i=1 Sti)×Stenv be the set of all possible global states generated by a modular interpreted system S. DEFINITION 6. The unfolding of a MIS S for initial states Q ⊆ Stg to a CEGS cegs(S, Q) = Agt , St , Π , π , Act , d , o , ∼1, ..., ∼k is defined as follows: • Agt = {1, ..., k} and Act = Act, • St is the set of global states from Stg which are reachable from some state in Q via the transition relation defined by o (below), • Π = Sk i=1 Πi ∪ Πenv, • For each q = q1, . . . , qk, qenv ∈ St and i = 1, ..., k, env, we define q ∈ π (p) iff p ∈ Πi and qi ∈ πi(p), • d (i, q) = di(qi) for global state q = q1, ..., qk, qenv , • The transition function is constructed as follows. Let q = q1, ..., qk, qenv ∈ St , and α = α1, ..., αk be an action profile s.t. αi ∈ d (i, q). We define inputi(q, α) = ini ` qi, out1(q1, α1), . . . , outi−1(qi−1, αi−1), outi+1(qi+1, αi+1), . . . , outk(qk, αk), outenv(qenv) ´ for each agent i = 1, . . . , k, and inputenv(q, α) = inenv ` qenv, out1(q1, α1), . . . , outk(qk, αk) ´ . Then, o (q, α) = o1( q1, α1 , input1(q, α)), . . . , ok( qk, αk , inputk(q, α)), oenv(qenv, inputenv(q, α)) ; The Sixth Intl. Joint Conf. on Autonomous Agents and Multi-Agent Systems (AAMAS 07) 901 • For each i = 1, ..., k: q1, ..., qk, qenv ∼i q1, ..., qk, qenv iff qi = qi .6 REMARK 2. Note that MISs can be used as representations of CGSs too. In that case, epistemic relations ∼i are simply omitted in the unfolding. We denote the unfolding of a MIS S for initial states Q into a CGS by cgs(S, Q). Propositions 3 and 5 state that modular interpreted systems can be used as representations for explicit models of multi-agent systems. On the other hand, these representations are not always compact, as demonstrated by Propositions 7 and 8. PROPOSITION 3. For every CEGS M, there is a MIS SM and a set of global states Q of SM such that cegs(SM , Q) is isomorphic to M.7 PROOF. Let M = {1, . . . , k}, St, Act, d, o, Π, π, ∼1, . . . , ∼k be a CEGS. We construct a MIS SM = {a1, . . . , ak}, env, Act, In with agents ai = Sti, di, outi, ini, oi, Πi, πi and environment env = Stenv, outenv, inenv, oenv, Πenv, πenv , plus a set Q ⊆ Stg of global states, as follows. • In = Act ∪ St ∪ (Actk−1 × St), • Sti = {[q]∼i | q ∈ St} for 1 ≤ i ≤ k (i.e., Sti is the set of i"s indistinguishability classes in M), • Stenv = St, • di([q]∼i ) = d(i, q) for 1 ≤ i ≤ k (this is well-defined since d(i, q) = d(i, q ) whenever q ∼i q ), • outi([q]∼i , αi) = αi for 1 ≤ i ≤ k; outenv(q) = q, • ini([q]∼i , α1, . . . , αi−1, αi+1, . . . , αk, qenv) = α1, . . . , αi−1, αi+1, . . . , αk, qenv for i ∈ {1, . . . , k}; inenv(q, α1 . . . , αk) = α1, . . . , αk ; ini(x) and inenv(x) are arbitrary for other arguments x, • oi( [q]∼i , αi , α1, . . . , αi−1, αi+1, . . . , αk, qenv ) = [o(qenv, α1, . . . , αk)]∼i for 1 ≤ i ≤ k and αi ∈ di([q]∼i ); oenv(q, α1, . . . , αk ) = o(q, α1, . . . , αk); oi and oenv are arbitrary for other arguments, • Πi = ∅ for 1 ≤ i ≤ k, and Πenv = Π, • πenv(p) = π(p) • Q = { [q]∼1 , . . . , [q]∼k , q : q ∈ St} Let M = cegs(SM , Q) = Agt , St , Act , d , o , Π , π , ∼1, . . . , ∼k . We argue that M and M are isomorphic by establishing a oneto-one correspondence between the respective sets of states, and showing that the other parts of the structures agree on corresponding states. First we show that, for any ˆq = [q ]∼1 , . . . , [q ]∼k , q ∈ Q and any α = α1, . . . , αk such that αi ∈ d (i, ˆq ), we have o (ˆq , α) = [q]∼1 , . . . , [q]∼k , q where q = o(q , α) (1) Let ˆq = o (ˆq , α). Now, for any i: inputi(ˆq , α) = ini([q ]∼i , out1([q ]∼1 , α1), ..., outi−1([q ]∼i−1 , αi−1), outi+1([q ]∼i+1 , αi+1), . . . , outk([q ]∼k , αk), outenv(q )) = ini([q ]∼i , α1, . . . , αi−1, αi+1, 6 This shows another difference between the environment and the agents: the environment does not possess knowledge. 7 We say that two CEGS are isomorphic if they only differ in the names of states and/or actions. . . . , αk, q ) = α1, . . . , αi−1, αi+1, . . . , αk, q . Similarly, we get that inputenv(ˆq , α) = α1, . . . , αk . Thus we get that o (ˆq , α) = o1( [q ]∼1 , α1 , input1(ˆq , α)), . . . , ok( [q ]∼k , αk , inputk(ˆq , α)), oenv(q , inputenv(ˆq , α)) = [o(q , α1, . . . , αk)]∼1 , . . . , [o(q , α1, . . . , αk)]∼k , o(q , α1, . . . , αk) . Thus, ˆq = [q]∼1 , . . . , [q]∼k , q for q = o(q , α1, . . . , αk), which completes the proof of (1). We now argue that St = Q. Clearly, Q ⊆ St . Let ˆq ∈ St ; we must show that ˆq ∈ Q. The argument is on induction on the length of the least o path from Q to ˆq. The base case, ˆq ∈ Q, is immediate. For the inductive step, ˆq = o (ˆq , α) for some ˆq ∈ Q, and then we have that ˆq ∈ Q by (1). Thus, St = Q. Now we have a one-to-one correspondence between St and St : r ∈ St corresponds to [r]∼1 , . . . , [r]∼k , r ∈ St . It remains to be shown that the other parts of the structures M and M agree on corresponding states: • Agt = Agt, • Act = Act, • Π = Sk i=1 Πi ∪ Πenv = Π, • For p ∈ Π = Π: [q ]∼1 , . . . , [q ]∼k , q ∈ π (p) iff q ∈ πenv(p) iff q ∈ π(p) (same valuations at corresponding states), • d (i, [q ]∼1 , . . . , [q ]∼k , q ) = di([q ]∼i ) = d(i, q), • It follows immediately from (1), and the fact that Q = St , that o ( [q ]∼1 , . . . , [q ]∼k , q , α) = [r ]∼1 , . . . , [r ]∼k , r iff o(q , α) = r (transitions on the same joint action in corresponding states lead to corresponding states), • [q ]∼1 , . . . , [q ]∼k , q ∼i [r ]∼1 , . . . , [r ]∼k , r iff [q ]∼i = [r ]∼i iff q ∼i r (the accessibility relations relate corresponding states), which completes the proof. COROLLARY 4. For every CEGS M, there is an ATLir-equivalent MIS S with initial states Q, that is, for every state q in M there is a state q in cegs(S, Q) satisfying exactly the same ATLir formulae, and vice versa. PROPOSITION 5. For every CGS M, there is a MIS SM and a set of global states Q of SM such that cgs(SM , Q) is isomorphic to M. PROOF. Let M = Agt, St, Act, d, o, Π, π be given. Now, let ˆM = Agt, St, Act, d, o, Π, π, ∼1, . . . , ∼k for some arbitrary accessibility relations ∼i over St. By Proposition 3, there exists a MIS S ˆM with global states Q such that ˆM = cegs(S ˆM , Q) is isomorphic to ˆM. Let M be the CGS obtained by removing the accessibility relations from ˆM . Clearly, M is isomorphic to M. COROLLARY 6. For every CGS M, there is an ATL-equivalent MIS S with initial states Q. That is, for every state q in M there is a state q in cgs(S, Q) satisfying exactly the same ATL formulae, and vice versa. PROPOSITION 7. The local state spaces in a MIS are not always compact with respect to the underlying concurrent epistemic game structure. PROOF. Take a CEGS M in which agent i has always perfect information about the current global state of the system. When constructing a modular interpreted system S such that M = cegs(S, Q), we have that Sti must be isomorphic with St. 902 The Sixth Intl. Joint Conf. on Autonomous Agents and Multi-Agent Systems (AAMAS 07) The above property is a part of the interpreted systems heritage. The next proposition stems from the fact that explicit models (and interpreted systems) allow for intensive interaction between agents. PROPOSITION 8. The size of In in S is, in general, exponential with respect to the number of local states and local actions. This is the case even when epistemic relations are not relevant (i.e., when S is taken as a representation of an ordinary CGS). PROOF. Consider a CGS M with agents Agt = {1, ..., k}, global states St = Qk i=1{qi 0, ..., qi i}, and actions Act = {0, 1}, all enabled everywhere. The transition function is defined as o( q1 j1 , ..., qk jk , α1, ..., αk) = q1 l1 , ..., qk lk , where li = (ji + α1 + ... + αk) mod i. Note that M can be represented as a modular interpreted system with succinct local state spaces Sti = {qi 0, ..., qi i}. Still, the current actions of all agents are relevant to determine the resulting local transition of agent i. We will call items In, outi, ini the interaction layer of a modular interpreted system S; the other elements of S constitute the local layer of the MIS. In this paper we are ultimately interested in model checking complexity with respect to the size of the local layer. To this end, we will assume that the size of interaction layer is polynomial in the number of local states and actions. Note that, by Propositions 7 and 8, not every explicit model submits to compact representation with a MIS. Still, as we declared at the beginning of Section 4, we are mainly interested in a modeling framework for systems of loosely coupled components, where interaction is essential, but most processing is done locally anyway. More importantly, the framework of MIS allows for separating the interaction of agents from their local structure to a larger extent. Moreover, we can control and measure the complexity of each layer in a finer way than before. First, we can try to abstract from the complexity of a layer (e.g. like in this paper, by assuming that the other layer is kept within certain complexity bounds). Second, we can also measure separately the interaction complexity of different agents. 4.2 Modular Interpreted Systems vs. Simple Reactive Modules In this section we show that simple reactive modules are (as we already suggested) a specific (and somewhat limited) implementation of modular interpreted systems. First, we define our (quite strong) notion of equivalence of representations. DEFINITION 7. Two representations are equivalent if they unfold to isomorphic concurrent epistemic game structures. They are CGS-equivalent if they unfold to the same CGS. PROPOSITION 9. For any SRML there is a CGS-equivalent MIS. PROOF. Consider an SRML R with k modules and n variables. We construct S = Agt, Act, In with Agt = {a1, ..., ak}, Act = { 1, ..., n, ⊥1, ..., ⊥n}, and In = Sk i=1 Sti × Sti (the local state spaces Sti will be defined in a moment). Let us assume without loss of generality that ctri = {x1, ..., xr}. Also, we consider all guarded commands of i to be of type γi,ψ : ψ ; xi := , or γ⊥ i,ψ : ψ ; xi := ⊥. Now, agent ai in S has the following components: Sti = P(ctri) (i.e., local states of ai are valuations of variables controlled by i); di(qi) = { 1, ..., r, ⊥1, ..., ⊥r}; outi(qi, α) = qi, qi ; ini(qi, q1, q1 , ..., qi−1, qi−1 , qi+1, qi+1 , qk, qk ) = {xi ∈ ctri | q1, ..., qk |= W γi,ψ ψ}, {xi ∈ ctri | q1, ..., qk |= W γ⊥ i,ψ ψ} . To define local transitions, we consider three cases. If t = f = ∅ (no update is enabled), then oi(qi, α, t, f ) = qi for every action α. If t = ∅, we take any arbitrary ˆx ∈ t, and define oi(qi, j, t, f ) = qi ∪ {xj} if xj ∈ t, and qi ∪ {ˆx} otherwise; oi(qi, ⊥j, t, f ) = qi \ {xj} if xj ∈ f, and qi ∪ {ˆx} otherwise. Moreover, if t = ∅ = f, we take any arbitrary ˆx ∈ f, and define oi(qi, j, t, f ) = qi ∪ {xj} if xj ∈ t, and qi \ {ˆx} otherwise; oi(qi, ⊥j, t, f ) = qi \{xj} if xj ∈ f, and qi \{ˆx} otherwise. Finally, Πi = ctri, and qi ∈ πi(xj) iff xj ∈ qi. The above construction shows that SRML have more compact representation of states than MIS: ri local variables of agent i give rise to 2ri local states. In a way, reactive modules (both simple and full) are two-level representations: first, the system is represented as a product of modules; next, each module can be seen as a product of its variables (together with their update operations). Note, however, that specification of updates with respect to a single variable in an SRML may require guarded commands of total length O(2 Pk i=1 ri ). Thus, the representation of transitions in SRML is (in the worst case) no more compact than in MIS, despite the two-level structure of SRML. We observe finally that MIS are more general, because in SRML the current actions of other agents have no influence on the outcome of agent i"s current action (although the outcome can be influenced by other agents" current local states). 4.3 Model Checking Modular Interpreted Systems One of our main aims was to study the complexity of symbolic model checking ATLir in a meaningful way. Following the reviewers" remarks, we state our complexity results only as conjectures. Preliminary proofs can be found in [14]. CONJECTURE 10. Model checking ATL for modular interpreted systems is EXPTIME-complete. CONJECTURE 11. Model checking ATLir for the class of modular interpreted systems is PSPACE-complete. A summary of complexity results for model checking temporal and strategic logics (with and without epistemic component) is given in the table below. The table presents completeness results for various models and settings of input parameters. Symbols n, k, m stand for the number of states, agents and transitions in an explicit model; l is the length of the formula, and nlocal is the number of local states in a concurrent program or modular interpreted system. The new results, conjectured in this paper, are printed in italics. Note that the result for model checking ATL against modular interpreted systems is an extension of the result from [22]. m, l n, k, l nlocal, k, l CTL P [5] P [5] PSPACE [15] CTLK P [5, 8] P [5, 8] PSPACE [20] ATL P [3] ΔP 3 [12, 16] EXPTIME ATLir ΔP 2 [21, 13] ΔP 3 [13] PSPACE If we are right, then the results for ATL and ATLir form an intriguing pattern. When we compare model checking agents with perfect vs. imperfect information, the first problem appears to be much easier against explicit models measured with the number of transitions; next, we get the same complexity class against explicit models measured with the number of states and agents; finally, model checking imperfect information turns out to be easier than model checking perfect information for modular interpreted systems. Why can it be so? First, a MIS unfolds into CEGS and CGS in a different way. In the first case, the MIS is assumed to encode the epistemic relations explicitly (which makes it explode when we model agents with perfect, or almost perfect information). In the latter case, the epistemic The Sixth Intl. Joint Conf. on Autonomous Agents and Multi-Agent Systems (AAMAS 07) 903 aspect is ignored, which gives some extra room for encoding the transition relation more efficiently. Another crucial factor is the number of available strategies (relative to the size of input parameters). The number of all strategies is exponential in the number of global states; for uniform strategies, there are usually much less of them but still exponentially many in general. Thus, the fact that perfect information strategies can be synthesized incrementally has a substantial impact on the complexity of the problem. However, measured in terms of local states and agents, the number of all strategies is doubly exponential, while there are only exponentially many uniform strategies - which settles the results in favor of imperfect information. 5. CONCLUSIONS We have presented a new class of representations for open multiagent systems. Our representations, called modular interpreted systems, are: modular, in the sense that components can be changed, replaced, removed or added, with as little changes to the whole representation as possible; more compact than traditional explicit representations; and grounded, in the sense that the correspondences between the primitives of the model and the entities being modeled are more immediate, giving a methodology for designing and implementing systems. We also conjecture that the complexity of model checking strategic ability for our representations is higher if we assume perfect information than if we assume imperfect information. The solutions, proposed in this paper, are not necessarily perfect (for example, the impression functions ini seem to be the main source of non-modularity in MIS, and can be perhaps improved), but we believe them to be a step in the right direction. We also do not mean to claim that our representations should replace more elaborate modeling languages like Promela or reactive modules. We only suggest that there is a need for compact, modular and reasonably grounded models that are more expressive than concurrent (epistemic) programs, and still allow for easier theoretical analysis than reactive modules. We also suggest that MIS might be better suited for modeling simple multi-agent domains, especially for human-oriented (as opposed to computer-oriented) design. 6. ACKNOWLEDGMENTS We thank the anonymous reviewers and Andrzej Tarlecki for their helpful remarks. Thomas Ågotnes" work on this paper was supported by grants 166525/V30 and 176853/S10 from the Research Council of Norway. 7. REFERENCES [1] R. Alur and T. A. Henzinger. Reactive modules. Formal Methods in System Design, 15(1):7-48, 1999. [2] R. Alur, T. A. Henzinger, and O. Kupferman. Alternating-time Temporal Logic. Lecture Notes in Computer Science, 1536:23-60, 1998. [3] R. Alur, T. A. Henzinger, and O. Kupferman. Alternating-time Temporal Logic. Journal of the ACM, 49:672-713, 2002. [4] R. Alur, T.A. Henzinger, F.Y.C. Mang, S. Qadeer, S.K. Rajamani, and S. Tasiran. MOCHA user manual. In Proceedings of CAV"98, volume 1427 of Lecture Notes in Computer Science, pages 521-525, 1998. [5] E.M. Clarke, E.A. Emerson, and A.P. Sistla. Automatic verification of finite-state concurrent systems using temporal logic specifications. ACM Transactions on Programming Languages and Systems, 8(2):244-263, 1986. [6] E.A. Emerson and J.Y. Halpern. "sometimes" and "not never" revisited: On branching versus linear time temporal logic. In Proceedings of the Annual ACM Symposium on Principles of Programming Languages, pages 151-178, 1982. [7] R. Fagin, J. Y. Halpern, Y. Moses, and M. Y. Vardi. Reasoning about Knowledge. MIT Press: Cambridge, MA, 1995. [8] M. Franceschet, A. Montanari, and M. de Rijke. Model checking for combined logics. In Proceedings of the 3rd International Conference on Temporal Logic (ICTL), 2000. [9] V. Goranko and W. Jamroga. Comparing semantics of logics for multi-agent systems. Synthese, 139(2):241-280, 2004. [10] J. Y. Halpern. Reasoning about knowledge: a survey. In D. M. Gabbay, C. J. Hogger, and J. A. Robinson, editors, The Handbook of Logic in Artificial Intelligence and Logic Programming, Volume IV, pages 1-34. Oxford University Press, 1995. [11] J.Y. Halpern and R. Fagin. Modelling knowledge and action in distributed systems. Distributed Computing, 3(4):159-177, 1989. [12] W. Jamroga and J. Dix. Do agents make model checking explode (computationally)? In M. P˘echou˘cek, P. Petta, and L.Z. Varga, editors, Proceedings of CEEMAS 2005, volume 3690 of Lecture Notes in Computer Science, pages 398-407. Springer Verlag, 2005. [13] W. Jamroga and J. Dix. Model checking abilities of agents: A closer look. Submitted, 2006. [14] W. Jamroga and T. Ågotnes. Modular interpreted systems: A preliminary report. Technical Report IfI-06-15, Clausthal University of Technology, 2006. [15] O. Kupferman, M.Y. Vardi, and P. Wolper. An automata-theoretic approach to branching-time model checking. Journal of the ACM, 47(2):312-360, 2000. [16] F. Laroussinie, N. Markey, and G. Oreiby. Expressiveness and complexity of ATL. Technical Report LSV-06-03, CNRS & ENS Cachan, France, 2006. [17] K.L. McMillan. Symbolic Model Checking: An Approach to the State Explosion Problem. Kluwer Academic Publishers, 1993. [18] K.L. McMillan. Applying SAT methods in unbounded symbolic model checking. In Proceedings of CAV"02, volume 2404 of Lecture Notes in Computer Science, pages 250-264, 2002. [19] W. Penczek and A. Lomuscio. Verifying epistemic properties of multi-agent systems via bounded model checking. In Proceedings of AAMAS"03, pages 209-216, New York, NY, USA, 2003. ACM Press. [20] F. Raimondi and A. Lomuscio. The complexity of symbolic model checking temporal-epistemic logics. In L. Czaja, editor, Proceedings of CS&P"05, 2005. [21] P. Y. Schobbens. Alternating-time logic with imperfect recall. Electronic Notes in Theoretical Computer Science, 85(2), 2004. [22] W. van der Hoek, A. Lomuscio, and M. Wooldridge. On the complexity of practical ATL model checking. In P. Stone and G. Weiss, editors, Proceedings of AAMAS"06, pages 201-208, 2006. 904 The Sixth Intl. Joint Conf. on Autonomous Agents and Multi-Agent Systems (AAMAS 07)

temporal and strategic logic;synchronous concurrent program;reactive module;model methodology;multi-agent system;higher level representation language;kripke structure;modeling methodology;open computational system;branching time;alternating-time temporal logic;model checking;computation tree logic ctl;model check;modular interpreted system

train_I-47

Operational Semantics of Multiagent Interactions

The social stance advocated by institutional frameworks and most multi-agent system methodologies has resulted in a wide spectrum of organizational and communicative abstractions which have found currency in several programming frameworks and software platforms. Still, these tools and frameworks are designed to support a limited range of interaction capabilities that constrain developers to a fixed set of particular, pre-defined abstractions. The main hypothesis motivating this paper is that the variety of multi-agent interaction mechanisms - both, organizational and communicative, share a common semantic core. In the realm of software architectures, the paper proposes a connector-based model of multi-agent interactions which attempts to identify the essential structure underlying multi-agent interactions. Furthermore, the paper also provides this model with a formal execution semantics which describes the dynamics of social interactions. The proposed model is intended as the abstract machine of an organizational programming language which allows programmers to accommodate an open set of interaction mechanisms.

1. INTRODUCTION The suitability of agent-based computing to manage the complex patterns of interactions naturally occurring in the development of large scale, open systems, has become one of its major assets over the last few years [26, 24, 15]. Particularly, the organizational or social stance advocated by institutional frameworks [2] and most multi-agent system (MAS) methodologies [26, 10], provides an excellent basis to deal with the complexity and dynamism of the interactions among system components. This approach has resulted in a wide spectrum of organizational and communicative abstractions, such as institutions, normative positions, power relationships, organizations, groups, scenes, dialogue games, communicative actions (CAs), etc., to effectively model the interaction space of MAS. This wealth of computational abstractions has found currency in several programming frameworks and software platforms (AMELI [9], MadKit [13], INGENIAS toolkit [18], etc.), which leverage multi-agent middlewares built upon raw ACL-based interaction mechanism [14], and minimize the gap between organizational metamodels and target implementation languages. Still, these tools and frameworks are designed to support a limited range of interaction capabilities that constrain developers to a fixed set of particular, pre-defined abstractions. The main hypothesis motivating this paper is that the variety of multi-agent interaction mechanisms - both, organizational and communicative, share a common semantic core. This paper thus focuses on the fundamental building blocks of multi-agent interactions: those which may be composed, extended or refined in order to define more complex organizational or communicative types of interactions. Its first goal is to carry out a principled analysis of multiagent interactions, departing from general features commonly ascribed to agent-based computing: autonomy, situatedness and sociality [26]. To approach this issue, we draw on the notion of connector, put forward within the field of software architectures [1, 17]. The outcome of this analysis will be a connector-based model of multi-agent interactions between autonomous social and situated components, i.e. agents, attempting to identify their essential structure. Furthermore, the paper also provides this model with a formal execution semantics which describes the dynamics of multi-agent (or social) interactions. Structural Operational Semantics (SOS)[21], a common technique to specify the operational semantics of programming languages, is used for this purpose. The paper is structured as follows: first, the major entities and relationships which constitute the structure of social interactions are introduced. Next, the dynamics of social interactions will show how these entities and relationships evolve. Last, relevant work in the literature is discussed 889 978-81-904262-7-5 (RPS) c 2007 IFAAMAS with respect to the proposal, limitations are addressed, and current and future work is described. 2. SOCIAL INTERACTION STRUCTURE From an architectural point of view, interactions between software components are embodied in software connectors: first-class entities defined on the basis of the different roles played by software components and the protocols that regulate their behaviour [1]. The roles of a connector represent its participants, such as the caller and callee roles of an RPC connector, or the sender and receiver roles in a message passing connector. The attachment operation binds a component to the role of a given connector. The analysis of social interactions introduced in this section gives rise to a new kind of social connector. It refines the generic model in several respects, attending to the features commonly ascribed to agent-based computing: • According to the autonomy feature, we may distinguish a first kind of participant (i.e. role) in a social interaction, so-called agents. Basically, agents are those software components which will be regarded as autonomous within the scope of the interaction1 . • A second group of participants, so-called environmental resources, may be identified from the situatedness feature. Unlike agents, resources represent those nonautonomous components whose state may be externally controlled by other components (agents or resources) within the interaction. Moreover, the participation of resources in an interaction is not mandatory. • Last, according to the sociality of agents, the specification of social connector protocols - the glue linking agents among themselves and with resources, will rely on normative concepts such as permissions, obligations and empowerments [23]. Besides agents, resources and social protocols, two other kinds of entities are of major relevance in our analysis of social interactions: actions, which represent the way in which agents alter the environmental and social state of the interaction; and events, which represent the changes in the interaction resulting from the performance of actions or the activity of environmental resources. In the following, we describe the basic entities involved in social interactions. Each kind of entity T will be specified as a record type T l1 : T1, . . . ln : Tn , possibly followed by a number of invariants, definitions, and the actions affecting their state. Instances or values v of a record type T will be represented as v = v1, . . . , vn : T. The type SetT represents a collection of values drawn from type T. The type QueueT represents a queue of values v : T waiting to be processed. The value v in the expression [v| ] : Queue[T] represents the head of the queue. The type Enum {v1, . . . , vn} 1 Note that we think of the autonomy feature in a relative, rather than absolute, perspective. Basically, this means that software components counting as agents in a social interaction may behave non-autonomously in other contexts, e.g. in their interactions through human-user interfaces. This conceptualization of agenthood resembles the way in which objects are understood in CORBA: as any kind of software component (C, Prolog, Cobol, etc.) attached to an ORB. represents an enumeration type whose values are v1, . . . , vn. Given some value v : T, the term vl refers to the value of the field l of a record type T. Given some labels l1, l2, . . . , the expression vl1,l2,... is syntactic sugar for ((vl1 )l2 ) . . .. The special term nil will be used to represent the absence of proper value for an optional field, so that vl = nil will be true in those cases and false otherwise. The formal model will be illustrated with several examples drawn from the design of a virtual organization to aid in the management of university courses. 2.1 Social Interactions Social interactions shall be considered as composite connectors [17], structured in terms of a tree of nested subinteractions. Let"s consider an interaction representing a university course (e.g. on data structures). On the one hand, this interaction is actually a complex one, made up of lower-level interactions. For instance, within the scope of the course agents will participate in programming assignment groups, lectures, tutoring meetings, examinations and so on. Assignment groups, in turn, may hold a number of assignment submissions and test requests interactions. A test request may also be regarded as a complex interaction, ultimately decomposed in the atomic, or bottom-level interactions represented by communicative actions (e.g. request, agree, refuse, . . . ). On the other hand, courses are run within the scope of a particular degree (e.g. computer science), a higher-level interaction. Traversing upwards from a degree to its ancestors, we find its faculty, the university and, finally, the multi-agent community or agent society. The community is thus the top-level interaction which subsumes any other kind of multi-agent interaction2 . The organizational and communicative interaction types identified above clearly differ in many ways. However, we may identify four major components in all of them: the participating agents, the resources that agents manipulate, the protocol regulating the agent activities and the subinteraction space. Accordingly, we may specify the type I of social interactions, ranged over by the meta-variable i, as follows: I state : SI, ini : A, mem : Set A, env : Set R, sub : Set I, prot : P, ch : CH def. : (1) icontext = i1 ⇔ i ∈ isub 1 inv. : (2) iini = nil ⇔ icontext = nil act. : setUp, join, create, destroy where the member and environment fields represent the agents (A) and local resources (R) participating in the interaction; the sub-interaction field, its set of inner interactions; and the protocol field the rules that govern the interaction (P). The event channel, to be described in the next section, allows the dispatching of local events to external interactions. The context of some interaction is defined as its super-interaction (def. 1), so that the context of the toplevel interaction is nil. The type SI Enum {open, closing, closed} represents the possible execution states of the interaction. Any interaction, but the top-level one, is set up within the context of another interaction by an initiator agent. The initiator is 2 In the context of this application, a one-to-one mapping between human users and software components attached to the community as agents would be a right choice. 890 The Sixth Intl. Joint Conf. on Autonomous Agents and Multi-Agent Systems (AAMAS 07) thus a mandatory feature for any interaction different to the community (inv. 2). The life-cycle of the interaction begins in the open state. Its sets of agent and resource participants, initially empty, vary as agents join and leave the interaction, and as they create and destroy resources from its local environment. Eventually, the interaction may come to an end (according to the protocol"s rules), or be explicitly closed by some agent, thus prematurely disabling the activity of its participants. The transient closing state will be described in the next section. 2.2 Agents Components attach themselves as agents in social interactions with the purpose of achieving something. The purpose declared by some agent when it joins an interaction shall be regarded as the institutional goal that it purports to satisfy within that context3 . The types of agents participating in a given interaction are primarily identified from their purposes. For instance, students are those agents participating in a course who purport to obtain a certificate in the course"s subject. Other members of the course include lecturers and teaching assistants. The type A of agents, ranged over by meta-variable a, is defined as follows: A state : SA, player : A, purp : F, att : Queue ACT , ev : Queue E, obl : Set O def. : (3) acontext = i ⇔ a ∈ imem (4) a1 ∈ aroles ⇔ aplayer 1 = a (5) i ∈ apartIn ⇔ a1 ∈ imem ∧ a1 ∈ aroles act. : see where the purpose is represented as a well-formed boolean formula, of a generic type F, which evaluates to true if the purpose is satisfied and false otherwise. The context of some agent is defined as the interaction in which it participates (def. 3). The type SA Enum {playing, leaving, succ, unsuc} represents the execution state of the agent. Its life-cycle begins in the playing state when its player agent joins the interaction, or some software component is attached as an agent to the multi-agent system (in this latter case, the player value is nil). The derived roles and partIn features represent the roles played by the agent and the contexts in which these roles are played (def. 4, 5)4 . An agent may play roles at interactions within or outside the scope of its context. For instance, students of a course are played by student agents belonging to the (undergraduate) degree, whereas lecturers may be played by teachers of a given department and the assistant role may be played by students of a Ph.D degree (both, the department and the Ph.D. degrees, are modelled as sub-interactions of the faculty). Components will normally attempt to perform different actions (e.g. to set up sub-interactions) in order to satisfy their purposes within some interaction. Moreover, components need to be aware of the current state of the interaction, so that they will also be capable of observing certain events from the interaction. Both, the visibility of the interaction 3 Thus, it may or may not correspond to actual internal goals or intentions of the component. 4 Free variables in the antecedents/consequents of implications shall be understood as universally/existentially quantified. and the attempts of members, are subject to the rules governing the interaction. The attempts and events fields of the agent structure represent the queues of attempts to execute some actions (ACT ), and the events (E) received by the agent which have not been observed yet. An agent may update its event queue by seeing the state of some entity of the community. The last field of the structure represents the obligations (O) of agents, to be described later. Eventually, the participation of some agent in the interaction will be over. This may either happen when certain conditions are met (specified by the protocol rules), or when the agent takes the explicit decision of leaving the interaction. In either case, the final state of the agent will be successful if its purpose was satisfied; unsuccessful otherwise. The transient leaving state will be described in the next section. 2.3 Resources Resources are software components which may represent different types of non-autonomous informational or computational entities. For instance, objectives, topics, assignments, grades and exams are different kinds of informational resources created by lecturers and assistants in the context of the course interaction. Students may also create programs to satisfy the requirements of some assignment. Other types of computational resources put at the disposal of students by teachers include compilers and interpreters. The type R of resources, ranged over by meta-variable r, can be specified by the following record type: R cr : A, owners : Set A, op : Set OP def. : (6) rcontext = i ⇔ r ∈ ienv act. : take, share, give, invoke Essentially, resources can be regarded as objects deployed in a social setting. This means that resources are created, accessed and manipulated by agents in a social interaction context (def. 6), according to the rules specified by its protocol. The mandatory feature creator represents the agent who created this resource. Moreover, resources may have owners. The ownership relationship between members and resources is considered as a normative device aimed at the simplification of the protocol"s rules that govern the interaction of agents and the environment. Members may gain ownership of some resource by taking it, and grant ownership to other agents by giving or sharing their own properties. For instance, the ownership of programs may be shared by several students if the assignment can be performed by groups of two or more students. The last operations feature represents the interface of the resource, consisting of a set of operations. A resource is structured around several public operations that participants may invoke, in accordance to the rules specified by the interaction"s protocol. The set of operations of a resource makes up its interface. 2.4 Protocols The protocol of any interaction is made up of the rules which govern its overall state and dynamics. The present specification abstracts away the particular formalism used to specify these rules, and focuses instead on several requirements concerning the structure and interface of protocols. Accordingly, the type P of protocols, ranged over by metaThe Sixth Intl. Joint Conf. on Autonomous Agents and Multi-Agent Systems (AAMAS 07) 891 variable p, is defined as follows5 : P emp : A × ACT → Boolean, perm : A × ACT → Boolean, obl :→ Set (A × Set O × Set E), monitor : E → Set A, finish :→ Boolean, over : A → Boolean def. : (7) pcontext = i ⇔ p = iprot inv. : (8) pfinish() ∧ s ∈ pcontext,sub ⇒ sprot,finish() (9) pfinish() ∧ a ∈ pcontext,mem ⇒ pover(a) (10) pover(a) ∧ ai ∈ aroles ⇒ acontext,prot,over i (ai) (11) αadd ∪ {a} ⊆ pmonitor( a, α, ) act. : Close, Leave We demand from protocols four major kinds of functions. Firstly, protocols shall include rules to identify the empowerments and permissions of any agent attempting to alter the state of the interaction (e.g. its members, the environment, etc.) through the execution of some action (e.g. join, create, etc.). Empowerments shall be regarded as the institutional capabilities which some agent possesses in order to satisfy its purpose. Corresponding rules, encapsulated by the empowered function field, shall allow to determine whether some agent is capable to perform a given action over the interaction. Empowerments may only be exercised under certain circumstances - that permissions specify. Permission rules shall allow to determine whether the attempt of an empowered agent to perform some particular action is satisfied or not (cf. permitted field). For instance, the course"s protocol specifies that the agents empowered to join the interaction as students are those students of the degree who have payed the fee established for the course"s subject, and own the certificates corresponding to its prerequisite subjects. Permission rules, in turn, specify that those students may only join the course in the admission stage. Hence, even if some student has paid the fee, the attempt to join the course will fail if the course has not entered the corresponding stage6 . Secondly, protocols shall allow to determine the obligations of agents towards the interaction. Obligations represent a normative device of social enforcement, fully compatible with the autonomy of agents, used to bias their behaviour in a certain direction. These kinds of rules shall allow to determine whether some agent must perform an action of a given type, as well as if some obligation was fulfilled, violated or needs to be revoked. The function obligations of the protocol structure thus identifies the agents whose obligation set must be updated. Moreover, it returns for each agent a collection of events representing the changes in the obligation set. For instance, the course"s protocol establishes that members of departments must join the course as teachers whenever they are assigned to the course"s subject. Thirdly, the protocol shall allow to specify monitoring rules for the different events originating within the interaction. Corresponding rules shall establish the set of agents that must be awared of some event. For instance, this func5 The formalization assumes that protocol"s functions implicitly recieve as input the interaction being regulated. 6 The hasPaidFee relationship between (degree) students and subject resources is represented by an additional, application-dependent field of the agent structure for this kind of roles. Similarly, the admission stage is an additional boolean field of the structure for school interactions. The generic types I, A, R and P are thus extendable. tionality is exploited by teachers in order to monitor the enrollment of students to the course. Last, the protocol shall allow to control the state of the interaction as well as the states of its members. Corresponding rules identify the conditions under which some interaction will be automatically finished, and whether the participation of some member agent will be automatically over. Thus, the function field finish returns true if the regulated interaction must finish its execution. If so happens, a well-defined set of protocols must ensure that its sub-interactions and members are finished as well (inv. 8,9). Similarly, the function over returns true if the participation of the specified member must be over. Well-formed protocols must ensure the consistency between these functions across playing roles (inv. 10)7 . For instance, the course"s protocol establishes that the participation of students is over when they gain ownership of the course"s certificate or the chances to get it are exhausted. It also establishes that the course must be finished when the admission stage has passed and all the students finished their participation. 3. SOCIAL INTERACTION DYNAMICS The dynamics of the multi-agent community is influenced by the external actions executed by software components and the protocols governing their interactions. This section focuses on the dynamics resulting from a particular kind of external action: the attempt of some component, attached to the community as an agent, to execute a given (internal) action. The description of other external actions concerning agents (e.g. observe the events from its event queue, enter or exit from the community) and resources (e.g. a timer resource may signal the pass of time) will be skipped. The processing of some attempt may give rise to changes in the scope of the target interaction, such as the instantiation of new participants (agents or resources) or the setting up of new sub-interactions. These resulting events may cause further changes in the state of other interactions (the target one included), namely, in its execution state as well as in the execution state, obligations and visibility of their members. This section will also describe the way in which these events are processed. The resulting dynamics described bellow allows for actions and events corresponding to different agents and interactions to be processed simultaneously. Due to lack of space, we only include some of the operational rules that formalise the execution semantics. 3.1 Attempt processing An attempt is defined by the structure AT T perf : A, act : ACT , where the performer represents the agent in charge of executing the specified action. This action is intended to alter the state of some target interaction (possibly, the performer"s context itself), and notify a collection of addressees of the changes resulting from a successful execution. Accordingly, the type ACT of actions, ranged over by meta-variable α, is specified as follows: ACT state : SACT , target : I, add : Set A def. : (12) αperf = a ⇔ α ∈ aatt 7 The close and leave actions update the finish and over function fields as explained in the next section. Additional actions, such as permit, forbid, empower, etc., to update other protocol"s fields are yet to be identified in future work. 892 The Sixth Intl. Joint Conf. on Autonomous Agents and Multi-Agent Systems (AAMAS 07) where: the performer is formally defined as the agent who stores the action in its queue of attempts, and the state field represents the current phase of processing. This process goes through four major phases, as specified by the enumeration type SACT Enum {emp, perm, exec} : empowerment checking, permission checking and action execution, described in the sequel. 3.1.1 Empowerment checking The post-condition of an attempt consists of inserting the action in the queue of attempts of the specified performer. As rule 1 specifies8 , this will only be possible if the performer is empowered to execute that action according to the rules that govern the state of the target interaction. If this condition is not met, the attempt will simply be ignored. Moreover, the performer agent must be in the playing state (this pre-condition is also required for any rule concerning the processing of attempts). If these pre-conditions are satisfied the rule is fired and the processing of the action continues in the permission checking stage. For instance, when the software component attached as a student in a degree attempts to join as a student the course in which some subject is teached, the empowerment rules of the course interaction are checked. If the (degree) student has passed the course"s prerequisite subjects the join action will be inserted in its queue of attempts and considered for execution. αtarget,prot,emp(a, α) a = playing, , , qACT , , a,α :AT T −→ playing, , , qACT , , (1) W here : (α )state = perm (qACT ) = insert(α , qACT ) 3.1.2 Permissions checking The processing of the action resumes when the possible preceding actions in the performer"s queue of attempts are fully processed and removed from the queue. Moreover, there should be no pending events to be processed in the interaction, for these events may cause the member or the interaction to be finished (as will be shortly explained in the next sub-section). If these conditions are met the permissions to execute the given action (and notify the specified addressees) are checked (e.g. it will be checked whether the student paid the fee for the course"s subject). If the protocol of the target interaction grants permission, the processing of the attempt moves to the action execution stage (rule 2). Otherwise, the action is discharged and removed from the queue. Unlike unempowered attempts, a forbidden one will cause an event to be generated and transfered to the event channel for further processing. αstate = perm ∧ acontext,ch,in,ev = ∅ ∧ αtarget,prot,perm(a, α) a = playing, , , [α| ], , −→ playing, , , [α | ], , (2) W here : (α )state = exec 8 Labels of record instances are omitted to allow for more compact specifications. Moreover, note that record updates in where clauses only affect the specified fields. 3.1.3 Action execution The transitions fired in this stage are classified according to the different types of actions to be executed. The intended effects of some actions may directly be achieved in a single step, while others will required an indirect approach and possibly several execution steps. Actions of the first kind are constructive ones such as set up and join. The second group of actions include those, such as close and leave, whose effects are indirectly achieved by updating the interaction protocol. As an example of constructive action, let"s consider the execution of a set up action, whose type is defined as follows9 : SetUp ACT · new : I inv. : (13) αnew,mem = αnew,res = αnew,sub = ∅ (14) αnew,state = open where the new field represents the new interaction to be initiated. Its sets of participants (agents and resources) and sub-interactions must be empty (inv. 13) and its state must be open (inv. 14). The setting up of the new interaction may thus affect its protocol and possible application-dependent fields (e.g. the subject of a course interaction). According to rule 3, the outcome of the execution is threefold: firstly, the performer"s attempt queue is updated so that the executing action is removed; secondly, the new interaction is added to the target"s set of sub-interactions (moreover, its initiator field is set to the performer agent); last, the event representing this change (which includes a description of the change, the agent that caused it and the action performed) is inserted in the output port of the target"s event channel. αstate = exec ∧ α : SetUp ∧ αnew = i a = playing, , , [α|qACT ], , −→ playing, , , qACT , , αtarget = open, , , , , sI , c −→ open, , , , , sI ∪ i , c (3) W here : (i )ini = a (c )out,ev = insert( a, α, sub(αtarget , i ) , cout,ev ) Let"s consider now the case of a close action. This action represents an attempt by the performer to force some interaction to finish, thus bypassing its current protocol rules (those concerning the finish function). The way to achieve this effect is to cause an update on the protocol so that the finish function returns true afterwards10 . Accordingly, we may specify this type of action as follows: Close ACT · upd : (→ Bool) → (→ Bool) inv. : (15) αtarget,state = open (16) αtarget,context = nil (17) αupd(αtarget,prot,finish)() where the inherited target field represents the interaction to be closed (which must be open and different to the topinteraction, according to invariants 15 and 16) and the new 9 The resulting type consists of the fields of the ACT record extended with an additional new field. 10 This strategy is also followed in the definition of leave and may also be used in the definition of other types of actions such as fire, permit, forbid, etc. The Sixth Intl. Joint Conf. on Autonomous Agents and Multi-Agent Systems (AAMAS 07) 893 update field represents a proper higher-order function to update the target"s protocol (inv. 17). The transition which models the execution of this action, specified by rule 4, defines two effects in the target interaction: its protocol is updated and the event representing this change is inserted in its output port. This event will actually trigger the closing process of the interaction as described in the next subsection. αstate = exec ∧ α : Close a = playing, , , [α|qACT ], , −→ playing, , , qACT , , αtarget = open, , , , , p, c −→ open, , , , , p , c (4) W here : (p )finish = αupd (pfinish ) (c )out,ev = insert( a, α, finish(αtarget ) , cout,ev ) 3.2 Event Processing The processing of events is encapsulated in the event channels of interactions. Channels, ranged over by meta-variable c, are defined by two input and output ports, according to the following definition: CH out : OutP, in : InP inv. : (18) ccontext ∈ cout,disp( , , finish(ccontext) ) (19) ccontext ∈ cout,disp( , , over(a) ) (20) ccontext,sub ⊆ cout,disp(closing(ccontext)) (21) apartsIn ⊆ cout,disp(leaving(a)) (22) ccontext ∈ cout,disp(closed(i)) (23) {ccontext, aplayer,context} ⊆ cout,disp(left(a)) OutP ev : Queue E, disp : E → Set I, int : Set I, ag : Set A InP ev : Queue E, stage : Enum {int, mem, obl}, ag : Set A The output port stores and processes the events originated within the scope of the channel"s interaction. Its first purpose is to dispatch the local events to the agents identified by the protocol"s monitoring function. Moreover, since these events may influence the results of the finishing, over and obligation functions of certain protocols, they will also be dispatched to the input ports of the interactions identified through a dispatching function - whose invariants will be explained later on. Thus, input ports serve as a coordination mechanism which activate the re-evaluation of the above functios whenever some event is received11 . Accordingly, the processing of some event goes through four major stages: event dispatching, interaction state update, member state update and obligations update. The first one takes place in the output port of the interaction in which the event originated, whereas the other ones execute in separate control threads associated to the input ports of the interactions to which the event was dispatched. 3.2.1 Event dispatching The processing of some event stored in the output port is triggered when all its preceding events have been dispatched. As a first step, the auxiliary int and ag fields are initialised 11 Alternatively, we may have assumed that interactions are fully aware of any change in the multi-agent community. In this scenario, interactions would trigger themselves without requiring any explicit notification. On the contrary, we adhere to the more realistic assumption of limited awareness. with the returned values of the dispatching and protocol"s monitoring functions, respectively (rule 5). Then, additional rules simply iterate over these collections until all agents and interactions have been notified (i.e., both sets are empty). Last, the event is removed from the queue and the auxiliary fields are re-set to nil. The dispatching function shall identify the set of interactions (possibly, empty) that may be affected by the event (which may include the channel"s interaction itself)12 . For instance, according to the finishing rule of university courses mentioned in the last section, the event representing the end of the admission stage, originated within the scope of the school interaction, will be dispatched to every course of the school"s degrees. Concerning the monitoring function, according to invariant 11 of protocols, if the event is generated as the result of an action performance, the agents to be notified will include the performer and addressees of that action. Thus, according to the monitoring rule of university courses, if a student of some degree joins a certain course and specifies a colleague as addressee of that action, the course"s teachers and itself will also be notified of the successful execution. ccontext,state s = open ∧ ccontext,prot,monitor s = mon cs = [e| ], d, nil, nil , −→ [e| ], , d(e), mon(e) , (5) 3.2.2 Interaction state update Input port activity is triggered when a new event is received. Irrespective of the kind of incoming event, the first processing action is to check whether the channel"s interaction must be finished. Thus, the dispatching of the finish event resulting from a close action (inv. 18) serves as a trigger of the closing procedure. If the interaction has not to be finished, the input port stage field is set to the member state update stage and the auxiliary ag field is initialised to the interaction members. Otherwise, we can consider two possible scenarios. In the first one, the interaction has no members and no sub-interactions. In this case, the interaction can be inmediately closed down. As rule 6 shows, the interaction is closed, removed from the context"s set of sub-interactions and a closed event is inserted in its output channel. According to invariant 22, this event will be later inserted to its input channel to allow for further treatment. cin,ev 1 = ∅ ∧ cin,stage 1 = int ∧ pfinish() , , , , {i} ∪ sI , , c −→ , , , , sI , , c i = , , ∅, , ∅, p, c1 −→ closed, , , , , , (6) W here : (c )out,ev = insert(closed(i), cout,ev ) In the second scenario, the interaction has some member or sub-interaction. In this case, clean-up is required prior to the disposal of the interaction (e.g. if the admission period ends and no student has matriculated for the course, teachers has to be finished before finishing the course itself). As rule 7 shows, the interaction is moved to the transient closing state and a corresponding event is inserted in the output port. According to invariant 20, the closing event will be dispatched to every sub-interaction in order to activate its closing procedure (guaranteed by invariant 8). Moreover, 12 This is essentially determined by the protocol rules of these interactions. The way in which the dispatching function is initialised and updated is out of the scope of this paper. 894 The Sixth Intl. Joint Conf. on Autonomous Agents and Multi-Agent Systems (AAMAS 07) the stage and ag fields are properly initialised so that the process goes on in the next member state update stage. This stage will further initiate the leaving process of the members (according to invariant 9). cin,ev = ∅ ∧ cin,stage = int ∧ pfinish() ∧ (sA = ∅ ∨ sI = ∅) i = open, , sA, , sI , p, c −→ closing, , sA, , sI , p, c (7) W here : (c )out,ev = insert(closing(i), cout,ev ) (c )in,stage = mem (c )in,ag = sA Eventually, every member will leave the interaction and every sub-interaction will be closed. Corresponding events will be received by the interaction (according to invariants 23 and 22) so that the conditions of the first scenario will hold. 3.2.3 Member state update This stage simply iterates over the members of the interaction to check whether they must be finished according to the protocol"s over function. When all members have been checked, the stage field will be set to the next obligation update stage and the auxiliary ag field will be initalised with the agents identified by the protocol"s obligation update function. If some member has to end its participation in the interaction and it is not playing any role, it will be inmediately abandoned (successfully or unsuccessfully, according to the satisfaction of its purpose). The corresponding event will be forwarded to its interaction and to the interaction of its player agent to account for further changes (inv. 23). Otherwise, the member enters the transient leaving state, thus preventing any action performance. Then, it waits for the completion of the leaving procedures of its played roles, triggered by proper dispatching of the leaving event (inv. 21). 3.2.4 Obligations update In this stage, the obligations of agents (not necessaryly members of the interaction) towards the interaction are updated accordingly. When all the identified agents have been updated, the event is removed from the input queue and the stage field is set back to the interaction state update. For instance, when a course interaction receives an event representing the assignment of some department member to its subject, an obligation to join the course as a teacher is created for that member. Moreover, the event representing this change is added to the output channel of the department interaction. 4. DISCUSSION This paper has attempted to expose a possible semantic core underlying the wide spectrum of interaction types between autonomous, social and situated software components. In the realm of software architectures, this core has been formalised as an operational model of social connectors, intended to describe both the basic structure and dynamics of multi-agent interactions, from the largest (the agent society itself) down to the smallest ones (communicative actions). Thus, top-level interactions may represent the kind of agent-web pursued by large-scale initiatives such as the Agentcities/openNet one [25]. Large-scale interactions, modelling complex aggregates of agent interactions such as those represented by e-institutions or virtual organizations [2, 26], are also amenable to be conceptualised as particular kinds of first-level social interactions. The last levels of the interaction tree may represent small-scale multiagent interactions such as those represented by interaction protocols [11], dialogue games [16], or scenes [2]. Finally, bottom-level interactions may represent communicative actions. From this perspective, the member types of a CA include the speaker and possibly many listeners. The purpose of the speaker coincides with the illocutionary purpose of the CA [22], whereas the purpose of any listener is to declare that it (actually, the software component) successfully processed the meaning of the CA. The analysis of social interactions put forward in this paper draws upon current proposals of the literature in several general respects, such as the institutional and organizational character of multi-agent systems [2, 26, 10, 7] and the normative perspective on multi-agent protocols [12, 23, 20]. These proposals as well as others focusing in relevant abstractions such as power relationships, contracts, trust and reputation mechanisms in organizational settings, etc., could be further exploited in order to characterize more accurately the organizational character of some multi-agent interactions. Similarly, the conceptualization of communicative actions as atomic interactions may similarly benefit from public semantics of communicative actions such as the one introduced in [3]. Last, the abstract model of protocols may be refined taking into account existing operational models of norms [12, 6]. These analyses shall result in new organizational and communicative abstractions obtained through a refinement and/or extension of the general model of social interactions. Thus, the proposed model is not intended to capture every organizational or communicative feature of multi-agent interactions, but to reveal their roots in basic interaction mechanisms. In turn, this would allow for the exploitation of common formalisms, particularly concerning protocols. Unlike the development of individual agents, which has greatly benefited from the design of several agent programming languages [4], societal features of multi-agent systems are mostly implemented in terms of visual modelling [8, 18] and a fixed set of interaction abstractions. We argue that the current field of multi-agent system programming may greatly benefit from multi-agent programming languages that allow programmers to accommodate an open set of interaction mechanisms. The model of social interactions put forward in this paper is intended as the abstract machine of a language of this type. This abstract machine would be independent of particular agent architectures and languages (i.e. software components may be programmed in a BDI language such as Jason [5] or in a non-agent oriented language). On top of the presented execution semantics, current and future work aims at the specification of the type system [19] which allows to program the abstract machine, the specification of the corresponding surface syntaxes (both textual and visual) and the design and implementation of a virtual machine over existing middleware technologies such as FIPA platforms or Web services. We also plan to study particular refinements and limitations to the proposed model, particularly with respect to the dispatching of events, semantics The Sixth Intl. Joint Conf. on Autonomous Agents and Multi-Agent Systems (AAMAS 07) 895 of obligations, dynamic updates of protocols and rule formalisms. In this latter aspect, we plan to investigate the use of Answer Set Programming to specify the rules of protocols, attending to the role that incompleteness (rules may only specify either necessary or sufficient conditions, for instance), explicit negation (e.g. prohibitions) and defaults play in this domain. 5. ACKNOWLEDGMENTS The authors thank anonymous reviewers for their comments and suggestions. Research sponsored by the Spanish Ministry of Science and Education (MEC), project TIN200615455-C03-03. 6. REFERENCES [1] R. Allen and D. Garlan. A Formal Basis for Architectural Connection. ACM Transactions on Software Engineering and Methodology, 6(3):213-249, June 1997. [2] J. L. Arcos, M. Esteva, P. Noriega, J. A. Rodr´ıguez, and C. Sierra. Engineering open environments with electronic institutions. Journal on Engineering Applications of Artificial Intelligence, 18(2):191-204, 2005. [3] G. Boella, R. Damiano, J. Hulstijn, and L. W. N. van der Torre. Role-based semantics for agent communication: embedding of the "mental attitudes" and "social commitments" semantics. In AAMAS, pages 688-690, 2006. [4] R. H. Bordini, L. Braubach, M. Dastani, A. E. F. Seghrouchni, J. J. G. Sanz, J. Leite, G. O"Hare, A. Pokahr, and A. Ricci. A survey of programming languages and platforms for multi-agent systems. Informatica, 30:33-44, 2006. [5] R. H. Bordini, J. F. H¨ubner, and R. Vieira. Jason and the golden fleece of agent-oriented programming. In R. H. Bordini, D. M., J. Dix, and A. El Fallah Seghrouchni, editors, Multi-Agent Programming: Languages, Platforms and Applications, chapter 1. Springer-Verlag, 2005. [6] O. Cliffe, M. D. Vos, and J. A. Padget. Specifying and analysing agent-based social institutions using answer set programming. In EUMAS, pages 476-477, 2005. [7] V. Dignum, J. V´azquez-Salceda, and F. Dignum. Omni: Introducing social structure, norms and ontologies into agent organizations. In R. Bordini, M. Dastani, J. Dix, and A. Seghrouchni, editors, Programming Multi-Agent Systems Second International Workshop ProMAS 2004, volume 3346 of LNAI, pages 181-198. Springer, 2005. [8] M. Esteva, D. de la Cruz, and C. Sierra. ISLANDER: an electronic institutions editor. In M. Gini, T. Ishida, C. Castelfranchi, and W. L. Johnson, editors, Proceedings of the First International Joint Conference on Autonomous Agents and Multiagent Systems (AAMAS"02), pages 1045-1052. ACM Press, July 2002. [9] M. Esteva, B. Rosell, J. A. Rodr´ıguez-Aguilar, and J. L. Arcos. AMELI: An agent-based middleware for electronic institutions. In Proceedings of the Third International Joint Conference on Autonomous Agents and Multiagent Systems, volume 1, pages 236-243, 2004. [10] J. Ferber, O. Gutknecht, and F. Michel. From agents to organizations: An organizational view of multi-agent systems. In AOSE, pages 214-230, 2003. [11] Foundation for Intelligent Physical Agents. FIPA Interaction Protocol Library Specification. http://www.fipa.org/repository/ips.html, 2003. [12] A. Garc´ıa-Camino, J. A. Rodr´ıguez-Aguilar, C. Sierra, and W. Vasconcelos. Norm-oriented programming of electronic institutions. In AAMAS, pages 670-672, 2006. [13] O. Gutknecht and J. Ferber. The MadKit agent platform architecture. Lecture Notes in Computer Science, 1887:48-55, 2001. [14] JADE. The JADE project home page. http://jade.cselt.it, 2005. [15] M. Luck, P. McBurney, O. Shehory, and S. Willmott. Agent Technology: Computing as Interaction - A Roadmap for Agent-Based Computing. AgentLink III, 2005. [16] P. McBurney and S. Parsons. A formal framework for inter-agent dialogues. In J. P. M¨uller, E. Andre, S. Sen, and C. Frasson, editors, Proceedings of the Fifth International Conference on Autonomous Agents, pages 178-179, Montreal, Canada, May 2001. ACM Press. [17] N. R. Mehta, N. Medvidovic, and S. Phadke. Towards a taxonomy of software connectors. In Proceedings of the 22nd International Conference on Software Engineering, pages 178-187. ACM Press, June 2000. [18] J. Pav´on and J. G´omez-Sanz. Agent oriented software engineering with ingenias. In V. Marik, J. Muller, and M. Pechoucek, editors, Proceedings of the 3rd International Central and Eastern European Conference on Multi-Agent Systems. Springer Verlag, 2003. [19] B. C. Pierce. Types and Programming Languages. The MIT Press, Cambridge, MA, 2002. [20] J. Pitt, L. Kamara, M. Sergot, and A. Artikis. Voting in multi-agent systems. Feb. 27 2006. [21] G. Plotkin. A structural approach to operational semantics. Technical Report DAIMI FN-19, Aarhus University, Sept. 1981. [22] J. Searle. Speech Acts. Cambridge University Press, 1969. [23] M. Sergot. A computational theory of normative positions. ACM Transactions on Computational Logic, 2(4):581-622, Oct. 2001. [24] M. P. Singh. Agent-based abstractions for software development. In F. Bergenti, M.-P. Gleizes, and F. Zambonelli, editors, Methodologies and Software Engineering for Agent Systems, chapter 1, pages 5-18. Kluwer, 2004. [25] S. Willmot and al. Agentcities / opennet testbed. http://x-opennet.net, 2004. [26] F. Zambonelli, N. R. Jennings, and M. Wooldridge. Developing multiagent systems: The Gaia methodology. ACM Transactions on Software Engineering and Methodology, 12(3):317-370, July 2003. 896 The Sixth Intl. Joint Conf. on Autonomous Agents and Multi-Agent Systems (AAMAS 07)

social interaction;operational semantics;organizational and communicative abstraction;organizational programming language;structural operational semantics;multi-agent interaction connector-based model;connector-based model of multi-agent interaction;institutional framework;pre-defined abstraction;formal execution semantics;multiagent interaction;software architecture;software connector

train_I-48

Normative System Games

We develop a model of normative systems in which agents are assumed to have multiple goals of increasing priority, and investigate the computational complexity and game theoretic properties of this model. In the underlying model of normative systems, we use Kripke structures to represent the possible transitions of a multiagent system. A normative system is then simply a subset of the Kripke structure, which contains the arcs that are forbidden by the normative system. We specify an agent"s goals as a hierarchy of formulae of Computation Tree Logic (CTL), a widely used logic for representing the properties of Kripke structures: the intuition is that goals further up the hierarchy are preferred by the agent over those that appear further down the hierarchy. Using this scheme, we define a model of ordinal utility, which in turn allows us to interpret our Kripke-based normative systems as games, in which agents must determine whether to comply with the normative system or not. We then characterise the computational complexity of a number of decision problems associated with these Kripke-based normative system games; for example, we show that the complexity of checking whether there exists a normative system which has the property of being a Nash implementation is NP-complete.

1. INTRODUCTION Normative systems, or social laws, have proved to be an attractive approach to coordination in multi-agent systems [13, 14, 10, 15, 1]. Although the various approaches to normative systems proposed in the literature differ on technical details, they all share the same basic intuition that a normative system is a set of constraints on the behaviour of agents in the system; by imposing these constraints, it is hoped that some desirable objective will emerge. The idea of using social laws to coordinate multi-agent systems was proposed by Shoham and Tennenholtz [13, 14]; their approach was extended by van der Hoek et al. to include the idea of specifying a desirable global objective for a social law as a logical formula, with the idea being that the normative system would be regarded as successful if, after implementing it (i.e., after eliminating all forbidden actions), the objective formula was guaranteed to be satisfied in the system [15]. However, this model did not take into account the preferences of individual agents, and hence neglected to account for possible strategic behaviour by agents when deciding whether to comply with the normative system or not. This model of normative systems was further extended by attributing to each agent a single goal in [16]. However, this model was still too impoverished to capture the kinds of decision making that take place when an agent decides whether or not to comply with a social law. In reality, strategic considerations come into play: an agent takes into account not just whether the normative system would be beneficial for itself, but also whether other agents will rationally choose to participate. In this paper, we develop a model of normative systems in which agents are assumed to have multiple goals, of increasing priority. We specify an agent"s goals as a hierarchy of formulae of Computation Tree Logic (CTL), a widely used logic for representing the properties of Kripke structures [8]: the intuition is that goals further up the hierarchy are preferred by the agent over those that appear further down the hierarchy. Using this scheme, we define a model of ordinal utility, which in turn allows us to interpret our Kripkebased normative systems as games, in which agents must determine whether to comply with the normative system or not. We thus provide a very natural bridge between logical structures and languages and the techniques and concepts of game theory, which have proved to be very powerful for analysing social contract-style scenarios such as normative systems [3, 4]. We then characterise the computational complexity of a number of decision problems associated with these Kripke-based normative system games; for example, we show that the complexity of checking whether there exists a normative system which has the property of being a Nash implementation is NP-complete. 2. KRIPKE STRUCTURES AND CTL We use Kripke structures as our basic semantic model for multiagent systems [8]. A Kripke structure is essentially a directed graph, with the vertex set S corresponding to possible states of the system being modelled, and the relation R ⊆ S × S capturing the 881 978-81-904262-7-5 (RPS) c 2007 IFAAMAS possible transitions of the system; intuitively, these transitions are caused by agents in the system performing actions, although we do not include such actions in our semantic model (see, e.g., [13, 2, 15] for related models which include actions as first class citizens). We let S0 denote the set of possible initial states of the system. Our model is intended to correspond to the well-known interleaved concurrency model from the reactive systems literature: thus an arc corresponds to the execution of an atomic action by one of the processes in the system, which we call agents. It is important to note that, in contrast to such models as [2, 15], we are therefore here not modelling synchronous action. This assumption is not in fact essential for our analysis, but it greatly simplifies the presentation. However, we find it convenient to include within our model the agents that cause transitions. We therefore assume a set A of agents, and we label each transition in R with the agent that causes the transition via a function α : R → A. Finally, we use a vocabulary Φ = {p, q, . . .} of Boolean variables to express the properties of individual states S: we use a function V : S → 2Φ to label each state with the Boolean variables true (or satisfied) in that state. Collecting these components together, an agent-labelled Kripke structure (over Φ) is a 6-tuple: K = S, S0 , R, A, α, V , where: • S is a finite, non-empty set of states, • S0 ⊆ S (S0 = ∅) is the set of initial states; • R ⊆ S × S is a total binary relation on S, which we refer to as the transition relation1 ; • A = {1, . . . , n} is a set of agents; • α : R → A labels each transition in R with an agent; and • V : S → 2Φ labels each state with the set of propositional variables true in that state. In the interests of brevity, we shall hereafter refer to an agentlabelled Kripke structure simply as a Kripke structure. A path over a transition relation R is an infinite sequence of states π = s0, s1, . . . which must satisfy the property that ∀u ∈ N: (su , su+1) ∈ R. If u ∈ N, then we denote by π[u] the component indexed by u in π (thus π[0] denotes the first element, π[1] the second, and so on). A path π such that π[0] = s is an s-path. Let ΠR(s) denote the set of s-paths over R; since it will usually be clear from context, we often omit reference to R, and simply write Π(s). We will sometimes refer to and think of an s-path as a possible computation, or system evolution, from s. EXAMPLE 1. Our running example is of a system with a single non-sharable resource, which is desired by two agents. Consider the Kripke structure depicted in Figure 1. We have two states, s and t, and two corresponding Boolean variables p1 and p2, which are 1 In the branching time temporal logic literature, a relation R ⊆ S × S is said to be total iff ∀s ∃s : (s, s ) ∈ R. Note that the term total relation is sometimes used to refer to relations R ⊆ S × S such that for every pair of elements s, s ∈ S we have either (s, s ) ∈ R or (s , s) ∈ R; we are not using the term in this way here. It is also worth noting that for some domains, other constraints may be more appropriate than simple totality. For example, one might consider the agent totality requirement, that in every state, every agent has at least one possible transition available: ∀s∀i ∈ A∃s : (s, s ) ∈ R and α(s, s ) = i. 2p t p 2 2 1 s 1 1 Figure 1: The resource control running example. mutually exclusive. Think of pi as meaning agent i has currently control over the resource. Each agent has two possible actions, when in possession of the resource: either give it away, or keep it. Obviously there are infinitely many different s-paths and t-paths. Let us say that our set of initial states S0 equals {s, t}, i.e., we don"t make any assumptions about who initially has control over the resource. 2.1 CTL We now define Computation Tree Logic (CTL), a branching time temporal logic intended for representing the properties of Kripke structures [8]. Note that since CTL is well known and widely documented in the literature, our presentation, though complete, will be somewhat terse. We will use CTL to express agents" goals. The syntax of CTL is defined by the following grammar: ϕ ::= | p | ¬ϕ | ϕ ∨ ϕ | E fϕ | E(ϕ U ϕ) | A fϕ | A(ϕ U ϕ) where p ∈ Φ. We denote the set of CTL formula over Φ by LΦ; since Φ is understood, we usually omit reference to it. The semantics of CTL are given with respect to the satisfaction relation |=, which holds between pairs of the form K, s, (where K is a Kripke structure and s is a state in K), and formulae of the language. The satisfaction relation is defined as follows: K, s |= ; K, s |= p iff p ∈ V (s) (where p ∈ Φ); K, s |= ¬ϕ iff not K, s |= ϕ; K, s |= ϕ ∨ ψ iff K, s |= ϕ or K, s |= ψ; K, s |= A fϕ iff ∀π ∈ Π(s) : K, π[1] |= ϕ; K, s |= E fϕ iff ∃π ∈ Π(s) : K, π[1] |= ϕ; K, s |= A(ϕ U ψ) iff ∀π ∈ Π(s), ∃u ∈ N, s.t. K, π[u] |= ψ and ∀v, (0 ≤ v < u) : K, π[v] |= ϕ K, s |= E(ϕ U ψ) iff ∃π ∈ Π(s), ∃u ∈ N, s.t. K, π[u] |= ψ and ∀v, (0 ≤ v < u) : K, π[v] |= ϕ The remaining classical logic connectives (∧, →, ↔) are assumed to be defined as abbreviations in terms of ¬, ∨, in the conventional manner. The remaining CTL temporal operators are defined: A♦ϕ ≡ A( U ϕ) E♦ϕ ≡ E( U ϕ) A ϕ ≡ ¬E♦¬ϕ E ϕ ≡ ¬A♦¬ϕ We say ϕ is satisfiable if K, s |= ϕ for some Kripke structure K and state s in K; ϕ is valid if K, s |= ϕ for all Kripke structures K and states s in K. The problem of checking whether K, s |= ϕ for given K, s, ϕ (model checking) can be done in deterministic polynomial time, while checking whether a given ϕ is satisfiable or whether ϕ is valid is EXPTIME-complete [8]. We write K |= ϕ if K, s0 |= ϕ for all s0 ∈ S0 , and |= ϕ if K |= ϕ for all K. 882 The Sixth Intl. Joint Conf. on Autonomous Agents and Multi-Agent Systems (AAMAS 07) 3. NORMATIVE SYSTEMS For our purposes, a normative system is simply a set of constraints on the behaviour of agents in a system [1]. More precisely, a normative system defines, for every possible system transition, whether or not that transition is considered to be legal or not. Different normative systems may differ on whether or not a transition is legal. Formally, a normative system η (w.r.t. a Kripke structure K = S, S0 , R, A, α, V ) is simply a subset of R, such that R \ η is a total relation. The requirement that R\η is total is a reasonableness constraint: it prevents normative systems which lead to states with no successor. Let N (R) = {η : (η ⊆ R) & (R \ η is total)} be the set of normative systems over R. The intended interpretation of a normative system η is that (s, s ) ∈ η means transition (s, s ) is forbidden in the context of η; hence R \ η denotes the legal transitions of η. Since it is assumed η is reasonable, we are guaranteed that a legal outward transition exists for every state. We denote the empty normative system by η∅, so η∅ = ∅. Note that the empty normative system η∅ is reasonable with respect to any transition relation R. The effect of implementing a normative system on a Kripke structure is to eliminate from it all transitions that are forbidden according to this normative system (see [15, 1]). If K is a Kripke structure, and η is a normative system over K, then K † η denotes the Kripke structure obtained from K by deleting transitions forbidden in η. Formally, if K = S, S0 , R, A, α, V , and η ∈ N (R), then let K†η = K be the Kripke structure K = S , S0 , R , A , α , V where: • S = S , S0 = S0 , A = A , and V = V ; • R = R \ η; and • α is the restriction of α to R : α (s, s ) = j α(s, s ) if (s, s ) ∈ R undefined otherwise. Notice that for all K, we have K † η∅ = K. EXAMPLE 1. (continued) When thinking in terms of fairness, it seems natural to consider normative systems η that contain (s, s) or (t, t). A normative system with (s, t) would not be fair, in the sense that A♦A ¬p1 ∨ A♦A ¬p2 holds: in all paths, from some moment on, one agent will have control forever. Let us, for later reference, fix η1 = {(s, s)}, η2 = {(t, t)}, and η3 = {(s, s), (t, t)}. Later, we will address the issue of whether or not agents should rationally choose to comply with a particular normative system. In this context, it is useful to define operators on normative systems which correspond to groups of agents defecting from the normative system. Formally, let K = S, S0 ,R, A,α, V be a Kripke structure, let C ⊆ A be a set of agents over K, and let η be a normative system over K. Then: • η C denotes the normative system that is the same as η except that it only contains the arcs of η that correspond to the actions of agents in C. We call η C the restriction of η to C, and it is defined as: η C = {(s, s ) : (s, s ) ∈ η & α(s, s ) ∈ C}. Thus K † (η C) is the Kripke structure that results if only the agents in C choose to comply with the normative system. • η C denotes the normative system that is the same as η except that it only contains the arcs of η that do not correspond to actions of agents in C. We call η C the exclusion of C from η, and it is defined as: η C = {(s, s ) : (s, s ) ∈ η & α(s, s ) ∈ C}. Thus K † (η C) is the Kripke structure that results if only the agents in C choose not to comply with the normative system (i.e., the only ones who comply are those in A \ C). Note that we have η C = η (A\C) and η C = η (A\C). EXAMPLE 1. (Continued) We have η1 {1} = η1 = {(s, s)}, while η1 {1} = η∅ = η1 {2}. Similarly, we have η3 {1} = {(s, s)} and η3 {1} = {(t, t)}. 4. GOALS AND UTILITIES Next, we want to be able to capture the goals that agents have, as these will drive an agent"s strategic considerations - particularly, as we will see, considerations about whether or not to comply with a normative system. We will model an agent"s goals as a prioritised list of CTL formulae, representing increasingly desired properties that the agent wishes to hold. The intended interpretation of such a goal hierarchy γi for agent i ∈ A is that the further up the hierarchy a goal is, the more it is desired by i. Note that we assume that if an agent can achieve a goal at a particular level in its goal hierarchy, then it is unconcerned about goals lower down the hierarchy. Formally, a goal hierarchy, γ, (over a Kripke structure K) is a finite, non-empty sequence of CTL formulae γ = (ϕ0, ϕ1, . . . , ϕk ) in which, by convention, ϕ0 = . We use a natural number indexing notation to extract the elements of a goal hierarchy, so if γ = (ϕ0, ϕ1, . . . , ϕk ) then γ[0] = ϕ0, γ[1] = ϕ1, and so on. We denote the largest index of any element in γ by |γ|. A particular Kripke structure K is said to satisfy a goal at index x in goal hierarchy γ if K |= γ[x], i.e., if γ[x] is satisfied in all initial states S0 of K. An obvious potential property of goal hierarchies is monotonicity: where goals at higher levels in the hierarchy logically imply those at lower levels in the hierarchy. Formally, a goal hierarchy γ is monotonic if for all x ∈ {1, . . . , |γ|} ⊆ N, we have |= γ[x] → γ[x − 1]. The simplest type of monotonic goal hierarchy is where γ[x + 1] = γ[x] ∧ ψx+1 for some ψx+1, so at each successive level of the hierarchy, we add new constraints to the goal of the previous level. Although this is a natural property of many goal hierarchies, it is not a property we demand of all goal hierarchies. EXAMPLE 1. (continued) Suppose the agents have similar, but opposing goals: each agent i wants to keep the source as often and long as possible for himself. Define each agent"s goal hierarchy as: γi = ( ϕi 0 = , ϕi 1 = E♦pi , ϕi 2 = E E♦pi , ϕi 3 = E♦E pi , ϕi 4 = A E♦pi , ϕi 5 = E♦A pi ϕi 6 = A A♦pi , ϕi 7 = A (A♦pi ∧ E pi ), ϕi 8 = A pi ) The most desired goal of agent i is to, in every computation, always have the resource, pi (this is expressed in ϕi 8). Thanks to our reasonableness constraint, this goal implies ϕi 7 which says that, no matter how the computation paths evolve, it will always be that all The Sixth Intl. Joint Conf. on Autonomous Agents and Multi-Agent Systems (AAMAS 07) 883 continuations will hit a point in which pi , and, moreover, there is a continuation in which pi always holds. Goal ϕi 6 is a fairness constraint implied by it. Note that A♦pi says that every computation eventually reaches a pi state. This may mean that after pi has happened, it will never happen again. ϕi 6 circumvents this: it says that, no matter where you are, there should be a future pi state. The goal ϕi 5 is like the strong goal ϕi 8 but it accepts that this is only achieved in some computation, eventually. ϕi 4 requires that in every path, there is always a continuation that eventually gives pi . Goal ϕi 3 says that pi should be true on some branch, from some moment on. It implies ϕi 2 which expresses that there is a computation such that everywhere during it, it is possible to choose a continuation that eventually satisfies pi . This implies ϕi 1, which says that pi should at least not be impossible. If we even drop that demand, we have the trivial goal ϕi 0. We remark that it may seem more natural to express a fairness constraint ϕi 6 as A ♦pi . However, this is not a proper CTL formula. It is in fact a formula in CTL ∗ [9], and in this logic, the two expressions would be equivalent. However, our basic complexity results in the next sections would not hold for the richer language CTL ∗2 , and the price to pay for this is that we have to formulate our desired goals in a somewhat more cumbersome manner than we might ideally like. Of course, our basic framework does not demand that goals are expressed in CTL; they could equally well be expressed in CTL ∗ or indeed ATL [2] (as in [15]). We comment on the implications of alternative goal representations at the conclusion of the next section. A multi-agent system collects together a Kripke structure (representing the basic properties of a system under consideration: its state space, and the possible state transitions that may occur in it), together with a goal hierarchy, one for each agent, representing the aspirations of the agents in the system. Formally, a multi-agent system, M , is an (n + 1)-tuple: M = K, γ1, . . . , γn where K is a Kripke structure, and for each agent i in K, γi is a goal hierarchy over K. 4.1 The Utility of Normative Systems We can now define the utility of a Kripke structure for an agent. The idea is that the utility of a Kripke structure is the highest index of any goal that is guaranteed for that agent in the Kripke structure. We make this precise in the function ui (·): ui (K) = max{j : 0 ≤ j ≤ |γi | & K |= γi [j ]} Note that using these definitions of goals and utility, it never makes sense to have a goal ϕ at index n if there is a logically weaker goal ψ at index n + k in the hierarchy: by definition of utility, it could never be n for any structure K. EXAMPLE 1. (continued) Let M = K, γ1, γ2 be the multiagent system of Figure 1, with γ1 and γ2 as defined earlier in this example. Recall that we have defined S0 as {s, t}. Then, u1(K) = u2(K) = 4: goal ϕ4 is true in S0 , but ϕ5 is not. To see that ϕ2 4 = A E♦p2 is true in s for instance: note that on ever path it is always the case that there is a transition to t, in which p2 is true. Notice that since for any goal hierarchy γi we have γ[0] = , then for all Kripke structures, ui (K) is well defined, with ui (K) ≥ 2 CTL ∗ model checking is PSPACE-complete, and hence much worse (under standard complexity theoretic assumptions) than model checking CTL [8]. η δ1(K, η) δ2(K, η) η∅ 0 0 η1 0 3 η2 3 0 η3 2 2 C D C (2, 2) (0, 3) D (3, 0) (0, 0) Figure 2: Benefits of implementing a normative system η (left) and pay-offs for the game ΣM . 0. Note that this is an ordinal utility measure: it tells us, for any given agent, the relative utility of different Kripke structures, but utility values are not on some standard system-wide scale. The fact that ui (K1) > ui (K2) certainly means that i strictly prefers K1 over K2, but the fact that ui (K) > uj (K) does not mean that i values K more highly than j . Thus, it does not make sense to compare utility values between agents, and so for example, some system wide measures of utility, (notably those measures that aggregate individual utilities, such as social welfare), do not make sense when applied in this setting. However, as we shall see shortly, other measures - such as Pareto efficiency - can be usefully applied. There are other representations for goals, which would allow us to define cardinal utilities. The simplest would be to specify goals γ for an agent as a finite, non-empty, one-to-one relation: γ ⊆ L×R. We assume that the x values in pairs (ϕ, x) ∈ γ are specified so that x for agent i means the same as x for agent j , and so we have cardinal utility. We then define the utility for i of a Kripke structure K asui (K) = max{x : (ϕ, x) ∈ γi & K |= ϕ}. The results of this paper in fact hold irrespective of which of these representations we actually choose; we fix upon the goal hierarchy approach in the interests of simplicity. Our next step is to show how, in much the same way, we can lift the utility function from Kripke structures to normative systems. Suppose we are given a multi-agent system M = K, γ1, . . . , γn and an associated normative system η over K. Let for agent i, δi (K, K ) be the difference in his utility when moving from K to K : δi (K, K ) = ui (K )− ui (K). Then the utility of η to agent i wrt K is δi (K, K † η). We will sometimes abuse notation and just write δi (K, η) for this, and refer to it as the benefit for agent i of implementing η in K. Note that this benefit can be negative. Summarising, the utility of a normative system to an agent is the difference between the utility of the Kripke structure in which the normative system was implemented and the original Kripke structure. If this value is greater than 0, then the agent would be better off if the normative system were imposed, while if it is less than 0 then the agent would be worse off if η were imposed than in the original system. We say η is individually rational for i wrt K if δi (K, η) > 0, and individually rational simpliciter if η is individually rational for every agent. A social system now is a pair Σ = M , η where M is a multi-agent system, and η is a normative system over M . EXAMPLE 1. The table at the left hand in Figure 2 displays the utilities δi (K, η) of implementing η in the Kripke structure of our running example, for the normative systems η = η∅, η1, η2 and η3, introduced before. Recall that u1(K) = u2(K) = 4. 4.2 Universal and Existential Goals 884 The Sixth Intl. Joint Conf. on Autonomous Agents and Multi-Agent Systems (AAMAS 07) Keeping in mind that a norm η restricts the possible transitions of the model under consideration, we make the following observation, borrowing from [15]. Some classes of goals are monotonic or anti-monotonic with respect to adding additional constraints to a system. Let us therefore define two fragments of the language of CTL: the universal language Lu with typical element μ, and the existential fragment Le with typical element ε. μ ::= | p | ¬p | μ ∨ μ | A fμ | A μ | A(μ U μ) ε ::= | p | ¬p | ε ∨ ε | E fε | E♦ε | E(ε U ε) Let us say, for two Kripke structures K1 = S, S0 , R1, A, α, V and K2 = S, S0 , R2, A, α, V that K1 is a subsystem of K2 and K2 is a supersystem of K1, written K1 K2 iff R1 ⊆ R2. Note that typically K † η K. Then we have (cf. [15]). THEOREM 1. Suppose K1 K2, and s ∈ S. Then ∀ε ∈ Le : K1, s |= ε ⇒ K2, s |= ε ∀μ ∈ Lu : K2, s |= μ ⇒ K1, s |= μ This has the following effect on imposing a new norm: COROLLARY 1. Let K be a structure, and η a normative system. Let γi denote a goal hierarchy for agent i. 1. Suppose agent i"s utility ui (K) is n, and γi [n] ∈ Lu , (i.e., γi [n] is a universal formula). Then, for any normative system η, δi (K, η) ≥ 0. 2. Suppose agent i"s utility ui (K † η) is n, and γi [n] is an existential formula ε. Then, δi (K † η, K) ≥ 0. Corollary 1"s first item says that an agent whose current maximal goal in a system is a universal formula, need never fear the imposition of a new norm η. The reason is that his current goal will at least remain true (in fact a goal higher up in the hierarchy may become true). It follows from this that an agent with only universal goals can only gain from the imposition of normative systems η. The opposite is true for existential goals, according to the second item of the corollary: it can never be bad for an agent to undo a norm η. Hence, an agent with only existential goals might well fear any norm η. However, these observations implicitly assume that all agents in the system will comply with the norm. Whether they will in fact do so, of course, is a strategic decision: it partly depends on what the agent thinks that other agents will do. This motivates us to consider normative system games. 5. NORMATIVE SYSTEM GAMES We now have a principled way of talking about the utility of normative systems for agents, and so we can start to apply the technical apparatus of game theory to analyse them. Suppose we have a multi-agent system M = K, γ1, . . . , γn and a normative system η over K. It is proposed to the agents in M that η should be imposed on K, (typically to achieve some coordination objective). Our agent - let"s say agent i - is then faced with a choice: should it comply with the strictures of the normative system, or not? Note that this reasoning takes place before the agent is in the system - it is a design time consideration. We can understand the reasoning here as a game, as follows. A game in strategic normal form (cf. [11, p.11]) is a structure: G = AG, S1, . . . , Sn , U1, . . . , Un where: • AG = {1, . . . , n} is a set of agents - the players of the game; • Si is the set of strategies for each agent i ∈ AG (a strategy for an agent i is nothing else than a choice between alternative actions); and • Ui : (S1 × · · · × Sn ) → R is the utility function for agent i ∈ AG, which assigns a utility to every combination of strategy choices for the agents. Now, suppose we are given a social system Σ = M , η where M = K, γ1, . . . , γn . Then we can associate a game - the normative system game - GΣ with Σ, as follows. The agents AG in GΣ are as in Σ. Each agent i has just two strategies available to it: • C - comply (cooperate) with the normative system; and • D - do not comply with (defect from) the normative system. If S is a tuple of strategies, one for each agent, and x ∈ {C, D}, then we denote by AGx S the subset of agents that play strategy x in S. Hence, for a social system Σ = M , η , the normative system η AGC S only implements the restrictions for those agents that choose to cooperate in GΣ. Note that this is the same as η AGD S : the normative system that excludes all the restrictions of agents that play D in GΣ. We then define the utility functions Ui for each i ∈ AG as: Ui (S) = δi (K, η AGC S ). So, for example, if SD is a collection of strategies in which every agent defects (i.e., does not comply with the norm), then Ui (SD ) = δi (K, (η AGD SD )) = ui (K † η∅) − ui (K) = 0. In the same way, if SC is a collection of strategies in which every agent cooperates (i.e., complies with the norm), then Ui (SC ) = δi (K, (η AGD SC )) = ui (K † (η ∅)) = ui (K † η). We can now start to investigate some properties of normative system games. EXAMPLE 1. (continued) For our example system, we have displayed the different U values for our multi agent system with the norm η3, i.e., {(s, s), (t, t)} as the second table of Figure 2. For instance, the pair (0, 3) in the matrix under the entry S = C, D is obtained as follows. U1( C, D ) = δ1(K, η3 AGC C,D ) = u1(K † η3 AGC C,D ) − u1(K). The first term of this is the utility of 1 in the system K where we implement η3 for the cooperating agent, i.e., 1, only. This means that the transitions are R \ {(s, s)}. In this system, still ϕ1 4 = A E♦p1 is the highest goal for agent 1. This is the same utility for 1 as in K, and hence, δ1(K, η3 AGC C,D ) = 0. Agent 2 of course benefits if agent 1 complies with η3 while 2 does not. His utility would be 3, since η3 AGC C,D is in fact η1. 5.1 Individually Rational Normative Systems A normative system is individually rational if every agent would fare better if the normative system were imposed than otherwise. This is a necessary, although not sufficient condition on a norm to expect that everybody respects it. Note that η3 of our example is individually rational for both 1 and 2, although this is not a stable situation: given that the other plays C, i is better of by playing D. We can easily characterise individually rationality with respect to the corresponding game in strategic form, as follows. Let Σ = M , η be a social system. Then the following are equivalent: The Sixth Intl. Joint Conf. on Autonomous Agents and Multi-Agent Systems (AAMAS 07) 885 f(xk) ... s0 s1 s2 s3 s4 s(2k−1) s2k t(x1) f(x1) t(x2) f(x2) t(xk) Figure 3: The Kripke structure produced in the reduction of Theorem 2; all transitions are associated with agent 1, the only initial state is s0. 1. η is individually rational in M ; 2. ∀i ∈ AG, Ui (SC ) > Ui (SD ) in the game GΣ. The decision problem associated with individually rational normative systems is as follows: INDIVIDUALLY RATIONAL NORMATIVE SYSTEM (IRNS): Given: Multi-agent system M . Question: Does there exist an individually rational normative system for M ? THEOREM 2. IRNS is NP-complete, even in one-agent systems. PROOF. For membership of NP, guess a normative system η, and verify that it is individually rational. Since η ⊆ R, we will be able to guess it in nondeterministic polynomial time. To verify that it is individually rational, we check that for all i, we have ui (K † η) > ui (K); computing K † η is just set subtraction, so can be done in polynomial time, while determining the value of ui (K) for any K can be done with a polynomial number of model checking calls, each of which requires only time polynomial in the K and γ. Hence verifying that ui (K † η) > ui (K) requires only polynomial time. For NP-hardness, we reduce SAT [12, p.77]. Given a SAT instance ϕ over Boolean variables x1, . . . , xk , we produce an instance of IRNS as follows. First, we define a single agent A = {1}. For each Boolean variable xi in the SAT instance, we create two Boolean variables t(xi ) and f (xi ) in the IRNS instance. We then create a Kripke structure Kϕ with 2k + 1 states, as shown in Figure 3: arcs in this graph correspond to transitions in Kϕ. Let ϕ∗ be the result of systematically substituting for every Boolean variable xi in ϕ the CTL expression (E ft(xi )). Next, consider the following formulae: k^ i=1 E f(t(xi ) ∨ f (xi )) (1) k^ i=1 ¬((E ft(xi )) ∧ (E ff (xi ))) (2) We then define the goal hierarchy for all agent 1 as follows: γ1[0] = γ1[1] = (1) ∧ (2) ∧ ϕ∗ We claim there is an individually rational normative system for the instance so constructed iff ϕ is satisfiable. First, notice that any individually rational normative system must force γ1[1] to be true, since in the original system, we do not have γ1[1]. For the ⇒ direction, if there is an individually rational normative system η, then we construct a satisfying assignment for ϕ by considering the arcs that are forbidden by η: formula (1) ensures that we must forbid an arc to either a t(xi ) or a f (xi ) state for all variables xi , but (2) ensures that we cannot forbid arcs to both. So, if we forbid an arc to a t(xi ) state then in the corresponding valuation for ϕ we make xi false, while if we forbid an arc to a f (xi ) state then we make xi true. The fact that ϕ∗ is part of the goal ensures that the normative system is indeed a valuation for ϕ. For ⇐, note that for any satisfying valuation for ϕ we can construct an individually rational normative system η, as follows: if the valuation makes xi true, we forbid the arc to the f (xi ) state, while if the valuation makes xi false, we forbid the arc to the t(xi ) state. The resulting normative system ensures γ1[1], and is thus individually rational. Notice that the Kripke structure constructed in the reduction contains just a single agent, and so the Theorem is proven. 5.2 Pareto Efficient Normative Systems Pareto efficiency is a basic measure of how good a particular outcome is for a group of agents [11, p.7]. Intuitively, an outcome is Pareto efficient if there is no other outcome that makes every agent better off. In our framework, suppose we are given a social system Σ = M , η , and asked whether η is Pareto efficient. This amounts to asking whether or not there is some other normative system η such that every agent would be better off under η than with η. If η makes every agent better off than η, then we say η Pareto dominates η. The decision problem is as follows: PARETO EFFICIENT NORMATIVE SYSTEM (PENS): Given: Multi-agent system M and normative system η over M . Question: Is η Pareto efficient for M ? THEOREM 3. PENS is co-NP-complete, even for one-agent systems. PROOF. Let M and η be as in the Theorem. We show that the complement problem to PENS, which we refer to as PARETO DOMINATED, is NP-complete. In this problem, we are given M and η, and we are asked whether η is Pareto dominated, i.e., whether or not there exists some η over M such that η makes every agent better off than η. For membership of NP, simply guess a normative system η , and verify that for all i ∈ A, we have ui (K † η ) > ui (K † η) - verifying requires a polynomial number of model checking problems, each of which takes polynomial time. Since η ⊆ R, the normative system can be guessed in non-deterministic polynomial time. For NP-hardness, we reduce IRNS, which we know to be NPcomplete from Theorem 2. Given an instance M of IRNS, we let M in the instance of PARETO DOMINATED be as in the IRNS instance, and define the normative system for PARETO DOMINATED to be η∅, the empty normative system. Now, it is straightforward that there exists a normative system η which Pareto dominates η∅ in M iff there exist an individually rational normative system in M . Since the complement problem is NP-complete, it follows that PENS is co-NP-complete. 886 The Sixth Intl. Joint Conf. on Autonomous Agents and Multi-Agent Systems (AAMAS 07) η0 η1 η2 η3 η4 η5 η6 η7 η8 u1(K † η) 4 4 7 6 5 0 0 8 0 u2(K † η) 4 7 4 6 0 5 8 0 0 Table 1: Utilities for all possible norms in our example How about Pareto efficient norms for our toy example? Settling this question amounts to finding the dominant normative systems among η0 = η∅, η1, η2, η3 defined before, and η4 = {(s, t)}, η5 = {(t, s)}, η6 = {(s, s), (t, s)}, η7 = {(t, t), (s, t)} and η8 = {(s, t), (t, s)}. The utilities for each system are given in Table 1. From this, we infer that the Pareto efficient norms are η1, η2, η3, η6 and η7. Note that η8 prohibits the resource to be passed from one agent to another, and this is not good for any agent (since we have chosen S0 = {s, t}, no agent can be sure to ever get the resource, i.e., goal ϕi 1 is not true in K † η8). 5.3 Nash Implementation Normative Systems The most famous solution concept in game theory is of course Nash equilibrium [11, p.14]. A collection of strategies, one for each agent, is said to form a Nash equilibrium if no agent can benefit by doing anything other than playing its strategy, under the assumption that the other agents play theirs. Nash equilibria are important because they provide stable solutions to the problem of what strategy an agent should play. Note that in our toy example, although η3 is individually rational for each agent, it is not a Nash equilibrium, since given this norm, it would be beneficial for agent 1 to deviate (and likewise for 2). In our framework, we say a social system Σ = M , η (where η = η∅) is a Nash implementation if SC (i.e., everyone complying with the normative system) forms a Nash equilibrium in the game GΣ. The intuition is that if Σ is a Nash implementation, then complying with the normative system is a reasonable solution for all concerned: there can be no benefit to deviating from it, indeed, there is a positive incentive for all to comply. If Σ is not a Nash implementation, then the normative system is unlikely to succeed, since compliance is not rational for some agents. (Our choice of terminology is deliberately chosen to reflect the way the term Nash implementation is used in implementation theory, or mechanism design [11, p.185], where a game designer seeks to achieve some outcomes by designing the rules of the game such that these outcomes are equilibria.) NASH IMPLEMENTATION (NI) : Given: Multi-agent system M . Question: Does there exist a non-empty normative system η over M such that M , η forms a Nash implementation? Verifying that a particular social system forms a Nash implementation can be done in polynomial time - it amounts to checking: ∀i ∈ A : ui (K † η) ≥ ui (K † (η {i})). This, clearly requires only a polynomial number of model checking calls, each of which requires only polynomial time. THEOREM 4. The NI problem is NP-complete, even for twoagent systems. PROOF. For membership of NP, simply guess a normative system η and check that it forms a Nash implementation; since η ⊆ R, guessing can be done in non-deterministic polynomial time, and as s(2k+1) 1 1 1 1 1 1 11 1 1 11 2 2 2 2 2 2 2 2 2 2 2 t(x1) f(x1) t(x2) f(x2) t(xk) f(xk) 2 2 t(x1) f(x1) t(x2) f(x2) t(xk) f(xk) ...... s0 Figure 4: Reduction for Theorem 4. we argued above, verifying that it forms a Nash implementation can be done in polynomial time. For NP-hardness, we reduce SAT. Suppose we are given a SAT instance ϕ over Boolean variables x1, . . . , xk . Then we construct an instance of NI as follows. We create two agents, A = {1, 2}. For each Boolean variable xi we create two Boolean variables, t(xi ) and f (xi ), and we then define a Kripke structure as shown in Figure 4, with s0 being the only initial state; the arc labelling in Figure 4 gives the α function, and each state is labelled with the propositions that are true in that state. For each Boolean variable xi , we define the formulae xi and x⊥ i as follows: xi = E f(t(xi ) ∧ E f((E f(t(xi ))) ∧ A f(¬f (xi )))) x⊥ i = E f(f (xi ) ∧ E f((E f(f (xi ))) ∧ A f(¬t(xi )))) Let ϕ∗ be the formula obtained from ϕ by systematically substituting xi for xi . Each agent has three goals: γi [0] = for both i ∈ {1, 2}, while γ1[1] = k^ i=1 ((E f(t(xi ))) ∧ (E f(f (xi )))) γ2[1] = E fE f k^ i=1 ((E f(t(xi ))) ∧ (E f(f (xi )))) and finally, for both agents, γi [2] being the conjunction of the following formulae: The Sixth Intl. Joint Conf. on Autonomous Agents and Multi-Agent Systems (AAMAS 07) 887 k^ i=1 (xi ∨ x⊥ i ) (3) k^ i=1 ¬(xi ∧ x⊥ i ) (4) k^ i=1 ¬(E f(t(xi )) ∧ E f(f (xi ))) (5) ϕ∗ (6) We denote the multi-agent system so constructed by Mϕ. Now, we prove that the SAT instance ϕ is satisfiable iff Mϕ has a Nash implementation normative system: For the ⇒ direction, suppose ϕ is satisfiable, and let X be a satisfying valuation, i.e., a set of Boolean variables making ϕ true. We can extract from X a Nash implementation normative system η as follows: if xi ∈ X , then η includes the arc from s0 to the state in which f (xi ) is true, and also includes the arc from s(2k + 1) to the state in which f (xi ) is true; if xi ∈ X , then η includes the arc from s0 to the state in which t(xi ) is true, and also includes the arc from s(2k + 1) to the state in which t(xi ) is true. No other arcs, apart from those so defined, as included in η. Notice that η is individually rational for both agents: if they both comply with the normative system, then they will have their γi [2] goals achieved, which they do not in the basic system. To see that η forms a Nash implementation, observe that if either agent defects from η, then neither will have their γi [2] goals achieved: agent 1 strictly prefers (C, C) over (D, C), and agent 2 strictly prefers (C, C) over (C, D). For the ⇐ direction, suppose there exists a Nash implementation normative system η, in which case η = ∅. Then ϕ is satisfiable; for suppose not. Then the goals γi [2] are not achievable by any normative system, (by construction). Now, since η must forbid at least one transition, then at least one agent would fail to have its γi [1] goal achieved if it complied, so at least one would do better by defecting, i.e., not complying with η. But this contradicts the assumption that η is a Nash implementation, i.e., that (C, C) forms a Nash equilibrium. This result is perhaps of some technical interest beyond the specific concerns of the present paper, since it is related to two problems that are of wider interest: the complexity of mechanism design [5], and the complexity of computing Nash equilibria [6, 7] 5.4 Richer Goal Languages It is interesting to consider what happens to the complexity of the problems we consider above if we allow richer languages for goals: in particular, CTL ∗ [9]. The main difference is that determining ui (K) in a given multi-agent system M when such a goal language is used involves solving a PSPACE-complete problem (since model checking for CTL ∗ is PSPACE-complete [8]). In fact, it seems that for each of the three problems we consider above, the corresponding problem under the assumption of a CTL ∗ representation for goals is also PSPACE-complete. It cannot be any easier, since determining the utility of a particular Kripke structure involves solving a PSPACE-complete problem. To see membership in PSPACE we can exploit the fact that PSPACE = NPSPACE [12, p.150], and so we can guess the desired normative system, applying a PSPACE verification procedure to check that it has the desired properties. 6. CONCLUSIONS Social norms are supposed to restrict our behaviour. Of course, such a restriction does not have to be bad: the fact that an agent"s behaviour is restricted may seem a limitation, but there may be benefits if he can assume that others will also constrain their behaviour. The question then, for an agent is, how to be sure that others will comply with a norm. And, for a system designer, how to be sure that the system will behave socially, that is, according to its norm. Game theory is a very natural tool to analyse and answer these questions, which involve strategic considerations, and we have proposed a way to translate key questions concerning logic-based normative systems to game theoretical questions. We have proposed a logical framework to reason about such scenarios, and we have given some computational costs for settling some of the main questions about them. Of course, our approach is in many senses open for extension or enrichment. An obvious issue is to consider is the complexity of the questions we give for more practical representations of models (cf. [1]), and to consider other classes of allowable goals. 7. REFERENCES [1] T. Agotnes, W. van der Hoek, J. A. Rodriguez-Aguilar, C. Sierra, and M. Wooldridge. On the logic of normative systems. In Proc. IJCAI-07, Hyderabad, India, 2007. [2] R. Alur, T. A. Henzinger, and O. Kupferman. Alternating-time temporal logic. Jnl. of the ACM, 49(5):672-713, 2002. [3] K. Binmore. Game Theory and the Social Contract Volume 1: Playing Fair. The MIT Press: Cambridge, MA, 1994. [4] K. Binmore. Game Theory and the Social Contract Volume 2: Just Playing. The MIT Press: Cambridge, MA, 1998. [5] V. Conitzer and T. Sandholm. Complexity of mechanism design. In Proc. UAI, Edmonton, Canada, 2002. [6] V. Conitzer and T. Sandholm. Complexity results about nash equilibria. In Proc. IJCAI-03, pp. 765-771, Acapulco, Mexico, 2003. [7] C. Daskalakis, P. W. Goldberg, and C. H. Papadimitriou. The complexity of computing a Nash equilibrium. In Proc. STOC, Seattle, WA, 2006. [8] E. A. Emerson. Temporal and modal logic. In Handbook of Theor. Comp. Sci. Vol. B, pages 996-1072. Elsevier, 1990. [9] E. A. Emerson and J. Y. Halpern. ‘Sometimes" and ‘not never" revisited: on branching time versus linear time temporal logic. Jnl. of the ACM, 33(1):151-178, 1986. [10] D. Fitoussi and M. Tennenholtz. Choosing social laws for multi-agent systems: Minimality and simplicity. Artificial Intelligence, 119(1-2):61-101, 2000. [11] M. J. Osborne and A. Rubinstein. A Course in Game Theory. The MIT Press: Cambridge, MA, 1994. [12] C. H. Papadimitriou. Computational Complexity. Addison-Wesley: Reading, MA, 1994. [13] Y. Shoham and M. Tennenholtz. On the synthesis of useful social laws for artificial agent societies. In Proc. AAAI, San Diego, CA, 1992. [14] Y. Shoham and M. Tennenholtz. On social laws for artificial agent societies: Off-line design. In Computational Theories of Interaction and Agency, pages 597-618. The MIT Press: Cambridge, MA, 1996. [15] W. van der Hoek, M. Roberts, and M. Wooldridge. Social laws in alternating time: Effectiveness, feasibility, and synthesis. Synthese, 2007. [16] M. Wooldridge and W. van der Hoek. On obligations and normative ability. Jnl. of Appl. Logic, 3:396-420, 2005. 888 The Sixth Intl. Joint Conf. on Autonomous Agents and Multi-Agent Systems (AAMAS 07)

social law;constraint;goal;decision making;normative system game;normative system;desirable objective;complexity;kripke structure;logic;computational complexity;game;game theoretic property;ordinal utility;multi-agent system;computation tree logic;nash implementation;multiple goal of increasing priority

train_I-49

A Multilateral Multi-issue Negotiation Protocol

In this paper, we present a new protocol to address multilateral multi-issue negotiation in a cooperative context. We consider complex dependencies between multiple issues by modelling the preferences of the agents with a multi-criteria decision aid tool, also enabling us to extract relevant information on a proposal assessment. This information is used in the protocol to help in accelerating the search for a consensus between the cooperative agents. In addition, the negotiation procedure is defined in a crisis management context where the common objective of our agents is also considered in the preferences of a mediator agent.

1. INTRODUCTION Multi-issue negotiation protocols represent an important field of study since negotiation problems in the real world are often complex ones involving multiple issues. To date, most of previous work in this area ([2, 3, 19, 13]) dealt almost exclusively with simple negotiations involving independent issues. However, real-world negotiation problems involve complex dependencies between multiple issues. When one wants to buy a car, for example, the value of a given car is highly dependent on its price, consumption, comfort and so on. The addition of such interdependencies greatly complicates the agents utility functions and classical utility functions, such as the weighted sum, are not sufficient to model this kind of preferences. In [10, 9, 17, 14, 20], the authors consider inter-dependencies between issues, most often defined with boolean values, except for [9], while we can deal with continuous and discrete dependent issues thanks to the modelling power of the Choquet integral. In [17], the authors deal with bilateral negotiation while we are interested in a multilateral negotiation setting. Klein et al. [10] present an approach similar to ours, using a mediator too and information about the strength of the approval or rejection that an agent makes during the negotiation. In our protocol, we use more precise information to improve the proposals thanks to the multi-criteria methodology and tools used to model the preferences of our agents. Lin, in [14, 20], also presents a mediation service but using an evolutionary algorithm to reach optimal solutions and as explained in [4], players in the evolutionary models need to repeatedly interact with each other until the stable state is reached. As the population size increases, the time it takes for the population to stabilize also increases, resulting in excessive computation, communication, and time overheads that can become prohibitive, and for one-to-many and many-to-many negotiations, the overheads become higher as the number of players increases. In [9], the authors consider a non-linear utility function by using constraints on the domain of the issues and a mediation service to find a combination of bids maximizing the social welfare. Our preference model, a nonlinear utility function too, is more complex than [9] one since the Choquet integral takes into account the interactions and the importance of each decision criteria/issue, not only the dependencies between the values of the issues, to determine the utility. We also use an iterative protocol enabling us to find a solution even when no bid combination is possible. In this paper, we propose a negotiation protocol suited for multiple agents with complex preferences and taking into account, at the same time, multiple interdependent issues and recommendations made by the agents to improve a proposal. Moreover, the preferences of our agents are modelled using a multi-criteria methodology and tools enabling us to take into account information about the improvements that can be made to a proposal, in order to help in accelerating the search for a consensus between the agents. Therefore, we propose a negotiation protocol consisting of solving our decision problem using a MAS with a multi-criteria decision aiding modelling at the agent level and a cooperation-based multilateral multi-issue negotiation protocol. This protocol is studied under a non-cooperative approach and it is shown 943 978-81-904262-7-5 (RPS) c 2007 IFAAMAS that it has subgame perfect equilibria, provided that agents behave rationally in the sense of von Neumann and Morgenstern. The approach proposed in this paper has been first introduced and presented in [8]. In this paper, we present our first experiments, with some noteworthy results, and a more complex multi-agent system with representatives to enable us to have a more robust system. In Section 2, we present our application, a crisis management problem. Section 3 deals with the general aspect of the proposed approach. The preference modelling is described in sect. 4, whereas the motivations of our protocol are considered in sect. 5 and the agent/multiagent modelling in sect. 6. Section 7 presents the formal modelling and properties of our protocol before presenting our first experiments in sect. 8. Finally, in Section 9, we conclude and present the future work. 2. CASE STUDY This protocol is applied to a crisis management problem. Crisis management is a relatively new field of management and is composed of three types of activities: crisis prevention, operational preparedness and management of declared crisis. The crisis prevention aims to bring the risk of crisis to an acceptable level and, when possible, avoid that the crisis actually happens. The operational preparedness includes strategic advanced planning, training and simulation to ensure availability, rapid mobilisation and deployment of resources to deal with possible emergencies. The management of declared crisis is the response to - including the evacuation, search and rescue - and the recovery from the crisis by minimising the effects of the crises, limiting the impact on the community and environment and, on a longer term, by bringing the community"s systems back to normal. In this paper, we focus on the response part of the management of declared crisis activity, and particularly on the evacuation of the injured people in disaster situations. When a crisis is declared, the plans defined during the operational preparedness activity are executed. For disasters, master plans are executed. These plans are elaborated by the authorities with the collaboration of civil protection agencies, police, health services, non-governmental organizations, etc. When a victim is found, several actions follow. First, a rescue party is assigned to the victim who is examined and is given first aid on the spot. Then, the victims can be placed in an emergency centre on the ground called the medical advanced post. For all victims, a sorter physician - generally a hospital physician - examines the seriousness of their injuries and classifies the victims by pathology. The evacuation by emergency health transport if necessary can take place after these clinical examinations and classifications. Nowadays, to evacuate the injured people, the physicians contact the emergency call centre to pass on the medical assessments of the most urgent cases. The emergency call centre then searches for available and appropriate spaces in the hospitals to care for these victims. The physicians are informed of the allocations, so they can proceed to the evacuations choosing the emergency health transports according to the pathologies and the transport modes provided. In this context, we can observe that the evacuation is based on three important elements: the examination and classification of the victims, the search for an allocation and the transport. In the case of the 11 March 2004 Madrid attacks, for instance, some injured people did not receive the appropriate health care because, during the search for space, the emergency call centre did not consider the transport constraints and, in particular, the traffic. Therefore, for a large scale crisis management problem, there is a need to support the emergency call centre and the physicians in the dispatching to take into account the hospitals and the transport constraints and availabilities. 3. PROPOSED APPROACH To accept a proposal, an agent has to consider several issues such as, in the case of the crisis management problem, the availabilities in terms of number of beds by unit, medical and surgical staffs, theatres and so on. Therefore, each agent has its own preferences in correlation with its resource constraints and other decision criteria such as, for the case study, the level of congestion of a hospital. All the agents also make decisions by taking into account the dependencies between these decision criteria. The first hypothesis of our approach is that there are several parties involved in and impacted by the decision, and so they have to decide together according to their own constraints and decision criteria. Negotiation is the process by which a group facing a conflict communicates with one another to try and come to a mutually acceptable agreement or decision and so, the agents have to negotiate. The conflict we have to resolve is finding an acceptable solution for all the parties by using a particular protocol. In our context, multilateral negotiation is a negotiation protocol type that is the best suited for this type of problem : this type of protocol enables the hospitals and the physicians to negotiate together. The negotiation also deals with multiple issues. Moreover, an other hypothesis is that we are in a cooperative context where all the parties have a common objective which is to provide the best possible solution for everyone. This implies the use of a negotiation protocol encouraging the parties involved to cooperate as satisfying its preferences. Taking into account these aspects, a Multi-Agent System (MAS) seems to be a reliable method in the case of a distributed decision making process. Indeed, a MAS is a suitable answer when the solution has to combine, at least, distribution features and reasoning capabilities. Another motivation for using MAS lies in the fact that MAS is well known for facilitating automated negotiation at the operative decision making level in various applications. Therefore, our approach consists of solving a multiparty decision problem using a MAS with • The preferences of the agents are modelled using a multi-criteria decision aid tool, MYRIAD, also enabling us to consider multi-issue problems by evaluating proposals on several criteria. • A cooperation-based multilateral and multi-issue negotiation protocol. 4. THE PREFERENCE MODEL We consider a problem where an agent has several decision criteria, a set Nk = {1, . . . , nk} of criteria for each agent k involved in the negotiation protocol. These decision criteria enable the agents to evaluate the set of issues that are negotiated. The issues correspond directly or not to the decision criteria. However, for the example of the crisis management 944 The Sixth Intl. Joint Conf. on Autonomous Agents and Multi-Agent Systems (AAMAS 07) problem, the issues are the set of victims to dispatch between the hospitals. These issues are translated to decision criteria enabling the hospital to evaluate its congestion and so to an updated number of available beds, medical teams and so on. In order to take into account the complexity that exists between the criteria/issues, we use a multi-criteria decision aiding (MCDA) tool named MYRIAD [12] developed at Thales for MCDA applications based on a two-additive Choquet integral which is a good compromise between versatility and ease to understand and model the interactions between decision criteria [6]. The set of the attributes of Nk is denoted by Xk 1 , . . . , Xk nk . All the attributes are made commensurate thanks to the introduction of partial utility functions uk i : Xk i → [0, 1]. The [0, 1] scale depicts the satisfaction of the agent k regarding the values of the attributes. An option x is identified to an element of Xk = Xk 1 × · · · × Xk nk , with x = (x1, . . . , xnk ). Then the overall assessment of x is given by Uk(x) = Hk(uk 1 (x1), . . . , uh nk (xnk )) (1) where Hk : [0, 1]nk → [0, 1] is the aggregation function. The overall preference relation over Xk is then x y ⇐⇒ Uk(x) ≥ Uk(y) . The two-additive Choquet integral is defined for (z1, . . . , znk ) ∈ [0, 1]nk by [7] Hk(z1, . . . , znk ) = X i∈Nk 0 @vk i − 1 2 X j=i |Ik i,j| 1 A zi + X Ik i,j >0 Ik i,j zi ∧ zj + X Ii,j <0 |Ii,j| zi ∨ zj (2) where vk i is the relative importance of criterion i for agent k and Ik i,j is the interaction between criteria i and j, ∧ and ∨ denote the min and max functions respectively. Assume that zi < zj. A positive interaction between criteria i and j depicts complementarity between these criteria (positive synergy) [7]. Hence, the lower score of z on criterion i conceals the positive effect of the better score on criterion j to a larger extent on the overall evaluation than the impact of the relative importance of the criteria taken independently of the other ones. In other words, the score of z on criterion j is penalized by the lower score on criterion i. Conversely, a negative interaction between criteria i and j depicts substitutability between these criteria (negative synergy) [7]. The score of z on criterion i is then saved by a better score on criterion j. In MYRIAD, we can also obtain some recommendations corresponding to an indicator ωC (H, x) measuring the worth to improve option x w.r.t. Hk on some criteria C ⊆ Nk as follows ωC (Hk, x)= Z 1 0 Hk ` (1 − τ)xC + τ, xNk\C ´ − Hk(x) EC (τ, x) dτ where ((1−τ)xC +τ, xNk\C ) is the compound act that equals (1 − τ)xi + τ if i ∈ C and equals xi if i ∈ Nk \ C. Moreover, EC (τ, x) is the effort to go from the profile x to the profile ((1 − τ)xC + τ, xNk\C ). Function ωC (Hk, x) depicts the average improvement of Hk when the criteria of coalition A range from xC to 1C divided by the average effort needed for this improvement. We generally assume that EC is of order 1, that is EC (τ, x) = τ P i∈C (1 − xi). The expression of ωC (Hk, x) when Hk is a Choquet integral, is given in [11]. The agent is then recommended to improve of coalition C for which ωC (Hk, x) is maximum. This recommendation is very useful in a negotiation protocol since it helps the agents to know what to do if they want an offer to be accepted while not revealing their own preference model. 5. PROTOCOL MOTIVATIONS For multi-issue problems, there are two approaches: a complete package approach where the issues are negotiated simultaneously in opposition to the sequential approach where the issues are negotiated one by one. When the issues are dependant, then it is the best choice to bargain simultaneously over all issues [5]. Thus, the complete package is the adopted approach so that an offer will be on the overall set of injured people while taking into account the other decision criteria. We have to consider that all the parties of the negotiation process have to agree on the decision since they are all involved in and impacted by this decision and so an unanimous agreement is required in the protocol. In addition, no party can leave the process until an agreement is reached, i.e. a consensus achieved. This makes sense since a proposal concerns all the parties. Moreover, we have to guarantee the availability of the resources needed by the parties to ensure that a proposal is realistic. To this end, the information about these availabilities are used to determine admissible proposals such that an offer cannot be made if one of the parties has not enough resources to execute/achieve it. At the beginning of the negotiation, each party provides its maximum availabilities, this defining the constraints that have to be satisfied for each offer submitted. The negotiation has also to converge quickly on an unanimous agreement. We decided to introduce in the negotiation protocol an incentive to cooperate taking into account the passed negotiation time. This incentive is defined on the basis of a time dependent penalty, the discounting factor as in [18] or a time-dependent threshold. This penalty has to be used in the accept/reject stage of our consensus procedure. In fact, in the case of a discounting factor, each party will accept or reject an offer by evaluating the proposal using its utility function deducted from the discounting factor. In the case of a time-dependent threshold, if the evaluation is greater or equal to this threshold, the offer is accepted, otherwise, in the next period, its threshold is reduced. The use of a penalty is not enough alone since it does not help in finding a solution. Some information about the assessments of the parties involved in the negotiation are needed. In particular, it would be helpful to know why an offer has been rejected and/or what can be done to make a proposal that would be accepted. MYRIAD provides an analysis that determines the flaws an option, here a proposal. In particular, it gives this type of information: which criteria of a proposal should be improved so as to reach the highest possible overall evaluation [11]. As we use this tool to model the parties involved in the negotiation, the information about the criteria to improve can be used by the mediator to elaborate the proposals. We also consider that the dual function can be used to take into account another type of information: on which criteria of a proposal, no improvement is necessary so that the overall evaluation of a proposal is still acceptable, do not decrease. Thus, all information is a constraint to be satisfied as much as possible by The Sixth Intl. Joint Conf. on Autonomous Agents and Multi-Agent Systems (AAMAS 07) 945 Figure 1: An illustration of some system. the parties to make a new proposal. We are in a cooperative context and revealing one"s opinion on what can be improved is not prohibited, on the contrary, it is useful and recommended here seeing that it helps in converging on an agreement. Therefore, when one of the parties refuses an offer, some information will be communicated. In order to facilitate and speed up the negotiation, we introduce a mediator. This specific entity is in charge of making the proposals to the other parties in the system by taking into account their public constraints (e.g. their availabilities) and the recommendations they make. This mediator can also be considered as the representative of the general interest we can have, in some applications, such as in the crisis management problem, the physician will be the mediator and will also have some more information to consider when making an offer (e.g. traffic state, transport mode and time). Each party in a negotiation N, a negotiator, can also be a mediator of another negotiation N , this party becoming the representative of N in the negotiation N, as illustrated by fig. 1 what can also help in reducing the communication time. 6. AGENTIFICATION How the problem is transposed in a MAS problem is a very important aspect when designing such a system. The agentification has an influence upon the systems efficiency in solving the problem. Therefore, in this section, we describe the elements and constraints taken into account during the modelling phase and for the model itself. However, for this negotiation application, the modelling is quite natural when one observes the negotiation protocol motivations and main properties. First of all, it seems obvious that there should be one agent for each player of our multilateral multi-issue negotiation protocol. The agents have the involved parties" information and preferences. These agents are: • Autonomous: they decide for themselves what, when and under what conditions actions should be performed; • Rational: they have a means-ends competence to fit its decisions, according to its knowledge, preferences and goal; • Self-interested: they have their own interests which may conflict with the interests of other agents. Moreover, their preferences are modelled and a proposal evaluated and analysed using MYRIAD. Each agent has private information and can access public information as knowledge. In fact, there are two types of agents: the mediator type for the agents corresponding to the mediator of our negotiation protocol, the delegated physician in our application, and the negotiator type for the agents corresponding to the other parties, the hospitals. The main behaviours that an agent of type mediator needs to negotiate in our protocol are the following: • convert_improvements: converts the information given by the other agents involved in the negotiation about the improvements to be done, into constraints on the next proposal to be made; • convert_no_decrease: converts the information given by the other agents involved in the negotiation about the points that should not be changed into constraints on the next proposal to be made; • construct_proposal: constructs a new proposal according to the constraints obtained with convert_improvements, convert_no_decrease and the agent preferences; The main behaviours that an agent of type negotiator needs to negotiate in our protocol are the following: • convert_proposal: converts a proposal to a MYRIAD option of the agent according to its preferences model and its private data; • convert_improvements_wc: converts the agent recommendations for the improvements of a MYRIAD option into general information on the proposal; • convert_no_decrease_wc: converts the agent recommendations about the criteria that should not be changed in the MYRIAD option into general information on the proposal; In addition to these behaviours, there are, for the two types of agents, access behaviours to MYRIAD functionalities such as the evaluation and improvement functions: • evaluate_option: evaluates the MYRIAD option obtained using the agent behaviour convert_proposal; • improvements: gets the agent recommendations to improve a proposal from the MYRIAD option; • no_decrease: gets the agent recommendations to not change some criteria from the MYRIAD option; Of course, before running the system with such agents, we must have defined each party preferences model in MYRIAD. This model has to be part of the agent so that it could be used to make the assessments and to retrieve the improvements. In addition to these behaviours, the communication acts between the agents is as follows. 1. mediator agent communication acts: 946 The Sixth Intl. Joint Conf. on Autonomous Agents and Multi-Agent Systems (AAMAS 07) m1 m 1 1 m inform1 m mediator negotiator accept−proposal l 1 accept−proposal m−l reject−proposal propose propose Figure 2: The protocol diagram in AUML, and where m is the number of negotiator agents and l is the number of agents refusing current proposal. (a) propose: sends a message containing a proposal to all negotiator agents; (b) inform: sends a message to all negotiator agents to inform them that an agreement has been reached and containing the consensus outcome. 2. negotiator agent communication acts: (a) accept-proposal: sends a message to the mediator agent containing the agent recommendations to improve the proposal and obtained with convert_improvements_wc; (b) reject-proposal: sends a message to the mediator agent containing the agent recommendations about the criteria that should not be changed and obtained with convert_no_decrease_wc. Such agents are interchangeable, in a case of failure, since they all have the same properties and represent a user with his preference model, not depending on the agent, but on the model defined in MYRIAD. When the issues and the decision criteria are different from each other, the information about the criteria improvement have to be pre-processed to give some instructions on the directions to take and about the negotiated issues. It is the same for the evaluation of a proposal: each agent has to convert the information about the issues to update its private information and to obtain the values of each attribute of the decision criteria. 7. OUR PROTOCOL Formally, we consider negotiations where a set of players A = {1, 2, . . . , m} and a player a are negotiating over a set Q of size q. The player a is the protocol mediator, the mediator agent of the agentification. The utility/preference function of a player k ∈ A ∪ {a} is Uk, defined using MYRIAD, as presented in Section 4, with a set Nk of criteria, Xk an option, and so on. An offer is a vector P = (P1, P2, · · · , Pm), a partition of Q, in which Pk is player k"s share of Q. We have P ∈ P where P is the set of admissible proposals, a finite set. Note that P is determined using all players general constraints on the proposals and Q. Moreover, let ˜P denote a particular proposal defined as a"s preferred proposal. We also have the following notation: δk is the threshold decrease factor of player k, Φk : Pk → Xk is player k"s function to convert a proposal to an option and Ψk is the function indicating which points P has to be improved, with Ψk its dual function - on which points no improvement is necessary. Ψk is obtained using the dual function of ωC (Hk, x): eωC (Hk, x)= Z 1 0 Hk(x) − Hk ` τ xC , xNk\C ´ eEC (τ, x) dτ Where eEC (τ, x) is the cost/effort to go from (τxC , xNk\C ) to x. In period t of our consensus procedure, player a proposes an agreement P. All players k ∈ A respond to a by accepting or rejecting P. The responses are made simultaneously. If all players k ∈ A accept the offer, the game ends. If any player k rejects P, then the next period t+1 begins: player a makes another proposal P by taking into account information provided by the players and the ones that have rejected P apply a penalty. Therefore, our negotiation protocol can be as follows: Protocol P1. • At the beginning, we set period t = 0 • a makes a proposal P ∈ P that has not been proposed before. • Wait that all players of A give their opinion Yes or No to the player a. If all players agree on P, this later is chosen. Otherwise t is incremented and we go back to previous point. • If there is no more offer left from P, the default offer ˜P will be chosen. • The utility of players regarding a given offer decreases over time. More precisely, the utility of player k ∈ A at period t regarding offer P is Uk(Φk(Pk), t) = ft(Uk(Φk(Pk))), where one can take for instance ft(x) = x.(δk)t or ft(x) = x − δk.t, as penalty function. Lemma 1. Protocol P1 has at least one subgame perfect equilibrium 1 . Proof : Protocol P1 is first transformed in a game in extensive form. To this end, one shall specify the order in which the responders A react to the offer P of a. However the order in which the players answer has no influence on the course of the game and in particular on their personal utility. Hence protocol P1 is strictly equivalent to a game in 1 A subgame perfect equilibrium is an equilibrium such that players" strategies constitute a Nash equilibrium in every subgame of the original game [18, 16]. A Nash equilibrium is a set of strategies, one for each player, such that no player has incentive to unilaterally change his/her action [15]. The Sixth Intl. Joint Conf. on Autonomous Agents and Multi-Agent Systems (AAMAS 07) 947 extensive form, considering any order of the players A. This game is clearly finite since P is finite and each offer can only be proposed once. Finally P1 corresponds to a game with perfect information. We end the proof by using a classical result stating that any finite game in extensive form with perfect information has at least one subgame perfect equilibrium (see e.g. [16]). Rational players (in the sense of von Neumann and Morgenstern) involved in protocol P1 will necessarily come up with a subgame perfect equilibrium. Example 1. Consider an example with A = {1, 2} and P = {P1 , P2 , P3 } where the default offer is P1 . Assume that ft(x) = x − 0.1 t. Consider the following table giving the utilities at t = 0. P1 P2 P3 a 1 0.8 0.7 1 0.1 0.7 0.5 2 0.1 0.3 0.8 It is easy to see that there is one single subgame perfect equilibrium for protocol P1 corresponding to these values. This equilibrium consists of the following choices: first a proposes P3 ; player 1 rejects this offer; a proposes then P2 and both players 1 and 2 accepts otherwise they are threatened to receive the worse offer P1 for them. Finally offer P2 is chosen. Option P1 is the best one for a but the two other players vetoed it. It is interesting to point out that, even though a prefers P2 to P3 , offer P3 is first proposed and this make P2 being accepted. If a proposes P2 first, then the subgame perfect equilibrium in this situation is P3 . To sum up, the worse preferred options have to be proposed first in order to get finally the best one. But this entails a waste of time. Analysing the previous example, one sees that the game outcome at the equilibrium is P2 that is not very attractive for player 2. Option P3 seems more balanced since no player judges it badly. It could be seen as a better solution as a consensus among the agents. In order to introduce this notion of balanceness in the protocol, we introduce a condition under which a player will be obliged to accept the proposal, reducing the autonomy of the agents but for increasing rationality and cooperation. More precisely if the utility of a player is larger than a given threshold then acceptance is required. The threshold decreases over time so that players have to make more and more concession. Therefore, the protocol becomes as follows. Protocol P2. • At the beginning we set period t = 0 • a makes a proposal P ∈ P that has not been proposed before. • Wait that all players of A give their opinion Yes or No to the player a. A player k must accept the offer if Uk(Φk(Pk)) ≥ ρk(t) where ρk(t) tends to zero when t grows. Moreover there exists T such that for all t ≥ T, ρk(t) = 0. If all players agree on P, this later is chosen. Otherwise t is incremented and we go back to previous point. • If there is no more offer left from P, the default offer ˜P will be chosen. One can show exactly as in Lemma 1 that protocol P2 has at least one subgame perfect equilibrium. We expect that protocol P2 provides a solution not to far from P , so it favours fairness among the players. Therefore, our cooperation-based multilateral multi-issue protocol is the following: Protocol P. • At the beginning we set period t = 0 • a makes a proposal P ∈ P that has not been proposed before, considering Ψk(Pt ) and Ψk(Pt ) for all players k ∈ A. • Wait that all players of A give their opinion (Yes , Ψk(Pt )) or (No , Ψk(Pt )) to the player a. A player k must accept the offer if Uk(Φk(Pk)) ≥ ρk(t) where ρk(t) tends to zero when t grows. Moreover there exists T such that for all t ≥ T, ρk(t) = 0. If all players agree on P, this later is chosen. Otherwise t is incremented and we go back to previous point. • If there is no more offer left from P, the default offer ˜P will be chosen. 8. EXPERIMENTS We developed a MAS using the widely used JADE agent platform [1]. This MAS is designed to be as general as possible (e.g. a general framework to specialise according to the application) and enable us to make some preliminary experiments. The experiments aim at verifying that our approach gives solutions as close as possible to the Maximin solution and in a small number of rounds and hopefully in a short time since our context is highly cooperative. We defined the two types of agents and their behaviours as introduced in section 6. The agents and their behaviours correspond to the main classes of our prototype, NegotiatorAgent and NegotiatorBehaviour for the negotiator agents, and MediatorAgent and MediatorBehaviour for the mediator agent. These classes extend JADE classes and integrate MYRIAD into the agents, reducing the amount of communications in the system. Some functionalities depending on the application have to be implemented according to the application by extending these classes. In particular, all conversion parts of the agents have to be specified according to the application since to convert a proposal into decision criteria, we need to know, first, this model and the correlations between the proposals and this model. First, to illustrate our protocol, we present a simple example of our dispatch problem. In this example, we have three hospitals, H1, H2 and H3. Each hospital can receive victims having a particular pathology in such a way that H1 can receive patients with the pathology burn, surgery or orthopedic, H2 can receive patients with the pathology surgery, orthopedic or cardiology and H3 can receive patients with the pathology burn or cardiology. All the hospitals have similar decision criteria reflecting their preferences on the level of congestion they can face for the overall hospital and the different services available, as briefly explained for hospital H1 hereafter. For hospital H1, the preference model, fig. 3, is composed of five criteria. These criteria correspond to the preferences on the pathologies the hospital can treat. In the case of 948 The Sixth Intl. Joint Conf. on Autonomous Agents and Multi-Agent Systems (AAMAS 07) Figure 3: The H1 preference model in MYRIAD. the pathology burn, the corresponding criterion, also named burn as shown in fig. 3, represents the preferences of H1 according to the value of Cburn which is the current capacity of burn. Therefore, the utility function of this criterion represents a preference such that the more there are patients of this pathology in the hospital, the less the hospital may satisfy them, and this with an initial capacity. In addition to reflecting this kind of viewpoint, the aggregation function as defined in MYRIAD introduces a veto on the criteria burn, surgery, orthopedic and EReceipt, where EReceipt is the criterion for the preferences about the capacity to receive a number of patients at the same time. In this simplified example, the physician have no particular preferences on the dispatch and the mediator agent chooses a proposal randomly in a subset of the set of admissibility. This subset have to satisfy as much as possible the recommendations made by the hospitals. To solve this problem, for this example, we decided to solve a linear problem with the availability constraints and the recommendations as linear constraints on the dispatch values. The set of admissibility is then obtained by solving this linear problem by the use of Prolog. Moreover, only the recommendations on how to improve a proposal are taken into account. The problem to solve is then to dispatch to hospital H1, H2 and H3, the set of victims composed of 5 victims with the pathology burn, 10 with surgery, 3 with orthopedic and 7 with cardiology. The availabilities of the hospitals are as presented in the following table. Available Overall burn surg. orthop. cardio. H1 11 4 8 10H2 25 - 3 4 10 H3 7 10 - - 3 We obtain a multiagent system with the mediator agent and three agents of type negotiator for the three hospital in the problem. The hospitals threshold are fixed approximatively to the level where an evaluation is considered as good. To start, the negotiator agents send their availabilities. The mediator agent makes a proposal chosen randomly in admissible set obtained with these availabilities as linear constraints. This proposal is the vector P0 = [[H1,burn, 3], [H1, surgery, 8], [H1, orthopaedic, 0], [H2, surgery, 2], [H2, orthopaedic, 3], [H2, cardiology, 6], [H3, burn, 2], [H3, cardiology, 1]] and the mediator sends propose(P0) to H1, H2 and H3 for approval. Each negotiator agent evaluates this proposal and answers back by accepting or rejecting P0: • Agent H1 rejects this offer since its evaluation is very far from the threshold (0.29, a bad score) and gives a recommendation to improve burn and surgery by sending the message reject_proposal([burn,surgery]); • Agent H2 accepts this offer by sending the message accept_proposal(), the proposal evaluation being good; • Agent H3 accepts P0 by sending the message accept_ proposal(), the proposal evaluation being good. Just with the recommendations provided by agent H1, the mediator is able to make a new proposal by restricting the value of burn and surgery. The new proposal obtained is then P1 = [[H1,burn, 0], [H1, surgery, 8], [H1, orthopaedic, 1], [H2, surgery, 2], [H2, orthopaedic, 2], [H2, cardiology, 6], [H3, burn, 5], [H3, cardiology, 1]]. The mediator sends propose(P1) the negotiator agents. H1, H2 and H3 answer back by sending the message accept_proposal(), P1 being evaluated with a high enough score to be acceptable, and also considered as a good proposal when using the explanation function of MYRIAD. An agreement is reached with P1. Note that the evaluation of P1 by H3 has decreased in comparison with P0, but not enough to be rejected and that this solution is the Pareto one, P∗ . Other examples have been tested with the same settings: issues in IN, three negotiator agents and the same mediator agent, with no preference model but selecting randomly the proposal. We obtained solutions either equal or close to the Maximin solution, the distance from the standard deviation being less than 0.0829, the evaluations not far from the ones obtained with P∗ and with less than seven proposals made. This shows us that we are able to solve this multi-issue multilateral negotiation problem in a simple and efficient way, with solutions close to the Pareto solution. 9. CONCLUSION AND FUTURE WORK This paper presents a new protocol to address multilateral multi-issue negotiation in a cooperative context. The first main contribution is that we take into account complex inter-dependencies between multiple issues with the use of a complex preference modelling. This contribution is reinforced by the use of multi-issue negotiation in a multilateral context. Our second contribution is the use of sharp recommendations in the protocol to help in accelerating the search of a consensus between the cooperative agents and in finding an optimal solution. We have also shown that the protocol has subgame perfect equilibria and these equilibria converge to the usual maximum solution. Moreover, we tested this protocol in a crisis management context where the negotiation aim is where to evacuate a whole set of injured people to predefined hospitals. We have already developed a first MAS, in particular integrating MYRIAD, to test this protocol in order to know more about its efficiency in terms of solution quality and quickness in finding a consensus. This prototype enabled us to solve some examples with our approach and the results we obtained are encouraging since we obtained quickly good agreements, close to the Pareto solution, in the light of the initial constraints of the problem: the availabilities. We still have to improve our MAS by taking into account The Sixth Intl. Joint Conf. on Autonomous Agents and Multi-Agent Systems (AAMAS 07) 949 the two types of recommendations and by adding a preference model to the mediator of our system. Moreover, a comparative study has to be done in order to evaluate the performance of our framework against the existing ones and against some variations on the protocol. 10. ACKNOWLEDGEMENT This work is partly funded by the ICIS research project under the Dutch BSIK Program (BSIK 03024). 11. REFERENCES [1] JADE. http://jade.tilab.com/. [2] P. Faratin, C. Sierra, and N. R. Jennings. Using similarity criteria to make issue trade-offs in automated negotiations. Artificial Intelligence, 142(2):205-237, 2003. [3] S. S. Fatima, M. Wooldridge, and N. R. Jennings. Optimal negotiation of multiple issues in incomplete information settings. In 3rd International Joint Conference on Autonomous Agents and Multiagent Systems (AAMAS"04), pages 1080-1087, New York, USA, 2004. [4] S. S. Fatima, M. Wooldridge, and N. R. Jennings. A comparative study of game theoretic and evolutionary models of bargaining for software agents. Artificial Intelligence Review, 23:185-203, 2005. [5] S. S. Fatima, M. Wooldridge, and N. R. Jennings. On efficient procedures for multi-issue negotiation. In 8th International Workshop on Agent-Mediated Electronic Commerce(AMEC"06), pages 71-84, Hakodate, Japan, 2006. [6] M. Grabisch. The application of fuzzy integrals in multicriteria decision making. European J. of Operational Research, 89:445-456, 1996. [7] M. Grabisch, T. Murofushi, and M. Sugeno. Fuzzy Measures and Integrals. Theory and Applications (edited volume). Studies in Fuzziness. Physica Verlag, 2000. [8] M. Hemaissia, A. El Fallah-Seghrouchni, C. Labreuche, and J. Mattioli. Cooperation-based multilateral multi-issue negotiation for crisis management. In 2th International Workshop on Rational, Robust and Secure Negotiation (RRS"06), pages 77-95, Hakodate, Japan, May 2006. [9] T. Ito, M. Klein, and H. Hattori. A negotiation protocol for agents with nonlinear utility functions. In AAAI, 2006. [10] M. Klein, P. Faratin, H. Sayama, and Y. Bar-Yam. Negotiating complex contracts. Group Decision and Negotiation, 12:111-125, March 2003. [11] C. Labreuche. Determination of the criteria to be improved first in order to improve as much as possible the overall evaluation. In IPMU 2004, pages 609-616, Perugia, Italy, 2004. [12] C. Labreuche and F. Le Hu´ed´e. MYRIAD: a tool suite for MCDA. In EUSFLAT"05, pages 204-209, Barcelona, Spain, 2005. [13] R. Y. K. Lau. Towards genetically optimised multi-agent multi-issue negotiations. In Proceedings of the 38th Annual Hawaii International Conference on System Sciences (HICSS"05), Big Island, Hawaii, 2005. [14] R. J. Lin. Bilateral multi-issue contract negotiation for task redistribution using a mediation service. In Agent Mediated Electronic Commerce VI (AMEC"04), New York, USA, 2004. [15] J. F. Nash. Non cooperative games. Annals of Mathematics, 54:286-295, 1951. [16] G. Owen. Game Theory. Academic Press, New York, 1995. [17] V. Robu, D. J. A. Somefun, and J. A. L. Poutr´e. Modeling complex multi-issue negotiations using utility graphs. In 4th International Joint Conference on Autonomous agents and multiagent systems (AAMAS"05), pages 280-287, 2005. [18] A. Rubinstein. Perfect equilibrium in a bargaining model. Econometrica, 50:97-109, jan 1982. [19] L.-K. Soh and X. Li. Adaptive, confidence-based multiagent negotiation strategy. In 3rd International Joint Conference on Autonomous agents and multiagent systems (AAMAS"04), pages 1048-1055, Los Alamitos, CA, USA, 2004. [20] H.-W. Tung and R. J. Lin. Automated contract negotiation using a mediation service. In 7th IEEE International Conference on E-Commerce Technology (CEC"05), pages 374-377, Munich, Germany, 2005. 950 The Sixth Intl. Joint Conf. on Autonomous Agents and Multi-Agent Systems (AAMAS 07)

multi-criterion decision make;decision making;automate negotiation;multi-agent system;cooperative agent;modelling;myriad;multilateral negotiation;negotiation protocol;automated negotiation;multi-issue negotiation;crisis management;negotiation strategy

train_I-50

Agents, Beliefs, and Plausible Behavior in a Temporal Setting

Logics of knowledge and belief are often too static and inflexible to be used on real-world problems. In particular, they usually offer no concept for expressing that some course of events is more likely to happen than another. We address this problem and extend CTLK (computation tree logic with knowledge) with a notion of plausibility, which allows for practical and counterfactual reasoning. The new logic CTLKP (CTLK with plausibility) includes also a particular notion of belief. A plausibility update operator is added to this logic in order to change plausibility assumptions dynamically. Furthermore, we examine some important properties of these concepts. In particular, we show that, for a natural class of models, belief is a KD45 modality. We also show that model checking CTLKP is PTIME-complete and can be done in time linear with respect to the size of models and formulae.

1. INTRODUCTION Notions like time, knowledge, and beliefs are very important for analyzing the behavior of agents and multi-agent systems. In this paper, we extend modal logics of time and knowledge with a concept of plausible behavior: this notion is added to the language of CTLK [19], which is a straightforward combination of the branching-time temporal logic CTL [4, 3] and standard epistemic logic [9, 5]. In our approach, plausibility can be seen as a temporal property of behaviors. That is, some behaviors of the system can be assumed plausible and others implausible, with the underlying idea that the latter should perhaps be ignored in practical reasoning about possible future courses of action. Moreover, behaviors can be formally understood as temporal paths in the Kripke structure modeling a multiagent system. As a consequence, we obtain a language to reason about what can (or must) plausibly happen. We propose a particular notion of beliefs (inspired by [20, 7]), defined in terms of epistemic relations and plausibility. The main intuition is that beliefs are facts that an agent would know if he assumed that only plausible things could happen. We believe that humans use such a concept of plausibility and practical beliefs quite often in their everyday reasoning. Restricting one"s reasoning to plausible possibilities is essential to make the reasoning feasible, as the space of all possibilities is exceedingly large in real life. We investigate some important properties of plausibility, knowledge, and belief in this new framework. In particular, we show that knowledge is an S5 modality, and that beliefs satisfy axioms K45 in general, and KD45 for the class of plausibly serial models. Finally, we show that the relationship between knowledge and belief for plausibly serial models is natural and reflects the initial intuition well. We also show how plausibility assumptions can be specified in the object language via a plausibility update operator, and we study properties of such updates. Finally, we show that model checking of the new logic is no more complex than model checking CTL and CTLK. Our ultimate goal is to come up with a logic that allows the study of strategies, time, knowledge, and plausible/rational behavior under both perfect and imperfect information. As combining all these dimensions is highly nontrivial (cf. [12, 14]) it seems reasonable to split this task. While this paper deals with knowledge, plausibility, and belief, the companion paper [11] proposes a general framework for multi-agent systems that regard game-theoretical rationality criteria like Nash equilibrium, Pareto optimality, etc. The latter approach is based on the more powerful logic ATL [1]. The paper is structured as follows. Firstly, we briefly present branching-time logic with knowledge, CTLK. In Section 3 we present our approach to plausibility and formally define CTLK with plausibility. We also show how 582 978-81-904262-7-5 (RPS) c 2007 IFAAMAS temporal formulae can be used to describe plausible paths, and we compare our logic with existing related work. In Section 4, properties of knowledge, belief, and plausibility are explored. Finally, we present verification complexity results for CTLKP in Section 5. 2. BRANCHING TIME AND KNOWLEDGE In this paper we develop a framework for agents" beliefs about how the world can (or must) evolve. Thus, we need a notion of time and change, plus a notion of what the agents are supposed to know in particular situations. CTLK [19] is a straightforward combination of the computation tree logic CTL [4, 3] and standard epistemic logic [9, 5]. CTL includes operators for temporal properties of systems: i.e., path quantifier E (there is a path), together with temporal operators: f(in the next state), 2 (always from now on) and U (until).1 Every occurrence of a temporal operator is preceded by exactly one path quantifier in CTL (this variant of the language is sometimes called vanilla CTL). Epistemic logic uses operators for representing agents" knowledge: Kaϕ is read as agent a knows that ϕ. Let Π be a set of atomic propositions with a typical element p, and Agt = {1, ..., k} be a set of agents with a typical element a. The language of CTLK consists of formulae ϕ, given as follows: ϕ ::= p | ¬ϕ | ϕ ∧ ϕ | Eγ | Kaϕ γ ::= fϕ | 2 ϕ | ϕU ϕ. We will sometimes refer to formulae ϕ as (vanilla) state formulae and to formulae γ as (vanilla) path formulae. The semantics of CTLK is based on Kripke models M = Q, R, ∼1, ..., ∼k, π , which include a nonempty set of states Q, a state transition relation R ⊆ Q × Q, epistemic indistinguishability relations ∼a⊆ Q × Q (one per agent), and a valuation of propositions π : Π → P(Q). We assume that relation R is serial and that all ∼a are equivalence relations. A path λ in M refers to a possible behavior (or computation) of system M, and can be represented as an infinite sequence of states that follow relation R, that is, a sequence q0q1q2... such that qiRqi+1 for every i = 0, 1, 2, ... We denote the ith state in λ by λ[i]. The set of all paths in M is denoted by ΛM (if the model is clear from context, M will be omitted). A q-path is a path that starts from q, i.e., λ[0] = q. A q-subpath is a sequence of states, starting from q, which is a subpath of some path in the model, i.e. a sequence q0q1... such that q = q0 and there are q0 , ..., qi such that q0 ...qi q0q1... ∈ ΛM.2 The semantics of CTLK is defined as follows: M, q |= p iff q ∈ π(p); M, q |= ¬ϕ iff M, q |= ϕ; M, q |= ϕ ∧ ψ iff M, q |= ϕ and M, q |= ψ; M, q |= E fϕ iff there is a q-path λ such that M, λ[1] |= ϕ; M, q |= E2 ϕ iff there is a q-path λ such that M, λ[i] |= ϕ for every i ≥ 0; 1 Additional operators A (for every path) and ♦ (sometime in the future) are defined in the usual way. 2 For CTLK models, λ is a q-subpath iff it is a q-path. It will not always be so when plausible paths are introduced. M, q |= EϕU ψ iff there is a q-path λ and i ≥ 0 such that M, λ[i] |= ψ, and M, λ[j] |= ϕ for every 0 ≤ j < i; M, q |= Kaϕ iff M, q |= ϕ for every q such that q ∼a q . 3. EXTENDING TIME AND KNOWLEDGE WITH PLAUSIBILITY AND BELIEFS In this section we discuss the central concept of this paper, i.e. the concept of plausibility. First, we outline the idea informally. Then, we extend CTLK with the notion of plausibility by adding plausible path operators Pl a and physical path operator Ph to the logic. Formula Pl aϕ has the intended meaning: according to agent a, it is plausible that ϕ holds; formula Ph ϕ reads as: ϕ holds in all physically possible scenarios (i.e., even in implausible ones). The plausible path operator restricts statements only to those paths which are defined to be sensible, whereas the physical path operator generates statements about all paths that may theoretically occur. Furthermore, we define beliefs on top of plausibility and knowledge, as the facts that an agent would know if he assumed that only plausible things could happen. Finally, we discuss related work [7, 8, 20, 18, 16], and compare it with our approach. 3.1 The Concept of Plausibility It is well known how knowledge (or beliefs) can be modeled with Kripke structures. However, it is not so obvious how we can capture knowledge and beliefs in a sensible way in one framework. Clearly, there should be a connection between these two notions. Our approach is to use the notion of plausibility for this purpose. Plausibility can serve as a primitive concept that helps to define the semantics of beliefs, in a similar way as indistinguishability of states (represented by relation ∼a) is the semantic concept that underlies knowledge. In this sense, our work follows [7, 20]: essentially, beliefs are what an agent would know if he took only plausible options into account. In our approach, however, plausibility is explicitly seen as a temporal property. That is, we do not consider states (or possible worlds) to be more plausible than others but rather define some behaviors to be plausible, and others implausible. Moreover, behaviors can be formally understood as temporal paths in the Kripke structure modeling a multi-agent system. An actual notion of plausibility (that is, a particular set of plausible paths) can emerge in many different ways. It may result from observations and learning; an agent can learn from its observations and see specific patterns of events as plausible (a lot of people wear black shoes if they wear a suit). Knowledge exchange is another possibility (e.g., an agent a can tell agent b that player c always bluffs when he is smiling). Game theory, with its rationality criteria (undominated strategies, maxmin, Nash equilibrium etc.) is another viable source of plausibility assumptions. Last but not least, folk knowledge can be used to establish plausibilityrelated classifications of behavior (players normally want to win a game, people want to live). In any case, restricting the reasoning to plausible possibilities can be essential if we want to make the reasoning feasible, as the space of all possibilities (we call them physical possibilities in the rest of the paper) is exceedingly large in real life. Of course, this does not exclude a more extensive analysis in special cases, e.g. when our plausibility assumptions do not seem accurate any more, or when the cost of The Sixth Intl. Joint Conf. on Autonomous Agents and Multi-Agent Systems (AAMAS 07) 583 inaccurate assumptions can be too high (as in the case of high-budget business decisions). But even in these cases, we usually do not get rid of plausibility assumptions completely - we only revise them to make them more cautious.3 To formalize this idea, we extend models of CTLK with sets of plausible paths and add plausibility operators Pl a, physical paths operator Ph , and belief operators Ba to the language of CTLK. Now, it is possible to make statements that refer to plausible paths only, as well as statements that regard all paths that may occur in the system. 3.2 CTLK with Plausibility In this section, we extend the logic of CTLK with plausibility; we call the resulting logic CTLKP. Formally, the language of CTLKP is defined as: ϕ ::= p | ¬ϕ | ϕ ∧ ϕ | Eγ | Pl aϕ | Ph ϕ | Kaϕ | Baϕ γ ::= fϕ | 2 ϕ | ϕU ϕ. For instance, we may claim it is plausible to assume that a shop is closed after the opening hours, though the manager may be physically able to open it at any time: Pl aA2 (late → ¬open) ∧ Ph E♦ (late ∧ open). The semantics of CTLKP extends that of CTLK as follows. Firstly, we augment the models with sets of plausible paths. A model with plausibility is given as M = Q, R, ∼1, ..., ∼k, Υ1, ..., Υk, π , where Q, R, ∼1, ..., ∼k, π is a CTLK model, and Υa ⊆ ΛM is the set of paths in M that are plausible according to agent a. If we want to make it clear that Υa is taken from model M, we will write ΥM a . It seems worth emphasizing that this notion of plausibility is subjective and holistic. It is subjective because Υa represents agent a"s subjective view on what is plausible - and indeed, different agents may have different ideas on plausibility (i.e., Υa may differ from Υb). It is holistic because Υa represents agent a"s idea of the plausible behavior of the whole system (including the behavior of other agents). Remark 1. In our models, plausibility is also global, i.e., plausibility sets do not depend on the state of the system. Investigating systems, in which plausibility is relativized with respect to states (like in [7]), might be an interesting avenue of future work. However, such an approach - while obviously more flexible - allows for potentially counterintuitive system descriptions. For example, it might be the case that path λ is plausible in q = λ[0], but the set of plausible paths in q = λ[1] is empty. That is, by following plausible path λ we are bound to get to an implausible situation. But then, does it make sense to consider λ as plausible? Secondly, we use a non-standard satisfaction relation |=P , which we call plausible satisfaction. Let M be a CTLKP 3 That is, when planning to open an industrial plant in the UK, we will probably consider the possibility of our main contractor taking her life, but we will still not take into account the possibilities of: an invasion of UFO, England being destroyed by a meteorite, Fidel Castro becoming the British Prime Minister etc. Note that this is fundamentally different from using a probabilistic model in which all these unlikely scenarios are assigned very low probabilities: in that case, they also have a very small influence on our final decision, but we must process the whole space of physical possibilities to evaluate the options. model and P ⊆ ΛM be an arbitrary subset of paths in M (not necessarily any ΥM a ). |=P restricts the evaluation of temporal formulae to the paths given in P only. The absolute satisfaction relation |= is defined as |=ΛM . Let on(P) be the set of all states that lie on at least one path in P, i.e. on(P) = {q ∈ Q | ∃λ ∈ P∃i (λ[i] = q)}. Now, the semantics of CTLKP can be given through the following clauses: M, q |=P p iff q ∈ π(p); M, q |=P ¬ϕ iff M, q |=P ϕ; M, q |=P ϕ ∧ ψ iff M, q |=P ϕ and M, q |=P ψ; M, q |=P E fϕ iff there is a q-subpath λ ∈ P such that M, λ[1] |=P ϕ; M, q |=P E2 ϕ iff there is a q-subpath λ ∈ P such that M, λ[i] |=P ϕ for every i ≥ 0; M, q |=P EϕU ψ iff there is a q-subpath λ ∈ P and i ≥ 0 such that M, λ[i] |=P ψ, and M, λ[j] |=P ϕ for every 0 ≤ j < i; M, q |=P Pl aϕ iff M, q |=Υa ϕ; M, q |=P Ph ϕ iff M, q |= ϕ; M, q |=P Kaϕ iff M, q |= ϕ for every q such that q ∼a q ; M, q |=P Baϕ iff for all q ∈ on(Υa) with q ∼a q , we have that M, q |=Υa ϕ. One of the main reasons for using the concept of plausibility is that we want to define agents" beliefs out of more primitive concepts - in our case, these are plausibility and indistinguishability - in a way analogous to [20, 7]. If an agent knows that ϕ, he must be sure about it. However, beliefs of an agent are not necessarily about reliable facts. Still, they should make sense to the agent; if he believes that ϕ, then the formula should at least hold in all futures that he envisages as plausible. Thus, beliefs of an agent may be seen as things known to him if he disregards all non-plausible possibilities. We say that ϕ is M-true (M |= ϕ) if M, q |= ϕ for all q ∈ QM. ϕ is valid (|= ϕ) if M |= ϕ for all models M. ϕ is M-strongly true (M |≡ ϕ) if M, q |=P ϕ for all q ∈ QM and all P ⊆ ΛM. ϕ is strongly valid ( |≡ ϕ) if M |≡ ϕ for all models M. Proposition 2. Strong truth and strong validity imply truth and validity, respectively. The reverse does not hold. Ultimately, we are going to be interested in normal (not strong) validity, as parameterizing the satisfaction relation with a set P is just a technical device for propagating sets of plausible paths Υa into the semantics of nested formulae. The importance of strong validity, however, lies in the fact that |≡ ϕ ↔ ψ makes ϕ and ψ completely interchangeable, while the same is not true for normal validity. Proposition 3. Let Φ[ϕ/ψ] denote formula Φ in which every occurrence of ψ was replaced by ϕ. Also, let |≡ ϕ ↔ ψ. Then for all M, q, P: M, q |=P Φ iff M, q |=P Φ[ϕ/ψ] (in particular, M, q |= Φ iff M, q |= Φ[ϕ/ψ]). Note that |= ϕ ↔ ψ does not even imply that M, q |= Φ iff M, q |= Φ[ϕ/ψ]. 584 The Sixth Intl. Joint Conf. on Autonomous Agents and Multi-Agent Systems (AAMAS 07) Figure 1: Guessing Robots game Example 1 (Guessing Robots). Consider a simple game with two agents a and b, shown in Figure 1. First, a chooses a real number r ∈ [0, 1] (without revealing the number to b); then, b chooses a real number r ∈ [0, 1]. The agents win the game (and collect EUR 1, 000, 000) if both chose 1, otherwise they lose. Formally, we model the game with a CTLKP model M, in which the set of states Q includes qs for the initial situation, states qr, r ∈ [0, 1], for the situations after a has chosen number r, and final states qw, ql for the winning and the losing situation, respectively. The transition relation is as follows: qsRqr and qrRql for all r ∈ [0, 1]; q1Rqw, qwRqw, and qlRql. Moreover, π(one) = {q1} and π(win) = {qw}. Player a has perfect information in the game (i.e., q ∼a q iff q = q ), but player b does not distinguish between states qr (i.e., qr ∼b qr for all r, r ∈ [0, 1]). Obviously, the only sensible thing to do for both agents is to choose 1 (using game-theoretical vocabulary, these strategies are strongly dominant for the respective players). Thus, there is only one plausible course of events if we assume that our players are rational, and hence Υa = Υb = {qsq1qwqw . . .}. Note that, in principle, the outcome of the game is uncertain: M, qs |= ¬A♦ win∧¬A2 ¬win. However, assuming rationality of the players makes it only plausible that the game must end up with a win: M, qs |= Pla A♦ win ∧ Plb A♦ win, and the agents believe that this will be the case: M, qs |= BaA♦ win ∧ BbA♦ win. Note also that, in any of the states qr, agent b believes that a (being rational) has played 1: M, qr |= Bbone for all r ∈ [0, 1]. 3.3 Defining Plausible Paths with Formulae So far, we have assumed that sets of plausible paths are somehow given in models. In this section we present a dynamic approach where an actual notion of plausibility can be specified in the object language. Note that we want to specify (usually infinite) sets of infinite paths, and we need a finite representation of these structures. One logical solution is given by using path formulae γ. These formulae describe properties of paths; therefore, a specific formula can be used to characterize a set of paths. For instance, think about a country in Africa where it has never snowed. Then, plausible paths might be defined as ones in which it never snows, i.e., all paths that satisfy 2 ¬snows. Formally, let γ be a CTLK path formula. We define |γ|M to be the set of paths that satisfy γ in model M: | fϕ|M = {λ | M, λ[1] |= ϕ} |2 ϕ|M = {λ | ∀i (M, λ[i] |= ϕ)} |ϕ1U ϕ2|M = {λ | ∃i ` M, λ[i] |= ϕ2 ∧ ∀j(0 ≤ j < i ⇒ M, λ[j] |= ϕ1) ´ }. Moreover, we define the plausible paths model update as follows. Let M = Q, R, ∼1, ..., ∼k, Υ1, ..., Υk, π be a CTLKP model, and let P ⊆ ΛM be a set of paths. Then Ma,P = Q, R, ∼1, ..., ∼k, Υ1, ..., Υa−1, P, Υa+1, ..., Υk, π denotes model M with a"s set of plausible paths reset to P. Now we can extend the language of CTLKP with formulae (set-pla γ)ϕ with the intuitive reading: suppose that γ exactly characterizes the set of plausible paths, then ϕ holds, and formal semantics given below: M, q |=P (set-pla γ)ϕ iff Ma,|γ|M , q |=P ϕ. We observe that this update scheme is similar to the one proposed in [13]. 3.4 Comparison to Related Work Several modal notions of plausibility were already discussed in the existing literature [7, 8, 20, 18, 16]. In these papers, like in ours, plausibility is used as a primitive semantic concept that helps to define beliefs on top of agents" knowledge. A similar idea was introduced by Moses and Shoham in [18]. Their work preceded both [7, 8] and [20]and although Moses and Shoham do not explicitly mention the term plausibility, it seems appropriate to summarize their idea first. Moses and Shoham: Beliefs as Conditional Knowledge In [18], beliefs are relativized with respect to a formula α (which can be seen as a plausibility assumption expressed in the object language). More precisely, worlds that satisfy α can be considered as plausible. This concept is expressed via symbols Bα i ϕ; the index i ∈ {1, 2, 3} is used to distinguish between three different implementations of beliefs. The first version is given by Bα 1 ϕ ≡ K(α → ϕ).4 A drawback of this version is that if α is false, then everything will be believed with respect to α. The second version overcomes this problem: Bα 2 ϕ ≡ K(α → ϕ) ∧ (K¬α → Kϕ); now ϕ is only believed if it is known that ϕ follows from assumption α, and ϕ must be known if assumption α is known to be false. Finally, Bα 3 ϕ ≡ K(α → ϕ) ∧ ¬K¬α: if the assumption α is known to be false, nothing should be believed with respect to α. The strength of these different notions is given as follows: Bα 3 ϕ implies Bα 2 ϕ, and Bα 2 ϕ implies Bα 1 ϕ. In this approach, belief is strongly connected to knowledge in the sense that belief is knowledge with respect to a given assumption. Friedman and Halpern: Plausibility Spaces The work of Friedman and Halpern [7] extends the concepts of knowledge and belief with an explicit notion of plausibility; i.e., some worlds are more plausible for an agent than others. To implement this idea, Kripke models are extended with function P which assigns a plausibility space P(q, a) = (Ω(q,a), (q,a)) to every state, or more generally every possible world q, and agent a. The plausibility space 4 Unlike in most approaches, K is interpreted over all worlds and not only over the indistinguishable worlds. The Sixth Intl. Joint Conf. on Autonomous Agents and Multi-Agent Systems (AAMAS 07) 585 is just a partially ordered subset of states/worlds; that is, Ω(q, a) ⊆ Q, and (q,a)⊆ Q ×Q is a reflexive and transitive relation. Let S, T ⊆ Ω(q,a) be finite subsets of states; now, T is defined to be plausible given S with respect to P(q, a), denoted by S →P (q,a) T, iff all minimal points/states in S (with respect to (q,a)) are also in T.5 Friedman and Halpern"s view to modal plausibility is closely related to probability and, more generally, plausibility measures. Logics of plausibility can be seen as a qualitative description of agents preferences/knowledge; logics of probability [6, 15], on the other hand, offer a quantitative description. The logic from [7] is defined by the following grammar: ϕ ::= p | ϕ∧ϕ | ¬ϕ | Kaϕ | ϕ →a ϕ, where the semantics of all operators except →a is given as usual, and formulae ϕ →a ψ have the meaning that ψ is true in the most plausible worlds in which ϕ holds. Formally, the semantics for →a is given as: M, q |= ϕ →a ψ iff Sϕ P (q,a) →P(q,a) Sψ P (q,a), where Sϕ (q,a) = {q ∈ Ω(q,a) | M, q |= ϕ} are the states in Ω(q,a) that satisfy ϕ. The idea of defining beliefs is given by the assumption that an agent believes in something if he knows that it is true in the most plausible worlds of Ω(q,a); formally, this can be stated as Baϕ ≡ Ka( →a ϕ). Friedman and Halpern have shown that the KD45 axioms are valid for operator Ba if plausibility spaces satisfy consistency (for all states q ∈ Q it holds that Ω(q,a) ⊆ { q ∈ Q | q ∼a q }) and normality (for all states q ∈ Q it holds that Ω(q,a) = ∅).6 A temporal extension of the language (mentioned briefly in [7], and discussed in more detail in [8]) uses the interpreted systems approach [10, 5]. A system R is given by runs, where a run r : N → Q is a function from time moments (modeled by N) to global states, and a time point (r, i) is given by a time point i ∈ N and a run r. A global state is a combination of local states, one per agent. An interpreted system M = (R, π) is given by a system R and a valuation of propositions π. Epistemic relations are defined over time points, i.e., (r , m ) ∼a (r, m) iff agent a"s local states ra(m ) and ra(m) of (r , m ) and (r, m) are equal. Formulae are interpreted in a straightforward way with respect to interpreted systems, e.g. M, r, m |= Kaϕ iff M, r , m |= ϕ for all (r , m ) ∼a (r, m). Now, these are time points that play the role of possible worlds; consequently, plausibility spaces P(r,m,a) are assigned to each point (r, m) and agent a. Su et al.: KBC Logic Su et al. [20] have developed a multi-modal, computationally grounded logic with modalities K, B, and C (knowledge, belief, and certainty). The computational model consists of (global) states q = (qvis , qinv , qper , Qpls ) where the environment is divided into a visible (qvis ) and an invisible part (qinv ), and qper captures the agent"s perception of the visible part of the environment. External sources may provide the agent with information about the invisible part of a state, which results in a set of states Qpls that are plausible for the agent. Given a global state q, we additionally define V is(q) = qvis , Inv(q) = qinv , Per(q) = qper , and Pls(q) = Qpls . The 5 When there are infinite chains . . . q3 q2 a q1, the definition is much more sophisticated. An interested reader is referred to [7] for more details. 6 Note that this normality is essentially seriality of states wrt plausibility spaces. semantics is given by an extension of interpreted systems [10, 5], here, it is called interpreted KBC systems. KBC formulae are defined as ϕ ::= p | ¬ϕ | ϕ ∧ ϕ | Kϕ | Bϕ | Cϕ. The epistemic relation ∼vis is captured in the following way: (r, i) ∼vis (r , i ) iff V is(r(i)) = V is(r (i )). The semantic clauses for belief and certainty are given below. M, r, i |= Bϕ iff M, r , i |= ϕ for all (r , i ) with V is(r (i )) = Per(r(i)) and Inv(r (i )) ∈ Pls(r(i)) M, r, i |= Cϕ iff M, r , i |= ϕ for all (r (i )) with V is(r (i )) = Per(r(i)) Thus, an agent believes ϕ if, and only if, ϕ is true in all states which look like what he sees now and seem plausible in the current state. Certainty is stronger: if an agent is certain about ϕ, the formula must hold in all states with a visible part equal to the current perception, regardless of whether the invisible part is plausible or not. The logic does not include temporal formulae, although it might be extended with temporal operators, as time is already present in KBC models. What Are the Differences to Our Logic? In our approach, plausibility is explicitly seen as a temporal property, i.e., it is a property of temporal paths rather than states. In the object language, this is reflected by the fact that plausibility assumptions are specified through path formulae. In contrast, the approach of [18] and [20] is static: not only the logics do not include operators for talking about time and/or change, but these are states that are assumed plausible or not in their semantics. The differences to [7, 8] are more subtle. Firstly, the framework of Friedman and Halpern is static in the sense that plausibility is taken as a property of (abstract) possible worlds. This formulation is flexible enough to allow for incorporating time; still, in our approach, time is inherent to plausibility rather than incidental. Secondly, our framework is more computationally oriented. The implementation of temporal plausibility in [7, 8] is based on the interpreted systems approach with time points (r, m) being subject to plausibility. As runs are included in time points, they can also be defined plausible or implausible.7 However, it also means that time points serve the role of possible worlds in the basic formulation, which yields Kripke structures with uncountable possible world spaces in all but the most trivial cases. Thirdly, [7, 8] build on linear time: a run (more precisely, a time moment (r, m)) is fixed when a formula is interpreted. In contrast, we use branching time with explicit quantification over temporal paths.8 We believe that branching time is more suitable for non-deterministic domains (cf. e.g. [4]), of which multi-agent systems are a prime example. Note that branching time makes our notion of belief different from Friedman and Halpern"s. Most notably, property Kϕ → Bϕ is valid in their approach, but not in ours: an agent may 7 Friedman and Halpern even briefly mention how plausibility of runs can be embedded in their framework. 8 To be more precise, time in [7] does implicitly branch at epistemic states. This is because (r, m) ∼a (r , m ) iff a"s local state corresponding to both time points is the same (ra(m) = ra(m )). In consequence, the semantics of Kaϕ can be read as for every run, and every moment on this run that yields the same local state as now, ϕ holds. 586 The Sixth Intl. Joint Conf. on Autonomous Agents and Multi-Agent Systems (AAMAS 07) know that some course of events is in principle possible, without believing that it can really become the case (see Section 4.2). As Proposition 13 suggests, such a subtle distinction between knowledge and beliefs is possible in our approach because branching time logics allow for existential quantification over runs. Fourthly, while Friedman and Halpern"s models are very flexible, they also enable system descriptions that may seem counterintuitive. Suppose that (r, m) is plausible in itself (formally: (r, m) is minimal wrt (r,m,a)), but (r, m + 1) is not plausible in (r, m + 1). This means that following the plausible path makes it implausible (cf. Remark 1), which is even stranger in the case of linear time. Combining the argument with computational aspects, we suggest that our approach can be more natural and straightforward for many applications. Last but not least, our logic provides a mechanism for specifying (and updating) sets of plausible paths in the object language. Thus, plausibility sets can be specified in a succinct way, which is another feature that makes our framework computation-friendly. The model checking results from Section 5 are especially encouraging in this light. 4. PLAUSIBILITY, KNOWLEDGE, AND BELIEFS IN CTLKP In this section we study some relevant properties of plausibility, knowledge, and beliefs; in particular, axioms KDT45 are examined. But first, we identify two important subclasses of models with plausibility. A CTLKP model is plausibly serial (or p-serial) for agent a if every state of the system is part of a plausible path according to a, i.e. on(Υa) = Q. As we will see further, a weaker requirement is sometimes sufficient. We call a model weakly p-serial if every state has at least one indistinguishable counterpart which lies on a plausible path, i.e. for each q ∈ Q there is a q ∈ Q such that q ∼a q and q ∈ on(Υa). Obviously, p-seriality implies weak p-seriality. We get the following characterization of both model classes. Proposition 4. M is plausibly serial for agent a iff formula Pl aE f is valid in M. M is weakly p-serial for agent a iff ¬KaPl aA f⊥ is valid in M. 4.1 Axiomatic Properties Theorem 5. Axioms K, D, 4, and 5 for knowledge are strongly valid, and axiom T is valid. That is, modalities Ka form system S5 in the sense of normal validity, and KD45 in the sense of strong validity. We do not include proofs here due to lack of space. The interested reader is referred to [2], where detailed proofs are given. Proposition 6. Axioms K, 4, and 5 for beliefs are strongly valid. That is, we have: |≡ (Baϕ ∧ Ba(ϕ → ψ)) → Baψ, |≡ (Baϕ → BaBaϕ), and |≡ (¬Baϕ → Ba¬Baϕ). The next proposition concerns the consistency axiom D: Baϕ → ¬Ba¬ϕ. It is easy to see that the axiom is not valid in general: as we have no restrictions on plausibility sets Υa, it may be as well that Υa = ∅. In that case we have Baϕ ∧ Ba¬ϕ for all formulae ϕ, because the set of states to be considered becomes empty. However, it turns out that D is valid for a very natural class of models. Proposition 7. Axiom D for beliefs is not valid in the class of all CTLKP models. However, it is strongly valid in the class of weak p-serial models (and therefore also in the class of p-serial models). Moreover, as one may expect, beliefs do not have to be always true. Proposition 8. Axiom T for beliefs is not valid; i.e., |= (Baϕ → ϕ). The axiom is not even valid in the class of p-serial models. Theorem 9. Belief modalities Ba form system K45 in the class of all models, and KD45 in the class of weakly plausibly serial models (in the sense of both normal and strong validity). Axiom T is not even valid for p-serial models. 4.2 Plausibility, Knowledge, and Beliefs First, we investigate the relationship between knowledge and plausibility/physicality operators. Then, we look at the interaction between knowledge and beliefs. Proposition 10. Let ϕ be a CTLKP formula, and M be a CTLKP model. We have the following strong validities: (i) |≡ Pl aKaϕ ↔ Kaϕ (ii) |≡ Ph Kaϕ ↔ KaPh ϕ and |≡ KaPh ϕ ↔ Kaϕ We now want to examine the relationship between knowledge and belief. For instance, if agent a believes in something, he knows that he believes it. Or, if he knows a fact, he also believes that he knows it. On the other hand, for instance, an agent does not necessarily believe in all the things he knows. For example, we may know that an invasion from another galaxy is in principle possible (KaE♦ invasion), but if we do not take this possibility as plausible (¬Pl aE♦ invasion), then we reject the corresponding belief in consequence (¬BaE♦ invasion). Note that this property reflects the strong connection between belief and plausibility in our framework. Proposition 11. The following formulae are strongly valid: (i) Baϕ → KaBaϕ, (ii) KaBaϕ → Baϕ, (iii) Kaϕ → BaKaϕ. The following formulae are not valid: (iv) Baϕ → BaKaϕ, (v) Kaϕ → Baϕ The last invalidity is especially important: it is not the case that knowing something implies believing in it. This emphasizes that we study a specific concept of beliefs here. Note that its specific is not due to the plausibility-based definition of beliefs. The reason lies rather in the fact that we investigate knowledge, beliefs and plausibility in a temporal framework, as Proposition 12 shows. Proposition 12. Let ϕ be a CTLKP formula that does not include any temporal operators. Then Kaϕ → Baϕ is strongly valid, and in the class of p-serial models we have even that |≡ Kaϕ ↔ Baϕ. The Sixth Intl. Joint Conf. on Autonomous Agents and Multi-Agent Systems (AAMAS 07) 587 Moreover, it is important that we use branching time with explicit quantification over paths; this observation is formalized in Proposition 13. Definition 1. We define the universal sublanguage of CTLK in a way similar to [21]: ϕu ::= p | ¬p | ϕu ∧ ϕu | ϕu ∨ ϕu | Aγu | Kaϕu, γu ::= fϕu | 2 ϕu | ϕuU ϕu. We call such ϕu universal formulae, and γu universal path formulae. Proposition 13. Let ϕu be a universal CTLK formula. Then |≡ Kaϕu → Baϕu. The following two theorems characterize the relationship between knowledge and beliefs: first for the class of p-serial models, and then, finally, for all models. Theorem 14. The following formulae are strongly valid in the class of plausibly serial CTLKP models: (i) Baϕ ↔ KaPl aϕ, (ii) Kaϕ ↔ BaPh ϕ. Theorem 15. Formula Baϕ ↔ KaPl a(E f → ϕ) is strongly valid. Note that this characterization has a strong commonsense reading: believing in ϕ is knowing that ϕ plausibly holds in all plausibly imaginable situations. 4.3 Properties of the Update The first notable property of plausibility update is that it influences only formulae in which plausibility plays a role, i.e. ones in which belief or plausibility modalities occur. Proposition 16. Let ϕ be a CTLKP formula that does not include operators Pl a and Ba, and γ be a CTLKP path formula. Then, we have |≡ ϕ ↔ (set-pla γ)ϕ. What can be said about the result of an update? At first sight, formula (set-pla γ)Pl aAγ seems a natural characterization; however, it is not valid. This is because, by leaving the other (implausible) paths out of scope, we may leave out of |γ| some paths that were needed to satisfy γ (see the example in Section 4.2). We propose two alternative ways out: the first one restricts the language of the update similarly to [21]; the other refers to physical possibilities, in a way analogous to [13]. Proposition 17. The CTLKP formula (set-pla γ)Pl aAγ is not valid. However, we have the following validities: (i) |≡ (set-pla γu)Pl aAγu, where γu is a universal CTLK path formula from Definition 1. (ii) If ϕ, ϕ1, ϕ2 are arbitrary CTLK formulae, then: |≡ (set-pla fϕ)Pl aA f(Ph ϕ), |≡ (set-pla 2 ϕ)Pl aA2 (Ph ϕ), and |≡ (set-pla ϕ1U ϕ2)Pl aA(Ph ϕ1)U (Ph ϕ2). 5. VERIFICATION OF PLAUSIBILITY, TIME AND BELIEFS In this section we report preliminary results on model checking CTLKP formulae. Clearly, verifying CTLKP properties directly against models with plausibility does not make much sense, since these models are inherently infinite; what we need is a finite representation of plausibility sets. One such representation has been discussed in Section 3.3: plausibility sets can be defined by path formulae and the update operator (set-pla γ). We follow this idea here, studying the complexity of model checking CTLKP formulae against CTLK models (which can be seen as a compact representation of CTLKP models in which all the paths are assumed plausible), with the underlying idea that plausibility sets, when needed, must be defined explicitly in the object language. Below we sketch an algorithm that model-checks CTLKP formulae in time linear wrt the size of the model and the length of the formula. This means that we have extended CTLK to a more expressive language with no computational price to pay. First of all, we get rid of the belief operators (due to Theorem 15), replacing every occurrence of Baϕ with KaPl a(E f → ϕ). Now, let −→γ = γ1, ..., γk be a vector of vanilla path formulae (one per agent), with the initial vector −→γ0 = , ..., , and −→γ [γ /a] denoting vector −→γ , in which −→γ [a] is replaced with γ . Additionally, we define −→γ [0] = . We translate the resulting CTLKP formulae to ones without plausibility via function tr(ϕ) = tr−→γ0,0(ϕ), defined as follows: tr−→γ ,i(p) = p, tr−→γ ,i(ϕ1 ∧ ϕ2) = tr−→γ ,i(ϕ1) ∧ tr−→γ ,i(ϕ2), tr−→γ ,i(¬ϕ) = ¬tr−→γ ,i(ϕ), tr−→γ ,i(Kaϕ) = Ka tr−→γ ,0(ϕ), tr−→γ ,i(Pla ϕ) = tr−→γ ,a(ϕ), tr−→γ ,i((set-pla γ )ϕ) = tr−→γ [γ /a],i(ϕ), tr−→γ ,i(Ph ϕ) = tr−→γ ,0(ϕ), tr−→γ ,i( fϕ) = ftr−→γ ,i(ϕ), tr−→γ ,i(2 ϕ) = 2 tr−→γ ,i(ϕ), tr−→γ ,i(ϕ1U ϕ2) = tr−→γ ,i(ϕ1)U tr−→γ ,i(ϕ2), tr−→γ ,i(Eγ ) = E(−→γ [i] ∧ tr−→γ ,i(γ )). Note that the resulting sentences belong to the logic of CTLK+, that is CTL+ (where each path quantifier can be followed by a Boolean combination of vanilla path formulae)9 with epistemic modalities. The following proposition justifies the translation. Proposition 18. For any CTLKP formula ϕ without Ba, we have that M, q |=CTLKP ϕ iff M, q |=CTLK+ tr(ϕ). In general, model checking CTL+ (and also CTLK+) is ΔP 2 -complete. However, in our case, the Boolean combinations of path subformulae are always conjunctions of at most two non-negated elements, which allows us to propose the following model checking algorithm. First, subformulae are evaluated recursively: for every subformula ψ of ϕ, the set of states in M that satisfy ψ is computed and labeled with a new proposition pψ. Now, it is enough to define checking M, q |= ϕ for ϕ in which all (state) subformulae are propositions, with the following cases: Case M, q |= E(2 p ∧ γ): If M, q |= p, then return no. Otherwise, remove from M all the states that do not satisfy p (yielding a sparser model M ), and check the CTL formula Eγ in M , q with any CTL model-checker. Case M, q |= E( fp ∧ γ): Create M by adding a copy q of state q, in which only the transitions to states satisfying p are kept (i.e., M, q |= r iff M, q |= r; and q Rq iff qRq and M, q |= p). Then, check Eγ in M , q . 9 For the semantics of CTL+, and discussion of model checking complexity, cf. [17]. 588 The Sixth Intl. Joint Conf. on Autonomous Agents and Multi-Agent Systems (AAMAS 07) Case M, q |= E(p1U p2 ∧ p3U p4): Note that this is equivalent to checking E(p1 ∧ p3)U (p2 ∧ Ep3U p4) ∨ E(p1 ∧ p3)U (p4 ∧ Ep1U p2), which is a CTL formula. Other cases: The above cases cover all possible formulas that begin with a path quantifier. For other cases, standard CTLK model checking can be used. Theorem 19. Model checking CTLKP against CTLK models is PTIME-complete, and can be done in time O(ml), where m is the number of transitions in the model, and l is the length of the formula to be checked. That is, the complexity is no worse than for CTLK itself. 6. CONCLUSIONS In this paper a notion of plausible behavior is considered, with the underlying idea that implausible options should be usually ignored in practical reasoning about possible future courses of action. We add the new notion of plausibility to the logic of CTLK [19], and obtain a language which enables reasoning about what can (or must) plausibly happen. As a technical device to define the semantics of the resulting logic, we use a non-standard satisfaction relation |=P that allows to propagate the current set of plausible paths into subformulae. Furthermore, we propose a non-standard notion of beliefs, defined in terms of indistinguishability and plausibility. We also propose how plausibility assumptions can be specified in the object language via a plausibility update operator (in a way similar to [13]). We use this new framework to investigate some important properties of plausibility, knowledge, beliefs, and updates. In particular, we show that knowledge is an S5 modality, and that beliefs satisfy axioms K45 in general, and KD45 for the class of plausibly serial models. We also prove that believing in ϕ is knowing that ϕ plausibly holds in all plausibly possible situations. That is, the relationship between knowledge and beliefs is very natural and reflects the initial intuition precisely. Moreover, the model checking results from Section 5 show that verification for CTLKP is no more complex than for CTL and CTLK. We would like to stress that we do not see this contribution as a mere technical exercise in formal logic. Human agents use a similar concept of plausibility and practical beliefs in their everyday reasoning in order to reduce the search space and make the reasoning feasible. As a consequence, we suggest that the framework we propose may prove suitable for modeling, design, and analysis resource-bounded agents in general. We would like to thank Juergen Dix for fruitful discussions, useful comments and improvements. 7. REFERENCES [1] R. Alur, T. A. Henzinger, and O. Kupferman. Alternating-time Temporal Logic. Journal of the ACM, 49:672-713, 2002. [2] N. Bulling and W. Jamroga. Agents, beliefs and plausible behavior in a temporal setting. Technical Report IfI-06-05, Clausthal Univ. of Technology, 2006. [3] E. A. Emerson. Temporal and modal logic. In J. van Leeuwen, editor, Handbook of Theoretical Computer Science, volume B, pages 995-1072. Elsevier, 1990. [4] E.A. Emerson and J.Y. Halpern. sometimes and not never revisited: On branching versus linear time temporal logic. Journal of the ACM, 33(1):151-178, 1986. [5] R. Fagin, J. Y. Halpern, Y. Moses, and M. Y. Vardi. Reasoning about Knowledge. MIT Press: Cambridge, MA, 1995. [6] R. Fagin and J.Y. Halpern. Reasoning about knowledge and probability. Journal of ACM, 41(2):340-367, 1994. [7] N. Friedman and J.Y. Halpern. A knowledge-based framework for belief change, Part I: Foundations. In Proceedings of TARK, pages 44-64, 1994. [8] N. Friedman and J.Y. Halpern. A knowledge-based framework for belief change, Part II: Revision and update. In Proceedings of KR"94, 1994. [9] J.Y. Halpern. Reasoning about knowledge: a survey. In Handbook of Logic in Artificial Intelligence and Logic Programming. Vol. 4: Epistemic and Temporal Reasoning, pages 1-34. Oxford University Press, Oxford, 1995. [10] J.Y. Halpern and R. Fagin. Modelling knowledge and action in distributed systems. Distributed Computing, 3(4):159-177, 1989. [11] W. Jamroga and N. Bulling. A general framework for reasoning about rational agents. In Proceedings of AAMAS"07, 2007. Short paper. [12] W. Jamroga and W. van der Hoek. Agents that know how to play. Fundamenta Informaticae, 63(2-3):185-219, 2004. [13] W. Jamroga, W. van der Hoek, and M. Wooldridge. Intentions and strategies in game-like scenarios. In Progress in Artificial Intelligence: Proceedings of EPIA 2005, volume 3808 of LNAI, pages 512-523. Springer Verlag, 2005. [14] W. Jamroga and Thomas ˚Agotnes. Constructive knowledge: What agents can achieve under incomplete information. Technical Report IfI-05-10, Clausthal University of Technology, 2005. [15] B.P. Kooi. Probabilistic dynamic epistemic logic. Journal of Logic, Language and Information, 12(4):381-408, 2003. [16] P. Lamarre and Y. Shoham. Knowledge, certainty, belief, and conditionalisation (abbreviated version). In Proceedings of KR"94, pages 415-424, 1994. [17] F. Laroussinie, N. Markey, and Ph. Schnoebelen. Model checking CTL+ and FCTL is hard. In Proceedings of FoSSaCS"01, volume 2030 of LNCS, pages 318-331. Springer, 2001. [18] Y. Moses and Y. Shoham. Belief as defeasible knowledge. Artificial Intelligence, 64(2):299-321, 1993. [19] W. Penczek and A. Lomuscio. Verifying epistemic properties of multi-agent systems via bounded model checking. In Proceedings of AAMAS"03, pages 209-216, New York, NY, USA, 2003. ACM Press. [20] K. Su, A. Sattar, G. Governatori, and Q. Chen. A computationally grounded logic of knowledge, belief and certainty. In Proceedings of AAMAS"05, pages 149-156. ACM Press, 2005. [21] W. van der Hoek, M. Roberts, and M. Wooldridge. Social laws in alternating time: Effectiveness, feasibility and synthesis. Synthese, 2005. The Sixth Intl. Joint Conf. on Autonomous Agents and Multi-Agent Systems (AAMAS 07) 589

belief notion;notion of plausibility;semantics;notion of belief;belief;plausibility notion;plausibility;reasoning;indistinguishability;logic;temporal logic;plausibility update operator;framework;multi-agent system;computation tree logic

train_I-51

Learning and Joint Deliberation through Argumentation in Multi-Agent Systems

In this paper we will present an argumentation framework for learning agents (AMAL) designed for two purposes: (1) for joint deliberation, and (2) for learning from communication. The AMAL framework is completely based on learning from examples: the argument preference relation, the argument generation policy, and the counterargument generation policy are case-based techniques. For join deliberation, learning agents share their experience by forming a committee to decide upon some joint decision. We experimentally show that the argumentation among committees of agents improves both the individual and joint performance. For learning from communication, an agent engages into arguing with other agents in order to contrast its individual hypotheses and receive counterexamples; the argumentation process improves their learning scope and individual performance.

1. INTRODUCTION Argumentation frameworks for multi-agent systems can be used for different purposes like joint deliberation, persuasion, negotiation, and conflict resolution. In this paper we will present an argumentation framework for learning agents, and show that it can be used for two purposes: (1) joint deliberation, and (2) learning from communication. Argumentation-based joint deliberation involves discussion over the outcome of a particular situation or the appropriate course of action for a particular situation. Learning agents are capable of learning from experience, in the sense that past examples (situations and their outcomes) are used to predict the outcome for the situation at hand. However, since individual agents experience may be limited, individual knowledge and prediction accuracy is also limited. Thus, learning agents that are capable of arguing their individual predictions with other agents may reach better prediction accuracy after such an argumentation process. Most existing argumentation frameworks for multi-agent systems are based on deductive logic or some other deductive logic formalism specifically designed to support argumentation, such as default logic [3]). Usually, an argument is seen as a logical statement, while a counterargument is an argument offered in opposition to another argument [4, 13]; agents use a preference relation to resolve conflicting arguments. However, logic-based argumentation frameworks assume agents with preloaded knowledge and preference relation. In this paper, we focus on an Argumentation-based Multi-Agent Learning (AMAL) framework where both knowledge and preference relation are learned from experience. Thus, we consider a scenario with agents that (1) work in the same domain using a shared ontology, (2) are capable of learning from examples, and (3) communicate using an argumentative framework. Having learning capabilities allows agents effectively use a specific form of counterargument, namely the use of counterexamples. Counterexamples offer the possibility of agents learning during the argumentation process. Moreover, learning agents allow techniques that use learnt experience to generate adequate arguments and counterarguments. Specifically, we will need to address two issues: (1) how to define a technique to generate arguments and counterarguments from examples, and (2)how to define a preference relation over two conflicting arguments that have been induced from examples. This paper presents a case-based approach to address both issues. The agents use case-based reasoning (CBR) [1] to learn from past cases (where a case is a situation and its outcome) in order to predict the outcome of a new situation. We propose an argumentation protocol inside the AMAL framework at supports agents in reaching a joint prediction over a specific situation or problem - moreover, the reasoning needed to support the argumentation process will also be based on cases. In particular, we present two case-based measures, one for generating the arguments and counterarguments adequate to a particular situation and another for determining preference relation among arguments. Finally, we evaluate (1) if argumentation between learning agents can produce a joint prediction that improves over individual learning performance and (2) if learning from the counterexamples conveyed during the argumentation process increases the individual performance with precisely those cases being used while arguing among them. The paper is structured as follows. Section 2 discusses the relation among argumentation, collaboration and learning. Then Section 3 introduces our multi-agent CBR (MAC) framework and the notion of justified prediction. After that, Section 4 formally defines our argumentation framework. Sections 5 and 6 present our case-based preference relation and argument generation policies respectively. Later, Section 7 presents the argumentation protocol in our AMAL framework. After that, Section 8 presents an exemplification of the argumentation framework. Finally, Section 9 presents an empirical evaluation of our two main hypotheses. The paper closes with related work and conclusions sections. 2. ARGUMENTATION,COLLABORATION AND LEARNING Both learning and collaboration are ways in which an agent can improve individual performance. In fact, there is a clear parallelism between learning and collaboration in multi-agent systems, since both are ways in which agents can deal with their shortcomings. Let us show which are the main motivations that an agent can have to learn or to collaborate. • Motivations to learn: - Increase quality of prediction, - Increase efficiency, - Increase the range of solvable problems. • Motivations to collaborate: - Increase quality of prediction, - Increase efficiency, - Increase the range of solvable problems, - Increase the range of accessible resources. Looking at the above lists of motivation, we can easily see that learning and collaboration are very related in multi-agent systems. In fact, with the exception of the last item in the motivations to collaborate list, they are two extremes of a continuum of strategies to improve performance. An agent may choose to increase performance by learning, by collaborating, or by finding an intermediate point that combines learning and collaboration in order to improve performance. In this paper we will propose AMAL, an argumentation framework for learning agents, and will also also show how AMAL can be used both for learning from communication and for solving problems in a collaborative way: • Agents can solve problems in a collaborative way via engaging an argumentation process about the prediction for the situation at hand. Using this collaboration, the prediction can be done in a more informed way, since the information known by several agents has been taken into account. • Agents can also learn from communication with other agents by engaging an argumentation process. Agents that engage in such argumentation processes can learn from the arguments and counterexamples received from other agents, and use this information for predicting the outcomes of future situations. In the rest of this paper we will propose an argumentation framework and show how it can be used both for learning and for solving problems in a collaborative way. 3. MULTI-AGENT CBR SYSTEMS A Multi-Agent Case Based Reasoning System (MAC) M = {(A1, C1), ..., (An, Cn)} is a multi-agent system composed of A = {Ai, ..., An}, a set of CBR agents, where each agent Ai ∈ A possesses an individual case base Ci. Each individual agent Ai in a MAC is completely autonomous and each agent Ai has access only to its individual and private case base Ci. A case base Ci = {c1, ..., cm} is a collection of cases. Agents in a MAC system are able to individually solve problems, but they can also collaborate with other agents to solve problems. In this framework, we will restrict ourselves to analytical tasks, i.e. tasks like classification, where the solution of a problem is achieved by selecting a solution class from an enumerated set of solution classes. In the following we will note the set of all the solution classes by S = {S1, ..., SK }. Therefore, a case c = P, S is a tuple containing a case description P and a solution class S ∈ S. In the following, we will use the terms problem and case description indistinctly. Moreover, we will use the dot notation to refer to elements inside a tuple; e.g., to refer to the solution class of a case c, we will write c.S. Therefore, we say a group of agents perform joint deliberation, when they collaborate to find a joint solution by means of an argumentation process. However, in order to do so, an agent has to be able to justify its prediction to the other agents (i.e. generate an argument for its predicted solution that can be examined and critiqued by the other agents). The next section addresses this issue. 3.1 Justified Predictions Both expert systems and CBR systems may have an explanation component [14] in charge of justifying why the system has provided a specific answer to the user. The line of reasoning of the system can then be examined by a human expert, thus increasing the reliability of the system. Most of the existing work on explanation generation focuses on generating explanations to be provided to the user. However, in our approach we use explanations (or justifications) as a tool for improving communication and coordination among agents. We are interested in justifications since they can be used as arguments. For that purpose, we will benefit from the ability of some machine learning methods to provide justifications. A justification built by a CBR method after determining that the solution of a particular problem P was Sk is a description that contains the relevant information from the problem P that the CBR method has considered to predict Sk as the solution of P. In particular, CBR methods work by retrieving similar cases to the problem at hand, and then reusing their solutions for the current problem, expecting that since the problem and the cases are similar, the solutions will also be similar. Thus, if a CBR method has retrieved a set of cases C1, ..., Cn to solve a particular problem P the justification built will contain the relevant information from the problem P that made the CBR system retrieve that particular set of cases, i.e. it will contain the relevant information that P and C1, ..., Cn have in common. For example, Figure 1 shows a justification build by a CBR system for a toy problem (in the following sections we will show justifications for real problems). In the figure, a problem has two attributes (Traffic_light, and Cars_passing), the retrieval mechanism of the CBR system notices that by considering only the attribute Traffic_light, it can retrieve two cases that predict the same solution: wait. Thus, since only this attribute has been used, it is the only one appearing in the justification. The values of the rest of attributes are irrelevant, since whatever their value the solution class would have been the same. 976 The Sixth Intl. Joint Conf. on Autonomous Agents and Multi-Agent Systems (AAMAS 07) Problem Traffic_light: red Cars_passing: no Case 1 Traffic_light: red Cars_passing: no Solution: wait Case 3 Traffic_light: red Cars_passing: yes Solution: wait Case 4 Traffic_light: green Cars_passing: yes Solution: wait Case 2 Traffic_light: green Cars_passing: no Solution: cross Retrieved cases Solution: wait Justification Traffic_light: red Figure 1: An example of justification generation in a CBR system. Notice that, since the only relevant feature to decide is Traffic_light (the only one used to retrieve cases), it is the only one appearing in the justification. In general, the meaning of a justification is that all (or most of) the cases in the case base of an agent that satisfy the justification (i.e. all the cases that are subsumed by the justification) belong to the predicted solution class. In the rest of the paper, we will use to denote the subsumption relation. In our work, we use LID [2], a CBR method capable of building symbolic justifications such as the one exemplified in Figure 1. When an agent provides a justification for a prediction, the agent generates a justified prediction: DEFINITION 3.1. A Justified Prediction is a tuple J = A, P, S, D where agent A considers S the correct solution for problem P, and that prediction is justified a symbolic description D such that J.D J.P. Justifications can have many uses for CBR systems [8, 9]. In this paper, we are going to use justifications as arguments, in order to allow learning agents to engage in argumentation processes. 4. ARGUMENTS AND COUNTERARGUMENTS For our purposes an argument α generated by an agent A is composed of a statement S and some evidence D supporting S as correct. In the remainder of this section we will see how this general definition of argument can be instantiated in specific kind of arguments that the agents can generate. In the context of MAC systems, agents argue about predictions for new problems and can provide two kinds of information: a) specific cases P, S , and b) justified predictions: A, P, S, D . Using this information, we can define three types of arguments: justified predictions, counterarguments, and counterexamples. A justified prediction α is generated by an agent Ai to argue that Ai believes that the correct solution for a given problem P is α.S, and the evidence provided is the justification α.D. In the example depicted in Figure 1, an agent Ai may generate the argument α = Ai, P, Wait, (Traffic_light = red) , meaning that the agent Ai believes that the correct solution for P is Wait because the attribute Traffic_light equals red. A counterargument β is an argument offered in opposition to another argument α. In our framework, a counterargument consists of a justified prediction Aj, P, S , D generated by an agent Aj with the intention to rebut an argument α generated by another agent Ai, that endorses a solution class S different from that of α.S for the problem at hand and justifies this with a justification D . In the example in Figure 1, if an agent generates the argument α = Ai, P, Walk, (Cars_passing = no) , an agent that thinks that the correct solution is Wait might answer with the counterargument β = Aj, P, Wait, (Cars_passing = no ∧ Traffic_light = red) , meaning that, although there are no cars passing, the traffic light is red, and the street cannot be crossed. A counterexample c is a case that contradicts an argument α. Thus a counterexample is also a counterargument, one that states that a specific argument α is not always true, and the evidence provided is the case c. Specifically, for a case c to be a counterexample of an argument α, the following conditions have to be met: α.D c and α.S = c.S, i.e. the case must satisfy the justification α.D and the solution of c must be different than the predicted by α. By exchanging arguments and counterarguments (including counterexamples), agents can argue about the correct solution of a given problem, i.e. they can engage a joint deliberation process. However, in order to do so, they need a specific interaction protocol, a preference relation between contradicting arguments, and a decision policy to generate counterarguments (including counterexamples). In the following sections we will present these elements. 5. PREFERENCE RELATION A specific argument provided by an agent might not be consistent with the information known to other agents (or even to some of the information known by the agent that has generated the justification due to noise in training data). For that reason, we are going to define a preference relation over contradicting justified predictions based on cases. Basically, we will define a confidence measure for each justified prediction (that takes into account the cases owned by each agent), and the justified prediction with the highest confidence will be the preferred one. The idea behind case-based confidence is to count how many of the cases in an individual case base endorse a justified prediction, and how many of them are counterexamples of it. The more the endorsing cases, the higher the confidence; and the more the counterexamples, the lower the confidence. Specifically, to assess the confidence of a justified prediction α, an agent obtains the set of cases in its individual case base that are subsumed by α.D. With them, an agent Ai obtains the Y (aye) and N (nay) values: • Y Ai α = |{c ∈ Ci| α.D c.P ∧ α.S = c.S}| is the number of cases in the agent"s case base subsumed by the justification α.D that belong to the solution class α.S, • NAi α = |{c ∈ Ci| α.D c.P ∧ α.S = c.S}| is the number of cases in the agent"s case base subsumed by justification α.D that do not belong to that solution class. The Sixth Intl. Joint Conf. on Autonomous Agents and Multi-Agent Systems (AAMAS 07) 977 + + + + + + - - + Figure 2: Confidence of arguments is evaluated by contrasting them against the case bases of the agents. An agent estimates the confidence of an argument as: CAi (α) = Y Ai α 1 + Y Ai α + NAi α i.e. the confidence on a justified prediction is the number of endorsing cases divided by the number of endorsing cases plus counterexamples. Notice that we add 1 to the denominator, this is to avoid giving excessively high confidences to justified predictions whose confidence has been computed using a small number of cases. Notice that this correction follows the same idea than the Laplace correction to estimate probabilities. Figure 2 illustrates the individual evaluation of the confidence of an argument, in particular, three endorsing cases and one counterexample are found in the case base of agents Ai, giving an estimated confidence of 0.6 Moreover, we can also define the joint confidence of an argument α as the confidence computed using the cases present in the case bases of all the agents in the group: C(α) = i Y Ai α 1 + i Y Ai α + NAi α Notice that, to collaboratively compute the joint confidence, the agents only have to make public the aye and nay values locally computed for a given argument. In our framework, agents use this joint confidence as the preference relation: a justified prediction α is preferred over another one β if C(α) ≥ C(β). 6. GENERATION OF ARGUMENTS In our framework, arguments are generated by the agents from cases, using learning methods. Any learning method able to provide a justified prediction can be used to generate arguments. For instance, decision trees and LID [2] are suitable learning methods. Specifically, in the experiments reported in this paper agents use LID. Thus, when an agent wants to generate an argument endorsing that a specific solution class is the correct solution for a problem P, it generates a justified prediction as explained in Section 3.1. For instance, Figure 3 shows a real justification generated by LID after solving a problem P in the domain of marine sponges identification. In particular, Figure 3 shows how when an agent receives a new problem to solve (in this case, a new sponge to determine its order), the agent uses LID to generate an argument (consisting on a justified prediction) using the cases in the case base of the agent. The justification shown in Figure 3 can be interpreted saying that the predicted solution is hadromerida because the smooth form of the megascleres of the spiculate skeleton of the sponge is of type tylostyle, the spikulate skeleton of the sponge has no uniform length, and there is no gemmules in the external features of the sponge. Thus, the argument generated will be α = A1, P, hadromerida, D1 . 6.1 Generation of Counterarguments As previously stated, agents may try to rebut arguments by generating counterargument or by finding counterexamples. Let us explain how they can be generated. An agent Ai wants to generate a counterargument β to rebut an argument α when α is in contradiction with the local case base of Ai. Moreover, while generating such counterargument β, Ai expects that β is preferred over α. For that purpose, we will present a specific policy to generate counterarguments based on the specificity criterion [10]. The specificity criterion is widely used in deductive frameworks for argumentation, and states that between two conflicting arguments, the most specific should be preferred since it is, in principle, more informed. Thus, counterarguments generated based on the specificity criterion are expected to be preferable (since they are more informed) to the arguments they try to rebut. However, there is no guarantee that such counterarguments will always win, since, as we have stated in Section 5, agents in our framework use a preference relation based on joint confidence. Moreover, one may think that it would be better that the agents generate counterarguments based on the joint confidence preference relation; however it is not obvious how to generate counterarguments based on joint confidence in an efficient way, since collaboration is required in order to evaluate joint confidence. Thus, the agent generating the counterargument should constantly communicate with the other agents at each step of the induction algorithm used to generate counterarguments (presently one of our future research lines). Thus, in our framework, when an agent wants to generate a counterargument β to an argument α, β has to be more specific than α (i.e. α.D < β.D). The generation of counterarguments using the specificity criterion imposes some restrictions over the learning method, although LID or ID3 can be easily adapted for this task. For instance, LID is an algorithm that generates a description starting from scratch and heuristically adding features to that term. Thus, at every step, the description is made more specific than in the previous step, and the number of cases that are subsumed by that description is reduced. When the description covers only (or almost only) cases of a single solution class LID terminates and predicts that solution class. To generate a counterargument to an argument α LID just has to use as starting point the description α.D instead of starting from scratch. In this way, the justification provided by LID will always be subsumed by α.D, and thus the resulting counterargument will be more specific than α. However, notice that LID may sometimes not be able to generate counterarguments, since LID may not be able to specialize the description α.D any further, or because the agent Ai has no case inCi that is subsumed by α.D. Figure 4 shows how an agent A2 that disagreed with the argument shown in Figure 3, generates a counterargument using LID. Moreover, Figure 4 shows the generation of a counterargument β1 2 for the argument α0 1 (in Figure 3) that is a specialization of α0 1. 978 The Sixth Intl. Joint Conf. on Autonomous Agents and Multi-Agent Systems (AAMAS 07) Solution: hadromerida Justification: D1 Sponge Spikulate skeleton External features External features Gemmules: no Spikulate Skeleton Megascleres Uniform length: no Megascleres Smooth form: tylostyle Case Base of A1 LID New sponge P Figure 3: Example of a real justification generated by LID in the marine sponges data set. Specifically, in our experiments, when an agent Ai wants to rebut an argument α, uses the following policy: 1. Agent Ai uses LID to try to find a counterargument β more specific than α; if found, β is sent to the other agent as a counterargument of α. 2. If not found, then Ai searches for a counterexample c ∈ Ci of α. If a case c is found, then c is sent to the other agent as a counterexample of α. 3. If no counterexamples are found, then Ai cannot rebut the argument α. 7. ARGUMENTATION-BASED MULTI-AGENT LEARNING The interaction protocol of AMAL allows a group of agents A1, ..., An to deliberate about the correct solution of a problem P by means of an argumentation process. If the argumentation process arrives to a consensual solution, the joint deliberation ends; otherwise a weighted vote is used to determine the joint solution. Moreover, AMAL also allows the agents to learn from the counterexamples received from other agents. The AMAL protocol consists on a series of rounds. In the initial round, each agent states which is its individual prediction for P. Then, at each round an agent can try to rebut the prediction made by any of the other agents. The protocol uses a token passing mechanism so that agents (one at a time) can send counterarguments or counterexamples if they disagree with the prediction made by any other agent. Specifically, each agent is allowed to send one counterargument or counterexample each time he gets the token (notice that this restriction is just to simplify the protocol, and that it does not restrict the number of counterargument an agent can sent, since they can be delayed for subsequent rounds). When an agent receives a counterargument or counterexample, it informs the other agents if it accepts the counterargument (and changes its prediction) or not. Moreover, agents have also the opportunity to answer to counterarguments when they receive the token, by trying to generate a counterargument to the counterargument. When all the agents have had the token once, the token returns to the first agent, and so on. If at any time in the protocol, all the agents agree or during the last n rounds no agent has generated any counterargument, the protocol ends. Moreover, if at the end of the argumentation the agents have not reached an agreement, then a voting mechanism that uses the confidence of each prediction as weights is used to decide the final solution (Thus, AMAL follows the same mechanism as human committees, first each individual member of a committee exposes his arguments and discuses those of the other members (joint deliberation), and if no consensus is reached, then a voting mechanism is required). At each iteration, agents can use the following performatives: • assert(α): the justified prediction held during the next round will be α. An agent can only hold a single prediction at each round, thus is multiple asserts are send, only the last one is considered as the currently held prediction. • rebut(β, α): the agent has found a counterargument β to the prediction α. We will define Ht = αt 1, ..., αt n as the predictions that each of the n agents hold at a round t. Moreover, we will also define contradict(αt i) = {α ∈ Ht|α.S = αt i.S} as the set of contradicting arguments for an agent Ai in a round t, i.e. the set of arguments at round t that support a different solution class than αt i. The protocol is initiated because one of the agents receives a problem P to be solved. After that, the agent informs all the other agents about the problem P to solve, and the protocol starts: 1. At round t = 0, each one of the agents individually solves P, and builds a justified prediction using its own CBR method. Then, each agent Ai sends the performative assert(α0 i ) to the other agents. Thus, the agents know H0 = α0 i , ..., α0 n . Once all the predictions have been sent the token is given to the first agent A1. 2. At each round t (other than 0), the agents check whether their arguments in Ht agree. If they do, the protocol moves to step 5. Moreover, if during the last n rounds no agent has sent any counterexample or counterargument, the protocol also moves to step 5. Otherwise, the agent Ai owner of the token tries to generate a counterargument for each of the opposing arguments in contradict(αt i) ⊆ Ht (see Section 6.1). Then, the counterargument βt i against the prediction αt j with the lowest confidence C(αt j) is selected (since αt j is the prediction more likely to be successfully rebutted). • If βt i is a counterargument, then, Ai locally compares αt i with βt i by assessing their confidence against its individual case base Ci (see Section 5) (notice that Ai is comparing its previous argument with the counterargument that Ai itself has just generated and that is about The Sixth Intl. Joint Conf. on Autonomous Agents and Multi-Agent Systems (AAMAS 07) 979 Sponge Spikulate skeleton External features External features Gemmules: no Growing: Spikulate Skeleton Megascleres Uniform length: no Megascleres Smooth form: tylostyle Growing Grow: massive Case Base of A2 LID Solution: astrophorida Justification: D2 Figure 4: Generation of a counterargument using LID in the sponges data set. to send to Aj). If CAi (βt i ) > CAi (αt i), then Ai considers that βt i is stronger than its previous argument, changes its argument to βt i by sending assert(βt i ) to the rest of the agents (the intuition behind this is that since a counterargument is also an argument, Ai checks if the newly counterargument is a better argument than the one he was previously holding) and rebut(βt i , αt j) to Aj. Otherwise (i.e. CAi (βt i ) ≤ CAi (αt i)), Ai will send only rebut(βt i , αt j) to Aj. In any of the two situations the protocol moves to step 3. • If βt i is a counterexample c, then Ai sends rebut(c, αt j) to Aj. The protocol moves to step 4. • If Ai cannot generate any counterargument or counterexample, the token is sent to the next agent, a new round t + 1 starts, and the protocol moves to state 2. 3. The agent Aj that has received the counterargument βt i , locally compares it against its own argument, αt j, by locally assessing their confidence. If CAj (βt i ) > CAj (αt j), then Aj will accept the counterargument as stronger than its own argument, and it will send assert(βt i ) to the other agents. Otherwise (i.e. CAj (βt i ) ≤ CAj (αt j)), Aj will not accept the counterargument, and will inform the other agents accordingly. Any of the two situations start a new round t + 1, Ai sends the token to the next agent, and the protocol moves back to state 2. 4. The agent Aj that has received the counterexample c retains it into its case base and generates a new argument αt+1 j that takes into account c, and informs the rest of the agents by sending assert(αt+1 j ) to all of them. Then, Ai sends the token to the next agent, a new round t + 1 starts, and the protocol moves back to step 2. 5. The protocol ends yielding a joint prediction, as follows: if the arguments in Ht agree then their prediction is the joint prediction, otherwise a voting mechanism is used to decide the joint prediction. The voting mechanism uses the joint confidence measure as the voting weights, as follows: S = arg max Sk∈S αi∈Ht|αi.S=Sk C(αi) Moreover, in order to avoid infinite iterations, if an agent sends twice the same argument or counterargument to the same agent, the message is not considered. 8. EXEMPLIFICATION Let us consider a system composed of three agents A1, A2 and A3. One of the agents, A1 receives a problem P to solve, and decides to use AMAL to solve it. For that reason, invites A2 and A3 to take part in the argumentation process. They accept the invitation, and the argumentation protocol starts. Initially, each agent generates its individual prediction for P, and broadcasts it to the other agents. Thus, all of them can compute H0 = α0 1, α0 2, α0 3 . In particular, in this example: • α0 1 = A1, P, hadromerida, D1 • α0 2 = A2, P, astrophorida, D2 • α0 3 = A3, P, axinellida, D3 A1 starts owning the token and tries to generate counterarguments for α0 2 and α0 3, but does not succeed, however it has one counterexample c13 for α0 3. Thus, A1 sends the the message rebut( c13, α0 3) to A3. A3 incorporates c13 into its case base and tries to solve the problem P again, now taking c13 into consideration. A3 comes up with the justified prediction α1 3 = A3, P, hadromerida, D4 , and broadcasts it to the rest of the agents with the message assert(α1 3). Thus, all of them know the new H1 = α0 1, α0 2, α1 3 . Round 1 starts and A2 gets the token. A2 tries to generate counterarguments for α0 1 and α1 3 and only succeeds to generate a counterargument β1 2 = A2, P, astrophorida, D5 against α1 3. The counterargument is sent to A3 with the message rebut(β1 2 , α1 3). Agent A3 receives the counterargument and assesses its local confidence. The result is that the individual confidence of the counterargument β1 2 is lower than the local confidence of α1 3. Therefore, A3 does not accept the counterargument, and thus H2 = α0 1, α0 2, α1 3 . Round 2 starts and A3 gets the token. A3 generates a counterargument β2 3 = A3, P, hadromerida, D6 for α0 2 and sends it to A2 with the message rebut(β2 3 , α0 2). Agent A2 receives the counterargument and assesses its local confidence. The result is that the local confidence of the counterargument β2 3 is higher than the local confidence of α0 2. Therefore, A2 accepts the counterargument and informs the rest of the agents with the message assert(β2 3 ). After that, H3 = α0 1, β2 3 , α1 3 . At Round 3, since all the agents agree (all the justified predictions in H3 predict hadromerida as the solution class) The protocol ends, and A1 (the agent that received the problem) considers hadromerida as the joint solution for the problem P. 9. EXPERIMENTAL EVALUATION 980 The Sixth Intl. Joint Conf. on Autonomous Agents and Multi-Agent Systems (AAMAS 07) SPONGE 75 77 79 81 83 85 87 89 91 2 3 4 5 AMAL Voting Individual SOYBEAN 55 60 65 70 75 80 85 90 2 3 4 5 AMAL Voting Individual Figure 5: Individual and joint accuracy for 2 to 5 agents. In this section we empirically evaluate the AMAL argumentation framework. We have made experiments in two different data sets: soybean (from the UCI machine learning repository) and sponge (a relational data set). The soybean data set has 307 examples and 19 solution classes, while the sponge data set has 280 examples and 3 solution classes. In an experimental run, the data set is divided in 2 sets: the training set and the test set. The training set examples are distributed among 5 different agents without replication, i.e. there is no example shared by two agents. In the testing stage, problems in the test set arrive randomly to one of the agents, and their goal is to predict the correct solution. The experiments are designed to test two hypotheses: (H1) that argumentation is a useful framework for joint deliberation and can improve over other typical methods such as voting; and (H2) that learning from communication improves the individual performance of a learning agent participating in an argumentation process. Moreover, we also expect that the improvement achieved from argumentation will increase as the number of agents participating in the argumentation increases (since more information will be taken into account). Concerning H1 (argumentation is a useful framework for joint deliberation), we ran 4 experiments, using 2, 3, 4, and 5 agents respectively (in all experiments each agent has a 20% of the training data, since the training is always distributed among 5 agents). Figure 5 shows the result of those experiments in the sponge and soybean data sets. Classification accuracy is plotted in the vertical axis, and in the horizontal axis the number of agents that took part in the argumentation processes is shown. For each number of agents, three bars are shown: individual, Voting, and AMAL. The individual bar shows the average accuracy of individual agents predictions; the voting bar shows the average accuracy of the joint prediction achieved by voting but without any argumentation; and finally the AMAL bar shows the average accuracy of the joint prediction using argumentation. The results shown are the average of 5 10-fold cross validation runs. Figure 5 shows that collaboration (voting and AMAL) outperforms individual problem solving. Moreover, as we expected, the accuracy improves as more agents collaborate, since more information is taken into account. We can also see that AMAL always outperforms standard voting, proving that joint decisions are based on better information as provided by the argumentation process. For instance, the joint accuracy for 2 agents in the sponge data set is of 87.57% for AMAL and 86.57% for voting (while individual accuracy is just 80.07%). Moreover, the improvement achieved by AMAL over Voting is even larger in the soybean data set. The reason is that the soybean data set is more difficult (in the sense that agents need more data to produce good predictions). These experimental results show that AMAL effectively exploits the opportunity for improvement: the accuracy is higher only because more agents have changed their opinion during argumentation (otherwise they would achieve the same result as Voting). Concerning H2 (learning from communication in argumentation processes improves individual prediction ), we ran the following experiment: initially, we distributed a 25% of the training set among the five agents; after that, the rest of the cases in the training set is sent to the agents one by one; when an agent receives a new training case, it has several options: the agent can discard it, the agent can retain it, or the agent can use it for engaging an argumentation process. Figure 6 shows the result of that experiment for the two data sets. Figure 6 contains three plots, where NL (not learning) shows accuracy of an agent with no learning at all; L (learning), shows the evolution of the individual classification accuracy when agents learn by retaining the training cases they individually receive (notice that when all the training cases have been retained at 100%, the accuracy should be equal to that of Figure 5 for individual agents); and finally LFC (learning from communication) shows the evolution of the individual classification accuracy of learning agents that also learn by retaining those counterexamples received during argumentation (i.e. they learn both from training examples and counterexamples). Figure 6 shows that if an agent Ai learns also from communication, Ai can significantly improve its individual performance with just a small number of additional cases (those selected as relevant counterexamples for Ai during argumentation). For instance, in the soybean data set, individual agents have achieved an accuracy of 70.62% when they also learn from communication versus an accuracy of 59.93% when they only learn from their individual experience. The number of cases learnt from communication depends on the properties of the data set: in the sponges data set, agents have retained only very few additional cases, and significantly improved individual accuracy; namely they retain 59.96 cases in average (compared to the 50.4 cases retained if they do not learn from communication). In the soybean data set more counterexamples are learnt to significantly improve individual accuracy, namely they retain 87.16 cases in average (compared to 55.27 cases retained if they do not learn from communication). Finally, the fact that both data sets show a significant improvement points out the adaptive nature of the argumentation-based approach to learning from communication: the useful cases are selected as counterexamples (and no more than those needed), and they have the intended effect. 10. RELATED WORK Concerning CBR in a multi-agent setting, the first research was on negotiated case retrieval [11] among groups of agents. Our work on multi-agent case-based learning started in 1999 [6]; later Mc Ginty and Smyth [7] presented a multi-agent collaborative CBR approach (CCBR) for planning. Finally, another interesting approach is multi-case-base reasoning (MCBR) [5], that deals with The Sixth Intl. Joint Conf. on Autonomous Agents and Multi-Agent Systems (AAMAS 07) 981 SPONGE 60 65 70 75 80 85 25% 40% 55% 70% 85% 100% LFC L NL SOYBEAN 20 30 40 50 60 70 80 90 25% 40% 55% 70% 85% 100% LFC L NL Figure 6: Learning from communication resulting from argumentation in a system composed of 5 agents. distributed systems where there are several case bases available for the same task and addresses the problems of cross-case base adaptation. The main difference is that our MAC approach is a way to distribute the Reuse process of CBR (using a voting system) while Retrieve is performed individually by each agent; the other multiagent CBR approaches, however, focus on distributing the Retrieve process. Research on MAS argumentation focus on several issues like a) logics, protocols and languages that support argumentation, b) argument selection and c) argument interpretation. Approaches for logic and languages that support argumentation include defeasible logic [4] and BDI models [13]. Although argument selection is a key aspect of automated argumentation (see [12] and [13]), most research has been focused on preference relations among arguments. In our framework we have addressed both argument selection and preference relations using a case-based approach. 11. CONCLUSIONS AND FUTURE WORK In this paper we have presented an argumentation-based framework for multi-agent learning. Specifically, we have presented AMAL, a framework that allows a group of learning agents to argue about the solution of a given problem and we have shown how the learning capabilities can be used to generate arguments and counterarguments. The experimental evaluation shows that the increased amount of information provided to the agents by the argumentation process increases their predictive accuracy, and specially when an adequate number of agents take part in the argumentation. The main contributions of this work are: a) an argumentation framework for learning agents; b) a case-based preference relation over arguments, based on computing an overall confidence estimation of arguments; c) a case-based policy to generate counterarguments and select counterexamples; and d) an argumentation-based approach for learning from communication. Finally, in the experiments presented here a learning agent would retain all counterexamples submitted by the other agent; however, this is a very simple case retention policy, and we will like to experiment with more informed policies - with the goal that individual learning agents could significantly improve using only a small set of cases proposed by other agents. Finally, our approach is focused on lazy learning, and future works aims at incorporating eager inductive learning inside the argumentative framework for learning from communication. 12. REFERENCES [1] Agnar Aamodt and Enric Plaza. Case-based reasoning: Foundational issues, methodological variations, and system approaches. Artificial Intelligence Communications, 7(1):39-59, 1994. [2] E. Armengol and E. Plaza. Lazy induction of descriptions for relational case-based learning. In ECML"2001, pages 13-24, 2001. [3] Gerhard Brewka. Dynamic argument systems: A formal model of argumentation processes based on situation calculus. Journal of Logic and Computation, 11(2):257-282, 2001. [4] Carlos I. Chesñevar and Guillermo R. Simari. Formalizing Defeasible Argumentation using Labelled Deductive Systems. Journal of Computer Science & Technology, 1(4):18-33, 2000. [5] D. Leake and R. Sooriamurthi. Automatically selecting strategies for multi-case-base reasoning. In S. Craw and A. Preece, editors, ECCBR"2002, pages 204-219, Berlin, 2002. Springer Verlag. [6] Francisco J. Martín, Enric Plaza, and Josep-Lluis Arcos. Knowledge and experience reuse through communications among competent (peer) agents. International Journal of Software Engineering and Knowledge Engineering, 9(3):319-341, 1999. [7] Lorraine McGinty and Barry Smyth. Collaborative case-based reasoning: Applications in personalized route planning. In I. Watson and Q. Yang, editors, ICCBR, number 2080 in LNAI, pages 362-376. Springer-Verlag, 2001. [8] Santi Ontañón and Enric Plaza. Justification-based multiagent learning. In ICML"2003, pages 576-583. Morgan Kaufmann, 2003. [9] Enric Plaza, Eva Armengol, and Santiago Ontañón. The explanatory power of symbolic similarity in case-based reasoning. Artificial Intelligence Review, 24(2):145-161, 2005. [10] David Poole. On the comparison of theories: Preferring the most specific explanation. In IJCAI-85, pages 144-147, 1985. [11] M V Nagendra Prassad, Victor R Lesser, and Susan Lander. Retrieval and reasoning in distributed case bases. Technical report, UMass Computer Science Department, 1995. [12] K. Sycara S. Kraus and A. Evenchik. Reaching agreements through argumentation: a logical model and implementation. Artificial Intelligence Journal, 104:1-69, 1998. [13] N. R. Jennings S. Parsons, C. Sierra. Agents that reason and negotiate by arguing. Journal of Logic and Computation, 8:261-292, 1998. [14] Bruce A. Wooley. Explanation component of software systems. ACM CrossRoads, 5.1, 1998. 982 The Sixth Intl. Joint Conf. on Autonomous Agents and Multi-Agent Systems (AAMAS 07)

group;case-base reason;learning agent;multi-agent learn;case-based policy;argumentation framework;predictive accuracy;argumentation;learning from communication;argumentation protocol;multi-agent system;collaboration;joint deliberation

train_I-52

A Unified and General Framework for Argumentation-based Negotiation

This paper proposes a unified and general framework for argumentation-based negotiation, in which the role of argumentation is formally analyzed. The framework makes it possible to study the outcomes of an argumentation-based negotiation. It shows what an agreement is, how it is related to the theories of the agents, when it is possible, and how this can be attained by the negotiating agents in this case. It defines also the notion of concession, and shows in which situation an agent will make one, as well as how it influences the evolution of the dialogue.

1. INTRODUCTION Roughly speaking, negotiation is a process aiming at finding some compromise or consensus between two or several agents about some matters of collective agreement, such as pricing products, allocating resources, or choosing candidates. Negotiation models have been proposed for the design of systems able to bargain in an optimal way with other agents for example, buying or selling products in ecommerce. Different approaches to automated negotiation have been investigated, including game-theoretic approaches (which usually assume complete information and unlimited computation capabilities) [11], heuristic-based approaches which try to cope with these limitations [6], and argumentation-based approaches [2, 3, 7, 8, 9, 12, 13] which emphasize the importance of exchanging information and explanations between negotiating agents in order to mutually influence their behaviors (e.g. an agent may concede a goal having a small priority), and consequently the outcome of the dialogue. Indeed, the two first types of settings do not allow for the addition of information or for exchanging opinions about offers. Integrating argumentation theory in negotiation provides a good means for supplying additional information and also helps agents to convince each other by adequate arguments during a negotiation dialogue. Indeed, an offer supported by a good argument has a better chance to be accepted by an agent, and can also make him reveal his goals or give up some of them. The basic idea behind an argumentationbased approach is that by exchanging arguments, the theories of the agents (i.e. their mental states) may evolve, and consequently, the status of offers may change. For instance, an agent may reject an offer because it is not acceptable for it. However, the agent may change its mind if it receives a strong argument in favor of this offer. Several proposals have been made in the literature for modeling such an approach. However, the work is still preliminary. Some researchers have mainly focused on relating argumentation with protocols. They have shown how and when arguments in favor of offers can be computed and exchanged. Others have emphasized on the decision making problem. In [3, 7], the authors argued that selecting an offer to propose at a given step of the dialogue is a decision making problem. They have thus proposed an argumentationbased decision model, and have shown how such a model can be related to the dialogue protocol. In most existing works, there is no deep formal analysis of the role of argumentation in negotiation dialogues. It is not clear how argumentation can influence the outcome of the dialogue. Moreover, basic concepts in negotiation such as agreement (i.e. optimal solutions, or compromise) and concession are neither defined nor studied. This paper aims to propose a unified and general framework for argumentation-based negotiation, in which the role of argumentation is formally analyzed, and where the existing systems can be restated. In this framework, a negotiation dialogue takes place between two agents on a set O of offers, whose structure is not known. The goal of a negotiation is to find among elements of O, an offer that satisfies more or less 967 978-81-904262-7-5 (RPS) c 2007 IFAAMAS the preferences of both agents. Each agent is supposed to have a theory represented in an abstract way. A theory consists of a set A of arguments whose structure and origin are not known, a function specifying for each possible offer in O, the arguments of A that support it, a non specified conflict relation among the arguments, and finally a preference relation between the arguments. The status of each argument is defined using Dung"s acceptability semantics. Consequently, the set of offers is partitioned into four subsets: acceptable, rejected, negotiable and non-supported offers. We show how an agent"s theory may evolve during a negotiation dialogue. We define formally the notions of concession, compromise, and optimal solution. Then, we propose a protocol that allows agents i) to exchange offers and arguments, and ii) to make concessions when necessary. We show that dialogues generated under such a protocol terminate, and even reach optimal solutions when they exist. This paper is organized as follows: Section 2 introduces the logical language that is used in the rest of the paper. Section 3 defines the agents as well as their theories. In section 4, we study the properties of these agents" theories. Section 5 defines formally an argumentation-based negotiation, shows how the theories of agents may evolve during a dialogue, and how this evolution may influence the outcome of the dialogue. Two kinds of outcomes: optimal solution and compromise are defined, and we show when such outcomes are reached. Section 6 illustrates our general framework through some examples. Section 7 compares our formalism with existing ones. Section 8 concludes and presents some perspectives. Due to lack of space, the proofs are not included. These last are in a technical report that we will make available online at some later time. 2. THE LOGICAL LANGUAGE In what follows, L will denote a logical language, and ≡ is an equivalence relation associated with it. From L, a set O = {o1, . . . , on} of n offers is identified, such that oi, oj ∈ O such that oi ≡ oj. This means that the offers are different. Offers correspond to the different alternatives that can be exchanged during a negotiation dialogue. For instance, if the agents try to decide the place of their next meeting, then the set O will contain different towns. Different arguments can be built from L. The set Args(L) will contain all those arguments. By argument, we mean a reason in believing or of doing something. In [3], it has been argued that the selection of the best offer to propose at a given step of the dialogue is a decision problem. In [4], it has been shown that in an argumentation-based approach for decision making, two kinds of arguments are distinguished: arguments supporting choices (or decisions), and arguments supporting beliefs. Moreover, it has been acknowledged that the two categories of arguments are formally defined in different ways, and they play different roles. Indeed, an argument in favor of a decision, built both on an agent"s beliefs and goals, tries to justify the choice; whereas an argument in favor of a belief, built only from beliefs, tries to destroy the decision arguments, in particular the beliefs part of those decision arguments. Consequently, in a negotiation dialogue, those two kinds of arguments are generally exchanged between agents. In what follows, the set Args(L) is then divided into two subsets: a subset Argso(L) of arguments supporting offers, and a subset Argsb(L) of arguments supporting beliefs. Thus, Args(L) = Argso(L) ∪ Argsb(L). As in [5], in what follows, we consider that the structure of the arguments is not known. Since the knowledge bases from which arguments are built may be inconsistent, the arguments may be conflicting too. In what follows, those conflicts will be captured by the relation RL, thus RL ⊆ Args(L) × Args(L). Three assumptions are made on this relation: First the arguments supporting different offers are conflicting. The idea behind this assumption is that since offers are exclusive, an agent has to choose only one at a given step of the dialogue. Note that, the relation RL is not necessarily symmetric between the arguments of Argsb(L). The second hypothesis says that arguments supporting the same offer are also conflicting. The idea here is to return the strongest argument among these arguments. The third condition does not allow an argument in favor of an offer to attack an argument supporting a belief. This avoids wishful thinking. Formally: Definition 1. RL ⊆ Args(L) × Args(L) is a conflict relation among arguments such that: • ∀a, a ∈ Argso(L), s.t. a = a , a RL a • a ∈ Argso(L) and a ∈ Argsb(L) such that a RL a Note that the relation RL is not symmetric. This is due to the fact that arguments of Argsb(L) may be conflicting but not necessarily in a symmetric way. In what follows, we assume that the set Args(L) of arguments is finite, and each argument is attacked by a finite number of arguments. 3. NEGOTIATING AGENTS THEORIES AND REASONING MODELS In this section we define formally the negotiating agents, i.e. their theories, as well as the reasoning model used by those agents in a negotiation dialogue. 3.1 Negotiating agents theories Agents involved in a negotiation dialogue, called negotiating agents, are supposed to have theories. In this paper, the theory of an agent will not refer, as usual, to its mental states (i.e. its beliefs, desires and intentions). However, it will be encoded in a more abstract way in terms of the arguments owned by the agent, a conflict relation among those arguments, a preference relation between the arguments, and a function that specifies which arguments support offers of the set O. We assume that an agent is aware of all the arguments of the set Args(L). The agent is even able to express a preference between any pair of arguments. This does not mean that the agent will use all the arguments of Args(L), but it encodes the fact that when an agent receives an argument from another agent, it can interpret it correctly, and it can also compare it with its own arguments. Similarly, each agent is supposed to be aware of the conflicts between arguments. This also allows us to encode the fact that an agent can recognize whether the received argument is in conflict or not with its arguments. However, in its theory, only the conflicts between its own arguments are considered. Definition 2 (Negotiating agent theory). Let O be a set of n offers. A negotiating agent theory is a tuple A, F, , R, Def such that: • A ⊆ Args(L). 968 The Sixth Intl. Joint Conf. on Autonomous Agents and Multi-Agent Systems (AAMAS 07) • F: O → 2A s.t ∀i, j with i = j, F(oi) ∩ F(oj) = ∅. Let AO = ∪F(oi) with i = 1, . . . , n. • ⊆ Args(L) × Args(L) is a partial preorder denoting a preference relation between arguments. • R ⊆ RL such that R ⊆ A × A • Def ⊆ A × A such that ∀ a, b ∈ A, a defeats b, denoted a Def b iff: - a R b, and - not (b a) The function F returns the arguments supporting offers in O. In [4], it has been argued that any decision may have arguments supporting it, called arguments PRO, and arguments against it, called arguments CONS. Moreover, these two types of arguments are not necessarily conflicting. For simplicity reasons, in this paper we consider only arguments PRO. Moreover, we assume that an argument cannot support two distinct offers. However, it may be the case that an offer is not supported at all by arguments, thus F(oi) may be empty. Example 1. Let O = {o1, o2, o3} be a set of offers. The following theory is the theory of agent i: • A = {a1, a2, a3, a4} • F(o1) = {a1}, F(o2) = {a2}, F(o3) = ∅. Thus, Ao = {a1, a2} • = {(a1, a2), (a2, a1), (a3, a2), (a4, a3)} • R = {a1, a2), (a2, a1), (a3, a2), (a4, a3)} • Def = {(a4, a3), (a3, a2)} From the above definition of agent theory, the following hold: Property 1. • Def ⊆ R • ∀a, a ∈ F(oi), a R a 3.2 The reasoning model From the theory of an agent, one can define the argumentation system used by that agent for reasoning about the offers and the arguments, i.e. for computing the status of the different offers and arguments. Definition 3 (Argumentation system). Let A, F, , R, Def be the theory of an agent. The argumentation system of that agent is the pair A, Def . In [5], different acceptability semantics have been introduced for computing the status of arguments. These are based on two basic concepts, defence and conflict-free, defined as follows: Definition 4 (Defence/conflict-free). Let S ⊆ A. • S defends an argument a iff each argument that defeats a is defeated by some argument in S. • S is conflict-free iff there exist no a, a in S such that a Def a . Definition 5 (Acceptability semantics). Let S be a conflict-free set of arguments, and let T : 2A → 2A be a function such that T (S) = {a | a is defended by S}. • S is a complete extension iff S = T (S). • S is a preferred extension iff S is a maximal (w.r.t set ⊆) complete extension. • S is a grounded extension iff it is the smallest (w.r.t set ⊆) complete extension. Let E1, . . . , Ex denote the different extensions under a given semantics. Note that there is only one grounded extension. It contains all the arguments that are not defeated, and those arguments that are defended directly or indirectly by nondefeated arguments. Theorem 1. Let A, Def the argumentation system defined as shown above. 1. It may have x ≥ 1 preferred extensions. 2. The grounded extensions is S = i≥1 T (∅). Note that when the grounded extension (or the preferred extension) is empty, this means that there is no acceptable offer for the negotiating agent. Example 2. In example 1, there is one preferred extension, E = {a1, a2, a4}. Now that the acceptability semantics is defined, we are ready to define the status of any argument. Definition 6 (Argument status). Let A, Def be an argumentation system, and E1, . . . , Ex its extensions under a given semantics. Let a ∈ A. 1. a is accepted iff a ∈ Ei, ∀Ei with i = 1, . . . , x. 2. a is rejected iff Ei such that a ∈ Ei. 3. a is undecided iff a is neither accepted nor rejected. This means that a is in some extensions and not in others. Note that A = {a|a is accepted} ∪ {a|a is rejected} ∪ {a|a is undecided}. Example 3. In example 1, the arguments a1, a2 and a4 are accepted, whereas the argument a3 is rejected. As said before, agents use argumentation systems for reasoning about offers. In a negotiation dialogue, agents propose and accept offers that are acceptable for them, and reject bad ones. In what follows, we will define the status of an offer. According to the status of arguments, one can define four statuses of the offers as follows: Definition 7 (Offers status). Let o ∈ O. • The offer o is acceptable for the negotiating agent iff ∃ a ∈ F(o) such that a is accepted. Oa = {oi ∈ O, such that oi is acceptable}. • The offer o is rejected for the negotiating agent iff ∀ a ∈ F(o), a is rejected. Or = {oi ∈ O, such that oi is rejected}. The Sixth Intl. Joint Conf. on Autonomous Agents and Multi-Agent Systems (AAMAS 07) 969 • The offer o is negotiable iff ∀ a ∈ F(o), a is undecided. On = {oi ∈ O, such that oi is negotiable}. • The offer o is non-supported iff it is neither acceptable, nor rejected or negotiable. Ons = {oi ∈ O, such that oi is non-supported offers}. Example 4. In example 1, the two offers o1 and o2 are acceptable since they are supported by accepted arguments, whereas the offer o3 is non-supported since it has no argument in its favor. From the above definitions, the following results hold: Property 2. Let o ∈ O. • O = Oa ∪ Or ∪ On ∪ Ons. • The set Oa may contain more than one offer. From the above partition of the set O of offers, a preference relation between offers is defined. Let Ox and Oy be two subsets of O. Ox Oy means that any offer in Ox is preferred to any offer in the set Oy. We can write also for two offers oi, oj, oi oj iff oi ∈ Ox, oj ∈ Oy and Ox Oy. Definition 8 (Preference between offers). Let O be a set of offers, and Oa, Or, On, Ons its partition. Oa On Ons Or. Example 5. In example 1, we have o1 o3, and o2 o3. However, o1 and o2 are indifferent. 4. THE STRUCTURE OF NEGOTIATION THEORIES In this section, we study the properties of the system developed above. We first show that in the particular case where A = AO (ie. all of the agent"s arguments refer to offers), the corresponding argumentation system will return at least one non-empty preferred extension. Theorem 2. Let A, Def an argumentation system such that A = AO. Then the system returns at least one extension E, such that |E| ≥ 1. We now present some results that demonstrate the importance of indifference in negotiating agents, and more specifically its relation to acceptable outcomes. We first show that the set Oa may contain several offers when their corresponding accepted arguments are indifferent w.r.t the preference relation . Theorem 3. Let o1, o2 ∈ O. o1, o2 ∈ Oa iff ∃ a1 ∈ F(o1), ∃ a2 ∈ F(o2), such that a1 and a2 are accepted and are indifferent w.r.t (i.e. a b and b a). We now study acyclic preference relations that are defined formally as follows. Definition 9 (Acyclic relation). A relation R on a set A is acyclic if there is no sequence a1, a2, . . . , an ∈ A, with n > 1, such that (ai, ai+1) ∈ R and (an, a1) ∈ R, with 1 ≤ i < n. Note that acyclicity prohibits pairs of arguments a, b such that a b and b a, ie., an acyclic preference relation disallows indifference. Theorem 4. Let A be a set of arguments, R the attacking relation of A defined as R ⊆ A × A, and an acyclic relation on A. Then for any pair of arguments a, b ∈ A, such that (a, b) ∈ R, either (a, b) ∈ Def or (b, a) ∈ Def (or both). The previous result is used in the proof of the following theorem that states that acyclic preference relations sanction extensions that support exactly one offer. Theorem 5. Let A be a set of arguments, and an acyclic relation on A. If E is an extension of <A, Def>, then |E ∩ AO| = 1. An immediate consequence of the above is the following. Property 3. Let A be a set of arguments such that A = AO. If the relation on A is acyclic, then each extension Ei of <A, Def>, |Ei| = 1. Another direct consequence of the above theorem is that in acyclic preference relations, arguments that support offers can participate in only one preferred extension. Theorem 6. Let A be a set of arguments, and an acyclic relation on A. Then the preferred extensions of A, Def are pairwise disjoint w.r.t arguments of AO. Using the above results we can prove the main theorem of this section that states that negotiating agents with acyclic preference relations do not have acceptable offers. Theorem 7. Let A, F, R, , Def be a negotiating agent such that A = AO and is an acyclic relation. Then the set of accepted arguments w.r.t A, Def is emtpy. Consequently, the set of acceptable offers, Oa is empty as well. 5. ARGUMENTATION-BASED NEGOTIATION In this section, we define formally a protocol that generates argumentation-based negotiation dialogues between two negotiating agents P and C. The two agents negotiate about an object whose possible values belong to a set O. This set O is supposed to be known and the same for both agents. For simplicity reasons, we assume that this set does not change during the dialogue. The agents are equipped with theories denoted respectively AP , FP , P , RP , DefP , and AC , FC , C , RC , DefC . Note that the two theories may be different in the sense that the agents may have different sets of arguments, and different preference relations. Worst yet, they may have different arguments in favor of the same offers. Moreover, these theories may evolve during the dialogue. 5.1 Evolution of the theories Before defining formally the evolution of an agent"s theory, let us first introduce the notion of dialogue moves, or moves for short. Definition 10 (Move). A move is a tuple mi = pi, ai, oi, ti such that: • pi ∈ {P, C} • ai ∈ Args(L) ∪ θ1 1 In what follows θ denotes the fact that no argument, or no offer is given 970 The Sixth Intl. Joint Conf. on Autonomous Agents and Multi-Agent Systems (AAMAS 07) • oi ∈ O ∪ θ • ti ∈ N∗ is the target of the move, such that ti < i The function Player (resp. Argument, Offer, Target) returns the player of the move (i.e. pi) (resp. the argument of a move, i.e ai, the offer oi, and the target of the move, ti). Let M denote the set of all the moves that can be built from {P, C}, Arg(L), O . Note that the set M is finite since Arg(L) and O are assumed to be finite. Let us now see how an agent"s theory evolves and why. The idea is that if an agent receives an argument from another agent, it will add the new argument to its theory. Moreover, since an argument may bring new information for the agent, thus new arguments can emerge. Let us take the following example: Example 6. Suppose that an agent P has the following propositional knowledge base: ΣP = {x, y → z}. From this base one cannot deduce z. Let"s assume that this agent receives the following argument {a, a → y} that justifies y. It is clear that now P can build an argument, say {a, a → y, y → z} in favor of z. In a similar way, if a received argument is in conflict with the arguments of the agent i, then those conflicts are also added to its relation Ri . Note that new conflicts may arise between the original arguments of the agent and the ones that emerge after adding the received arguments to its theory. Those new conflicts should also be considered. As a direct consequence of the evolution of the sets Ai and Ri , the defeat relation Defi is also updated. The initial theory of an agent i, (i.e. its theory before the dialogue starts), is denoted by Ai 0, Fi 0, i 0, Ri 0, Defi 0 , with i ∈ {P, C}. Besides, in this paper, we suppose that the preference relation i of an agent does not change during the dialogue. Definition 11 (Theory evolution). Let m1, . . ., mt, . . ., mj be a sequence of moves. The theory of an agent i at a step t > 0 is: Ai t, Fi t , i t, Ri t, Defi t such that: • Ai t = Ai 0 ∪ {ai, i = 1, . . . , t, ai = Argument(mi)} ∪ A with A ⊆ Args(L) • Fi t = O → 2Ai t • i t = i 0 • Ri t = Ri 0 ∪ {(ai, aj) | ai = Argument(mi), aj = Argument(mj), i, j ≤ t, and ai RL aj} ∪ R with R ⊆ RL • Defi t ⊆ Ai t × Ai t The above definition captures the monotonic aspect of an argument. Indeed, an argument cannot be removed. However, its status may change. An argument that is accepted at step t of the dialogue by an agent may become rejected at step t + i. Consequently, the status of offers also change. Thus, the sets Oa, Or, On, and Ons may change from one step of the dialogue to another. That means for example that some offers could move from the set Oa to the set Or and vice-versa. Note that in the definition of Rt, the relation RL is used to denote a conflict between exchanged arguments. The reason is that, such a conflict may not be in the set Ri of the agent i. Thus, in order to recognize such conflicts, we have supposed that the set RL is known to the agents. This allows us to capture the situation where an agent is able to prove an argument that it was unable to prove before, by incorporating in its beliefs some information conveyed through the exchange of arguments with another agent. This, unknown at the beginning of the dialogue argument, could give to this agent the possibility to defeat an argument that it could not by using its initial arguments. This could even lead to a change of the status of these initial arguments and this change would lead to the one of the associated offers" status. In what follows, Oi t,x denotes the set of offers of type x, where x ∈ {a, n, r, ns}, of the agent i at step t of the dialogue. In some places, we can use for short the notation Oi t to denote the partition of the set O at step t for agent i. Note that we have: not(Oi t,x ⊆ Oi t+1,x). 5.2 The notion of agreement As said in the introduction, negotiation is a process aiming at finding an agreement about some matters. By agreement, one means a solution that satisfies to the largest possible extent the preferences of both agents. In case there is no such solution, we say that the negotiation fails. In what follows, we will discuss the different kinds of solutions that may be reached in a negotiation. The first one is the optimal solution. An optimal solution is the best offer for both agents. Formally: Definition 12 (Optimal solution). Let O be a set of offers, and o ∈ O. The offer o is an optimal solution at a step t ≥ 0 iff o ∈ OP t,a ∩ OC t,a Such a solution does not always exist since agents may have conflicting preferences. Thus, agents make concessions by proposing/accepting less preferred offers. Definition 13 (Concession). Let o ∈ O be an offer. The offer o is a concession for an agent i iff o ∈ Oi x such that ∃Oi y = ∅, and Oi y Oi x. During a negotiation dialogue, agents exchange first their most preferred offers, and if these last are rejected, they make concessions. In this case, we say that their best offers are no longer defendable. In an argumentation setting, this means that the agent has already presented all its arguments supporting its best offers, and it has no counter argument against the ones presented by the other agent. Formally: Definition 14 (Defendable offer). Let Ai t, Fi t , i t, Ri t, Defi t be the theory of agent i at a step t > 0 of the dialogue. Let o ∈ O such that ∃j ≤ t with Player(mj) = i and offer(mj) = o. The offer o is defendable by the agent i iff: • ∃a ∈ Fi t (o), and k ≤ t s.t. Argument(mk) = a, or • ∃a ∈ At \Fi t (o) s.t. a Defi t b with - Argument(mk) = b, k ≤ t, and Player(mk) = i - l ≤ t, Argument(ml) = a The offer o is said non-defendable otherwise and NDi t is the set of non-defendable offers of agent i at a step t. The Sixth Intl. Joint Conf. on Autonomous Agents and Multi-Agent Systems (AAMAS 07) 971 5.3 Negotiation dialogue Now that we have shown how the theories of the agents evolve during a dialogue, we are ready to define formally an argumentation-based negotiation dialogue. For that purpose, we need to define first the notion of a legal continuation. Definition 15 (Legal move). A move m is a legal continuation of a sequence of moves m1, . . . , ml iff j, k < l, such that: • Offer(mj) = Offer(mk), and • Player(mj) = Player(mk) The idea here is that if the two agents present the same offer, then the dialogue should terminate, and there is no longer possible continuation of the dialogue. Definition 16 (Argumentation-based negotiation). An argumentation-based negotiation dialogue d between two agents P and C is a non-empty sequence of moves m1, . . . , ml such that: • pi = P iff i is even, and pi = C iff i is odd • Player(m1) = P, Argument(m1) = θ, Offer(m1) = θ, and Target(m1) = 02 • ∀ mi, if Offer(mi) = θ, then Offer(mi) oj, ∀ oj ∈ O\(O Player(mi) i,r ∪ ND Player(mi) i ) • ∀i = 1, . . . , l, mi is a legal continuation of m1, . . . , mi−1 • Target(mi) = mj such that j < i and Player(mi) = Player(mj) • If Argument(mi) = θ, then: - if Offer(mi) = θ then Argument(mi) ∈ F(Offer(mi)) - if Offer(mi) = θ then Argument(mi) Def Player(mi) i Argument(Target(mi)) • i, j ≤ l such that mi = mj • m ∈ M such that m is a legal continuation of m1, . . . , ml Let D be the set of all possible dialogues. The first condition says that the two agents take turn. The second condition says that agent P starts the negotiation dialogue by presenting an offer. Note that, in the first turn, we suppose that the agent does not present an argument. This assumption is made for strategical purposes. Indeed, arguments are exchanged as soon as a conflict appears. The third condition ensures that agents exchange their best offers, but never the rejected ones. This condition takes also into account the concessions that an agent will have to make if it was established that a concession is the only option for it at the current state of the dialogue. Of course, as we have shown in a previous section, an agent may have several good or acceptable offers. In this case, the agent chooses one of them randomly. The fourth condition ensures that the moves are legal. This condition allows to terminate the dialogue as soon as an offer is presented by both agents. The fifth condition allows agents to backtrack. The sixth 2 The first move has no target. condition says that an agent may send arguments in favor of offers, and in this case the offer should be stated in the same move. An agent can also send arguments in order to defeat arguments of the other agent. The next condition prevents repeating the same move. This is useful for avoiding loops. The last condition ensures that all the possible legal moves have been presented. The outcome of a negotiation dialogue is computed as follows: Definition 17 (Dialogue outcome). Let d = m1, . . ., ml be a argumentation-based negotiation dialogue. The outcome of this dialogue, denoted Outcome, is Outcome(d) = Offer(ml) iff ∃j < l s.t. Offer(ml) = Offer(mj), and Player(ml) = Player(mj). Otherwise, Outcome(d) = θ. Note that when Outcome(d) = θ, the negotiation fails, and no agreement is reached by the two agents. However, if Outcome(d) = θ, the negotiation succeeds, and a solution that is either optimal or a compromise is found. Theorem 8. ∀di ∈ D, the argumentation-based negotiation di terminates. The above result is of great importance, since it shows that the proposed protocol avoids loops, and dialogues terminate. Another important result shows that the proposed protocol ensures to reach an optimal solution if it exists. Formally: Theorem 9 (Completeness). Let d = m1, . . . , ml be a argumentation-based negotiation dialogue. If ∃t ≤ l such that OP t,a ∩ OC t,a = ∅, then Outcome(d) ∈ OP t,a ∩ OC t,a. We show also that the proposed dialogue protocol is sound in the sense that, if a dialogue returns a solution, then that solution is for sure a compromise. In other words, that solution is a common agreement at a given step of the dialogue. We show also that if the negotiation fails, then there is no possible solution. Theorem 10 (Soundness). Let d = m1, . . . , ml be a argumentation-based negotiation dialogue. 1. If Outcome(d) = o, (o = θ), then ∃t ≤ l such that o ∈ OP t,x ∩ OC t,y, with x, y ∈ {a, n, ns}. 2. If Outcome(d) = θ, then ∀t ≤ l, OP t,x ∩ OC t,y = ∅, ∀ x, y ∈ {a, n, ns}. A direct consequence of the above theorem is the following: Property 4. Let d = m1, . . . , ml be a argumentationbased negotiation dialogue. If Outcome(d) = θ, then ∀t ≤ l, • OP t,r = OC t,a ∪ OC t,n ∪ OC t,ns, and • OC t,r = OP t,a ∪ OP t,n ∪ OP t,ns. 6. ILLUSTRATIVE EXAMPLES In this section we will present some examples in order to illustrate our general framework. Example 7 (No argumentation). Let O = {o1, o2} be the set of all possible offers. Let P and C be two agents, equipped with the same theory: A, F, , R, Def such that A = ∅, F(o1) = F(o2) = ∅, = ∅, R = ∅, Def = ∅. In this case, it is clear that the two offers o1 and o2 are nonsupported. The proposed protocol (see Definition 16) will generate one of the following dialogues: 972 The Sixth Intl. Joint Conf. on Autonomous Agents and Multi-Agent Systems (AAMAS 07) P: m1 = P, θ, o1, 0 C: m2 = C, θ, o1, 1 This dialogue ends with o1 as a compromise. Note that this solution is not considered as optimal since it is not an acceptable offer for the agents. P: m1 = P, θ, o1, 0 C: m2 = C, θ, o2, 1 P: m3 = P, θ, o2, 2 This dialogue ends with o2 as a compromise. P: m1 = P, θ, o2, 0 C: m2 = C, θ, o2, 1 This dialogue also ends with o2 as a compromise. The last possible dialgue is the following that ends with o1 as a compromise. P: m1 = P, θ, o2, 0 C: m2 = C, θ, o1, 1 P: m3 = P, θ, o1, 2 Note that in the above example, since there is no exchange of arguments, the theories of both agents do not change. Let us now consider the following example. Example 8 (Static theories). Let O = {o1, o2} be the set of all possible offers. The theory of agent P is AP , FP , P , RP , DefP such that: AP = {a1, a2}, FP (o1) = {a1}, FP (o2) = {a2}, P = {(a1, a2)}, RP = {(a1, a2), (a2, a1)}, DefP = {a1, a2}. The argumentation system AP , DefP of this agent will return a1 as an accepted argument, and a2 as a rejected one. Consequently, the offer o1 is acceptable and o2 is rejected. The theory of agent C is AC , FC , C , RC , DefC such that: AC = {a1, a2}, FC (o1) = {a1}, FC (o2) = {a2}, C = {(a2, a1)}, RC = {(a1, a2), (a2, a1)}, DefC = {a2, a1}. The argumentation system AC , DefC of this agent will return a2 as an accepted argument, and a1 as a rejected one. Consequently, the offer o2 is acceptable and o1 is rejected. The only possible dialogues that may take place between the two agents are the following: P: m1 = P, θ, o1, 0 C: m2 = C, θ, o2, 1 P: m3 = P, a1, o1, 2 C: m4 = C, a2, o2, 3 The second possible dialogue is the following: P: m1 = P, θ, o1, 0 C: m2 = C, a2, o2, 1 P: m3 = P, a1, o1, 2 C: m4 = C, θ, o2, 3 Both dialogues end with failure. Note that in both dialogues, the theories of both agents do not change. The reason is that the exchanged arguments are already known to both agents. The negotiation fails because the agents have conflicting preferences. Let us now consider an example in which argumentation will allow agents to reach an agreement. Example 9 (Dynamic theories). Let O = {o1, o2} be the set of all possible offers. The theory of agent P is AP , FP , P , RP , DefP such that: AP = {a1, a2}, FP (o1) = {a1}, FP (o2) = {a2}, P = {(a1, a2), (a3, a1)}, RP = {(a1, a2), (a2, a1)}, DefP = {(a1, a2)}. The argumentation system AP , DefP of this agent will return a1 as an accepted argument, and a2 as a rejected one. Consequently, the offer o1 is acceptable and o2 is rejected. The theory of agent C is AC , FC , C , RC , DefC such that: AC = {a1, a2, a3}, FC (o1) = {a1}, FC (o2) = {a2}, C = {(a1, a2), (a3, a1)}, RC = {(a1, a2), (a2, a1), (a3, a1)}, DefC = {(a1, a2), (a3, a1)}. The argumentation system AC , DefC of this agent will return a3 and a2 as accepted arguments, and a1 as a rejected one. Consequently, the offer o2 is acceptable and o1 is rejected. The following dialogue may take place between the two agents: P: m1 = P, θ, o1, 0 C: m2 = C, θ, o2, 1 P: m3 = P, a1, o1, 2 C: m4 = C, a3, θ, 3 C: m5 = P, θ, o2, 4 At step 4 of the dialogue, the agent P receives the argument a3 from P. Thus, its theory evolves as follows: AP = {a1, a2, a3}, RP = {(a1, a2), (a2, a1), (a3, a1)}, DefP = {(a1, a2), (a3, a1)}. At this step, the argument a1 which was accepted will become rejected, and the argument a2 which was at the beginning of the dialogue rejected will become accepted. Thus, the offer o2 will be acceptable for the agent, whereas o1 will become rejected. At this step 4, the offer o2 is acceptable for both agents, thus it is an optimal solution. The dialogue ends by returning this offer as an outcome. 7. RELATED WORK Argumentation has been integrated in negotiation dialogues at the early nineties by Sycara [12]. In that work, the author has emphasized the advantages of using argumentation in negotiation dialogues, and a specific framework has been introduced. In [8], the different types of arguments that are used in a negotiation dialogue, such as threats and rewards, have been discussed. Moreover, a particular framework for negotiation have been proposed. In [9, 13], different other frameworks have been proposed. Even if all these frameworks are based on different logics, and use different definitions of arguments, they all have at their heart an exchange of offers and arguments. However, none of those proposals explain when arguments can be used within a negotiation, and how they should be dealt with by the agent that receives them. Thus the protocol for handling arguments was missing. Another limitation of the above frameworks is the fact that the argumentation frameworks they The Sixth Intl. Joint Conf. on Autonomous Agents and Multi-Agent Systems (AAMAS 07) 973 use are quite poor, since they use a very simple acceptability semantics. In [2] a negotiation framework that fills the gap has been suggested. A protocol that handles the arguments was proposed. However, the notion of concession is not modeled in that framework, and it is not clear what is the status of the outcome of the dialogue. Moreover, it is not clear how an agent chooses the offer to propose at a given step of the dialogue. In [1, 7], the authors have focused mainly on this decision problem. They have proposed an argumentation-based decision framework that is used by agents in order to choose the offer to propose or to accept during the dialogue. In that work, agents are supposed to have a beliefs base and a goals base. Our framework is more general since it does not impose any specific structure for the arguments, the offers, or the beliefs. The negotiation protocol is general as well. Thus this framework can be instantiated in different ways by creating, in such manner, different specific argumentation-based negotiation frameworks, all of them respecting the same properties. Our framework is also a unified one because frameworks like the ones presented above can be represented within this framework. For example the decision making mechanism proposed in [7] for the evaluation of arguments and therefore of offers, which is based on a priority relation between mutually attacked arguments, can be captured by the relation defeat proposed in our framework. This relation takes simultaneously into account the attacking and preference relations that may exist between two arguments. 8. CONCLUSIONS AND FUTURE WORK In this paper we have presented a unified and general framework for argumentation-based negotiation. Like any other argumentation-based negotiation framework, as it is evoked in (e.g. [10]), our framework has all the advantages that argumentation-based negotiation approaches present when related to the negotiation approaches based either on game theoretic models (see e.g. [11]) or heuristics ([6]). This work is a first attempt to formally define the role of argumentation in the negotiation process. More precisely, for the first time, it formally establishes the link that exists between the status of the arguments and the offers they support, it defines the notion of concession and shows how it influences the evolution of the negotiation, it determines how the theories of agents evolve during the dialogue and performs an analysis of the negotiation outcomes. It is also the first time where a study of the formal properties of the negotiation theories of the agents as well as of an argumentative negotiation dialogue is presented. Our future work concerns several points. A first point is to relax the assumption that the set of possible offers is the same to both agents. Indeed, it is more natural to assume that agents may have different sets of offers. During a negotiation dialogue, these sets will evolve. Arguments in favor of the new offers may be built from the agent theory. Thus, the set of offers will be part of the agent theory. Another possible extension of this work would be to allow agents to handle both arguments PRO and CONS offers. This is more akin to the way human take decisions. Considering both types of arguments will refine the evaluation of the offers status. In the proposed model, a preference relation between offers is defined on the basis of the partition of the set of offers. This preference relation can be refined. For instance, among the acceptable offers, one may prefer the offer that is supported by the strongest argument. In [4], different criteria have been proposed for comparing decisions. Our framework can thus be extended by integrating those criteria. Another interesting point to investigate is that of considering negotiation dialogues between two agents with different profiles. By profile, we mean the criterion used by an agent to compare its offers. 9. REFERENCES [1] L. Amgoud, S. Belabbes, and H. Prade. Towards a formal framework for the search of a consensus between autonomous agents. In Proceedings of the 4th International Joint Conference on Autonomous Agents and Multi-Agents systems, pages 537-543, 2005. [2] L. Amgoud, S. Parsons, and N. Maudet. Arguments, dialogue, and negotiation. In Proceedings of the 14th European Conference on Artificial Intelligence, 2000. [3] L. Amgoud and H. Prade. Reaching agreement through argumentation: A possibilistic approach. In 9 th International Conference on the Principles of Knowledge Representation and Reasoning, KR"2004, 2004. [4] L. Amgoud and H. Prade. Explaining qualitative decision under uncertainty by argumentation. In 21st National Conference on Artificial Intelligence, AAAI"06, pages 16 - 20, 2006. [5] P. M. Dung. On the acceptability of arguments and its fundamental role in nonmonotonic reasoning, logic programming and n-person games. Artificial Intelligence, 77:321-357, 1995. [6] N. R. Jennings, P. Faratin, A. R. Lumuscio, S. Parsons, and C. Sierra. Automated negotiation: Prospects, methods and challenges. International Journal of Group Decision and Negotiation, 2001. [7] A. Kakas and P. Moraitis. Adaptive agent negotiation via argumentation. In Proceedings of the 5th International Joint Conference on Autonomous Agents and Multi-Agents systems, pages 384-391, 2006. [8] S. Kraus, K. Sycara, and A. Evenchik. Reaching agreements through argumentation: a logical model and implementation. Artificial Intelligence, 104:1-69, 1998. [9] S. Parsons and N. R. Jennings. Negotiation through argumentation-a preliminary report. In Proceedings of the 2nd International Conference on Multi Agent Systems, pages 267-274, 1996. [10] I. Rahwan, S. D. Ramchurn, N. R. Jennings, P. McBurney, S. Parsons, and E. Sonenberg. Argumentation-based negotiation. Knowledge Engineering Review, 18 (4):343-375, 2003. [11] J. Rosenschein and G. Zlotkin. Rules of Encounter: Designing Conventions for Automated Negotiation Among Computers,. MIT Press, Cambridge, Massachusetts, 1994., 1994. [12] K. Sycara. Persuasive argumentation in negotiation. Theory and Decision, 28:203-242, 1990. [13] F. Tohm´e. Negotiation and defeasible reasons for choice. In Proceedings of the Stanford Spring Symposium on Qualitative Preferences in Deliberation and Practical Reasoning, pages 95-102, 1997. 974 The Sixth Intl. Joint Conf. on Autonomous Agents and Multi-Agent Systems (AAMAS 07)

outcome;belief;agent;argumentation-based negotiation;concession notion;argument;information;argumentation;decision making mechanism;negotiation;framework;solution;theory;notion of concession

train_I-53

A Randomized Method for the Shapley Value for the Voting Game

The Shapley value is one of the key solution concepts for coalition games. Its main advantage is that it provides a unique and fair solution, but its main problem is that, for many coalition games, the Shapley value cannot be determined in polynomial time. In particular, the problem of finding this value for the voting game is known to be #P-complete in the general case. However, in this paper, we show that there are some specific voting games for which the problem is computationally tractable. For other general voting games, we overcome the problem of computational complexity by presenting a new randomized method for determining the approximate Shapley value. The time complexity of this method is linear in the number of players. We also show, through empirical studies, that the percentage error for the proposed method is always less than 20% and, in most cases, less than 5%.

1. INTRODUCTION Coalition formation, a key form of interaction in multi-agent systems, is the process of joining together two or more agents so as to achieve goals that individuals on their own cannot, or to achieve them more efficiently [1, 11, 14, 13]. Often, in such situations, there is more than one possible coalition and a player"s payoff depends on which one it joins. Given this, a key problem is to ensure that none of the parties in a coalition has any incentive to break away from it and join another coalition (i.e., the coalitions should be stable). However, in many cases there may be more than one solution (i.e., a stable coalition). In such cases, it becomes difficult to select a single solution from among the possible ones, especially if the parties are self-interested (i.e., they have different preferences over stable coalitions). In this context, cooperative game theory deals with the problem of coalition formation and offers a number of solution concepts that possess desirable properties like stability, fair division of joint gains, and uniqueness [16, 14]. Cooperative game theory differs from its non-cooperative counterpart in that for the former the players are allowed to form binding agreements and so there is a strong incentive to work together to receive the largest total payoff. Also, unlike non-cooperative game theory, cooperative game theory does not specify a game through a description of the strategic environment (including the order of players" moves and the set of actions at each move) and the resulting payoffs, but, instead, it reduces this collection of data to the coalitional form where each coalition is represented by a single real number: there are no actions, moves, or individual payoffs. The chief advantage of this approach, at least in multiple-player environments, is its practical usefulness. Thus, many more real-life situations fit more easily into a coalitional form game, whose structure is more tractable than that of a non-cooperative game, whether that be in normal or extensive form and it is for this reason that we focus on such forms in this paper. Given these observations, a number of multiagent systems researchers have used and extended cooperative game-theoretic solutions to facilitate automated coalition formation [20, 21, 18]. Moreover, in this work, one of the most extensively studied solution concepts is the Shapley value [19]. A player"s Shapley value gives an indication of its prospects of playing the game - the higher the Shapley value, the better its prospects. The main advantage of the Shapley value is that it provides a solution that is both unique and fair (see Section 2.1 for a discussion of the property of fairness). However, while these are both desirable properties, the Shapley value has one major drawback: for many coalition games, it cannot be determined in polynomial time. For instance, finding this value for the weighted voting game is, in general, #P-complete [6]. A problem is #P-hard if solving it is as hard as counting satisfying assignments of propositional logic formulae [15, p442]. Since #P-completeness thus subsumes NP-completeness, this implies that computing the Shapley value for the weighted voting game will be intractable in general. In other words, it is practically infeasible to try to compute the exact Shapley value. However, the voting game has practical relevance to multi-agent systems as it is an important means of reaching consensus between multiple agents. Hence, our objective is to overcome the computational complexity of finding the Shapley value for this game. Specifically, we first show that there are some specific voting games for which the exact value can 959 978-81-904262-7-5 (RPS) c 2007 IFAAMAS be computed in polynomial time. By identifying such games, we show, for the first tme, when it is feasible to find the exact value and when it is not. For the computationally complex voting games, we present a new randomised method, along the lines of Monte-Carlo simulation, for computing the approximate Shapley value. The computational complexity of such games has typically been tackled using two main approaches. The first is to use generating functions [3]. This method trades time complexity for storage space. The second uses an approximation technique based on Monte Carlo simulation [12, 7]. However the method we propose is more general than either of these (see Section 6 for details). Moreover, no work has previously analysed the approximation error. The approximation error relates to how close the approximate is to the true Shapley value. Specifically, it is the difference between the true and the approximate Shapley value. It is important to determine this error because the performance of an approximation method is evaluated in terms of two criteria: its time complexity, and its approximation error. Thus, our contribution lies in also in providing, for the first time, an analysis of the percentage error in the approximate Shapley value. This analysis is carried out empirically. Our experiments show that the error is always less than 20%, and in most cases it is under 5%. Finally, our method has time complexity linear in the number of players and it does not require any arrays (i.e., it is economical in terms of both computing time and storage space). Given this, and the fact that software agents have limited computational resources and therefore cannot compute the true Shapley value, our results are especially relevant to such resource bounded agents. The rest of the paper is organised as follows. Section 2 defines the Shapley value and describes the weighted voting game. In Section 3 we describe voting games whose Shapley value can be found in polynomial time. In Section 4, we present a randomized method for finding the approximate Shapley value and analyse its performance in Section 5. Section 6 discusses related literature. Finally, Section 7 concludes. 2. BACKGROUND We begin by introducing coalition games and the Shapley value and then define the weighted voting game. A coalition game is a game where groups of players (coalitions) may enforce cooperative behaviour between their members. Hence the game is a competition between coalitions of players, rather than between individual players. Depending on how the players measure utility, coalition game theory is split into two parts. If the players measure utility or the payoff in the same units and there is a means of exchange of utility such as side payments, we say the game has transferable utility; otherwise it has non-transferable utility. More formally, a coalition game with transferable utility, N, v , consists of: 1. a finite set (N = {1, 2, . . . , n}) of players and 2. a function (v) that associates with every non-empty subset S of N (i.e., a coalition) a real number v(S) (the worth of S). For each coalition S, the number v(S) is the total payoff that is available for division among the members of S (i.e., the set of joint actions that coalition S can take consists of all possible divisions of v(S) among the members of S). Coalition games with nontransferable payoffs differ from ones with transferable payoffs in the following way. For the former, each coalition is associated with a set of payoff vectors that is not necessarily the set of all possible divisions of some fixed amount. The focus of this paper is on the weighted voting game (described in Section 2.1) which is a game with transferable payoffs. Thus, in either case, the players will only join a coalition if they expect to gain from it. Here, the players are allowed to form binding agreements, and so there is strong incentive to work together to receive the largest total payoff. The problem then is how to split the total payoff between or among the players. In this context, Shapley [19] constructed a solution using an axiomatic approach. Shapley defined a value for games to be a function that assigns to a game (N, v), a number ϕi(N, v) for each i in N. This function satisfies three axioms [17]: 1. Symmetry. This axiom requires that the names of players play no role in determining the value. 2. Carrier. This axiom requires that the sum of ϕi(N, v) for all players i in any carrier C equal v(C). A carrier C is a subset of N such that v(S) = v(S ∩ C) for any subset of players S ⊂ N. 3. Additivity. This axiom specifies how the values of different games must be related to one another. It requires that for any games ϕi(N, v) and ϕi(N, v ), ϕi(N, v)+ϕi(N, v ) = ϕi(N, v + v ) for all i in N. Shapley showed that there is a unique function that satisfies these three axioms. Shapley viewed this value as an index for measuring the power of players in a game. Like a price index or other market indices, the value uses averages (or weighted averages in some of its generalizations) to aggregate the power of players in their various cooperation opportunities. Alternatively, one can think of the Shapley value as a measure of the utility of risk neutral players in a game. We first introduce some notation and then define the Shapley value. Let S denote the set N − {i} and fi : S → 2N−{i} be a random variable that takes its values in the set of all subsets of N − {i}, and has the probability distribution function (g) defined as: g(fi(S) = S) = |S|!(n − |S| − 1)! n! The random variable fi is interpreted as the random choice of a coalition that player i joins. Then, a player"s Shapley value is defined in terms of its marginal contribution. Thus, the marginal contribution of player i to coalition S with i /∈ S is a function Δiv that is defined as follows: Δiv(S) = v(S ∪ {i}) − v(S) Thus a player"s marginal contribution to a coalition S is the increase in the value of S as a result of i joining it. DEFINITION 1. The Shapley value (ϕi) of the game N, v for player i is the expectation (E) of its marginal contribution to a coalition that is chosen randomly: ϕi(N, v) = E[Δiv ◦ fi] (1) The Shapley value is interpreted as follows. Suppose that all the players are arranged in some order, all orderings being equally likely. Then ϕi(N, v) is the expected marginal contribution, over all orderings, of player i to the set of players who precede him. The method for finding a player"s Shapley value depends on the definition of the value function (v). This function is different for different games, but here we focus specifically on the weighted voting game for the reasons outlined in Section 1. 960 The Sixth Intl. Joint Conf. on Autonomous Agents and Multi-Agent Systems (AAMAS 07) 2.1 The weighted voting game We adopt the definition of the voting game given in [6]. Thus, there is a set of n players that may, for example, represent shareholders in a company or members in a parliament. The weighted voting game is then a game G = N, v in which: v(S) = j 1 if w(S) ≥ q 0 otherwise for some q ∈ IR+ and wi ∈ IRN + , where: w(S) = X i∈S wi for any coalition S. Thus wi is the number of votes that player i has and q is the number of votes needed to win the game (i.e., the quota). Note that for this game (denoted q; w1, . . . , wn ), a player"s marginal contribution is either zero or one. This is because the value of any coalition is either zero or one. A coalition with value zero is called a losing coalition and with value one a winning coalition. If a player"s entry to a coalition changes it from losing to winning, then the player"s marginal contribution for that coalition is one; otherwise it is zero. The main advantage of the Shapley value is that it gives a solution that is both unique and fair. The property of uniqueness is desirable because it leaves no ambiguity. The property of fairness relates to how the gains from cooperation are split between the members of a coalition. In this case, a player"s Shapley value is proportional to the contribution it makes as a member of a coalition; the more contribution it makes, the higher its value. Thus, from a player"s perspective, both uniqueness and fairness are desirable properties. 3. VOTING GAMES WITH POLYNOMIAL TIME SOLUTIONS Here we describe those voting games for which the Shapley value can be determined in polynomial time. This is achieved using the direct enumeration approach (i.e., listing all possible coalitions and finding a player"s marginal contribution to each of them). We characterise such games in terms of the number of players and their weights. 3.1 All players have equal weight Consider the game q; j, . . . , j with m parties. Each party has j votes. If q ≤ j, then there would be no need for the players to form a coalition. On the other hand, if q = mj (m = |N| is the number of players), only the grand coalition is possible. The interesting games are those for which the quota (q) satisfies the constraint: (j + 1) ≤ q ≤ j(m − 1). For these games, the value of a coalition is one if the weight of the coalition is greater than or equal to q, otherwise it is zero. Let ϕ denote the Shapley value for a player. Consider any one player. This player can join a coalition as the ith member where 1 ≤ i ≤ m. However, the marginal contribution of the player is 1 only if it joins a coalition as the q/j th member. In all other cases, its marginal contribution is zero. Thus, the Shapley value for each player ϕ = 1/m. Since ϕ requires one division operation, it can be found in constant time (i.e., O(1)). 3.2 A single large party Consider a game in which there are two types of players: large (with weight wl > ws) and small (with weight ws). There is one large player and m small ones. The quota for this game is q; i.e., we have a game of the form q; wl, ws, ws, . . . , ws . The total number of players is (m + 1). The value of a coalition is one if the weight of the coalition is greater than or equal to q, otherwise it is zero. Let ϕl denote the Shapley value for the large player and ϕs that for each small player. We first consider ws = 1 and then ws > 1. The smallest possible value for q is wl + 1. This is because, if q ≤ wl, then the large party can win the election on its own without the need for a coalition. Thus, the quota for the game satisfies the constraint wl + 1 ≤ q ≤ m + wl − 1. Also, the lower and upper limits for wl are 2 and (q − 1) respectively. The lower limit is 2 because the weight of the large party has to be greater than each small one. Furthermore, the weight of the large party cannot be greater than q, since in that case, there would be no need for the large party to form a coalition. Recall that for our voting game, a player"s marginal contribution to a coalition can only be zero or one. Consider the large party. This party can join a coalition as the ith member where 1 ≤ i ≤ (m + 1). However, the marginal contribution of the large party is one if it joins a coalition as the ith member where (q − wl) ≤ i < q. In all the remaining cases, its marginal contribution is zero. Thus, out of the total (m + 1) possible cases, its marginal contribution is one in wl cases. Hence, the Shapley value of the large party is: ϕl = wl/(m + 1). In the same way, we obtain the Shapley value of the large party for the general case where ws > 1 as: ϕl = wl/ws /(m + 1) Now consider a small player. We know that the sum of the Shapley values of all the m+1 players is one. Also, since the small parties have equal weights, their Shapley values are the same. Hence, we get: ϕs = 1 − ϕl m Thus, both ϕl and ϕs can be computed in constant time. This is because both require a constant number of basic operations (addition, subtraction, multiplication, and division). In the same way, the Shapley value for a voting game with a single large party and multiple small parties can be determined in constant time. 3.3 Multiple large and small parties We now consider a voting game that has two player types: large and small (as in Section 3.2), but now there are multiple large and multiple small parties. The set of parties consists of ml large parties and ms small parties. The weight of each large party is wl and that of each small one is ws, where ws < wl. We show the computational tractability for this game by considering the following four possible scenarios: S1 q ≤ mlwl and q ≤ msws S2 q ≤ mlwl and q ≥ msws S3 q ≥ mlwl and q ≥ msws S4 q ≥ mlwl and q ≤ msws For the first scenario, consider a large player. In order to determine the Shapley value for this player, we need to consider the number of all possible coalitions that give it a marginal contribution of one. It is possible for the marginal contribution of this player to be one if it joins a coalition in which the number of large players is between zero and (q −1)/wl. In other words, there are (q −1)/wl +1 such cases and we now consider each of them. The Sixth Intl. Joint Conf. on Autonomous Agents and Multi-Agent Systems (AAMAS 07) 961 Consider a coalition such that when the large player joins in, there are i large players and (q − iwl − 1)/ws small players already in it, and the remaining players join after the large player. Such a coalition gives the large player unit marginal contribution. Let C2 l (i, q) denote the number of all such coalitions. To begin, consider the case i = 0: C2 l (0, q) = C „ ms, q − 1 ws « × FACTORIAL „ q − 1 ws « × FACTORIAL „ ml + ms − q − 1 ws − 1 « where C(y, x) denotes the number of possible combinations of x items from a set of y items. For i = 1, we get: C2 l (1, q) = C(ml, 1) × C „ ms, q − wl − 1 ws « × FACTORIAL „ q − wl − 1 ws « × FACTORIAL „ ml + ms − q − wl − 1 ws − 1 « In general, for i > 1, we get: C2 l (i, q) = C(ml, i) × C „ ms, q − iwl − 1 ws « × FACTORIAL „ q − iwl − 1 ws « × FACTORIAL „ ml + ms − q − wl − 1 ws − 1 « Thus the large player"s Shapley value is: ϕl = q−1 wlX i=0 C2 l (i, q)/FACTORIAL(ml + ms) For a given i, the time to find C2 l (i, q) is O(T) where T = (mlms(q − iwl − 1)(ml + ms))/ws Hence, the time to find the Shapley value is O(T q/wl). In the same way, a small player"s Shapley value is: ϕs = q−1 wsX i=0 C2 s (i, q)/FACTORIAL(ml + ms) and can be found in time O(T q/ws). Likewise, the remaining three scenarios (S2 to S4) can be shown to have the same time complexity. 3.4 Three player types We now consider a voting game that has three player types: 1, 2, and 3. The set of parties consists of m1 players of type 1 (each with weight w1), m2 players of type 2 (each with weight w2), and m3 players of type 3 (each with weight w3). For this voting game, consider a player of type 1. It is possible for the marginal contribution of this player to be one if it joins a coalition in which the number of type 1 players is between zero and (q − 1)/w1. In other words, there are (q − 1)/w1 + 1 such cases and we now consider each of them. Consider a coalition such that when the type 1 player joins in, there are i type 1 players already in it. The remaining players join after the type 1 player. Let C3 l (i, q) denote the number of all such coalitions that give a marginal contribution of one to the type 1 player where: C3 1 (i, q) = q−1 w1X i=0 q−iw1−1 w2X j=0 C2 1 (j, q − iw1) Therefore the Shapley value of the type 1 player is: ϕ1 = q−1 w1X i=0 C3 1 (i, q)/FACTORIAL(m1 + m2 + m3) The time complexity of finding this value is O(T q2 /w1w2) where: T = ( 3Y i=1 mi)(q − iwl − 1)( 3X i=1 mi)/(w2 + w3) Likewise, for the other two player types (2 and 3). Thus, we have identified games for which the exact Shapley value can be easily determined. However, the computational complexity of the above direct enumeration method increases with the number of player types. For a voting game with more than three player types, the time complexity of the above method is a polynomial of degree four or more. To deal with such situations, therefore, the following section presents a faster randomised method for finding the approximate Shapley value. 4. FINDING THE APPROXIMATE SHAPLEY VALUE We first give a brief introduction to randomized algorithms and then present our randomized method for finding the approximate Shapley value. Randomized algorithms are the most commonly used approach for finding approximate solutions to computationally hard problems. A randomized algorithm is an algorithm that, during some of its steps, performs random choices [2]. The random steps performed by the algorithm imply that by executing the algorithm several times with the same input we are not guaranteed to find the same solution. Now, since such algorithms generate approximate solutions, their performance is evaluated in terms of two criteria: their time complexity, and their error of approximation. The approximation error refers to the difference between the exact solution and its approximation. Against this background, we present a randomized method for finding the approximate Shapley value and empirically evaluate its error. We first describe the general voting game and then present our randomized algorithm. In its general form, a voting game has more than two types of players. Let wi denote the weight of player i. Thus, for m players and for quota q the game is of the form q; w1, w2, . . . , wm . The weights are specified in terms of a probability distribution function. For such a game, we want to find the approximate Shapley value. We let P denote a population of players. The players" weights in this population are defined by a probability distribution function. Irrespective of the actual probability distribution function, let μ be the mean weight for the population of players and ν the variance in the players" weights. From this population of players we randomly draw samples and find the sum of the players" weights in the sample using the following rule from Sampling Theory (see [8] p425): If w1, w2, . . . , wn is a random sample of size n drawn from any distribution with mean μ and variance ν, then the sample sum has an approximate Normal distribution with mean nμ and variance ν n (the larger the n the better the approximation). 962 The Sixth Intl. Joint Conf. on Autonomous Agents and Multi-Agent Systems (AAMAS 07) R-SHAPLEYVALUE (P, μ, ν, q, wi) P: Population of players μ: Mean weight of the population P ν: Variance in the weights for poulation P q: Quota for the voting game wi: Player i"s weight 1. Ti ← 0; a ← q − wi; b ← q − 2. For X from 1 to m repeatedly do the following 2.1. Select a random sample SX of size X from the population P 2.2. Evaluate expected marginal contribution (ΔX i ) of player i to SX as: ΔX i ← 1√ (2πν/X) R b a e−X (x−Xμ)2 2ν dx 2.3. Ti ← Ti + ΔX i 3. Evaluate Shapley value of player i as: ϕi ← Ti/m Table 1: Randomized algorithm to find the Shapley value for player i. We know from Definition 1, that the Shapley value for a player is the expectation (E) of its marginal contribution to a coalition that is chosen randomly. We use this rule to determine the Shapley value as follows. For player i with weight wi, let ϕi denote the Shapley value. Let X denote the size of a random sample drawn from a population in which the individual player weights have any distribution. The marginal contribution of player i to this random sample is one if the total weight of the X players in the sample is greater than or equal to a = q −wi but less than b = q − (where is an inifinitesimally small quantity). Otherwise, its marginal contribution is zero. Thus, the expected marginal contribution of player i (denoted ΔX i ) to the sample coalition is the area under the curve defined by N(Xμ, ν X ) in the interval [a, b]. This area is shown as the region B in Figure 1 (the dotted line in the figure is Xμ). Hence we get: ΔX i = 1 p (2πν/X) Z b a e−X (x−Xμ)2 2ν dx (2) and the Shapley value is: ϕi = 1 m mX X=1 ΔX i (3) The above steps are described in Table 1. In more detail, Step 1 does the initialization. In Step 2, we vary X between 1 and m and repeatedly do the following. In Step 2.1, we randomly select a sample SX of size X from the population P. Player i"s marginal contribution to the random coalition SX is found in Step 2.2. The average marginal contribution is found in Step 3 - and this is the Shapley value for player i. THEOREM 1. The time complexity of the proposed randomized method is linear in the number of players. PROOF. As per Equation 3, ΔX i must be computed m times. This is done in the for loop of Step 2 in Table 1. Hence, the time complexity of computing a player"s Shapley value is O(m). The following section analyses the approximation error for the proposed method. 5. PERFORMANCE OF THE RANDOMIZED METHOD We first derive the formula for measuring the error in the approximate Shapley value and then conduct experiments for evaluating this error in a wide range of settings. However, before doing so, we introduce the idea of error. The concept of error relates to a measurement made of a quantity which has an accepted value [22, 4]. Obviously, it cannot be determined exactly how far off a measurement is from the accepted value; if this could be done, it would be possible to just give a more accurate, corrected value. Thus, error has to do with uncertainty in measurements that nothing can be done about. If a measurement is repeated, the values obtained will differ and none of the results can be preferred over the others. However, although it is not possible to do anything about such error, it can be characterized. As described in Section 4, we make measurements on samples that are drawn randomly from a given population (P) of players. Now, there are statistical errors associated with sampling which are unavoidable and must be lived with. Hence, if the result of a measurement is to have meaning it cannot consist of the measured value alone. An indication of how accurate the result is must be included also. Thus, the result of any physical measurement has two essential components: 1. a numerical value giving the best estimate possible of the quantity measured, and 2. the degree of uncertainty associated with this estimated value. For example, if the estimate of a quantity is x and the uncertainty is e(x) the quantity would lie in x ± e(x). For sampling experiments, the standard error is by far the most common way of characterising uncertainty [22]. Given this, the following section defines this error and uses it to evaluate the performance of the proposed randomized method. 5.1 Approximation error The accuracy of the above randomized method depends on its sampling error which is defined as follows [22, 4]: DEFINITION 2. The sampling error (or standard error) is defined as the standard deviation for a set of measurements divided by the square root of the number of measurements. To this end, let e(σX ) be the sampling error in the sum of the weights for a sample of size X drawn from the distribution N(Xμ, ν X ) where: e(σX ) = p (ν/X)/ p (X) = p (ν)/X (4) Let e(ΔX i ) denote the error in the marginal contribution for player i (given in Equation 2). This error is obtained by propagating the error in Equation 4 to Equation 2. In Equation 2, a and b are the lower and upper limits for the sum of the players" weights for a The Sixth Intl. Joint Conf. on Autonomous Agents and Multi-Agent Systems (AAMAS 07) 963 b C B a − e (σX ) a A )X(σeb+ Sum of weights Z1 Z2 Figure 1: A normal distribution for the sum of players" weights in a coalition of size X. 40 60 80 100 0 50 100 0 5 10 15 20 25 QuotaWeight PercentageerrorintheShapleyvalue Figure 2: Performance of the randomized method for m = 10 players. coalition of size X. Since the error in this sum is e(σX ), the actual values of a and b lie in the interval a ± e(σX ) and b ± e(σX ) respectively. Hence, the error in Equation 2 is either the probability that the sum lies between the limits a − e(σX ) and a (i.e., the area under the curve defined by N(Xμ, ν X ) between a − e(σX ) and a, which is the shaded region A in Figure 1) or the probability that the sum of weights lies between the limits b and b+e(σX ) (i.e., the area under the curve defined by N(Xμ, ν X ) between b and b + e(σX ), which is the shaded region C in Figure 1). More specifically, the error is the maximum of these two probabilities: e(ΔX i ) = 1 p (2πν/X) × MAX „Z a a−e(σX ) e−X (x−Xμ)2 2ν dx, Z b+e(σX ) b e−X (x−Xμ)2 2ν dx « On the basis of the above error, we find the error in the Shapley value by using the following standard error propagation rules [22]: R1 If x and y are two random variables with errors e(x) and e(y) respectively, then the error in the random variable z = x + y is given by: e(z) = e(x) + e(y) R2 If x is a random variable with error e(x) and z = kx where 0 100 200 300 400 500 0 100 200 300 400 500 0 5 10 15 20 25 QuotaWeight PercentageerrorintheShapleyvalue Figure 3: Performance of the randomized method for m = 50 players. the constant k has no error, then the error in z is: e(z) = |k|e(x) Using the above rules, the error in the Shapley value (given in Equation 3) is obtained by propagating the error in Equation 4 to all coalitions between the sizes X = 1 and X = m. Let e(ϕi) denote this error where: e(ϕi) = 1 m mX X=1 e(ΔX i ) We analyze the performance of our method in terms of the percentage error PE in the approximate Shapley value which is defined as follows: PE = 100 × e(ϕi)/ϕi (5) 5.2 Experimental Results We now compute the percentage error in the Shapley value using the above equation for PE. Since this error depends on the parameters of the voting game, we evaluate it in a range of settings by systematically varying the parameters of the voting game. In particular, we conduct experiments in the following setting. For a player with weight w, the percentage error in a player"s Shapley value depends on the following five parameters (see Equation 3): 1. The number of parties (m). 2. The mean weight (μ). 3. The variance in the player"s weights (ν). 4. The quota for the voting game (q). 5. The given player"s weight (w). We fix μ = 10 and ν = 1. This is because, for the normal distribution, μ = 10 ensures that for almost all the players the weight is positive, and ν = 1 is used most commonly in statistical experiments (ν can be higher or lower but PE is increasing in νsee Equations 4 and 5). We then vary m, q, and w as follows. We vary m between 5 and 100 (since beyond 100 we found that the error is close to zero), for each m we vary q between 4μ and mμ (we impose these limits because they ensure that the size of the 964 The Sixth Intl. Joint Conf. on Autonomous Agents and Multi-Agent Systems (AAMAS 07) 0 200 400 600 800 1000 0 200 400 600 800 1000 0 5 10 15 20 25 QuotaWeight PercentageerrorintheShapleyvalue Figure 4: Performance of the randomized method for m = 100 players. winning coalition is more than one and less than m - see Section 3 for details), and for each q, we vary w between 1 and q−1 (because a winning coalition must contain at least two players). The results of these experiments are shown in Figures 2, 3, and 4. As seen in the figures, the maximum PE is around 20% and in most cases it is below 5%. We now analyse the effect of the three parameters: w, q, and m on the percentage error in more detail. - Effect of w. The PE depends on e(σX ) because, in Equation 5, the limits of integration depend on e(σX ). The interval over which the first integration in Equation 5 is done is a − a + e(σX ) = e(σX ), and the interval over which the second one is done is b + e(σX ) − b = e(σX ). Thus, the interval is the same for both integrations and it is independent of wi. Note that each of the two functions that are integrated in Equation 5 are the same as the function that is integrated in Equation 2. Only the limits of the integration are different. Also, the interval over which the integration for the marginal contribution of Equation 2 is done is b − a = wi − (see Figure 1). The error in the marginal contribution is either the area of the shaded region A (between a − e(σX ) and a) in Figure 1, or the shaded area C (between b and b + e(σX )). As per Equation 5, it is the maximum of these two areas. Since e(σX ) is independent of wi, as wi increases, e(σX ) remains unchanged. However, the area of the unshaded region B increases. Hence, as wi increases, the error in the marginal contribution decreases and PE also decreases. - Effect of q. For a given q, the Shapley value for player i is as given in Equation 3. We know that, for a sample of size X, the sum of the players" weights is distributed normally with mean Xμ and variance ν/X. Since 99% of a normal distribution lies within two standard deviations of its mean [8], player i"s marginal contribution to a sample of size X is almost zero if: a < Xμ + 2 p ν/X or b > Xμ − 2 p ν/X This is because the three regions A, B, and C (in Figure 1) lie either to the right of Z2 or to the left of Z1. However, player i"s marginal contribution is greater than zero for those X for which the following constraint is satisfied: Xμ − 2 p ν/X < a < b < Xμ + 2 p ν/X For this constraint, the three regions A, B, and C lie somewhere between Z1 and Z2. Since a = q −wi and b = q − , Equation 6 can also be written as: Xμ − 2 p ν/X < q − wi < q − < Xμ + 2 p ν/X The smallest X that satisfies the constraint in Equation 6 strictly increases with q. As X increases, the error in sum of weights in a sample (i.e., e(σX ) = p (ν)/X) decreases. Consequently, the error in a player"s marginal contribution (see Equation 5) also decreases. This implies that as q increases, the error in the marginal contribution (and consequently the error in the Shapley value) decreases. - Effect of m. It is clear from Equation 4 that the error e(σX ) is highest for X = 1 and it decreases with X. Hence, for small m, e(σ1 ) has a significant effect on PE. But as m increases, the effect of e(σ1 ) on PE decreases and, as a result, PE decreases. 6. RELATED WORK In order to overcome the computational complexity of finding the Shapley value, two main approaches have been proposed in the literature. One approach is to use generating functions [3]. This method is an exact procedure that overcomes the problem of time complexity, but its storage requirements are substantial - it requires huge arrays. It also has the limitation (not shared by other approaches) that it can only be applied to games with integer weights and quotas. The other method uses an approximation technique based on Monte Carlo simulation. In [12], for instance, the Shapley value is computed by considering a random sample from a large population of players. The method we propose differs from this in that they define the Shapley value by treating a player"s number of swings (if a player can change a losing coalition to a winning one, then, for the player, the coalition is counted as a swing) as a random variable, while we treat the players" weights as random variables. In [12], however, the question remains how to get the number of swings from the definition of a voting game and what is the time complexity of doing this. Since the voting game is defined in terms of the players" weights and the number of swings are obtained from these weights, our method corresponds more closely to the definition of the voting game. Our method also differs from [7] in that while [7] presents a method for the case where all the players" weights are distributed normally, our method applies to any type of distribution for these weights. Thus, as stated in Section 1, our method is more general than [3, 12, 7]. Also, unlike all the above mentioned work, we provide an analysis of the performance of our method in terms of the percentage error in the approximate Shapley value. A method for finding the Shapley value was also proposed in [5]. This method gives the exact Shapley value, but its time complexity is exponential. Furthermore, the method can be used only if the game is represented in a specific form (viz., the multi-issue representation), not otherwise. Finally, [9, 10] present a polynomial time method for finding the Shapley value. This method can be used if the coalition game is represented as a marginal contribution net. Furthermore, they assume that the Shapley value of a component of a given coalition game is given by an oracle, and on the basis of this assumption aggregate these values to find the value for the overall game. In contrast, our method is independent The Sixth Intl. Joint Conf. on Autonomous Agents and Multi-Agent Systems (AAMAS 07) 965 of the representation and gives an approximate Shapley value in linear time, without the need for an oracle. 7. CONCLUSIONS AND FUTURE WORK Coalition formation is an important form of interaction in multiagent systems. An important issue in such work is for the agents to decide how to split the gains from cooperation between the members of a coalition. In this context, cooperative game theory offers a solution concept called the Shapley value. The main advantage of the Shapley value is that it provides a solution that is both unique and fair. However, its main problem is that, for many coalition games, the Shapley value cannot be determined in polynomial time. In particular, the problem of finding this value for the voting game is #P-complete. Although this problem is, in general #P-complete, we show that there are some specific voting games for which the Shapley value can be determined in polynomial time and characterise such games. By doing so, we have shown when it is computationally feasible to find the exact Shapley value. For other complex voting games, we presented a new randomized method for determining the approximate Shapley value. The time complexity of the proposed method is linear in the number of players. We analysed the performance of this method in terms of the percentage error in the approximate Shapley value. Our experiments show that the percentage error in the Shapley value is at most 20. Furthermore, in most cases, the error is less than 5%. Finally, we analyse the effect of the different parameters of the voting game on this error. Our study shows that the error decreases as 1. a player"s weight increases, 2. the quota increases, and 3. the number of players increases. Given the fact that software agents have limited computational resources and therefore cannot compute the true Shapley value, our results are especially relevant to such resource bounded agents. In future, we will explore the problem of determining the Shapley value for other commonly occurring coalition games like the production economy and the market economy. 8. REFERENCES [1] R. Aumann. Acceptable points in general cooperative n-person games. In Contributions to theTheory of Games volume IV. Princeton University Press, 1959. [2] G. Ausiello, P. Crescenzi, G. Gambosi, V. Kann, A. Marchetti-Spaccamela, and M. Protasi. Complexity and approximation: Combinatorial optimization problems and their approximability properties. Springer, 2003. [3] J. M. Bilbao, J. R. Fernandez, A. J. Losada, and J. J. Lopez. Generating functions for computing power indices efficiently. Top 8, 2:191-213, 2000. [4] P. Bork, H. Grote, D. Notz, and M. Regler. Data Analysis Techniques in High Energy Physics Experiments. Cambridge University Press, 1993. [5] V. Conitzer and T. Sandholm. Computing Shapley values, manipulating value division schemes, and checking core membership in multi-issue domains. In Proceedings of the National Conference on Artificial Intelligence, pages 219-225, San Jose, California, 2004. [6] X. Deng and C. H. Papadimitriou. On the complexity of cooperative solution concepts. Mathematics of Operations Research, 19(2):257-266, 1994. [7] S. S. Fatima, M. Wooldridge, and N. R. Jennings. An analysis of the shapley value and its uncertainty for the voting game. In Proc. 7th Int. Workshop on Agent Mediated Electronic Commerce, pages 39-52, 2005. [8] A. Francis. Advanced Level Statistics. Stanley Thornes Publishers, 1979. [9] S. Ieong and Y. Shoham. Marginal contribution nets: A compact representation scheme for coalitional games. In Proceedings of the Sixth ACM Conference on Electronic Commerce, pages 193-202, Vancouver, Canada, 2005. [10] S. Ieong and Y. Shoham. Multi-attribute coalition games. In Proceedings of the Seventh ACM Conference on Electronic Commerce, pages 170-179, Ann Arbor, Michigan, 2006. [11] J. P. Kahan and A. Rapoport. Theories of Coalition Formation. Lawrence Erlbaum Associates Publishers, 1984. [12] I. Mann and L. S. Shapley. Values for large games iv: Evaluating the electoral college exactly. Technical report, The RAND Corporation, Santa Monica, 1962. [13] A. MasColell, M. Whinston, and J. R. Green. Microeconomic Theory. Oxford University Press, 1995. [14] M. J. Osborne and A. Rubinstein. A Course in Game Theory. The MIT Press, 1994. [15] C. H. Papadimitriou. Computational Complexity. Addison Wesley Longman, 1994. [16] A. Rapoport. N-person Game Theory : Concepts and Applications. Dover Publications, Mineola, NY, 2001. [17] A. E. Roth. Introduction to the shapley value. In A. E. Roth, editor, The Shapley value, pages 1-27. University of Cambridge Press, Cambridge, 1988. [18] T. Sandholm and V. Lesser. Coalitions among computationally bounded agents. Artificial Intelligence Journal, 94(1):99-137, 1997. [19] L. S. Shapley. A value for n person games. In A. E. Roth, editor, The Shapley value, pages 31-40. University of Cambridge Press, Cambridge, 1988. [20] O. Shehory and S. Kraus. A kernel-oriented model for coalition-formation in general environments: Implemetation and results. In In Proceedings of the National Conference on Artificial Intelligence (AAAI-96), pages 131-140, 1996. [21] O. Shehory and S. Kraus. Methods for task allocation via agent coalition formation. Artificial Intelligence Journal, 101(2):165-200, 1998. [22] J. R. Taylor. An introduction to error analysis: The study of uncertainties in physical measurements. University Science Books, 1982. 966 The Sixth Intl. Joint Conf. on Autonomous Agents and Multi-Agent Systems (AAMAS 07)

mean of reaching consensus;approximation;unique and fair solution;polynomial time;reaching consensus mean;generating function;coalition formation;shapley value;game-theory;cooperative game theory;interaction;multi-agent system;randomised method

train_I-54

Approximate and Online Multi-Issue Negotiation

This paper analyzes bilateral multi-issue negotiation between selfinterested autonomous agents. The agents have time constraints in the form of both deadlines and discount factors. There are m > 1 issues for negotiation where each issue is viewed as a pie of size one. The issues are indivisible (i.e., individual issues cannot be split between the parties; each issue must be allocated in its entirety to either agent). Here different agents value different issues differently. Thus, the problem is for the agents to decide how to allocate the issues between themselves so as to maximize their individual utilities. For such negotiations, we first obtain the equilibrium strategies for the case where the issues for negotiation are known a priori to the parties. Then, we analyse their time complexity and show that finding the equilibrium offers is an NP-hard problem, even in a complete information setting. In order to overcome this computational complexity, we then present negotiation strategies that are approximately optimal but computationally efficient, and show that they form an equilibrium. We also analyze the relative error (i.e., the difference between the true optimum and the approximate). The time complexity of the approximate equilibrium strategies is O(nm/ 2 ) where n is the negotiation deadline and the relative error. Finally, we extend the analysis to online negotiation where different issues become available at different time points and the agents are uncertain about their valuations for these issues. Specifically, we show that an approximate equilibrium exists for online negotiation and show that the expected difference between the optimum and the approximate is O( √ m) . These approximate strategies also have polynomial time complexity.

1. INTRODUCTION Negotiation is a key form of interaction in multiagent systems. It is a process in which disputing agents decide how to divide the gains from cooperation. Since this decision is made jointly by the agents themselves [20, 19, 13, 15], each party can only obtain what the other is prepared to allow them. Now, the simplest form of negotiation involves two agents and a single issue. For example, consider a scenario in which a buyer and a seller negotiate on the price of a good. To begin, the two agents are likely to differ on the price at which they believe the trade should take place, but through a process of joint decision-making they either arrive at a price that is mutually acceptable or they fail to reach an agreement. Since agents are likely to begin with different prices, one or both of them must move toward the other, through a series of offers and counter offers, in order to obtain a mutually acceptable outcome. However, before the agents can actually perform such negotiations, they must decide the rules for making offers and counter offers. That is, they must set the negotiation protocol [20]. On the basis of this protocol, each agent chooses its strategy (i.e., what offers it should make during the course of negotiation). Given this context, this work focuses on competitive scenarios with self-interested agents. For such cases, each participant defines its strategy so as to maximise its individual utility. However, in most bilateral negotiations, the parties involved need to settle more than one issue. For this case, the issues may be divisible or indivisible [4]. For the former, the problem for the agents is to decide how to split each issue between themselves [21]. For the latter, the individual issues cannot be divided. An issue, in its entirety, must be allocated to either of the two agents. Since the agents value different issues differently, they must come to terms about who will take which issue. To date, most of the existing work on multi-issue negotiation has focussed on the former case [7, 2, 5, 23, 11, 6]. However, in many real-world settings, the issues are indivisible. Hence, our focus here is on negotiation for indivisible issues. Such negotiations are very common in multiagent systems. For example, consider the case of task allocation between two agents. There is a set of tasks to be carried out and different agents have different preferences for the tasks. The tasks cannot be partitioned; a task must be carried out by one agent. The problem then is for the agents to negotiate about who will carry out which task. A key problem in the study of multi-issue negotiation is to determine the equilibrium strategies. An equally important problem, especially in the context of software agents, is to find the time complexity of computing the equilibrium offers. However, such computational issues have so far received little attention. As we will show, this is mainly due to the fact that existing work (describe in Section 5) has mostly focused on negotiation for divisible issues 951 978-81-904262-7-5 (RPS) c 2007 IFAAMAS and finding the equilibrium for this case is computationally easier than that for the case of indivisible issues. Our primary objective is, therefore, to answer the computational questions for the latter case for the types of situations that are commonly faced by agents in real-world contexts. Thus, we consider negotiations in which there is incomplete information and time constraints. Incompleteness of information on the part of negotiators is a common feature of most practical negotiations. Also, agents typically have time constraints in the form of both deadlines and discount factors. Deadlines are an essential element since negotiation cannot go on indefinitely, rather it must end within a reasonable time limit. Likewise, discount factors are essential since the goods may be perishable or their value may decline due to inflation. Moreover, the strategic behaviour of agents with deadlines and discount factors differs from those without (see [21] for single issue bargaining without deadlines and [23, 13] for bargaining with deadlines and discount factors in the context of divisible issues). Given this, we consider indivisible issues and first analyze the strategic behaviour of agents to obtain the equilibrium strategies for the case where all the issues for negotiation are known a priori to both agents. For this case, we show that the problem of finding the equilibrium offers is NP-hard, even in a complete information setting. Then, in order to overcome the problem of time complexity, we present strategies that are approximately optimal but computationally efficient, and show that they form an equilibrium. We also analyze the relative error (i.e., the difference between the true optimum and the approximate). The time complexity of the approximate equilibrium strategies is O(nm/ 2 ) where n is the negotiation deadline and the relative error. Finally, we extend the analysis to online negotiation where different issues become available at different time points and the agents are uncertain about their valuations for these issues. Specifically, we show that an approximate equilibrium exists for online negotiation and show that the expected difference between the optimum and the approximate is O( √ m) . These approximate strategies also have polynomial time complexity. In so doing, our contribution lies in analyzing the computational complexity of the above multi-issue negotiation problem, and finding the approximate and online equilibria. No previous work has determined these equilibria. Since software agents have limited computational resources, our results are especially relevant to such resource bounded agents. The remainder of the paper is organised as follows. We begin by giving a brief overview of single-issue negotiation in Section 2. In Section 3, we obtain the equilibrium for multi-issue negotiation and show that finding equilibrium offers is an NP-hard problem. We then present an approximate equilibrium and evaluate its approximation error. Section 4 analyzes online multi-issue negotiation. Section 5 discusses the related literature and Section 6 concludes. 2. SINGLE-ISSUE NEGOTIATION We adopt the single issue model of [27] because this is a model where, during negotiation, the parties are allowed to make offers from a set of discrete offers. Since our focus is on indivisible issues (i.e., parties are allowed to make one of two possible offers: zero or one), our scenario fits in well with [27]. Hence we use this basic single issue model and extend it to multiple issues. Before doing so, we give an overview of this model and its equilibrium strategies. There are two strategic agents: a and b. Each agent has time constraints in the form of deadlines and discount factors. The two agents negotiate over a single indivisible issue (i). This issue is a ‘pie" of size 1 and the agents want to determine who gets the pie. There is a deadline (i.e., a number of rounds by which negotiation must end). Let n ∈ N+ denote this deadline. The agents use an alternating offers protocol (as the one of Rubinstein [18]), which proceeds through a series of time periods. One of the agents, say a, starts negotiation in the first time period (i.e., t = 1) by making an offer (xi = 0 or 1) to b. Agent b can either accept or reject the offer. If it accepts, negotiation ends in an agreement with a getting xi and b getting yi = 1 − xi. Otherwise, negotiation proceeds to the next time period, in which agent b makes a counter-offer. This process of making offers continues until one of the agents either accepts an offer or quits negotiation (resulting in a conflict). Thus, there are three possible actions an agent can take during any time period: accept the last offer, make a new counter-offer, or quit the negotiation. An essential feature of negotiations involving alternating offers is that the agents" utilities decrease with time [21]. Specifically, the decrease occurs at each step of offer and counteroffer. This decrease is represented with a discount factor denoted 0 < δi ≤ 1 for both1 agents. Let [xt i, yt i ] denote the offer made at time period t where xt i and yt i denote the share for agent a and b respectively. Then, for a given pie, the set of possible offers is: {[xt i, yt i ] : xt i = 0 or 1, yt i = 0 or 1, and xt i + yt i = 1} At time t, if a and b receive a share of xt i and yt i respectively, then their utilities are: ua i (xt i, t) = j xt i × δt−1 if t ≤ n 0 otherwise ub i (yt i , t) = j yt i × δt−1 if t ≤ n 0 otherwise The conflict utility (i.e., the utility received in the event that no deal is struck) is zero for both agents. For the above setting, the agents reason as follows in order to determine what to offer at t = 1. We let A(1) (B(1)) denote a"s (b"s) equilibrium offer for the first time period. Let agent a denote the first mover (i.e., at t = 1, a proposes to b who should get the pie). To begin, consider the case where the deadline for both agents is n = 1. If b accepts, the division occurs as agreed; if not, neither agent gets anything (since n = 1 is the deadline). Here, a is in a powerful position and is able to propose to keep 100 percent of the pie and give nothing to b 2 . Since the deadline is n = 1, b accepts this offer and agreement takes place in the first time period. Now, consider the case where the deadline is n = 2. In order to decide what to offer in the first round, a looks ahead to t = 2 and reasons backwards. Agent a reasons that if negotiation proceeds to the second round, b will take 100 percent of the pie by offering [0, 1] and leave nothing for a. Thus, in the first time period, if a offers b anything less than the whole pie, b will reject the offer. Hence, during the first time period, agent a offers [0, 1]. Agent b accepts this and an agreement occurs in the first time period. In general, if the deadline is n, negotiation proceeds as follows. As before, agent a decides what to offer in the first round by looking ahead as far as t = n and then reasoning backwards. Agent a"s 1 Having a different discount factor for different agents only makes the presentation more involved without leading to any changes in the analysis of the strategic behaviour of the agents or the time complexity of finding the equilibrium offers. Hence we have a single discount factor for both agents. 2 It is possible that b may reject such a proposal. However, irrespective of whether b accepts or rejects the proposal, it gets zero utility (because the deadline is n = 1). Thus, we assume that b accepts a"s offer. 952 The Sixth Intl. Joint Conf. on Autonomous Agents and Multi-Agent Systems (AAMAS 07) offer for t = 1 depends on who the offering agent is for the last time period. This, in turn, depends on whether n is odd or even. Since a makes an offer at t = 1 and the agents use the alternating offers protocol, the offering agent for the last time period is b if n is even and it is a if n is odd. Thus, depending on whether n is odd or even, a makes the following offer at t = 1: A(1) = j OFFER [1, 0] IF ODD n ACCEPT IF b"s TURN B(1) = j OFFER [0, 1] IF EVEN n ACCEPT IF a"s TURN Agent b accepts this offer and negotiation ends in the first time period. Note that the equilibrium outcome depends on who makes the first move. Since we have two agents and either of them could move first, we get two possible equilibrium outcomes. On the basis of the above equilibrium for single-issue negotiation with complete information, we first obtain the equilibrium for multiple issues and then show that computing these offers is a hard problem. We then present a time efficient approximate equilibrium. 3. MULTI-ISSUE NEGOTIATION We first analyse the complete information setting. This section forms the base which we extend to the case of information uncertainty in Section 4. Here a and b negotiate over m > 1 indivisible issues. These issues are m distinct pies and the agents want to determine how to distribute the pies between themselves. Let S = {1, 2, . . . , m} denote the set of m pies. As before, each pie is of size 1. Let the discount factor for issue c, where 1 ≤ c ≤ m, be 0 < δc ≤ 1. For each issue, let n denote each agent"s deadline. In the offer for time period t (where 1 ≤ t ≤ n), agent a"s (b"s) share for each of the m issues is now represented as an m element vector xt ∈ Bm (yt ∈ Bm ) where B denotes the set {0, 1}. Thus, if agent a"s share for issue c at time t is xt c, then agent b"s share is yt c = (1−xt c). The shares for a and b are together represented as the package [xt , yt ]. As is traditional in multi-issue utility theory, we define an agent"s cumulative utility using the standard additive form [12]. The functions Ua : Bm × Bm × N+ → R and Ub : Bm × Bm × N+ → R give the cumulative utilities for a and b respectively at time t. These are defined as follows: Ua ([xt , yt ], t) = ( Σm c=1ka c ua c (xt c, t) if t ≤ n 0 otherwise (1) Ub ([xt , yt ], t) = ( Σm c=1kb cub c(yt c, t) if t ≤ n 0 otherwise (2) where ka ∈ Nm + denotes an m element vector of constants for agent a and kb ∈ Nm + that for b. Here N+ denotes the set of positive integers. These vectors indicate how the agents value different issues. For example, if ka c > ka c+1, then agent a values issue c more than issue c + 1. Likewise for agent b. In other words, the m issues are perfect substitutes (i.e., all that matters to an agent is its total utility for all the m issues and not that for any subset of them). In all the settings we study, the issues will be perfect substitutes. To begin each agent has complete information about all negotiation parameters (i.e., n, m, ka c , kb c, and δc for 1 ≤ c ≤ m). Now, multi-issue negotiation can be done using different procedures. Broadly speaking, there are three key procedures for negotiating multiple issues [19]: 1. the package deal procedure where all the issues are settled together as a bundle, 2. the sequential procedure where the issues are discussed one after another, and 3. the simultaneous procedure where the issues are discussed in parallel. Between these three procedures, the package deal is known to generate Pareto optimal outcomes [19, 6]. Hence we adopt it here. We first give a brief description of the procedure and then determine the equilibrium strategies for it. 3.1 The package deal procedure In this procedure, the agents use the same protocol as for singleissue negotiation (described in Section 2). However, an offer for the package deal includes a proposal for each issue under negotiation. Thus, for m issues, an offer includes m divisions, one for each issue. Agents are allowed to either accept a complete offer (i.e., all m issues) or reject a complete offer. An agreement can therefore take place either on all m issues or on none of them. As per the single-issue negotiation, an agent decides what to offer by looking ahead and reasoning backwards. However, since an offer for the package deal includes a share for all the m issues, the agents can now make tradeoffs across the issues in order to maximise their cumulative utilities. For 1 ≤ c ≤ m, the equilibrium offer for issue c at time t is denoted as [at c, bt c] where at c and bt c denote the shares for agent a and b respectively. We denote the equilibrium package at time t as [at , bt ] where at ∈ Bm (bt ∈ Bm ) is an m element vector that denotes a"s (b"s) share for each of the m issues. Also, for 1 ≤ c ≤ m, δc is the discount factor for issue c. The symbols 0 and 1 denote m element vectors of zeroes and ones respectively. Note that for 1 ≤ t ≤ n, at c + bt c = 1 (i.e., the sum of the agents" shares (at time t) for each pie is one). Finally, for time period t (for 1 ≤ t ≤ n) we let A(t) (respectively B(t)) denote the equilibrium strategy for agent a (respectively b). 3.2 Equilibrium strategies As mentioned in Section 1, the package deal allows agents to make tradeoffs. We let TRADEOFFA (TRADEOFFB) denote agent a"s (b"s) function for making tradeoffs. We let P denote a set of parameters to the procedure TRADEOFFA (TRADEOFFB) where P = {ka , kb , δ, m}. Given this, the following theorem characterises the equilibrium for the package deal procedure. THEOREM 1. For the package deal procedure, the following strategies form a Nash equilibrium. The equilibrium strategy for t = n is: A(n) = j OFFER [1, 0] IF a"s TURN ACCEPT IF b"s TURN B(n) = j OFFER [0, 1] IF b"s TURN ACCEPT IF a"s TURN For all preceding time periods t < n, if [xt , yt ] denotes the offer made at time t, then the equilibrium strategies are defined as follows: A(t) = 8 < : OFFER TRADEOFFA(P, UB(t), t) IF a"s TURN If (Ua ([xt , yt ], t) ≥ UA(t)) ACCEPT else REJECT IF b"s TURN B(t) = 8 < : OFFER TRADEOFFB(P, UA(t), t) IF b"s TURN If (Ub ([xt , yt ], t) ≥ UB(t)) ACCEPT else REJECT IF a"s TURN The Sixth Intl. Joint Conf. on Autonomous Agents and Multi-Agent Systems (AAMAS 07) 953 where UA(t) = Ua ([at+1 , bt+1 ], t + 1) and UB(t) = Ub ([at+1 , bt+1 ], t + 1). PROOF. We look ahead to the last time period (i.e., t = n) and then reason backwards. To begin, if negotiation reaches the deadline (n), then the agent whose turn it is takes everything and leaves nothing for its opponent. Hence, we get the strategies A(n) and B(n) as given in the statement of the theorem. In all the preceding time periods (t < n), the offering agent proposes a package that gives its opponent a cumulative utility equal to what the opponent would get from its own equilibrium offer for the next time period. During time period t, either a or b could be the offering agent. Consider the case where a makes an offer at t. The package that a offers at t gives b a cumulative utility of Ub ([at+1 , bt+1 ], t + 1). However, since there is more than one issue, there is more than one package that gives b this cumulative utility. From among these packages, a offers the one that maximises its own cumulative utility (because it is a utility maximiser). Thus, the problem for a is to find the package [at , bt ] so as to: maximize mX c=1 ka c (1 − bt c)δt−1 c (3) such that mX c=1 bt ckb c ≥ UB(t) bt c = 0 or 1 for 1 ≤ c ≤ m where UB(t), δt−1 c , ka c , and kb c are constants and bt c (1 ≤ c ≤ m) is a variable. Assume that the function TRADEOFFA takes parameters P, UB(t), and t, to solve the maximisation problem given in Equation 3 and returns the corresponding package. If there is more than one package that solves Equation 3, then TRADEOFFA returns any one of them (because agent a gets equal utility from all such packages and so does agent b). The function TRADEOFFB for agent b is analogous to that for a. On the other hand, the equilibrium strategy for the agent that receives an offer is as follows. For time period t, let b denote the receiving agent. Then, b accepts [xt , yt ] if UB(t) ≤ Ub ([xt , yt ], t), otherwise it rejects the offer because it can get a higher utility in the next time period. The equilibrium strategy for a as receiving agent is defined analogously. In this way, we reason backwards and obtain the offers for the first time period. Thus, we get the equilibrium strategies (A(t) and B(t)) given in the statement of the theorem. The following example illustrates how the agents make tradeoffs using the above equilibrium strategies. EXAMPLE 1. Assume there are m = 2 issues for negotiation, the deadline for both issues is n = 2, and the discount factor for both issues for both agents is δ = 1/2. Let ka 1 = 3, ka 2 = 1, kb 1 = 1, and kb 2 = 5. Let agent a be the first mover. By using backward reasoning, a knows that if negotiation reaches the second time period (which is the deadline), then b will get a hundred percent of both the issues. This gives b a cumulative utility of UB(2) = 1/2 + 5/2 = 3. Thus, in the first time period, if b gets anything less than a utility of 3, it will reject a"s offer. So, at t = 1, a offers the package where it gets issue 1 and b gets issue 2. This gives a cumulative utility of 3 to a and 5 to b. Agent b accepts the package and an agreement takes place in the first time period. The maximization problem in Equation 3 can be viewed as the 0-1 knapsack problem3 . In the 0-1 knapsack problem, we have a set 3 Note that for the case of divisible issues this is the fractional knapof m items where each item has a profit and a weight. There is a knapsack with a given capacity. The objective is to fill the knapsack with items so as to maximize the cumulative profit of the items in the knapsack. This problem is analogous to the negotiation problem we want to solve (i.e., the maximization problem of Equation 3). Since ka c and δt−1 c are constants, maximizing Pm c=1 ka c (1−bt c)δt−1 c is the same as minimizing Pm c=1 ka c bt c. Hence Equation 3 can be written as: minimize mX c=1 ka c bt c (4) such that mX c=1 bt ckb c ≥ UB(t) bt c = 0 or 1 for 1 ≤ c ≤ m Equation 4 is a minimization version of the standard 0-1 knapsack problem4 with m items where ka c represents the profit for item c, kb c the weight for item c, and UB(t) the knapsack capacity. Example 1 was for two issues and so it was easy to find the equilibrium offers. But, in general, it is not computationally easy to find the equilibrium offers of Theorem 1. The following theorem proves this. THEOREM 2. For the package deal procedure, the problem of finding the equilibrium offers given in Theorem 1 is NP-hard. PROOF. Finding the equilibrium offers given in Theorem 1 requires solving the 0-1 knapsack problem given in Equation 4. Since the 0-1 knapsack problem is NP-hard [17], the problem of finding equilibrium for the package deal is also NP-hard. 3.3 Approximate equilibrium Researchers in the area of algorithms have found time efficient methods for computing approximate solutions to 0-1 knapsack problems [10]. Hence we use these methods to find a solution to our negotiation problem. At this stage, we would like to point out the main difference between solving the 0-1 knapsack problem and solving our negotiation problem. The 0-1 knapsack problem involves decision making by a single agent regarding which items to place in the knapsack. On the other hand, our negotiation problem involves two players and they are both strategic. Hence, in our case, it is not enough to just find an approximate solution to the knapsack problem, we must also show that such an approximation forms an equilibrium. The traditional approach for overcoming the computational complexity in finding an equilibrium has been to use an approximate equilibrium (see [14, 26] for example). In this approach, a strategy profile is said to form an approximate Nash equilibrium if neither agent can gain more than the constant by deviating. Hence, our aim is to use the solution to the 0-1 knapsack problem proposed in [10] and show that it forms an approximate equilibrium to our negotiation problem. Before doing so, we give a brief overview of the key ideas that underlie approximation algorithms. There are two key issues in the design of approximate algorithms [1]: sack problem. The factional knapsack problem is computationally easy; it can be solved in time polynomial in the number of items in the knapsack problem [17]. In contrast, the 0-1 knapsack problem is computationally hard. 4 Note that for the standard 0-1 knapsack problem the weights, profits and the capacity are positive integers. However a 0-1 knapsack problem with fractions and non positive values can easily be transformed to one with positive integers in time linear in m using the methods given in [8, 17]. 954 The Sixth Intl. Joint Conf. on Autonomous Agents and Multi-Agent Systems (AAMAS 07) 1. the quality of their solution, and 2. the time taken to compute the approximation. The quality of an approximate algorithm is determined by comparing its performance to that of the optimal algorithm and measuring the relative error [3, 1]. The relative error is defined as (z−z∗ )/z∗ where z is the approximate solution and z∗ the optimal one. In general, we are interested in finding approximate algorithms whose relative error is bounded from above by a certain constant , i.e., (z − z∗ )/z∗ ≤ (5) Regarding the second issue of time complexity, we are interested in finding fully polynomial approximation algorithms. An approximation algorithm is said to be fully polynomial if for any > 0 it finds a solution satisfying Equation 5 in time polynomially bounded by size of the problem (for the 0-1 knapsack problem, the problem size is equal to the number of items) and by 1/ [1]. For the 0-1 knapsack problem, Ibarra and Kim [10] presented a fully polynomial approximation method. This method is based on dynamic programming. It is a parametric method that takes as a parameter and for any > 0, finds a heuristic solution z with relative error at most , such that the time and space complexity grow polynomially with the number of items m and 1/ . More specifically, the space and time complexity are both O(m/ 2 ) and hence polynomial in m and 1/ (see [10] for the detailed approximation algorithm and proof of time and space complexity). Since the Ibarra and Kim method is fully polynomial, we use it to solve our negotiation problem. This is done as follows. For agent a, let APRX-TRADEOFFA(P, UB(t), t, ) denote a procedure that returns an approximate solution to Equation 4 using the Ibarra and Kim method. The procedure APRX-TRADEOFFB(P, UA(t), t, ) for agent b is analogous. For 1 ≤ c ≤ m, the approximate equilibrium offer for issue c at time t is denoted as [¯at c,¯bt c] where ¯at c and ¯bt c denote the shares for agent a and b respectively. We denote the equilibrium package at time t as [¯at ,¯bt ] where ¯at ∈ Bm (¯bt ∈ Bm ) is an m element vector that denotes a"s (b"s) share for each of the m issues. Also, as before, for 1 ≤ c ≤ m, δc is the discount factor for issue c. Note that for 1 ≤ t ≤ n, ¯at c + ¯bt c = 1 (i.e., the sum of the agents" shares (at time t) for each pie is one). Finally, for time period t (for 1 ≤ t ≤ n) we let ¯A(t) (respectively ¯B(t)) denote the approximate equilibrium strategy for agent a (respectively b).The following theorem uses this notation and characterizes an approximate equilibrium for multi-issue negotiation. THEOREM 3. For the package deal procedure, the following strategies form an approximate Nash equilibrium. The equilibrium strategy for t = n is: ¯A(n) = j OFFER [1, 0] IF a"s TURN ACCEPT IF b"s TURN ¯B(n) = j OFFER [0, 1] IF b"s TURN ACCEPT IF a"s TURN For all preceding time periods t < n, if [xt , yt ] denotes the offer made at time t, then the equilibrium strategies are defined as follows: ¯A(t) = 8 < : OFFER APRX-TRADEOFFA(P, UB(t), t, ) IF a"s TURN If (Ua ([xt , yt ], t) ≥ UA(t)) ACCEPT else REJECT IF b"s TURN ¯B(t) = 8 < : OFFER APRX-TRADEOFFB(P, UA(t), t, ) IF b"s TURN If (Ub ([xt , yt ], t) ≥ UB(t)) ACCEPT else REJECT IF a"s TURN where UA(t) = Ua ([¯at+1 ,¯bt+1 ], t + 1) and UB(t) = Ub ([¯at+1 , ¯bt+1 ], t + 1). An agreement takes place at t = 1. PROOF. As in the proof for Theorem 1, we use backward reasoning. We first obtain the strategies for the last time period t = n. It is straightforward to get these strategies; the offering agent gets a hundred percent of all the issues. Then for t = n − 1, the offering agent must solve the maximization problem of Equation 4 by substituting t = n−1 in it. For agent a (b), this is done by APPROX-TRADEOFFA (APPROX-TRADEOFFB). These two functions are nothing but the Ibarra and Kim"s approximation method for solving the 0-1 knapsack problem. These two functions take as a parameter and use the Ibarra and Kim"s approximation method to return a package that approximately maximizes Equation 4. Thus, the relative error for these two functions is the same as that for Ibarra and Kim"s method (i.e., it is at most where is given in Equation 5). Assume that a is the offering agent for t = n − 1. Agent a must offer a package that gives b a cumulative utility equal to what it would get from its own approximate equilibrium offer for the next time period (i.e., Ub ([¯at+1 ,¯bt+1 ], t + 1) where [¯at+1 ,¯bt+1 ] is the approximate equilibrium package for the next time period). Recall that for the last time period, the offering agent gets a hundred percent of all the issues. Since a is the offering agent for t = n − 1 and the agents use the alternating offers protocol, it is b"s turn at t = n. Thus Ub ([¯at+1 ,¯bt+1 ], t + 1) is equal to b"s cumulative utility from receiving a hundred percent of all the issues. Using this utility as the capacity of the knapsack, a uses APPROX-TRADEOFFA and obtains the approximate equilibrium package for t = n − 1. On the other hand, if b is the offering agent at t = n − 1, it uses APPROX-TRADEOFFB to obtain the approximate equilibrium package. In the same way for t < n − 1, the offering agent (say a) uses APPROX-TRADEOFFA to find an approximate equilibrium package that gives b a utility of Ub ([¯at+1 ,¯bt+1 ], t + 1). By reasoning backwards, we obtain the offer for time period t = 1. If a (b) is the offering agent, it proposes the offer APPROX-TRADEOFFA(P, UB(1), 1, ) (APPROX-TRADEOFFB(P, UA(1), 1, )). The receiving agent accepts the offer. This is because the relative error in its cumulative utility from the offer is at most . An agreement therefore takes place in the first time period. THEOREM 4. The time complexity of finding the approximate equilibrium offer for the first time period is O(nm/ 2 ). PROOF. The time complexity of APPROX-TRADEOFFA and APPROXTRADEOFFB is the same as the time complexity of the Ibarra and Kim method [10] i.e., O(m/ 2 )). In order to find the equilibrium offer for the first time period using backward reasoning, APPROXTRADEOFFA (or APPROX- TRADEOFFB) is invoked n times. Hence the time complexity of finding the approximate equilibrium offer for the first time period is O(nm/ 2 ). This analysis was done in a complete information setting. However an extension of this analysis to an incomplete information setting where the agents have probability distributions over some uncertain parameter is straightforward, as long as the negotiation is done offline; i.e., the agents know their preference for each individual issue before negotiation begins. For instance, consider the case where different agents have different discount factors, and each agent is uncertain about its opponent"s discount factor although it knows its own. This uncertainty is modelled with a probability distribution over the possible values for the opponent"s discount factor and having this distribution as common knowledge to the agents. All our analysis for the complete information setting still holds for The Sixth Intl. Joint Conf. on Autonomous Agents and Multi-Agent Systems (AAMAS 07) 955 this incomplete information setting, except for the fact that an agent must now use the given probability distribution and find its opponent"s expected utility instead of its actual utility. Hence, instead of analyzing an incomplete information setting for offline negotiation, we focus on online multi-issue negotiation. 4. ONLINE MULTI-ISSUE NEGOTIATION We now consider a more general and, arguably more realistic, version of multi-issue negotiation, where the agents are uncertain about the issues they will have to negotiate about in future. In this setting, when negotiating an issue, the agents know that they will negotiate more issues in the future, but they are uncertain about the details of those issues. As before, let m be the total number of issues that are up for negotiation. The agents have a probability distribution over the possible values of ka c and kb c. For 1 ≤ c ≤ m let ka c and kb c be uniformly distributed over [0,1]. This probability distribution, n, and m are common knowledge to the agents. However, the agents come to know ka c and kb c only just before negotiation for issue c begins. Once the agents reach an agreement on issue c, it cannot be re-negotiated. This scenario requires online negotiation since the agents must make decisions about an issue prior to having the information about the future issues [3]. We first give a brief introduction to online problems and then draw an analogy between the online knapsack problem and the negotiation problem we want to solve. In an online problem, data is given to the algorithm incrementally, one unit at a time [3]. The online algorithm must also produce the output incrementally: after seeing i units of input it must output the ith unit of output. Since decisions about the output are made with incomplete knowledge about the entire input, an online algorithm often cannot produce an optimal solution. Such an algorithm can only approximate the performance of the optimal algorithm that sees all the inputs in advance. In the design of online algorithms, the main aim is to achieve a performance that is close to that of the optimal offline algorithm on each input. An online algorithm is said to be stochastic if it makes decisions on the basis of the probability distributions for the future inputs. The performance of stochastic online algorithms is assessed in terms of the expected difference between the optimum and the approximate solution (denoted E[z∗ m −zm] where z∗ m is the optimal and zm the approximate solution). Note that the subscript m is used to indicate the fact that this difference depends on m. We now describe the protocol for online negotiation and then obtain an approximate equilibrium. The protocol is defined as follows. Let agent a denote the first mover (since we focus on the package deal procedure, the first mover is the same for all the m issues). Step 1. For c = 1, the agents are given the values of ka c and kb c. These two values are now common5 knowledge. Step 2. The agents settle issue c using the alternating offers protocol described in Section 2. Negotiation for issue c must end within n time periods from the start of negotiation on the issue. If an agreement is not reached within this time, then negotiation fails on this and on all remaining issues. Step 3. The above steps are repeated for issues c = 2, 3, . . . , m. Negotiation for issue c (2 ≤ c ≤ m) begins in the time period following an agreement on issue c − 1. 5 We assume common knowledge because it simplifies exposition. However, if ka c (kb c) is a"s (b"s) private knowledge, then our analysis will still hold but now an agent must find its opponent"s expected utility on the basis of the p.d.fs for ka c and kb c. Thus, during time period t, the problem for the offering agent (say a) is to find the optimal offer for issue c on the basis of ka c and kb c and the probability distribution for ka i and kb i (c < i ≤ m). In order to solve this online negotiation problem we draw analogy with the online knapsack problem. Before doing so, however, we give a brief overview of the online knapsack problem. In the online knapsack problem, there are m items. The agent must examine the m items one at a time according to the order they are input (i.e., as their profit and size coefficients become known). Hence, the algorithm is required to decide whether or not to include each item in the knapsack as soon as its weight and profit become known, without knowledge concerning the items still to be seen, except for their total number. Note that since the agents have a probability distribution over the weights and profits of the future items, this is a case of stochastic online knapsack problem. Our online negotiation problem is analogous to the online knapsack problem. This analogy is described in detail in the proof for Theorem 5. Again, researchers in algorithms have developed time efficient approximate solutions to the online knapsack problem [16]. Hence we use this solution and show that it forms an equilibrium. The following theorem characterizes an approximate equilibrium for online negotiation. Here the agents have to choose a strategy without knowing the features of the future issues. Because of this information incompleteness, the relevant equilibrium solution is that of a Bayes" Nash Equilibrium (BNE) in which each agent plays the best response to the other agents with respect to their expected utilities [18]. However, finding an agent"s BNE strategy is analogous to solving the online 0-1 knapsack problem. Also, the online knapsack can only be solved approximately [16]. Hence the relevant equilibrium solution concept is approximate BNE (see [26] for example). The following theorem finds this equilibrium using procedures ONLINE- TRADEOFFA and ONLINE-TRADEOFFB which are defined in the proof of the theorem. For a given time period, we let zm denote the approximately optimal solution generated by ONLINE-TRADEOFFA (or ONLINE-TRADEOFFB) and z∗ m the actual optimum. THEOREM 5. For the package deal procedure, the following strategies form an approximate Bayes" Nash equilibrium. The equilibrium strategy for t = n is: A(n) = j OFFER [1, 0] IF a"s TURN ACCEPT IF b"s TURN B(n) = j OFFER [0, 1] IF b"s TURN ACCEPT IF a"s TURN For all preceding time periods t < n, if [xt , yt ] denotes the offer made at time t, then the equilibrium strategies are defined as follows: A(t) = 8 < : OFFER ONLINE-TRADEOFFA(P, UB(t), t) IF a"s TURN If (Ua ([xt , yt ], t) ≥ UA(t)) ACCEPT else REJECT IF b"s TURN B(t) = 8 < : OFFER ONLINE-TRADEOFFB(P, UA(t), t) IF b"s TURN If (Ub ([xt , yt ], t) ≥ UB(t)) ACCEPT else REJECT IF a"s TURN where UA(t) = Ua ([¯at+1 ,¯bt+1 ], t + 1) and UB(t) = Ub ([¯at+1 , ¯bt+1 ], t + 1). An agreement on issue c takes place at t = c. For a given time period, the expected difference between the solution generated by the optimal strategy and that by the approximate strategy is E[z∗ m − zm] = O( √ m). 956 The Sixth Intl. Joint Conf. on Autonomous Agents and Multi-Agent Systems (AAMAS 07) PROOF. As in Theorem 1 we find the equilibrium offer for time period t = 1 using backward induction. Let a be the offering agent for t = 1 for all the m issues. Consider the last time period t = n (recall from Step 2 of the online protocol that n is the deadline for completing negotiation on the first issue). Since the first mover is the same for all the issues, and the agents make offers alternately, the offering agent for t = n is also the same for all the m issues. Assume that b is the offering agent for t = n. As in Section 3, the offering agent for t = n gets a hundred percent of all the m issues. Since b is the offering agent for t = n, his utility for this time period is: UB(n) = kb 1δn−1 1 + 1/2 mX i=2 δ i(n−1) i (6) Recall that ka i and kb i (for c < i ≤ m) are not known to the agents. Hence, the agents can only find their expected utilities from the future issues on the basis of the probability distribution functions for ka i and kb i . However, during the negotiation for issue c the agents know ka c but not kb c (see Step 1 of the online protocol). Hence, a computes UB(n) as follows. Agent b"s utility from issue c = 1 is kb 1δn−1 1 (which is the first term of Equation 6). Then, on the basis of the probability distribution functions for ka i and kb i , agent a computes b"s expected utility from each future issue i as δ i(n−1) i /2 (since ka i and kb i are uniformly distributed on [0, 1]). Thus, b"s expected cumulative utility from these m − c issues is 1/2 Pm i=2 δ i(n−1) i (which is the second term of Equation 6). Now, in order to decide what to offer for issue c = 1, the offering agent for t = n − 1 (i.e., agent a) must solve the following online knapsack problem: maximize Σm i=1ka i (1 − ¯bt i)δn−1 i (7) such that Σm i=1kb i ¯bt i ≥ UB(n) ¯bt i = 0 or 1 for 1 ≤ i ≤ m The only variables in the above maximization problem are ¯bt i. Now, maximizing Σm i=1ka i (1−¯bt i)δn−1 i is the same as minimizing Σm i=1ka i ¯bt i since δn−1 i and ka i are constants. Thus, we write Equation 7 as: minimize Σm i=1ka i ¯bt i (8) such that Σm i=1kb i ¯bt i ≥ UB(n) ¯bt i = 0 or 1 for 1 ≤ i ≤ m The above optimization problem is analogous to the online 0-1 knapsack problem. An algorithm to solve the online knapsack problem has already proposed in [16]. This algorithm is called the fixed-choice online algorithm. It has time complexity linear in the number of items (m) in the knapsack problem. We use this to solve our online negotiation problem. Thus, our ONLINE-TRADEOFFA algorithm is nothing but the fixed-choice online algorithm and therefore has the same time complexity as the latter. This algorithm takes the values of ka i and kb i one at a time and generates an approximate solution to the above knapsack problem. The expected difference between the optimum and approximate solution is E[z∗ m − zm] = O( √ m) [16] (see [16] for the detailed fixed-choice online algorithm and a proof for E[z∗ m − zm] = O( √ m)). The fixed-choice online algorithm of [16] is a generalization of the basic greedy algorithm for the offline knapsack problem; the idea behind it is as follows. A threshold value is determined on the basis of the information regarding weights and profits for the 0-1 knapsack problem. The method then includes into the knapsack all items whose profit density (profit density of an item is its profit per unit weight) exceeds the threshold until either the knapsack is filled or all the m items have been considered. In more detail, the algorithm ONLINE-TRADEOFFA works as follows. It first gets the values of ka 1 and kb 1 and finds ¯bt c. Since we have a 0-1 knapsack problem, ¯bt c can be either zero or one. Now, if ¯bt c = 1 for t = n, then ¯bt c must be one for 1 ≤ t < n (i.e., a must offer ¯bt c = 1 at t = 1). If ¯bt c = 1 for t = n, but a offers ¯bt c = 0 at t = 1, then agent b gets less utility than what it expects from a"s offer and rejects the proposal. Thus, if ¯bt c = 1 for t = n, then the optimal strategy for a is to offer ¯bt c = 1 at t = 1. Agent b accepts the offer. Thus, negotiation on the first issue starts at t = 1 and an agreement on it is also reached at t = 1. In the next time period (i.e., t = 2), negotiation proceeds to the next issue. The deadline for the second issue is n time periods from the start of negotiation on the issue. For c = 2, the algorithm ONLINE-TRADEOFFA is given the values of ka 2 and kb 2 and finds ¯bt c as described above. Agent offers bc at t = 2 and b accepts. Thus, negotiation on the second issue starts at t = 2 and an agreement on it is also reached at t = 2. This process repeats for the remaining issues c = 3, . . . , m. Thus, each issue is agreed upon in the same time period in which it starts. As negotiation for the next issue starts in the following time period (see step 3 of the online protocol), agreement on issue i occurs at time t = i. On the other hand, if b is the offering agent at t = 1, he uses the algorithm ONLINE-TRADEOFFB which is defined analogously. Thus, irrespective of who makes the first move, all the m issues are settled at time t = m. THEOREM 6. The time complexity of finding the approximate equilibrium offers of Theorem 5 is linear in m. PROOF. The time complexity of ONLINE-TRADEOFFA and ONLINETRADEOFFB is the same as the time complexity of the fixed-choice online algorithm of [16]. Since the latter has time complexity linear in m, the time complexity of ONLINE-TRADEOFFA and ONLINETRADEOFFB is also linear in m. It is worth noting that, for the 0-1 knapsack problem, the lower bound on the expected difference between the optimum and the solution found by any online algorithm is Ω(1) [16]. Thus, it follows that this lower bound also holds for our negotiation problem. 5. RELATED WORK Work on multi-issue negotiation can be divided into two main types: that for indivisible issues and that for divisible issues. We first describe the existing work for the case of divisible issues. Since Schelling [24] first noted that the outcome of negotiation depends on the choice of negotiation procedure, much research effort has been devoted to the study of different procedures for negotiating multiple issues. However, most of this work has focussed on the sequential procedure [7, 2]. For this procedure, a key issue is the negotiation agenda. Here the term agenda refers to the order in which the issues are negotiated. The agenda is important because each agent"s cumulative utility depends on the agenda; if we change the agenda then these utilities change. Hence, the agents must decide what agenda they will use. Now, the agenda can be decided before negotiating the issues (such an agenda is called exogenous) or it may be decided during the process of negotiation (such an agenda is called endogenous). For instance, Fershtman [7] analyze sequential negotiation with exogenous agenda. A number of researchers have also studied negotiations with an endogenous agenda [2]. In contrast to the above work that mainly deals with sequential negotiation, [6] studies the equilibrium for the package deal procedure. However, all the above mentioned work differs from ours in that we focus on indivisible issues while others focus on the case The Sixth Intl. Joint Conf. on Autonomous Agents and Multi-Agent Systems (AAMAS 07) 957 where each issue is divisible. Specifically, no previous work has determined an approximate equilibrium for multi-issue negotiation or for online negotiation. Existing work for the case of indivisible issues has mostly dealt with task allocation problems (for tasks that cannot be partioned) to a group of agents. The problem of task allocation has been previously studied in the context of coalitions involving more than two agents. For example [25] analyze the problem for the case where the agents act so as to maximize the benefit of the system as a whole. In contrast, our focus is on two agents where both of them are self-interested and want to maximize their individual utilities. On the other hand [22] focus on the use of contracts for task allocation to multiple self interested agents but this work concerns finding ways of decommitting contracts (after the initial allocation has been done) so as to improve an agent"s utility. In contrast, our focuses on negotiation regarding who will carry out which task. Finally, online and approximate mechanisms have been studied in the context of auctions [14, 9] but not for bilateral negotiations (which is the focus of our work). 6. CONCLUSIONS This paper has studied bilateral multi-issue negotiation between self-interested autonomous agents with time constraints. The issues are indivisible and different agents value different issues differently. Thus, the problem is for the agents to decide how to allocate the issues between themselves so as to maximize their individual utilities. Specifically, we first showed that finding the equilibrium offers is an NP-hard problem even in a complete information setting. We then presented approximately optimal negotiation strategies and showed that they form an equilibrium. These strategies have polynomial time complexity. We also analysed the difference between the true optimum and the approximate optimum. Finally, we extended the analysis to online negotiation where the issues become available at different time points and the agents are uncertain about the features of these issues. Specifically, we showed that an approximate equilibrium exists for online negotiation and analysed the approximation error. These approximate strategies also have polynomial time complexity. There are several interesting directions for future work. First, for online negotiation, we assumed that the constants ka c and kb c are both uniformly distributed. It will be interesting to analyze the case where ka c and kb c have other, possibly different, probability distributions. Apart from this, we treated the number of issues as being common knowledge to the agents. In future, it will be interesting to treat the number of issues as uncertain. 7. REFERENCES [1] G. Ausiello, P. Crescenzi, G. Gambosi, V. Kann, A. Marchetti-Spaccamela, and M. Protasi. Complexity and approximation: Combinatorial optimization problems and their approximability properties. Springer, 2003. [2] M. Bac and H. Raff. Issue-by-issue negotiations: the role of information and time preference. Games and Economic Behavior, 13:125-134, 1996. [3] A. Borodin and R. El-Yaniv. Online Computation and Competitive Analysis. Cambridge University Press, 1998. [4] S. J. Brams. Fair division: from cake cutting to dispute resolution. Cambridge University Press, 1996. [5] L. A. Busch and I. J. Horstman. Bargaining frictions, bargaining procedures and implied costs in multiple-issue bargaining. Economica, 64:669-680, 1997. [6] S. S. Fatima, M. Wooldridge, and N. R. Jennings. Multi-issue negotiation with deadlines. Journal of Artificial Intelligence Research, 27:381-417, 2006. [7] C. Fershtman. The importance of the agenda in bargaining. Games and Economic Behavior, 2:224-238, 1990. [8] F. Glover. A multiphase dual algorithm for the zero-one integer programming problem. Operations Research, 13:879-919, 1965. [9] M. T. Hajiaghayi, R. Kleinberg, and D. C. Parkes. Adaptive limited-supply online auctions. In ACM Conference on Electronic Commerce (ACMEC-04), pages 71-80, New York, 2004. [10] O. H. Ibarra and C. E. Kim. Fast approximation algorithms for the knapsack and sum of subset problems. Journal of ACM, 22:463-468, 1975. [11] R. Inderst. Multi-issue bargaining with endogenous agenda. Games and Economic Behavior, 30:64-82, 2000. [12] R. Keeney and H. Raiffa. Decisions with Multiple Objectives: Preferences and Value Trade-offs. New York: John Wiley, 1976. [13] S. Kraus. Strategic negotiation in multi-agent environments. The MIT Press, Cambridge, Massachusetts, 2001. [14] D. Lehman, L. I. O"Callaghan, and Y. Shoham. Truth revelation in approximately efficient combinatorial auctions. Journal of the ACM, 49(5):577-602, 2002. [15] A. Lomuscio, M. Wooldridge, and N. R. Jennings. A classification scheme for negotiation in electronic commerce. International Journal of Group Decision and Negotiation, 12(1):31-56, 2003. [16] A. Marchetti-Spaccamela and C. Vercellis. Stochastic online knapsack problems. Mathematical Programming, 68:73-104, 1995. [17] S. Martello and P. Toth. Knapsack problems: Algorithms and computer implementations. John Wiley and Sons, 1990. [18] M. J. Osborne and A. Rubinstein. A Course in Game Theory. The MIT Press, 1994. [19] H. Raiffa. The Art and Science of Negotiation. Harvard University Press, Cambridge, USA, 1982. [20] J. S. Rosenschein and G. Zlotkin. Rules of Encounter. MIT Press, 1994. [21] A. Rubinstein. Perfect equilibrium in a bargaining model. Econometrica, 50(1):97-109, January 1982. [22] T. Sandholm and V. Lesser. Levelled commitment contracts and strategic breach. Games and Economic Behavior: Special Issue on AI and Economics, 35:212-270, 2001. [23] T. Sandholm and N. Vulkan. Bargaining with deadlines. In AAAI-99, pages 44-51, Orlando, FL, 1999. [24] T. C. Schelling. An essay on bargaining. American Economic Review, 46:281-306, 1956. [25] O. Shehory and S. Kraus. Methods for task allocation via agent coalition formation. Artificial Intelligence Journal, 101(1-2):165-200, 1998. [26] S. Singh, V. Soni, and M. Wellman. Computing approximate Bayes Nash equilibria in tree games of incomplete information. In Proceedings of the ACM Conference on Electronic Commerce ACM-EC, pages 81-90, New York, May 2004. [27] I. Stahl. Bargaining Theory. Economics Research Institute, Stockholm School of Economics, Stockholm, 1972. 958 The Sixth Intl. Joint Conf. on Autonomous Agents and Multi-Agent Systems (AAMAS 07)

gain from cooperation;disputing agent;protocol;approximation;relative error;online computation;indivisible issue;game-theory;equilibrium;negotiation;strategy;key form of interaction;multiagent system;time constraint;interaction key form

train_I-55

Searching for Joint Gains in Automated Negotiations Based on Multi-criteria Decision Making Theory

It is well established by conflict theorists and others that successful negotiation should incorporate creating value as well as claiming value. Joint improvements that bring benefits to all parties can be realised by (i) identifying attributes that are not of direct conflict between the parties, (ii) tradeoffs on attributes that are valued differently by different parties, and (iii) searching for values within attributes that could bring more gains to one party while not incurring too much loss on the other party. In this paper we propose an approach for maximising joint gains in automated negotiations by formulating the negotiation problem as a multi-criteria decision making problem and taking advantage of several optimisation techniques introduced by operations researchers and conflict theorists. We use a mediator to protect the negotiating parties from unnecessary disclosure of information to their opponent, while also allowing an objective calculation of maximum joint gains. We separate out attributes that take a finite set of values (simple attributes) from those with continuous values, and we show that for simple attributes, the mediator can determine the Pareto-optimal values. In addition we show that if none of the simple attributes strongly dominates the other simple attributes, then truth telling is an equilibrium strategy for negotiators during the optimisation of simple attributes. We also describe an approach for improving joint gains on non-simple attributes, by moving the parties in a series of steps, towards the Pareto-optimal frontier.

1. INTRODUCTION Given that negotiation is perhaps one of the oldest activities in the history of human communication, it"s perhaps surprising that conducted experiments on negotiations have shown that negotiators more often than not reach inefficient compromises [1, 21]. Raiffa [17] and Sebenius [20] provide analyses on the negotiators" failure to achieve efficient agreements in practice and their unwillingness to disclose private information due to strategic reasons. According to conflict theorists Lax and Sebenius [13], most negotiation actually involves both integrative and distributive bargaining which they refer to as creating value and claiming value. They argue that negotiation necessarily includes both cooperative and competitive elements, and that these elements exist in tension. Negotiators face a dilemma in deciding whether to pursue a cooperative or a competitive strategy at a particular time during a negotiation. They refer to this problem as the Negotiator"s Dilemma. We argue that the Negotiator"s Dilemma is essentially informationbased, due to the private information held by the agents. Such private information contains both the information that implies the agent"s bottom lines (or, her walk-away positions) and the information that enforces her bargaining strength. For instance, when bargaining to sell a house to a potential buyer, the seller would try to hide her actual reserve price as much as possible for she hopes to reach an agreement at a much higher price than her reserve price. On the other hand, the outside options available to her (e.g. other buyers who have expressed genuine interest with fairly good offers) consist in the information that improves her bargaining strength about which she would like to convey to her opponent. But at the same time, her opponent is well aware of the fact that it is her incentive to boost her bargaining strength and thus will not accept every information she sends out unless it is substantiated by evidence. Coming back to the Negotiator"s Dilemma, it"s not always possible to separate the integrative bargaining process from the distributive bargaining process. In fact, more often than not, the two processes interplay with each other making information manipulation become part of the integrative bargaining process. This is because a negotiator could use the information about his opponent"s interests against her during the distributive negotiation process. That is, a negotiator may refuse to concede on an important conflicting issue by claiming that he has made a major concession (on another issue) to meet his opponent"s interests even though the concession he made could be insignificant to him. For instance, few buyers would start a bargaining with a dealer over a deal for a notebook computer by declaring that he is most interested in an extended warranty for the item and therefore prepared to pay a high price to get such an extended warranty. Negotiation Support Systems (NSSs) and negotiating software 508 978-81-904262-7-5 (RPS) c 2007 IFAAMAS agents (NSAs) have been introduced either to assist humans in making decisions or to enable automated negotiation to allow computer processes to engage in meaningful negotiation to reach agreements (see, for instance, [14, 15, 19, 6, 5]). However, because of the Negotiator"s Dilemma and given even bargaining power and incomplete information, the following two undesirable situations often arise: (i) negotiators reach inefficient compromises, or (ii) negotiators engage in a deadlock situation in which both negotiators refuse to act upon with incomplete information and at the same time do not want to disclose more information. In this paper, we argue for the role of a mediator to resolve the above two issues. The mediator thus plays two roles in a negotiation: (i) to encourage cooperative behaviour among the negotiators, and (ii) to absorb the information disclosure by the negotiators to prevent negotiators from using uncertainty and private information as a strategic device. To take advantage of existing results in negotiation analysis and operations research (OR) literatures [18], we employ multi-criteria decision making (MCDM) theory to allow the negotiation problem to be represented and analysed. Section 2 provides background on MCDM theory and the negotiation framework. Section 3 formulates the problem. In Section 4, we discuss our approach to integrative negotiation. Section 5 discusses the future work with some concluding remarks. 2. BACKGROUND 2.1 Multi-criteria decision making theory Let A denote the set of feasible alternatives available to a decision maker M. As an act, or decision, a in A may involve multiple aspects, we usually describe the alternatives a with a set of attributes j; (j = 1, . . . , m). (Attributes are also referred to as issues, or decision variables.) A typical decision maker also has several objectives X1, . . . , Xk. We assume that Xi, (i = 1, . . . , k), maps the alternatives to real numbers. Thus, a tuple (x1, . . . , xk) = (X1(a), . . . , Xk(a)) denotes the consequence of the act a to the decision maker M. By definition, objectives are statements that delineate the desires of a decision maker. Thus, M wishes to maximise his objectives. However, as discussed thoroughly by Keeney and Raiffa [8], it is quite likely that a decision maker"s objectives will conflict with each other in that the improved achievement with one objective can only be accomplished at the expense of another. For instance, most businesses and public services have objectives like minimise cost and maximise the quality of services. Since better services can often only be attained for a price, these objectives conflict. Due to the conflicting nature of a decision maker"s objectives, M usually has to settle at a compromise solution. That is, he may have to choose an act a ∈ A that does not optimise every objective. This is the topic of the multi-criteria decision making theory. Part of the solution to this problem is that M has to try to identify the Pareto frontier in the consequence space {(X1(a), . . . , Xk(a))}a∈A. DEFINITION 1. (Dominant) Let x = (x1, . . . , xk) and x = (x1, . . . , xk) be two consequences. x dominates x iff xi > xi for all i, and the inequality is strict for at least one i. The Pareto frontier in a consequence space then consists of all consequences that are not dominated by any other consequence. This is illustrated in Fig. 1 in which an alternative consists of two attributes d1 and d2 and the decision maker tries to maximise the two objectives X1 and X2. A decision a ∈ A whose consequence does not lie on the Pareto frontier is inefficient. While the Pareto 1x d2 a (X (a),X (a)) d1 1 x2 2 Alternative spaceA Pareto frontier Consequence space optimal consequenc Figure 1: The Pareto frontier frontier allows M to avoid taking inefficient decisions, M still has to decide which of the efficient consequences on the Pareto frontier is most preferred by him. MCDM theorists introduce a mechanism to allow the objective components of consequences to be normalised to the payoff valuations for the objectives. Consequences can then be ordered: if the gains in satisfaction brought about by C1 (in comparison to C2) equals to the losses in satisfaction brought about by C1 (in comparison to C2), then the two consequences C1 and C2 are considered indifferent. M can now construct the set of indifference curves1 in the consequence space (the dashed curves in Fig. 1). The most preferred indifference curve that intersects with the Pareto frontier is in focus: its intersection with the Pareto frontier is the sought after consequence (i.e., the optimal consequence in Fig. 1). 2.2 A negotiation framework A multi-agent negotiation framework consists of: 1. A set of two negotiating agents N = {1, 2}. 2. A set of attributes Att = {α1, . . . , αm} characterising the issues the agents are negotiating over. Each attribute α can take a value from the set V alα; 3. A set of alternative outcomes O. An outcome o ∈ O is represented by an assignment of values to the corresponding attributes in Att. 4. Agents" utility: Based on the theory of multiple-criteria decision making [8], we define the agents" utility as follows: • Objectives: Agent i has a set of ni objectives, or interests; denoted by j (j = 1, . . . , ni). To measure how much an outcome o fulfills an objective j to an agent i, we use objective functions: for each agent i, we define i"s interests using the objective vector function fi = [fij ] : O → Rni . • Value functions: Instead of directly evaluating an outcome o, agent i looks at how much his objectives are fulfilled and will make a valuation based on these more basic criteria. Thus, for each agent i, there is a value function σi : Rni → R. In particular, Raiffa [17] shows how to systematically construct an additive value function to each party involved in a negotiation. • Utility: Now, given an outcome o ∈ O, an agent i is able to determine its value, i.e., σi(fi(o)). However, a negotiation infrastructure is usually required to facilitate negotiation. This might involve other mechanisms and factors/parties, e.g., a mediator, a legal institution, participation fees, etc. The standard way to implement such a thing is to allow money 1 In fact, given the k-dimensional space, these should be called indifference surfaces. However, we will not bog down to that level of details. The Sixth Intl. Joint Conf. on Autonomous Agents and Multi-Agent Systems (AAMAS 07) 509 and side-payments. In this paper, we ignore those side-effects and assume that agent i"s utility function ui is normalised so that ui : O → [0, 1]. EXAMPLE 1. There are two agents, A and B. Agent A has a task T that needs to be done and also 100 units of a resource R. Agent B has the capacity to perform task T and would like to obtain at least 10 and at most 20 units of the resource R. Agent B is indifferent on any amount between 10 and 20 units of the resource R. The objective functions for both agents A and B are cost and revenue. And they both aim at minimising costs while maximising revenues. Having T done generates for A a revenue rA,T while doing T incurs a cost cB,T to B. Agent B obtains a revenue rB,R for each unit of the resource R while providing each unit of the resource R costs agent A cA,R. Assuming that money transfer between agents is possible, the set Att then contains three attributes: • T, taking values from the set {0, 1}, indicates whether the task T is assigned to agent B; • R, taking values from the set of non-negative integer, indicates the amount of resource R being allocated to agent B; and • MT, taking values from R, indicates the payment p to be transferred from A to B. Consider the outcome o = [T = 1, R = k, MT = p], i.e., the task T is assigned to B, and A allocates to B with k units of the resource R, and A transfers p dollars to B. Then, costA(o) = k.cA,R + p and revA(o) = rA,T ; and costB(o) = cB,T and revA(o) = j k.rB,R + p if 10 ≤ k ≤ 20 p otherwise. And, σi(costi(o), revi(o)) = revi(o) − costi(o), (i = A, B). 3. PROBLEM FORMALISATION Consider Example 1, assume that rA,T = $150 and cB,T = $100 and rB,R = $10 and cA,R = $7. That is, the revenues generated for A exceeds the costs incurred to B to do task T, and B values resource R more highly than the cost for A to provide it. The optimal solution to this problem scenario is to assign task T to agent B and to allocate 20 units of resource R (i.e., the maximal amount of resource R required by agent B) from agent A to agent B. This outcome regarding the resource and task allocation problems leaves payoffs of $10 to agent A and $100 to agent B.2 Any other outcome would leave at least one of the agents worse off. In other words, the presented outcome is Pareto-efficient and should be part of the solution outcome for this problem scenario. However, as the agents still have to bargain over the amount of money transfer p, neither agent would be willing to disclose their respective costs and revenues regarding the task T and the resource R. As a consequence, agents often do not achieve the optimal outcome presented above in practice. To address this issue, we introduce a mediator to help the agents discover better agreements than the ones they might try to settle on. Note that this problem is essentially the problem of searching for joint gains in a multilateral negotiation in which the involved parties hold strategic information, i.e., the integrative part in a negotiation. In order to help facilitate this process, we introduce the role of a neutral mediator. Before formalising the decision problems faced by the mediator and the 2 Certainly, without money transfer to compensate agent A, this outcome is not a fair one. negotiating agents, we discuss the properties of the solution outcomes to be achieved by the mediator. In a negotiation setting, the two typical design goals would be: • Efficiency: Avoid the agents from settling on an outcome that is not Pareto-optimal; and • Fairness: Avoid agreements that give the most of the gains to a subset of agents while leaving the rest with too little. The above goals are axiomatised in Nash"s seminal work [16] on cooperative negotiation games. Essentially, Nash advocates for the following properties to be satisfied by solution to the bilateral negotiation problem: (i) it produces only Pareto-optimal outcomes; (ii) it is invariant to affine transformation (to the consequence space); (iii) it is symmetric; and (iv) it is independent from irrelevant alternatives. A solution satisfying Nash"s axioms is called a Nash bargaining solution. It then turns out that, by taking the negotiators" utilities as its objectives the mediator itself faces a multi-criteria decision making problem. The issues faced by the mediator are: (i) the mediator requires access to the negotiators" utility functions, and (ii) making (fair) tradeoffs between different agents" utilities. Our methods allow the agents to repeatedly interact with the mediator so that a Nash solution outcome could be found by the parties. Informally, the problem faced by both the mediator and the negotiators is construction of the indifference curves. Why are the indifference curves so important? • To the negotiators, knowing the options available along indifference curves opens up opportunities to reach more efficient outcomes. For instance, consider an agent A who is presenting his opponent with an offer θA which she refuses to accept. Rather than having to concede, A could look at his indifference curve going through θA and choose another proposal θA. To him, θA and θA are indifferent but θA could give some gains to B and thus will be more acceptable to B. In other words, the outcome θA is more efficient than θA to these two negotiators. • To the mediator, constructing indifference curves requires a measure of fairness between the negotiators. The mediator needs to determine how much utility it needs to take away from the other negotiators to give a particular negotiator a specific gain G (in utility). In order to search for integrative solutions within the outcome space O, we characterise the relationship between the agents over the set of attributes Att. As the agents hold different objectives and have different capacities, it may be the case that changing between two values of a specific attribute implies different shifts in utility of the agents. However, the problem of finding the exact Paretooptimal set3 is NP-hard [2]. Our approach is thus to solve this optimisation problem in two steps. In the first steps, the more manageable attributes will be solved. These are attributes that take a finite set of values. The result of this step would be a subset of outcomes that contains the Pareto-optimal set. In the second step, we employ an iterative procedure that allows the mediator to interact with the negotiators to find joint improvements that move towards a Pareto-optimal outcome. This approach will not work unless the attributes from Att 3 The Pareto-optimal set is the set of outcomes whose consequences (in the consequence space) correspond to the Pareto frontier. 510 The Sixth Intl. Joint Conf. on Autonomous Agents and Multi-Agent Systems (AAMAS 07) are independent. Most works on multi-attribute, or multi-issue, negotiation (e.g. [17]) assume that the attributes or the issues are independent, resulting in an additive value function for each agent.4 ASSUMPTION 1. Let i ∈ N and S ⊆ Att. Denote by ¯S the set Att \ S. Assume that vS and vS are two assignments of values to the attributes of S and v1 ¯S, v2 ¯S are two arbitrary value assignments to the attributes of ¯S, then (ui([vS, v1 ¯S]) − ui([vS, v2 ¯S])) = (ui([vS, v1 ¯S])−ui([vS, v2 ¯S])). That is, the utility function of agent i will be defined on the attributes from S independently of any value assignment to other attributes. 4. MEDIATOR-BASED BILATERAL NEGOTIATIONS As discussed by Lax and Sebenius [13], under incomplete information the tension between creating and claiming values is the primary cause of inefficient outcomes. This can be seen most easily in negotiations involving two negotiators; during the distributive phase of the negotiation, the two negotiators"s objectives are directly opposing each other. We will now formally characterise this relationship between negotiators by defining the opposition between two negotiating parties. The following exposition will be mainly reproduced from [9]. Assuming for the moment that all attributes from Att take values from the set of real numbers R, i.e., V alj ⊆ R for all j ∈ Att. We further assume that the set O = ×j∈AttV alj of feasible outcomes is defined by constraints that all parties must obey and O is convex. Now, an outcome o ∈ O is just a point in the m-dimensional space of real numbers. Then, the questions are: (i) from the point of view of an agent i, is o already the best outcome for i? (ii) if o is not the best outcome for i then is there another outcome o such that o gives i a better utility than o and o does not cause a utility loss to the other agent j in comparison to o? The above questions can be answered by looking at the directions of improvement of the negotiating parties at o, i.e., the directions in the outcome space O into which their utilities increase at point o. Under the assumption that the parties" utility functions ui are differentiable concave, the set of all directions of improvement for a party at a point o can be defined in terms of his most preferred, or gradient, direction at that point. When the gradient direction ∇ui(o) of agent i at point o is outright opposing to the gradient direction ∇uj (o) of agent j at point o then the two parties strongly disagree at o and no joint improvements can be achieved for i and j in the locality surrounding o. Since opposition between the two parties can vary considerably over the outcome space (with one pair of outcomes considered highly antagonistic and another pair being highly cooperative), we need to describe the local properties of the relationship. We begin with the opposition at any point of the outcome space Rm . The following definition is reproduced from [9]: DEFINITION 2. 1. The parties are in local strict opposition at a point x ∈ Rm iff for all points x ∈ Rm that are sufficiently close to x (i.e., for some > 0 such that ∀x x −x < ), an increase of one utility can be achieved only at the expense of a decrease of the other utility. 2. The parties are in local non-strict opposition at a point x ∈ Rm iff they are not in local strict opposition at x, i.e., iff it is possible for both parties to raise their utilities by moving an infinitesimal distance from x. 4 Klein et al. [10] explore several implications of complex contracts in which attributes are possibly inter-dependent. 3. The parties are in local weak opposition at a point x ∈ Rm iff ∇u1(x).∇u2(x) ≥ 0, i.e., iff the gradients at x of the two utility functions form an acute or right angle. 4. The parties are in local strong opposition at a point x ∈ Rm iff ∇u1(x).∇u2(x) < 0, i.e., iff the gradients at x form an obtuse angle. 5. The parties are in global strict (nonstrict, weak, strong) opposition iff for every x ∈ Rm they are in local strict (nonstrict, weak, strong) opposition. Global strict and nonstrict oppositions are complementary cases. Essentially, under global strict opposition the whole outcome space O becomes the Pareto-optimal set as at no point in O can the negotiating parties make a joint improvement, i.e., every point in O is a Pareto-efficient outcome. In other words, under global strict opposition the outcome space O can be flattened out into a single line such that for each pair of outcomes x, y ∈ O, u1(x) < u1(y) iff u2(x) > u2(y), i.e., at every point in O, the gradient of the two utility functions point to two different ends of the line. Intuitively, global strict opposition implies that there is no way to obtain joint improvements for both agents. As a consequence, the negotiation degenerates to a distributive negotiation, i.e., the negotiating parties should try to claim as much shares from the negotiation issues as possible while the mediator should aim for the fairness of the division. On the other hand, global nonstrict opposition allows room for joint improvements and all parties might be better off trying to realise the potential gains by reaching Pareto-efficient agreements. Weak and strong oppositions indicate different levels of opposition. The weaker the opposition, the more potential gains can be realised making cooperation the better strategy to employ during negotiation. On the other hand, stronger opposition suggests that the negotiating parties tend to behave strategically leading to misrepresentation of their respective objectives and utility functions and making joint gains more difficult to realise. We have been temporarily making the assumption that the outcome space O is the subset of Rm . In many real-world negotiations, this assumption would be too restrictive. We will continue our exposition by lifting this restriction and allowing discrete attributes. However, as most negotiations involve only discrete issues with a bounded number of options, we will assume that each attribute takes values either from a finite set or from the set of real numbers R. In the rest of the paper, we will refer to attributes whose values are from finite sets as simple attributes and attributes whose values are from R as continuous attributes. The notions of local oppositions, i.e., strict, nonstrict, weak and strong, are not applicable to outcome spaces that contain simple attributes and nor are the notions of global weak and strong oppositions. However, the notions of global strict and nonstrict oppositions can be generalised for outcome spaces that contain simple attributes. DEFINITION 3. Given an outcome space O, the parties are in global strict opposition iff ∀x, y ∈ O, u1(x) < u1(y) iff u2(x) > u2(y). The parties are in global nonstrict opposition if they are not in global strict opposition. 4.1 Optimisation on simple attributes In order to extract the optimal values for a subset of attributes, in the first step of this optimisation process the mediator requests the negotiators to submit their respective utility functions over the set of simple attributes. Let Simp ⊆ Att denote the set of all simple attributes from Att. Note that, due to Assumption 1, agent i"s The Sixth Intl. Joint Conf. on Autonomous Agents and Multi-Agent Systems (AAMAS 07) 511 utility function can be characterised as follows: ui([vSimp, vSimp]) = wi 1 ∗ ui,1([vSimp]) + wi 2 ∗ ui,2([vSimp]), where Simp = Att \ Simp, and ui,1 and ui,2 are the utility components of ui over the sets of attributes Simp and Simp, respectively, and 0 < wi 1, wi 2 < 1 and wi 1 + wi 2 = 1. As attributes are independent of each other regarding the agents" utility functions, the optimisation problem over the attributes from Simp can be carried out by fixing ui([vSimp]) to a constant C, and then search for the optimal values within the set of attributes Simp. Now, how does the mediator determine the optimal values for the attributes in Simp? Several well-known optimisation strategies could be applicable here: • The utilitarian solution: The sum of the agents" utilities are maximised. Thus, the optimal values are the solution of the following optimisation problem: arg max v∈V alSimp X i∈N ui(v) • The Nash solution: The product of the agents" utilities are maximised. Thus, the optimal values are the solution of the following optimisation problem: arg max v∈V alSimp Y i∈N ui(v) • The egalitarian solution (aka. the maximin solution): The utility of the agent with minimum utility is maximised. Thus, the optimal values are the solution of the following optimisation problem: arg max v∈V alSimp min i∈N ui(v) The question now is of course whether a negotiator has the incentive to misrepresent his utility function. First of all, recall that the agents" utility functions are bounded, i.e., ∀o ∈ O.0 ≤ ui(o) ≤ 1. Thus, the agents have no incentive to overstate their utility regarding an outcome o: If o is the most preferred outcome to an agent i then he already assigns the maximal utility to o. On the other hand, if o is not the most preferred outcome to i then by overstating the utility he assigns to o, the agent i runs the risk of having to settle on an agreement which would give him less payoffs than he is supposed to receive. However, agents do have an incentive to understate their utility if the final settlement will be based on the above solutions alone. Essentially, the mechanism to avoid an agent to understate his utility regarding particular outcomes is to guarantee a certain measure of fairness for the final settlement. That is, the agents lose the incentive to be dishonest to obtain gains from taking advantage of the known solutions to determine the settlement outcome for they would be offset by the fairness maintenance mechanism. Firsts, we state an easy lemma. LEMMA 1. When Simp contains one single attributes, the agents have the incentive to understate their utility functions regarding outcomes that are not attractive to them. By way of illustration, consider the set Simp containing only one attribute that could take values from the finite set {A, B, C, D}. Assume that negotiator 1 assigns utilities of 0.4, 0.7, 0.9, and 1 to A, B, C, and D, respectively. Assume also that negotiator 2 assigns utilities of 1, 0.9, 0.7, and 0.4 to A, B, C, and D, respectively. If agent 1 misrepresents his utility function to the mediator by reporting utility 0 for all values A, B and C and utility 1 for value D then the agent 2 who plays honestly in his report to the mediator will obtain the worst outcome D given any of the above solutions. Note that agent 1 doesn"t need to know agent 2"s utility function, nor does he need to know the strategy employed by agent 2. As long as he knows that the mediator is going to employ one of the above three solutions, then the above misrepresentation is the dominant strategy for this game. However, when the set Simp contains more than one attribute and none of the attributes strongly dominate the other attributes then the above problem disminishes by itself thanks to the integrative solution. We of course have to define clearly what it means for an attribute to strongly dominate other attributes. Intuitively, if most of an agent"s utility concentrates on one of the attributes then this attribute strongly dominates other attributes. We again appeal to the Assumption 1 on additivity of utility functions to achieve a measure of fairness within this negotiation setting. Due to Assumption 1, we can characterise agent i"s utility component over the set of attributes Simp by the following equation: ui,1([vSimp]) = X j∈Simp wi j ∗ ui,j([vj]) (1) where P j∈Simp wj = 1. Then, an attribute ∈ Simp strongly dominates the rest of the attributes in Simp (for agent i) iff wi > P j∈(Simp− ) wi j . Attribute is said to be strongly dominant (for agent i) wrt. the set of simple attributes Simp. The following theorem shows that if the set of attributes Simp does not contain a strongly dominant attribute then the negotiators have no incentive to be dishonest. THEOREM 1. Given a negotiation framework, if for every agent the set of simple attributes doesn"t contain a strongly dominant attribute, then truth-telling is an equilibrium strategy for the negotiators during the optimisation of simple attributes. So far, we have been concentrating on the efficiency issue while leaving the fairness issue aside. A fair framework does not only support a more satisfactory distribution of utility among the agents, but also often a good measure to prevent misrepresentation of private information by the agents. Of the three solutions presented above, the utilitarian solution does not support fairness. On the other hand, Nash [16] proves that the Nash solution satisfies the above four axioms for the cooperative bargaining games and is considered a fair solution. The egalitarian solution is another mechanism to achieve fairness by essentially helping the worst off. The problem with these solutions, as discussed earlier, is that they are vulnerable to strategic behaviours when one of the attributes strongly dominates the rest of attributes. However, there is yet another solution that aims to guarantee fairness, the minimax solution. That is, the utility of the agent with maximum utility is minimised. It"s obvious that the minimax solution produces inefficient outcomes. However, to get around this problem (given that the Pareto-optimal set can be tractably computed), we can apply this solution over the Pareto-optimal set only. Let POSet ⊆ V alSimp be the Pareto-optimal subset of the simple outcomes, the minimax solution is defined to be the solution of the following optimisation problem. arg min v∈P OSet max i∈N ui(v) While overall efficiency often suffers under a minimax solution, i.e., the sum of all agents" utilities are often lower than under other solutions, it can be shown that the minimax solution is less vulnerable to manipulation. 512 The Sixth Intl. Joint Conf. on Autonomous Agents and Multi-Agent Systems (AAMAS 07) THEOREM 2. Given a negotiation framework, under the minimax solution, if the negotiators are uncertain about their opponents" preferences then truth-telling is an equilibrium strategy for the negotiators during the optimisation of simple attributes. That is, even when there is only one single simple attribute, if an agent is uncertain whether the other agent"s most preferred resolution is also his own most preferred resolution then he should opt for truth-telling as the optimal strategy. 4.2 Optimisation on continuous attributes When the attributes take values from infinite sets, we assume that they are continuous. This is similar to the common practice in operations research in which linear programming solutions/techniques are applied to integer programming problems. We denote the number of continuous attributes by k, i.e., Att = Simp ∪ Simp and |Simp| = k. Then, the outcome space O can be represented as follows: O = ( Q j∈Simp V alj) × ( Q l∈Simp V all), where Q l∈Simp V all ⊆ Rk is the continuous component of O. Let Oc denote the set Q l∈Simp V all. We"ll refer to Oc as the feasible set and assume that Oc is closed and convex. After carrying out the optimisation over the set of simple attributes, we are able to assign the optimal values to the simple attributes from Simp. Thus, we reduce the original problem to the problem of searching for optimal (and fair) outcomes within the feasible set Oc . Recall that, by Assumption 1, we can characterise agent i"s utility function as follows: ui([v∗ Simp, vSimp]) = C + wi 2 ∗ ui,2([vSimp]), where C is the constant wi 1 ∗ ui,1([v∗ Simp]) and v∗ Simp denotes the optimal values of the simple attributes in Simp. Hence, without loss of generality (albeit with a blatant abuse of notation), we can take the agent i"s utility function as ui : Rk → [0, 1]. Accordingly we will also take the set of outcomes under consideration by the agents to be the feasible set Oc . We now state another assumption to be used in this section: ASSUMPTION 2. The negotiators" utility functions can be described by continuously differentiable and concave functions ui : Rk → [0, 1], (i = 1, 2). It should be emphasised that we do not assume that agents explicitly know their utility functions. For the method to be described in the following to work, we only assume that the agents know the relevant information, e.g. at certain point within the feasible set Oc , the gradient direction of their own utility functions and some section of their respective indifference curves. Assume that a tentative agreement (which is a point x ∈ Rk ) is currently on the table, the process for the agents to jointly improve this agreement in order to reach a Pareto-optimal agreement can be described as follows. The mediator asks the negotiators to discretely submit their respective gradient directions at x, i.e., ∇u1(x) and ∇u2(x). Note that the goal of the process to be described here is to search for agreements that are more efficient than the tentative agreement currently on the table. That is, we are searching for points x within the feasible set Oc such that moving to x from the current tentative agreement x brings more gains to at least one of the agents while not hurting any of the agents. Due to the assumption made above, i.e. the feasible set Oc is bounded, the conditions for an alternative x ∈ Oc to be efficient vary depending on the position of x. The following results are proved in [9]: Let B(x) = 0 denote the equation of the boundary of Oc , defining x ∈ Oc iff B(x) ≥ 0. An alternative x∗ ∈ Oc is efficient iff, either A. x∗ is in the interior of Oc and the parties are in local strict opposition at x∗ , i.e., ∇u1(x∗ ) = −γ∇u2(x∗ ) (2) where γ > 0; or B. x∗ is on the boundary of Oc , and for some α, β ≥ 0: α∇u1(x∗ ) + β∇u2(x∗ ) = ∇B(x∗ ) (3) We are now interested in answering the following questions: (i) What is the initial tentative agreement x0? (ii) How to find the more efficient agreement xh+1, given the current tentative agreement xh? 4.2.1 Determining a fair initial tentative agreement It should be emphasised that the choice of the initial tentative agreement affects the fairness of the final agreement to be reached by the presented method. For instance, if the initial tentative agreement x0 is chosen to be the most preferred alternative to one of the agents then it is also a Pareto-optimal outcome, making it impossible to find any joint improvement from x0. However, if x0 will then be chosen to be the final settlement and if x0 turns out to be the worst alternative to the other agent then this outcome is a very unfair one. Thus, it"s important that the choice of the initial tentative agreement be sensibly made. Ehtamo et al [3] present several methods to choose the initial tentative agreement (called reference point in their paper). However, their goal is to approximate the Pareto-optimal set by systematically choosing a set of reference points. Once an (approximate) Pareto-optimal set is generated, it is left to the negotiators to decide which of the generated Pareto-optimal outcomes to be chosen as the final settlement. That is, distributive negotiation will then be required to settle the issue. We, on the other hand, are interested in a fair initial tentative agreement which is not necessarily efficient. Improving a given tentative agreement to yield a Pareto-optimal agreement is considered in the next section. For each attribute j ∈ Simp, an agent i will be asked to discretely submit three values (from the set V alj): the most preferred value, denoted by pvi,j, the least preferred value, denoted by wvi,j, and a value that gives i an approximately average payoff, denoted by avi,j. (Note that this is possible because the set V alj is bounded.) If pv1,j and pv2,j are sufficiently close, i.e., |pv1,j − pv2,j| < Δ for some pre-defined Δ > 0, then pv1,j and pv2,j are chosen to be the two core values, denoted by cv1 and cv2. Otherwise, between the two values pv1,j and av1,j, we eliminate the one that is closer to wv2,j, the remaining value is denoted by cv1. Similarly, we obtain cv2 from the two values pv2,j and av2,j. If cv1 = cv2 then cv1 is selected as the initial value for the attribute j as part of the initial tentative agreement. Otherwise, without loss of generality, we assume that cv1 < cv2. The mediator selects randomly p values mv1, . . . , mvp from the open interval (cv1, cv2), where p ≥ 1. The mediator then asks the agents to submit their valuations over the set of values {cv1, cv2, mv1, . . . , mvp}. The value whose the two valuations of two agents are closest is selected as the initial value for the attribute j as part of the initial tentative agreement. The above procedure guarantees that the agents do not gain by behaving strategically. By performing the above procedure on every attribute j ∈ Simp, we are able to identify the initial tentative agreement x0 such that x0 ∈ Oc . The next step is to compute a new tentative agreement from an existing tentative agreement so that the new one would be more efficient than the existing one. The Sixth Intl. Joint Conf. on Autonomous Agents and Multi-Agent Systems (AAMAS 07) 513 4.2.2 Computing new tentative agreement Our procedure is a combination of the method of jointly improving direction introduced by Ehtamo et al [4] and a method we propose in the coming section. Basically, the idea is to see how strong the opposition the parties are in. If the two parties are in (local) weak opposition at the current tentative agreement xh, i.e., their improving directions at xh are close to each other, then the compromise direction proposed by Ehtamo et al [4] is likely to point to a better agreement for both agents. However, if the two parties are in local strong opposition at the current point xh then it"s unclear whether the compromise direction would really not hurt one of the agents whilst bringing some benefit to the other. We will first review the method proposed by Ehtamo et al [4] to compute the compromise direction for a group of negotiators at a given point x ∈ Oc . Ehtamo et al define a a function T(x) that describes the mediator"s choice for a compromise direction at x. For the case of two-party negotiations, the following bisecting function, denoted by T BS , can be defined over the interior set of Oc . Note that the closed set Oc contains two disjoint subsets: Oc = Oc 0 ∪Oc B , where Oc 0 denotes the set of interior points of Oc and Oc B denotes the boundary of Oc . The bisecting compromise is defined by a function T BS : Oc 0 → R2 , T BS (x) = ∇u1(x) ∇u1(x) + ∇u2(x) ∇u2(x) , x ∈ Oc 0. (4) Given the current tentative agreement xh (h ≥ 0), the mediator has to choose a point xh+1 along d = T(xh) so that all parties gain. Ehtamo et al then define a mechanism to generate a sequence of points and prove that when the generated sequence is bounded and when all generated points (from the sequence) belong to the interior set Oc 0 then the sequence converges to a weakly Paretooptimal agreement [4, pp. 59-60].5 As the above mechanism does not work at the boundary points of Oc , we will introduce a procedure that works everywhere in an alternative space Oc . Let x ∈ Oc and let θ(x) denote the angle between the gradients ∇u1(x) and ∇u2(x) at x. That is, θ(x) = arccos( ∇u1(x).∇u2(x) ∇u1(x) . ∇u2(x) ) From Definition 2, it is obvious that the two parties are in local strict opposition (at x) iff θ(x) = π, and they are in local strong opposition iff π ≥ θ(x) > π/2, and they are in local weak opposition iff π/2 ≥ θ(x) ≥ 0. Note also that the two vectors ∇u1(x) and ∇u2(x) define a hyperplane, denoted by h∇(x), in the kdimensional space Rk . Furthermore, there are two indifference curves of agents 1 and 2 going through point x, denoted by IC1(x) and IC2(x), respectively. Let hT1(x) and hT2(x) denote the tangent hyperplanes to the indifference curves IC1(x) and IC2(x), respectively, at point x. The planes hT1(x) and hT2(x) intersect h∇(x) in the lines IS1(x) and IS2(x), respectively. Note that given a line L(x) going through the point x, there are two (unit) vectors from x along L(x) pointing to two opposite directions, denoted by L+ (x) and L− (x). We can now informally explain our solution to the problem of searching for joint gains. When it isn"t possible to obtain a compromise direction for joint improvements at a point x ∈ Oc either because the compromise vector points to the space outside of the feasible set Oc or because the two parties are in local strong opposition at x, we will consider to move along the indifference curve of one party while trying to improve the utility of the other party. As 5 Let S be the set of alternatives, x∗ is weakly Pareto optimal if there is no x ∈ S such that ui(x) > ui(x∗ ) for all agents i. the mediator does not know the indifference curves of the parties, he has to use the tangent hyperplanes to the indifference curves of the parties at point x. Note that the tangent hyperplane to a curve is a useful approximation of the curve in the immediate vicinity of the point of tangency, x. We are now describing an iteration step to reach the next tentative agreement xh+1 from the current tentative agreement xh ∈ Oc . A vector v whose tail is xh is said to be bounded in Oc if ∃λ > 0 such that xh +λv ∈ Oc . To start, the mediator asks the negotiators for their gradients ∇u1(xh) and ∇u2(xh), respectively, at xh. 1. If xh is a Pareto-optimal outcome according to equation 2 or equation 3, then the process is terminated. 2. If 1 ≥ ∇u1(xh).∇u2(xh) > 0 and the vector T BS (xh) is bounded in Oc then the mediator chooses the compromise improving direction d = T BS (xh) and apply the method described by Ehtamo et al [4] to generate the next tentative agreement xh+1. 3. Otherwise, among the four vectors ISσ i (xh), i = 1, 2 and σ = +/−, the mediator chooses the vector that (i) is bounded in Oc , and (ii) is closest to the gradient of the other agent, ∇uj (xh)(j = i). Denote this vector by T G(xh). That is, we will be searching for a point on the indifference curve of agent i, ICi(xh), while trying to improve the utility of agent j. Note that when xh is an interior point of Oc then the situation is symmetric for the two agents 1 and 2, and the mediator has the choice of either finding a point on IC1(xh) to improve the utility of agent 2, or finding a point on IC2(xh) to improve the utility of agent 1. To decide on which choice to make, the mediator has to compute the distribution of gains throughout the whole process to avoid giving more gains to one agent than to the other. Now, the point xh+1 to be generated lies somewhere on the intersection of ICi(xh) and the hyperplane defined by ∇ui(xh) and T G(xh). This intersection is approximated by T G(xh). Thus, the sought after point xh+1 can be generated by first finding a point yh along the direction of T G(xh) and then move from yh to the same direction of ∇ui(xh) until we intersect with ICi(xh). Mathematically, let ζ and ξ denote the vectors T G(xh) and ∇ui(xh), respectively, xh+1 is the solution to the following optimisation problem: max λ1,λ2∈L uj(xh + λ1ζ + λ2ξ) s.t. xh+λ1ζ+λ2ξ ∈ Oc , and ui(xh+λ1ζ+λ2ξ) = ui(xh), where L is a suitable interval of positive real numbers; e.g., L = {λ|λ > 0}, or L = {λ|a < λ ≤ b}, 0 ≤ a < b. Given an initial tentative agreement x0, the method described above allows a sequence of tentative agreements x1, x2, . . . to be iteratively generated. The iteration stops whenever a weakly Pareto optimal agreement is reached. THEOREM 3. If the sequence of agreements generated by the above method is bounded then the method converges to a point x∗ ∈ Oc that is weakly Pareto optimal. 5. CONCLUSION AND FUTURE WORK In this paper we have established a framework for negotiation that is based on MCDM theory for representing the agents" objectives and utilities. The focus of the paper is on integrative negotiation in which agents aim to maximise joint gains, or create value. 514 The Sixth Intl. Joint Conf. on Autonomous Agents and Multi-Agent Systems (AAMAS 07) We have introduced a mediator into the negotiation in order to allow negotiators to disclose information about their utilities, without providing this information to their opponents. Furthermore, the mediator also works toward the goal of achieving fairness of the negotiation outcome. That is, the approach that we describe aims for both efficiency, in the sense that it produces Pareto optimal outcomes (i.e. no aspect can be improved for one of the parties without worsening the outcome for another party), and also for fairness, which chooses optimal solutions which distribute gains amongst the agents in some appropriate manner. We have developed a two step process for addressing the NP-hard problem of finding a solution for a set of integrative attributes, which is within the Pareto-optimal set for those attributes. For simple attributes (i.e. those which have a finite set of values) we use known optimisation techniques to find a Paretooptimal solution. In order to discourage agents from misrepresenting their utilities to gain an advantage, we look for solutions that are least vulnerable to manipulation. We have shown that as long as one of the simple attributes does not strongly dominate the others, then truth telling is an equilibrium strategy for the negotiators during the stage of optimising simple attributes. For non-simple attributes we propose a mechanism that provides stepwise improvements to move the proposed solution in the direction of a Paretooptimal solution. The approach presented in this paper is similar to the ideas behind negotiation analysis [18]. Ehtamo et al [4] presents an approach to searching for joint gains in multi-party negotiations. The relation of their approach to our approach is discussed in the preceding section. Lai et al [12] provide an alternative approach to integrative negotiation. While their approach was clearly described for the case of two-issue negotiations, the generalisation to negotiations with more than two issues is not entirely clear. Zhang et at [22] discuss the use of integrative negotiation in agent organisations. They assume that agents are honest. Their main result is an experiment showing that in some situations, agents" cooperativeness may not bring the most benefits to the organisation as a whole, while giving no explanation. Jonker et al [7] consider an approach to multi-attribute negotiation without the use of a mediator. Thus, their approach can be considered a complement of ours. Their experimental results show that agents can reach Paretooptimal outcomes using their approach. The details of the approach have currently been shown only for bilateral negotiation, and while we believe they are generalisable to multiple negotiators, this work remains to be done. There is also future work to be done in more fully characterising the outcomes of the determination of values for the non-simple attributes. In order to provide a complete framework we are also working on the distributive phase using the mediator. Acknowledgement The authors acknowledge financial support by ARC Dicovery Grant (2006-2009, grant DP0663147) and DEST IAP grant (2004-2006, grant CG040014). The authors would like to thank Lawrence Cavedon and the RMIT Agents research group for their helpful comments and suggestions. 6. REFERENCES [1] F. Alemi, P. Fos, and W. Lacorte. A demonstration of methods for studying negotiations between physicians and health care managers. Decision Science, 21:633-641, 1990. [2] M. Ehrgott. Multicriteria Optimization. Springer-Verlag, Berlin, 2000. [3] H. Ehtamo, R. P. Hamalainen, P. Heiskanen, J. Teich, M. Verkama, and S. Zionts. Generating pareto solutions in a two-party setting: Constraint proposal methods. Management Science, 45(12):1697-1709, 1999. [4] H. Ehtamo, E. Kettunen, and R. P. Hmlinen. Searching for joint gains in multi-party negotiations. European Journal of Operational Research, 130:54-69, 2001. [5] P. Faratin. Automated Service Negotiation Between Autonomous Computational Agents. PhD thesis, University of London, 2000. [6] A. Foroughi. Minimizing negotiation process losses with computerized negotiation support systems. The Journal of Applied Business Research, 14(4):15-26, 1998. [7] C. M. Jonker, V. Robu, and J. Treur. An agent architecture for multi-attribute negotiation using incomplete preference information. J. Autonomous Agents and Multi-Agent Systems, (to appear). [8] R. L. Keeney and H. Raiffa. Decisions with Multiple Objectives: Preferences and Value Trade-Offs. John Wiley and Sons, Inc., New York, 1976. [9] G. Kersten and S. Noronha. Rational agents, contract curves, and non-efficient compromises. IEEE Systems, Man, and Cybernetics, 28(3):326-338, 1998. [10] M. Klein, P. Faratin, H. Sayama, and Y. Bar-Yam. Protocols for negotiating complex contracts. IEEE Intelligent Systems, 18(6):32-38, 2003. [11] S. Kraus, J. Wilkenfeld, and G. Zlotkin. Multiagent negotiation under time constraints. Artificial Intelligence Journal, 75(2):297-345, 1995. [12] G. Lai, C. Li, and K. Sycara. Efficient multi-attribute negotiation with incomplete information. Group Decision and Negotiation, 15:511-528, 2006. [13] D. Lax and J. Sebenius. The manager as negotiator: The negotiator"s dilemma: Creating and claiming value, 2nd ed. In S. Goldberg, F. Sander & N. Rogers, editors, Dispute Resolution, 2nd ed., pages 49-62. Little Brown & Co., 1992. [14] M. Lomuscio and N. Jennings. A classification scheme for negotiation in electronic commerce. In Agent-Mediated Electronic Commerce: A European Agentlink Perspective. Springer-Verlag, 2001. [15] R. Maes and A. Moukas. Agents that buy and sell. Communications of the ACM, 42(3):81-91, 1999. [16] J. Nash. Two-person cooperative games. Econometrica, 21(1):128-140, April 1953. [17] H. Raiffa. The Art and Science of Negotiation. Harvard University Press, Cambridge, USA, 1982. [18] H. Raiffa, J. Richardson, and D. Metcalfe. Negotiation Analysis: The Science and Art of Collaborative Decision Making. Belknap Press, Cambridge, MA, 2002. [19] T. Sandholm. Agents in electronic commerce: Component technologies for automated negotiation and coalition formation. JAAMAS, 3(1):73-96, 2000. [20] J. Sebenius. Negotiation analysis: A characterization and review. Management Science, 38(1):18-38, 1992. [21] L. Weingart, E. Hyder, and M. Pietrula. Knowledge matters: The effect of tactical descriptions on negotiation behavior and outcome. Tech. Report, CMU, 1995. [22] X. Zhang, V. R. Lesser, and T. Wagner. Integrative negotiation among agents situated in organizations. IEEE Trans. on Systems, Man, and Cybernetics, Part C, 36(1):19-30, 2006. The Sixth Intl. Joint Conf. on Autonomous Agents and Multi-Agent Systems (AAMAS 07) 515

concession;mediator;uncertainty;multi-criterion decision make;automate negotiation;integrative negotiation;inefficient compromise;dilemma;deadlock situation;creating value;claiming value;mcdm;incomplete information;negotiation

train_I-56

Unifying Distributed Constraint Algorithms in a BDI Negotiation Framework

This paper presents a novel, unified distributed constraint satisfaction framework based on automated negotiation. The Distributed Constraint Satisfaction Problem (DCSP) is one that entails several agents to search for an agreement, which is a consistent combination of actions that satisfies their mutual constraints in a shared environment. By anchoring the DCSP search on automated negotiation, we show that several well-known DCSP algorithms are actually mechanisms that can reach agreements through a common Belief-Desire-Intention (BDI) protocol, but using different strategies. A major motivation for this BDI framework is that it not only provides a conceptually clearer understanding of existing DCSP algorithms from an agent model perspective, but also opens up the opportunities to extend and develop new strategies for DCSP. To this end, a new strategy called Unsolicited Mutual Advice (UMA) is proposed. Performance evaluation shows that the UMA strategy can outperform some existing mechanisms in terms of computational cycles.

1. INTRODUCTION At the core of many emerging distributed applications is the distributed constraint satisfaction problem (DCSP) - one which involves finding a consistent combination of actions (abstracted as domain values) to satisfy the constraints among multiple agents in a shared environment. Important application examples include distributed resource allocation [1] and distributed scheduling [2]. Many important algorithms, such as distributed breakout (DBO) [3], asynchronous backtracking (ABT) [4], asynchronous partial overlay (APO) [5] and asynchronous weak-commitment (AWC) [4], have been developed to address the DCSP and provide the agent solution basis for its applications. Broadly speaking, these algorithms are based on two different approaches, either extending from classical backtracking algorithms [6] or introducing mediation among the agents. While there has been no lack of efforts in this promising research field, especially in dealing with outstanding issues such as resource restrictions (e.g., limits on time and communication) [7] and privacy requirements [8], there is unfortunately no conceptually clear treatment to prise open the model-theoretic workings of the various agent algorithms that have been developed. As a result, for instance, a deeper intellectual understanding on why one algorithm is better than the other, beyond computational issues, is not possible. In this paper, we present a novel, unified distributed constraint satisfaction framework based on automated negotiation [9]. Negotiation is viewed as a process of several agents searching for a solution called an agreement. The search can be realized via a negotiation mechanism (or algorithm) by which the agents follow a high level protocol prescribing the rules of interactions, using a set of strategies devised to select their own preferences at each negotiation step. Anchoring the DCSP search on automated negotiation, we show in this paper that several well-known DCSP algorithms [3] are actually mechanisms that share the same Belief-DesireIntention (BDI) interaction protocol to reach agreements, but use different action or value selection strategies. The proposed framework provides not only a clearer understanding of existing DCSP algorithms from a unified BDI agent perspective, but also opens up the opportunities to extend and develop new strategies for DCSP. To this end, a new strategy called Unsolicited Mutual Advice (UMA) is proposed. Our performance evaluation shows that UMA can outperform ABT and AWC in terms of the average number of computational cycles for both the sparse and critical coloring problems [6]. The rest of this paper is organized as follows. In Section 2, we provide a formal overview of DCSP. Section 3 presents a BDI negotiation model by which a DCSP agent reasons. Section 4 presents the existing algorithms ABT, AWC and DBO as different strategies formalized on a common protocol. A new strategy called Unsolicited Mutual Advice is proposed in Section 5; our empirical results and discussion attempt to highlight the merits of the new strategy over existing ones. Section 6 concludes the paper and points to some future work. 2. DCSP: PROBLEM FORMALIZATION The DCSP [4] considers the following environment. • There are n agents with k variables x0, x1, · · · , xk−1, n ≤ k, which have values in domains D1, D2, · · · , Dk, respectively. We define a partial function B over the productrange {0, 1, . . . , (n−1)}×{0, 1, . . . , (k −1)} such that, that variable xj belongs to agent i is denoted by B(i, j)!. The exclamation mark ‘!" means ‘is defined". • There are m constraints c0, c1, · · · cm−1 to be conjunctively satisfied. In a similar fashion as defined for B(i, j), we use E(l, j)!, (0 ≤ l < m, 0 ≤ j < k), to denote that xj is relevant to the constraint cl. The DCSP may be formally stated as follows. Problem Statement: ∀i, j (0 ≤ i < n)(0 ≤ j < k) where B(i, j)!, find the assignment xj = dj ∈ Dj such that ∀l (0 ≤ l < m) where E(l, j)!, cl is satisfied. A constraint may consist of different variables belonging to different agents. An agent cannot change or modify the assignment values of other agents" variables. Therefore, in cooperatively searching for a DCSP solution, the agents would need to communicate with one another, and adjust and re-adjust their own variable assignments in the process. 2.1 DCSP Agent Model In general, all DCSP agents must cooperatively interact, and essentially perform the assignment and reassignment of domain values to variables to resolve all constraint violations. If the agents succeed in their resolution, a solution is found. In order to engage in cooperative behavior, a DCSP agent needs five fundamental parameters, namely, (i) a variable [4] or a variable set [10], (ii) domains, (iii) priority, (iv) a neighbor list and (v) a constraint list. Each variable assumes a range of values called a domain. A domain value, which usually abstracts an action, is a possible option that an agent may take. Each agent has an assigned priority. These priority values help decide the order in which they revise or modify their variable assignments. An agent"s priority may be fixed (static) or changing (dynamic) when searching for a solution. If an agent has more than one variable, each variable can be assigned a different priority, to help determine which variable assignment the agent should modify first. An agent which shares the same constraint with another agent is called the latter"s neighbor. Each agent needs to refer to its list of neighbors during the search process. This list may also be kept unchanged or updated accordingly in runtime. Similarly, each agent maintains a constraint list. The agent needs to ensure that there is no violation of the constraints in this list. Constraints can be added or removed from an agent"s constraint list in runtime. As with an agent, a constraint can also be associated with a priority value. Constraints with a high priority are said to be more important than constraints with a lower priority. To distinguish it from the priority of an agent, the priority of a constraint is called its weight. 3. THE BDI NEGOTIATION MODEL The BDI model originates with the work of M. Bratman [11]. According to [12, Ch.1], the BDI architecture is based on a philosophical model of human practical reasoning, and draws out the process of reasoning by which an agent decides which actions to perform at consecutive moments when pursuing certain goals. Grounding the scope to the DCSP framework, the common goal of all agents is finding a combination of domain values to satisfy a set of predefined constraints. In automated negotiation [9], such a solution is called an agreement among the agents. Within this scope, we found that we were able to unearth the generic behavior of a DCSP agent and formulate it in a negotiation protocol, prescribed using the powerful concepts of BDI. Thus, our proposed negotiation model can be said to combine the BDI concepts with automated negotiation in a multiagent framework, allowing us to conceptually separate DCSP mechanisms into a common BDI interaction protocol and the adopted strategies. 3.1 The generic protocol Figure 1 shows the basic reasoning steps in an arbitrary round of negotiation that constitute the new protocol. The solid line indicates the common component or transition which always exists regardless of the strategy used. The dotted line indicates the Percept Belief Desire Intention Mediation Execution P B D I I I Info Message Info Message Negotiation Message Negotiation Message Negotiation Message Negotiation Message Negotiation Message Negotiation Message Negotiation Message Figure 1: The BDI interaction protocol component or transition which may or may not appear depending on the adopted strategy. Two types of messages are exchanged through this protocol, namely, the info message and the negotiation message. An info message perceived is a message sent by another agent. The message will contain the current selected values and priorities of the variables of that sending agent. The main purpose of this message is to update the agent about the current environment. Info message is sent out at the end of one negotiation round (also called a negotiation cycle), and received at the beginning of next round. A negotiation message is a message which may be sent within a round. This message is for mediation purposes. The agent may put different contents into this type of message as long as it is agreed among the group. The format of the negotiation message and when it is to be sent out are subject to the strategy. A negotiation message can be sent out at the end of one reasoning step and received at the beginning of the next step. Mediation is a step of the protocol that depends on whether the agent"s interaction with others is synchronous or asynchronous. In synchronous mechanism, mediation is required in every negotiation round. In an asynchronous one, mediation is needed only in a negotiation round when the agent receives a negotiation message. A more in-depth view of this mediation step is provided later in this section. The BDI protocol prescribes the skeletal structure for DCSP negotiation. We will show in Section 4 that several well-known DCSP mechanisms all inherit this generic model. The details of the six main reasoning steps for the protocol (see Figure 1) are described as follows for a DCSP agent. For a conceptually clearer description, we assume that there is only one variable per agent. • Percept. In this step, the agent receives info messages from its neighbors in the environment, and using its Percept function, returns an image P. This image contains the current values assigned to the variables of all agents in its neighbor list. The image P will drive the agent"s actions in subsequent steps. The agent also updates its constraint list C using some criteria of the adopted strategy. • Belief. Using the image P and constraint list C, the agent will check if there is any violated constraint. If there is no violation, the agent will believe it is choosing a correct option and therefore will take no action. The agent will do nothing if it is in a local stable state - a snapshot of the variables assignments of the agent and all its neighbors by which they satisfy their shared constraints. When all agents are in their local stable states, the whole environment is said to be in a global stable state and an agreeThe Sixth Intl. Joint Conf. on Autonomous Agents and Multi-Agent Systems (AAMAS 07) 525 ment is found. In case the agent finds its value in conflict with some of its neighbors", i.e., the combination of values assigned to the variables leads to a constraint violation, the agent will first try to reassign its own variable using a specific strategy. If it finds a suitable option which meets some criteria of the adopted strategy, the agent will believe it should change to the new option. However it does not always happen that an agent can successfully find such an option. If no option can be found, the agent will believe it has no option, and therefore will request its neighbors to reconsider their variable assignments. To summarize, there are three types of beliefs that a DCSP agent can form: (i) it can change its variable assignment to improve the current situation, (ii) it cannot change its variable assignment and some constraints violations cannot be resolved and (iii) it need not change its variable assignment as all the constraints are satisfied. Once the beliefs are formed, the agent will determine its desires, which are the options that attempt to resolve the current constraint violations. • Desire. If the agent takes Belief (i), it will generate a list of its own suitable domain values as its desire set. If the agent takes Belief (ii), it cannot ascertain its desire set, but will generate a sublist of agents from its neighbor list, whom it will ask to reconsider their variable assignments. How this sublist is created depends on the strategy devised for the agent. In this situation, the agent will use a virtual desire set that it determines based on its adopted strategy. If the agent takes Belief (iii), it will have no desire to revise its domain value, and hence no intention. • Intention. The agent will select a value from its desire set as its intention. An intention is the best desired option that the agent assigns to its variable. The criteria for selecting a desire as the agent"s intention depend on the strategy used. Once the intention is formed, the agent may either proceed to the execution step, or undergo mediation. Again, the decision to do so is determined by some criteria of the adopted strategy. • Mediation. This is an important function of the agent. Since, if the agent executes its intention without performing intention mediation with its neighbors, the constraint violation between the agents may not be resolved. Take for example, suppose two agents have variables, x1 and x2, associated with the same domain {1, 2}, and their shared constraint is (x1 + x2 = 3). Then if both the variables are initialized with value 1, they will both concurrently switch between the values 2 and 1 in the absence of mediation between them. There are two types of mediation: local mediation and group mediation. In the former, the agents exchange their intentions. When an agent receives another"s intention which conflicts with its own, the agent must mediate between the intentions, by either changing its own intention or informing the other agent to change its intention. In the latter, there is an agent which acts as a group mediator. This mediator will collect the intentions from the group - a union of the agent and its neighbors - and determine which intention is to be executed. The result of this mediation is passed back to the agents in the group. Following mediation, the agent may proceed to the next reasoning step to execute its intention or begin a new negotiation round. • Execution. This is the last step of a negotiation round. The agent will execute by updating its variable assignment if the intention obtained at this step is its own. Following execution, the agent will inform its neighbors about its new variable assignment and updated priority. To do so, the agent will send out an info message. 3.2 The strategy A strategy plays an important role in the negotiation process. Within the protocol, it will often determine the efficiency of the Percept Belief Desire Intention Mediation Execution P B D I Info Message Info Message Negotiation Message Negotiation Message Negotiation Message Figure 2: BDI protocol with Asynchronous Backtracking strategy search process in terms of computational cycles and message communication costs. The design space when devising a strategy is influenced by the following dimensions: (i) asynchronous or synchronous, (ii) dynamic or static priority, (iii) dynamic or static constraint weight, (iv) number of negotiation messages to be communicated, (v) the negotiation message format and (vi) the completeness property. In other words, these dimensions provide technical considerations for a strategy design. 4. DCSP ALGORITHMS: BDI PROTOCOL + STRATEGIES In this section, we apply the proposed BDI negotiation model presented in Section 3 to expose the BDI protocol and the different strategies used for three well-known algorithms, ABT, AWC and DBO. All these algorithms assume that there is only one variable per agent. Under our framework, we call the strategies applied the ABT, AWC and DBO strategies, respectively. To describe each strategy formally, the following mathematical notations are used: • n is the number of agents, m is the number of constraints; • xi denotes the variable held by agent i, (0 ≤ i < n); • Di denotes the domain of variable xi; Fi denotes the neighbor list of agent i; Ci denotes its constraint list; • pi denotes the priority of agent i; and Pi = {(xj = vj, pj = k) | agent j ∈ Fi, vj ∈ Dj is the current value assigned to xj and the priority value k is a positive integer } is the perception of agent i; • wl denotes the weight of constraint l, (0 ≤ l < m); • Si(v) is the total weight of the violated constraints in Ci when its variable has the value v ∈ Di. 4.1 Asynchronous Backtracking Figure 2 presents the BDI negotiation model incorporating the Asynchronous Backtracking (ABT) strategy. As mentioned in Section 3, for an asynchronous mechanism that ABT is, the mediation step is needed only in a negotiation round when an agent receives a negotiation message. For agent i, beginning initially with (wl = 1, (0 ≤ l < m); pi = i, (0 ≤ i < n)) and Fi contains all the agents who share the constraints with agent i, its BDI-driven ABT strategy is described as follows. Step 1 - Percept: Update Pi upon receiving the info messages from the neighbors (in Fi). Update Ci to be the list of 526 The Sixth Intl. Joint Conf. on Autonomous Agents and Multi-Agent Systems (AAMAS 07) constraints which only consists of agents in Fi that have equal or higher priority than this agent. Step 2 - Belief: The belief function GB (Pi,Ci) will return a value bi ∈ {0, 1, 2}, decided as follows: • bi = 0 when agent i can find an optimal option, i.e., if (Si(vi) = 0 or vi is in bad values list) and (∃a ∈ Di)(Si(a) = 0) and a is not in a list of domain values called bad values list. Initially this list is empty and it will be cleared when a neighbor of higher priority changes its variable assignment. • bi = 1 when it cannot find an optimal option, i.e., if (∀a ∈ Di)(Si(a) = 0) or a is in bad values list. • bi = 2 when its current variable assignment is an optimal option, i.e., if Si(vi) = 0 and vi is not in bad value list. Step 3 - Desire: The desire function GD (bi) will return a desire set denoted by DS, decided as follows: • If bi = 0, then DS = {a | (a = vi), (Si(a) = 0) and a is not in the bad value list }. • If bi = 1, then DS = ∅, the agent also finds agent k which is determined by {k | pk = min(pj) with agent j ∈ Fi and pk > pi }. • If bi = 2, then DS = ∅. Step 4 - Intention: The intention function GI (DS) will return an intention, decided as follows: • If DS = ∅, then select an arbitrary value (say, vi) from DS as the intention. • If DS = ∅, then assign nil as the intention (to denote its lack thereof). Step 5 - Execution: • If agent i has a domain value as its intention, the agent will update its variable assignment with this value. • If bi = 1, agent i will send a negotiation message to agent k, then remove k from Fi and begin its next negotiation round. The negotiation message will contain the list of variable assignments of those agents in its neighbor list Fi that have a higher priority than agent i in the current image Pi. Mediation: When agent i receives a negotiation message, several sub-steps are carried out, as follows: • If the list of agents associated with the negotiation message contains agents which are not in Fi, it will add these agents to Fi, and request these agents to add itself to their neighbor lists. The request is considered as a type of negotiation message. • Agent i will first check if the sender agent is updated with its current value vi. The agent will add vi to its bad values list if it is so, or otherwise send its current value to the sender agent. Following this step, agent i proceeds to the next negotiation round. 4.2 Asynchronous Weak Commitment Search Figure 3 presents the BDI negotiation model incorporating the Asynchronous Weak Commitment (AWC) strategy. The model is similar to that of incorporating the ABT strategy (see Figure 2). This is not surprising; AWC and ABT are found to be strategically similar, differing only in the details of some reasoning steps. The distinguishing point of AWC is that when the agent cannot find a suitable variable assignment, it will change its priority to the highest among its group members ({i} ∪ Fi). For agent i, beginning initially with (wl = 1, (0 ≤ l < m); pi = i, (0 ≤ i < n)) and Fi contains all the agents who share the constraints with agent i, its BDI-driven AWC strategy is described as follows. Step 1 - Percept: This step is identical to the Percept step of ABT. Step 2 - Belief: The belief function GB (Pi,Ci) will return a value bi ∈ {0, 1, 2}, decided as follows: Percept Belief Desire Intention Mediation Execution P B D I Info Message Info Message Negotiation Message Negotiation Message Negotiation Message Figure 3: BDI protocol with Asynchronous WeakCommitment strategy • bi = 0 when the agent can find an optimal option i.e., if (Si(vi) = 0 or the assignment xi = vi and the current variables assignments of the neighbors in Fi who have higher priority form a nogood [4]) stored in a list called nogood list and ∃a ∈ Di, Si(a) = 0 (initially the list is empty). • bi = 1 when the agent cannot find any optimal option i.e., if ∀a ∈ Di, Si(a) = 0. • bi = 2 when the current assignment is an optimal option i.e., if Si(vi) = 0 and the current state is not a nogood in nogood list. Step 3 - Desire: The desire function GD (bi) will return a desire set DS, decided as follows: • If bi = 0, then DS = {a | (a = vi), (Si(a) = 0) and the number of constraint violations with lower priority agents is minimized }. • If bi = 1, then DS = {a | a ∈ Di and the number of violations of all relevant constraints is minimized }. • If bi = 2, then DS = ∅. Following, if bi = 1, agent i will find a list Ki of higher priority neighbors, defined by Ki = {k | agent k ∈ Fi and pk > pi}. Step 4 - Intention: This step is similar to the Intention step of ABT. However, for this strategy, the negotiation message will contain the variable assignments (of the current image Pi) for all the agents in Ki. This list of assignment is considered as a nogood. If the same negotiation message had been sent out before, agent i will have nil intention. Otherwise, the agent will send the message and save the nogood in the nogood list. Step 5 - Execution: • If agent i has a domain value as its intention, the agent will update its variable assignment with this value. • If bi = 1, it will send the negotiation message to its neighbors in Ki, and set pi = max{pj} + 1, with agent j ∈ Fi. Mediation: This step is identical to the Mediation step of ABT, except that agent i will now add the nogood contained in the negotiation message received to its own nogood list. 4.3 Distributed Breakout Figure 4 presents the BDI negotiation model incorporating the Distributed Breakout (DBO) strategy. Essentially, by this synchronous strategy, each agent will search iteratively for improvement by reducing the total weight of the violated constraints. The iteration will continue until no agent can improve further, at which time if some constraints remain violated, the weights of The Sixth Intl. Joint Conf. on Autonomous Agents and Multi-Agent Systems (AAMAS 07) 527 Percept Belief Desire Intention Mediation Execution P B D I I A Info Message Info Message Negotiation Message Negotiation Message Figure 4: BDI protocol with Distributed Breakout strategy these constraints will be increased by 1 to help ‘breakout" from a local minimum. For agent i, beginning initially with (wl = 1, (0 ≤ l < m), pi = i, (0 ≤ i < n)) and Fi contains all the agents who share the constraints with agent i, its BDI-driven DBO strategy is described as follows. Step 1 - Percept: Update Pi upon receiving the info messages from the neighbors (in Fi). Update Ci to be the list of its relevant constraints. Step 2 - Belief: The belief function GB (Pi,Ci) will return a value bi ∈ {0, 1, 2}, decided as follows: • bi = 0 when agent i can find an option to reduce the number violations of the constraints in Ci, i.e., if ∃a ∈ Di, Si(a) < Si(vi). • bi = 1 when it cannot find any option to improve situation, i.e., if ∀a ∈ Di, a = vi, Si(a) ≥ Si(vi). • bi = 2 when its current assignment is an optimal option, i.e., if Si(vi) = 0. Step 3 - Desire: The desire function GD (bi) will return a desire set DS, decided as follows: • If bi = 0, then DS = {a | a = vi, Si(a) < Si(vi) and (Si(vi)−Si(a)) is maximized }. (max{(Si(vi)−Si(a))} will be referenced by hmax i in subsequent steps, and it defines the maximal reduction in constraint violations). • Otherwise, DS = ∅. Step 4 - Intention: The intention function GI (DS) will return an intention, decided as follows: • If DS = ∅, then select an arbitrary value (say, vi) from DS as the intention. • If DS = ∅, then assign nil as the intention. Following, agent i will send its intention to all its neighbors. In return, it will receive intentions from these agents before proceeding to Mediation step. Mediation: Agent i receives all the intentions from its neighbors. If it finds that the intention received from a neighbor agent j is associated with hmax j > hmax i , the agent will automatically cancel its current intention. Step 5 - Execution: • If agent i did not cancel its intention, it will update its variable assignment with the intended value. Percept Belief Desire Intention Mediation Execution P B D I I A Info Message Info Message Negotiation Message Negotiation Message Negotiation Message Negotiation Message Figure 5: BDI protocol with Unsolicited Mutual Advice strategy • If all intentions received and its own one are nil intention, the agent will increase the weight of each currently violated constraint by 1. 5. THE UMA STRATEGY Figure 5 presents the BDI negotiation model incorporating the Unsolicited Mutual Advice(UMA) strategy. Unlike when using the strategies of the previous section, a DCSP agent using UMA will not only send out a negotiation message when concluding its Intention step, but also when concluding its Desire step. The negotiation message that it sends out to conclude the Desire step constitutes an unsolicited advice for all its neighbors. In turn, the agent will wait to receive unsolicited advices from all its neighbors, before proceeding on to determine its intention. For agent i, beginning initially with (wl = 1, (0 ≤ l < m), pi = i, (0 ≤ i < n)) and Fi contains all the agents who share the constraints with agent i, its BDI-driven UMA strategy is described as follows. Step 1 - Percept: Update Pi upon receiving the info messages from the neighbors (in Fi). Update Ci to be the list of constraints relevant to agent i. Step 2 - Belief: The belief function GB (Pi,Ci) will return a value bi ∈ {0, 1, 2}, decided as follows: • bi = 0 when agent i can find an option to reduce the number violations of the constraints in Ci, i.e., if ∃a ∈ Di, Si(a) < Si(vi) and the assignment xi = a and the current variable assignments of its neighbors do not form a local state stored in a list called bad states list (initially this list is empty). • bi = 1 when it cannot find a value a such as a ∈ Di, Si(a) < Si(vi), and the assignment xi = a and the current variable assignments of its neighbors do not form a local state stored in the bad states list. • bi = 2 when its current assignment is an optimal option, i.e., if Si(vi) = 0. Step 3 - Desire: The desire function GD (bi) will return a desire set DS, decided as follows: • If bi = 0, then DS = {a | a = vi, Si(a) < Si(vi) and (Si(vi) − Si(a)) is maximized } and the assignment xi = a and the current variable assignments of agent i"s neighbors do not form a state in the bad states list. In this case, DS is called a set of voluntary desires. max{(Si(vi)−Si(a))} will be referenced by hmax i in subsequent steps, and it defines 528 The Sixth Intl. Joint Conf. on Autonomous Agents and Multi-Agent Systems (AAMAS 07) the maximal reduction in constraint violations. It is also referred to as an improvement). • If bi = 1, then DS = {a | a = vi, Si(a) is minimized } and the assignment xi = a and the current variable assignments of agent i"s neighbors do not form a state in the bad states list. In this case, DS is called a set of reluctant desires • If bi = 2, then DS = ∅. Following, if bi = 0, agent i will send a negotiation message containing hmax i to all its neighbors. This message is called a voluntary advice. If bi = 1, agent i will send a negotiation message called change advice to the neighbors in Fi who share the violated constraints with agent i. Agent i receives advices from all its neighbors and stores them in a list called A, before proceeding to the next step. Step 4 - Intention: The intention function GI (DS, A) will return an intention, decided as follows: • If there is a voluntary advice from an agent j which is associated with hmax j > hmax i , assign nil as the intention. • If DS = ∅, DS is a set of voluntary desires and hmax i is the biggest improvement among those associated with the voluntary advices received, select an arbitrary value (say, vi) from DS as the intention. This intention is called a voluntary intention. • If DS = ∅, DS is a set of reluctant desires and agent i receives some change advices, select an arbitrary value (say, vi) from DS as the intention. This intention is called reluctant intention. • If DS = ∅, then assign nil as the intention. Following, if the improvement hmax i is the biggest improvement and equal to some improvements associated with the received voluntary advices, agent i will send its computed intention to all its neighbors. If agent i has a reluctant intention, it will also send this intention to all its neighbors. In both cases, agent i will attach the number of received change advices in the current negotiation round with its intention. In return, agent i will receive the intentions from its neighbors before proceeding to Mediation step. Mediation: If agent i does not send out its intention before this step, i.e., the agent has either a nil intention or a voluntary intention with biggest improvement, it will proceed to next step. Otherwise, agent i will select the best intention among all the intentions received, including its own (if any). The criteria to select the best intention are listed, applied in descending order of importance as follows. • A voluntary intention is preferred over a reluctant intention. • A voluntary intention (if any) with biggest improvement is selected. • If there is no voluntary intention, the reluctant intention with the lowest number of constraint violations is selected. • The intention from an agent who has received a higher number of change advices in the current negotiation round is selected. • Intention from an agent with highest priority is selected. If the selected intention is not agent i"s intention, it will cancel its intention. Step 5 - Execution: If agent i does not cancel its intention, it will update its variable assignment with the intended value. Termination Condition: Since each agent does not have full information about the global state, it may not know when it has reached a solution, i.e., when all the agents are in a global stable state. Hence an observer is needed that will keep track of the negotiation messages communicated in the environment. Following a certain period of time when there is no more message communication (and this happens when all the agents have no more intention to update their variable assignments), the observer will inform the agents in the environment that a solution has been found. 1 2 3 4 5 6 7 8 9 10 Figure 6: Example problem 5.1 An Example To illustrate how UMA works, consider a 2-color graph problem [6] as shown in Figure 6. In this example, each agent has a color variable representing a node. There are 10 color variables sharing the same domain {Black, White}. The following records the outcome of each step in every negotiation round executed. Round 1: Step 1 - Percept: Each agent obtains the current color assignments of those nodes (agents) adjacent to it, i.e., its neighbors". Step 2 - Belief: Agents which have positive improvements are agent 1 (this agent believes it should change its color to White), agent 2 (this believes should change its color to White), agent 7 (this agent believes it should change its color to Black) and agent 10 (this agent believes it should change its value to Black). In this negotiation round, the improvements achieved by these agents are 1. Agents which do not have any improvements are agents 4, 5 and 8. Agents 3, 6 and 9 need not change as all their relevant constraints are satisfied. Step 3 - Desire: Agents 1, 2, 7 and 10 have the voluntary desire (White color for agents 1, 2 and Black color for agents 7, 10). These agents will send the voluntary advices to all their neighbors. Meanwhile, agents 4, 5 and 8 have the reluctant desires (White color for agent 4 and Black color for agents 5, 8). Agent 4 will send a change advice to agent 2 as agent 2 is sharing the violated constraint with it. Similarly, agents 5 and 8 will send change advices to agents 7 and 10 respectively. Agents 3, 6 and 9 do not have any desire to update their color assignments. Step 4 - Intention: Agents 2, 7 and 10 receive the change advices from agents 4, 5 and 8, respectively. They form their voluntary intentions. Agents 4, 5 and 8 receive the voluntary advices from agents 2, 7 and 10, hence they will not have any intention. Agents 3, 6 and 9 do not have any intention. Following, the intention from the agents will be sent to all their neighbors. Mediation: Agent 1 finds that the intention from agent 2 is better than its intention. This is because, although both agents have voluntary intentions with improvement of 1, agent 2 has received one change advice from agent 4 while agent 1 has not received any. Hence agent 1 cancels its intention. Agent 2 will keep its intention. Agents 7 and 10 keep their intentions since none of their neighbors has an intention. The rest of the agents do nothing in this step as they do not have any intention. Step 5 - Execution: Agent 2 changes its color to White. Agents 7 and 10 change their colors to Black. The new state after round 1 is shown in Figure 7. Round 2: Step 1 - Percept: The agents obtain the current color assignments of their neighbors. Step 2 - Belief: Agent 3 is the only agent who has a positive improvement which is 1. It believes it should change its The Sixth Intl. Joint Conf. on Autonomous Agents and Multi-Agent Systems (AAMAS 07) 529 1 2 3 4 5 6 7 8 9 10 Figure 7: The graph after round 1 color to Black. Agent 2 does not have any positive improvement. The rest of the agents need not make any change as all their relevant constraints are satisfied. They will have no desire, and hence no intention. Step 3 - Desire: Agent 3 desires to change its color to Black voluntarily, hence it sends out a voluntary advice to its neighbor, i.e., agent 2. Agent 2 does not have any value for its reluctant desire set as the only option, Black color, will bring agent 2 and its neighbors to the previous state which is known to be a bad state. Since agent 2 is sharing the constraint violation with agent 3, it sends a change advice to agent 3. Step 4 - Intention: Agent 3 will have a voluntary intention while agent 2 will not have any intention as it receives the voluntary advice from agent 3. Mediation: Agent 3 will keep its intention as its only neighbor, agent 2, does not have any intention. Step 5 - Execution: Agent 3 changes its color to Black. The new state after round 2 is shown in Figure 8. Round 3: In this round, every agent finds that it has no desire and hence no intention to revise its variable assignment. Following, with no more negotiation message communication in the environment, the observer will inform all the agents that a solution has been found. 2 3 4 5 6 7 8 91 10 Figure 8: The solution obtained 5.2 Performance Evaluation To facilitate credible comparisons with existing strategies, we measured the execution time in terms of computational cycles as defined in [4], and built a simulator that could reproduce the published results for ABT and AWC. The definition of a computational cycle is as follows. • In one cycle, each agent receives all the incoming messages, performs local computation and sends out a reply. • A message which is sent at time t will be received at time t + 1. The network delay is neglected. • Each agent has it own clock. The initial clock"s value is 0. Agents attach their clock value as a time-stamp in the outgoing message and use the time-stamp in the incoming message to update their own clock"s value. Four benchmark problems [6] were considered, namely, n-queens and node coloring for sparse, dense and critical graphs. For each problem, a finite number of test cases were generated for various problem sizes n. The maximum execution time was set to 0 200 400 600 800 1000 10 50 100 Number of queens Cycles Asynchronous Backtracking Asynchronous Weak Commitment Unsolicited Mutual Advice Figure 9: Relationship between execution time and problem size 10000 cycles for node coloring for critical graphs and 1000 cycles for other problems. The simulator program was terminated after this period and the algorithm was considered to fail a test case if it did not find a solution by then. In such a case, the execution time for the test was counted as 1000 cycles. 5.2.1 Evaluation with n-queens problem The n-queens problem is a traditional problem of constraint satisfaction. 10 test cases were generated for each problem size n ∈ {10, 50 and 100}. Figure 9 shows the execution time for different problem sizes when ABT, AWC and UMA were run. 5.2.2 Evaluation with graph coloring problem The graph coloring problem can be characterized by three parameters: (i) the number of colors k, the number of nodes/agents n and the number of links m. Based on the ratio m/n, the problem can be classified into three types [3]: (i) sparse (with m/n = 2), (ii) critical (with m/n = 2.7 or 4.7) and (iii) dense (with m/n = (n − 1)/4). For this problem, we did not include ABT in our empirical results as its failure rate was found to be very high. This poor performance of ABT was expected since the graph coloring problem is more difficult than the n-queens problem, on which ABT already did not perform well (see Figure 9). The sparse and dense (coloring) problem types are relatively easy while the critical type is difficult to solve. In the experiments, we fix k = 3. 10 test cases were created using the method described in [13] for each value of n ∈ {60, 90, 120}, for each problem type. The simulation results for each type of problem are shown in Figures 10 - 12. 0 40 80 120 160 200 60 90 120 150 Number of Nodes Cycles Asynchronous Weak Commitment Unsolicited Mutual Advice Figure 10: Comparison between AWC and UMA (sparse graph coloring) 5.3 Discussion 5.3.1 Comparison with ABT and AWC 530 The Sixth Intl. Joint Conf. on Autonomous Agents and Multi-Agent Systems (AAMAS 07) 0 1000 2000 3000 4000 5000 6000 60 90 120 Number of Nodes Cycles Asynchronous Weak Commitment Unsolicited Mutual Advice Figure 11: Comparison between AWC and UMA (critical graph coloring) 0 10 20 30 40 50 60 90 120 Number of Nodes Cycles Asynchronous Weak Commitment Unsolicited Mutual Advice Figure 12: Comparison between AWC and UMA (dense graph coloring) Figure 10 shows that the average performance of UMA is slightly better than AWC for the sparse problem. UMA outperforms AWC in solving the critical problem as shown in Figure 11. It was observed that the latter strategy failed in some test cases. However, as seen in Figure 12, both the strategies are very efficient when solving the dense problem, with AWC showing slightly better performance. The performance of UMA, in the worst (time complexity) case, is similar to that of all evaluated strategies. The worst case occurs when all the possible global states of the search are reached. Since only a few agents have the right to change their variable assignments in a negotiation round, the number of redundant computational cycles and info messages is reduced. As we observe from the backtracking in ABT and AWC, the difference in the ordering of incoming messages can result in a different number of computational cycles to be executed by the agents. 5.3.2 Comparison with DBO The computational performance of UMA is arguably better than DBO for the following reasons: • UMA can guarantee that there will be a variable reassignment following every negotiation round whereas DBO cannot. • UMA introduces one more communication round trip (that of sending a message and awaiting a reply) than DBO, which occurs due to the need to communicate unsolicited advices. Although this increases the communication cost per negotiation round, we observed from our simulations that the overall communication cost incurred by UMA is lower due to the significantly lower number of negotiation rounds. • Using UMA, in the worst case, an agent will only take 2 or 3 communication round trips per negotiation round, following which the agent or its neighbor will do a variable assignment update. Using DBO, this number of round trips is uncertain as each agent might have to increase the weights of the violated constraints until an agent has a positive improvement; this could result in a infinite loop [3]. 6. CONCLUSION Applying automated negotiation to DCSP, this paper has proposed a protocol that prescribes the generic reasoning of a DCSP agent in a BDI architecture. Our work shows that several wellknown DCSP algorithms, namely ABT, AWC and DBO, can be described as mechanisms sharing the same proposed protocol, and only differ in the strategies employed for the reasoning steps per negotiation round as governed by the protocol. Importantly, this means that it might furnish a unified framework for DCSP that not only provides a clearer BDI agent-theoretic view of existing DCSP approaches, but also opens up the opportunities to enhance or develop new strategies. Towards the latter, we have proposed and formulated a new strategy - the UMA strategy. Empirical results and our discussion suggest that UMA is superior to ABT, AWC and DBO in some specific aspects. It was observed from our simulations that UMA possesses the completeness property. Future work will attempt to formally establish this property, as well as formalize other existing DSCP algorithms as BDI negotiation mechanisms, including the recent endeavor that employs a group mediator [5]. The idea of DCSP agents using different strategies in the same environment will also be investigated. 7. REFERENCES [1] P. J. Modi, H. Jung, M. Tambe, W.-M. Shen, and S. Kulkarni, Dynamic distributed resource allocation: A distributed constraint satisfaction approach, in Lecture Notes in Computer Science, 2001, p. 264. [2] H. Schlenker and U. Geske, Simulating large railway networks using distributed constraint satisfaction, in 2nd IEEE International Conference on Industrial Informatics (INDIN-04), 2004, pp. 441- 446. [3] M. Yokoo, Distributed Constraint Satisfaction : Foundations of Cooperation in Multi-Agent Systems. Springer Verlag, 2000, springer Series on Agent Technology. [4] M. Yokoo, E. H. Durfee, T. Ishida, and K. Kuwabara, The distributed constraint satisfaction problem : Formalization and algorithms, IEEE Transactions on Knowledge and Data Engineering, vol. 10, no. 5, pp. 673-685, September/October 1998. [5] R. Mailler and V. Lesser, Using cooperative mediation to solve distributed constraint satisfaction problems, in Proceedings of the Third International Joint Conference on Autonomous Agents and Multiagent Systems (AAMAS-04), 2004, pp. 446-453. [6] E. Tsang, Foundation of Constraint Satisfaction. Academic Press, 1993. [7] R. Mailler, R. Vincent, V. Lesser, T. Middlekoop, and J. Shen, Soft Real-Time, Cooperative Negotiation for Distributed Resource Allocation, AAAI Fall Symposium on Negotiation Methods for Autonomous Cooperative Systems, November 2001. [8] M. Yokoo, K. Suzuki, and K. Hirayama, Secure distributed constraint satisfaction: Reaching agreement without revealing private information, Artificial Intelligence, vol. 161, no. 1-2, pp. 229-246, 2005. [9] J. S. Rosenschein and G. Zlotkin, Rules of Encounter. The MIT Press, 1994. [10] M. Yokoo and K. Hirayama, Distributed constraint satisfaction algorithm for complex local problems, in Proceedings of the Third International Conference on Multiagent Systems (ICMAS-98), 1998, pp. 372-379. [11] M. E. Bratman, Intentions, Plans and Practical Reason. Harvard University Press, Cambridge, M.A, 1987. [12] G. Weiss, Ed., Multiagent System : A Modern Approach to Distributed Artificial Intelligence. The MIT Press, London, U.K, 1999. [13] S. Minton, M. D. Johnson, A. B. Philips, and P. Laird, Minimizing conflicts: A heuristic repair method for constraint satisfaction and scheduling problems, Artificial Intelligence, vol. e58, no. 1-3, pp. 161-205, 1992. The Sixth Intl. Joint Conf. on Autonomous Agents and Multi-Agent Systems (AAMAS 07) 531

algorithm;agent negotiation;constraint;resource restriction;privacy requirement;shared environment;bdi;uma;mediation;backtracking;negotiation;dcsp;belief-desireintention model;distribute constraint satisfaction problem

train_I-57

Rumours and Reputation: Evaluating Multi-Dimensional Trust within a Decentralised Reputation System

In this paper we develop a novel probabilistic model of computational trust that explicitly deals with correlated multi-dimensional contracts. Our starting point is to consider an agent attempting to estimate the utility of a contract, and we show that this leads to a model of computational trust whereby an agent must determine a vector of estimates that represent the probability that any dimension of the contract will be successfully fulfilled, and a covariance matrix that describes the uncertainty and correlations in these probabilities. We present a formalism based on the Dirichlet distribution that allows an agent to calculate these probabilities and correlations from their direct experience of contract outcomes, and we show that this leads to superior estimates compared to an alternative approach using multiple independent beta distributions. We then show how agents may use the sufficient statistics of this Dirichlet distribution to communicate and fuse reputation within a decentralised reputation system. Finally, we present a novel solution to the problem of rumour propagation within such systems. This solution uses the notion of private and shared information, and provides estimates consistent with a centralised reputation system, whilst maintaining the anonymity of the agents, and avoiding bias and overconfidence.

1. INTRODUCTION The role of computational models of trust within multi-agent systems in particular, and open distributed systems in general, has recently generated a great deal of research interest. In such systems, agents must typically choose between interaction partners, and in this context trust can be viewed to provide a means for agents to represent and estimate the reliability with which these interaction partners will fulfill their commitments. To date, however, much of the work within this area has used domain specific or ad-hoc trust metrics, and has focused on providing heuristics to evaluate and update these metrics using direct experience and reputation reports from other agents (see [8] for a review). Recent work has attempted to place the notion of computational trust within the framework of probability theory [6, 11]. This approach allows many of the desiderata of computational trust models to be addressed through principled means. In particular: (i) it allows agents to update their estimates of the trustworthiness of a supplier as they acquire direct experience, (ii) it provides a natural framework for agents to express their uncertainty this trustworthiness, and, (iii) it allows agents to exchange, combine and filter reputation reports received from other agents. Whilst this approach is attractive, it is somewhat limited in that it has so far only considered single dimensional outcomes (i.e. whether the contract has succeeded or failed in its entirety). However, in many real world settings the success or failure of an interaction may be decomposed into several dimensions [7]. This presents the challenge of combining these multiple dimensions into a single metric over which a decision can be made. Furthermore, these dimensions will typically also exhibit correlations. For example, a contract within a supply chain may specify criteria for timeliness, quality and quantity. A supplier who is suffering delays may attempt a trade-off between these dimensions by supplying the full amount late, or supplying as much as possible (but less than the quantity specified within the contract) on time. Thus, correlations will naturally arise between these dimensions, and hence, between the probabilities that describe the successful fulfillment of each contract dimension. To date, however, no such principled framework exists to describe these multi-dimensional contracts, nor the correlations between these dimensions (although some ad-hoc models do exist - see section 2 for more details). To rectify this shortcoming, in this paper we develop a probabilistic model of computational trust that explicitly deals with correlated multi-dimensional contracts. The starting point for our work is to consider how an agent can estimate the utility that it will derive from interacting with a supplier. Here we use standard approaches from the literature of data fusion (since this is a well developed field where the notion of multi-dimensional correlated estimates is well established1 ) to show that this naturally leads to a trust model where the agent must estimate probabilities and correlations over 1 In this context, the multiple dimensions typically represent the physical coordinates of a target being tracked, and correlations arise through the operation and orientation of sensors. 1070 978-81-904262-7-5 (RPS) c 2007 IFAAMAS multiple dimensions. Building upon this, we then devise a novel trust model that addresses the three desiderata discussed above. In more detail, in this paper we extend the state of the art in four key ways: 1. We devise a novel multi-dimensional probabilistic trust model that enables an agent to estimate the expected utility of a contract, by estimating (i) the probability that each contract dimension will be successfully fulfilled, and (ii) the correlations between these estimates. 2. We present an exact probabilistic model based upon the Dirichlet distribution that allows agents to use their direct experience of contract outcomes to calculate the probabilities and correlations described above. We then benchmark this solution and show that it leads to good estimates. 3. We show that agents can use the sufficient statistics of this Dirichlet distribution in order to exchange reputation reports with one another. The sufficient statistics represent aggregations of their direct experience, and thus, express contract outcomes in a compact format with no loss of information. 4. We show that, while being efficient, the aggregation of contract outcomes can lead to double counting, and rumour propagation, in decentralised reputation systems. Thus, we present a novel solution based upon the idea of private and shared information. We show that it yields estimates consistent with a centralised reputation system, whilst maintaining the anonymity of the agents, and avoiding overconfidence. The remainder of this paper is organised as follows: in section 2 we review related work. In section 3 we present our notation for a single dimensional contract, before introducing our multi-dimensional trust model in section 4. In sections 5 and 6 we discuss communicating reputation, and present our solution to rumour propagation in decentralised reputation systems. We conclude in section 7. 2. RELATED WORK The need for a multi-dimensional trust model has been recognised by a number of researchers. Sabater and Sierra present a model of reputation, in which agents form contracts based on multiple variables (such as delivery date and quality), and define impressions as subjective evaluations of the outcome of these contracts. They provide heuristic approaches to combining these impressions to form a measure they call subjective reputation. Likewise, Griffiths decomposes overall trust into a number of different dimensions such as success, cost, timeliness and quality [4]. In his case, each dimension is scored as a real number that represents a comparative value with no strong semantic meaning. He develops an heuristic rule to update these values based on the direct experiences of the individual agent, and an heuristic function that takes the individual trust dimensions and generates a single scalar that is then used to select between suppliers. Whilst, he comments that the trust values could have some associated confidence level, heuristics for updating these levels are not presented. Gujral et al. take a similar approach and present a trust model over multiple domain specific dimensions [5]. They define multidimensional goal requirements, and evaluate an expected payoff based on a supplier"s estimated behaviour. These estimates are, however, simple aggregations over the direct experience of several agents, and there is no measure of the uncertainty. Nevertheless, they show that agents who select suppliers based on these multiple dimensions outperform those who consider just a single one. By contrast, a number of researchers have presented more principled computational trust models based on probability theory, albeit limited to a single dimension. Jøsang and Ismail describe the Beta Reputation System whereby the reputation of an agent is compiled from the positive and negative reports from other agents who have interacted with it [6]. The beta distribution represents a natural choice for representing these binary outcomes, and it provides a principled means of representing uncertainty. Moreover, they provide a number of extensions to this initial model including an approach to exchanging reputation reports using the sufficient statistics of the beta distribution, methods to discount the opinions of agents who themselves have low reputation ratings, and techniques to deal with reputations that may change over time. Likewise, Teacy et al. use the beta distribution to describe an agent"s belief in the probability that another agent will successfully fulfill its commitments [11]. They present a formalism using a beta distribution that allows the agent to estimate this probability based upon its direct experience, and again they use the sufficient statistics of this distribution to communicate this estimate to other agents. They provide a number of extensions to this initial model, and, in particular, they consider that agents may not always truthfully report their trust estimates. Thus, they present a principled approach to detecting and removing inconsistent reports. Our work builds upon these more principled approaches. However, the starting point of our approach is to consider an agent that is attempting to estimate the expected utility of a contract. We show that estimating this expected utility requires that an agent must estimate the probability with which the supplier will fulfill its contract. In the single-dimensional case, this naturally leads to a trust model using the beta distribution (as per Jøsang and Ismail and Teacy et al.). However, we then go on to extend this analysis to multiple dimensions, where we use the natural extension of the beta distribution, namely the Dirichlet distribution, to represent the agent"s belief over multiple dimensions. 3. SINGLE-DIMENSIONAL TRUST Before presenting our multi-dimensional trust model, we first introduce the notation and formalism that we will use by describing the more familiar single dimensional case. We consider an agent who must decide whether to engage in a future contract with a supplier. This contract will lead to some outcome, o, and we consider that o = 1 if the contract is successfully fulfilled, and o = 0 if not2 . In order for the agent to make a rational decision, it should consider the utility that it will derive from this contract. We assume that in the case that the contract is successfully fulfilled, the agent derives a utility u(o = 1), otherwise it receives no utility3 . Now, given that the agent is uncertain of the reliability with which the supplier will fulfill the contract, it should consider the expected utility that it will derive, E[U], and this is given by: E[U] = p(o = 1)u(o = 1) (1) where p(o = 1) is the probability that the supplier will successfully fulfill the contract. However, whilst u(o = 1) is known by the agent, p(o = 1) is not. The best the agent can do is to determine a distribution over possible values of p(o = 1) given its direct experience of previous contract outcomes. Given that it has been able to do so, it can then determine an estimate of the expected utility4 of the contract, E[E[U]], and a measure of its uncertainty in this expected utility, Var(E[U]). This uncertainty is important since a risk averse agent may make a decision regarding a contract, 2 Note that we only consider binary contract outcomes, although extending this to partial outcomes is part of our future work. 3 Clearly this can be extended to the case where some utility is derived from an unsuccessful outcome. 4 Note that this is often called the expected expected utility, and this is the notation that we adopt here [2]. The Sixth Intl. Joint Conf. on Autonomous Agents and Multi-Agent Systems (AAMAS 07) 1071 not only on its estimate of the expected utility of the contract, but also on the probability that the expected utility will exceed some minimum amount. These two properties are given by: E[E[U]] = ˆp(o = 1)u(o = 1) (2) Var(E[U]) = Var(p(o = 1))u(o = 1)2 (3) where ˆp(o = 1) and Var(p(o = 1)) are the estimate and uncertainty of the probability that a contract will be successfully fulfilled, and are calculated from the distribution over possible values of p(o = 1) that the agent determines from its direct experience. The utility based approach that we present here provides an attractive motivation for this model of Teacy et al. [11]. Now, in the case of binary contract outcomes, the beta distribution is the natural choice to represent the distribution over possible values of p(o = 1) since within Bayesian statistics this well known to be the conjugate prior for binomial observations [3]. By adopting the beta distribution, we can calculate ˆp(o = 1) and Var(p(o = 1)) using standard results, and thus, if an agent observed N previous contracts of which n were successfully fulfilled, then: ˆp(o = 1) = n + 1 N + 2 and: Var(p(o = 1)) = (n + 1)(N − n + 1) (N + 2)2(N + 3) Note that as expected, the greater the number of contracts the agent observes, the smaller the variance term Var(p(o = 1)), and, thus, the less the uncertainty regarding the probability that a contract will be successfully fulfilled, ˆp(o = 1). 4. MULTI-DIMENSIONAL TRUST We now extend the description above, to consider contracts between suppliers and agents that are represented by multiple dimensions, and hence the success or failure of a contract can be decomposed into the success or failure of each separate dimension. Consider again the example of the supply chain that specifies the timeliness, quantity, and quality of the goods that are to be delivered. Thus, within our trust model oa = 1 now indicates a successful outcome over dimension a of the contract and oa = 0 indicates an unsuccessful one. A contract outcome, X, is now composed of a vector of individual contract part outcomes (e.g. X = {oa = 1, ob = 0, oc = 0, . . .}). Given a multi-dimensional contract whose outcome is described by the vector X, we again consider that in order for an agent to make a rational decision, it should consider the utility that it will derive from this contract. To this end, we can make the general statement that the expected utility of a contract is given by: E[U] = p(X)U(X)T (4) where p(X) is a joint probability distribution over all possible contract outcomes: p(X) = ⎛ ⎜ ⎜ ⎜ ⎝ p(oa = 1, ob = 0, oc = 0, . . .) p(oa = 1, ob = 1, oc = 0, . . .) p(oa = 0, ob = 1, oc = 0, . . .) ... ⎞ ⎟ ⎟ ⎟ ⎠ (5) and U(X) is the utility derived from these possible outcomes: U(X) = ⎛ ⎜ ⎜ ⎜ ⎝ u(oa = 1, ob = 0, oc = 0, . . .) u(oa = 1, ob = 1, oc = 0, . . .) u(oa = 0, ob = 1, oc = 0, . . .) ... ⎞ ⎟ ⎟ ⎟ ⎠ (6) As before, whilst U(X) is known to the agent, the probability distribution p(X) is not. Rather, given the agent"s direct experience of the supplier, the agent can determine a distribution over possible values for p(X). In the single dimensional case, a beta distribution was the natural choice over possible values of p(o = 1). In the multi-dimensional case, where p(X) itself is a vector of probabilities, the corresponding natural choice is the Dirichlet distribution, since this is a conjugate prior for multinomial proportions [3]. Given this distribution, the agent is then able to calculate an estimate of the expected utility of a contract. As before, this estimate is itself represented by an expected value given by: E[E[U]] = ˆp(X)U(X)T (7) and a variance, describing the uncertainty in this expected utility: Var(E[U]) = U(X)Cov(p(X))U(X)T (8) where: Cov(p(X)) E[(p(X) − ˆp(X))(p(X) − ˆp(X))T ] (9) Thus, whilst the single dimensional case naturally leads to a trust model in which the agents attempt to derive an estimate of probability that a contract will be successfully fulfilled, ˆp(o = 1), along with a scalar variance that describes the uncertainty in this probability, Var(p(o = 1)), in this case, the agents must derive an estimate of a vector of probabilities, ˆp(X), along with a covariance matrix, Cov(p(X)), that represents the uncertainty in p(X) given the observed contractual outcomes. At this point, it is interesting to note that the estimate in the single dimensional case, ˆp(o = 1), has a clear semantic meaning in relation to trust; it is the agent"s belief in the probability of a supplier successfully fulfilling a contract. However, in the multi-dimensional case the agent must determine ˆp(X), and since this describes the probability of all possible contract outcomes, including those that are completely un-fulfilled, this direct semantic interpretation is not present. In the next section, we describe the exemplar utility function that we shall use in the remainder of this paper. 4.1 Exemplar Utility Function The approach described so far is completely general, in that it applies to any utility function of the form described above, and also applies to the estimation of any joint probability distribution. In the remainder of this paper, for illustrative purposes, we shall limit the discussion to the simplest possible utility function that exhibits a dependence upon the correlations between the contract dimensions. That is, we consider the case that expected utility is dependent only on the marginal probabilities of each contract dimension being successfully fulfilled, rather than the full joint probabilities: U(X) = ⎛ ⎜ ⎜ ⎜ ⎝ u(oa = 1) u(ob = 1) u(oc = 1) ... ⎞ ⎟ ⎟ ⎟ ⎠ (10) Thus, ˆp(X) is a vector estimate of the probability of each contract dimension being successfully fulfilled, and maintains the clear semantic interpretation seen in the single dimensional case: ˆp(X) = ⎛ ⎜ ⎜ ⎜ ⎝ ˆp(oa = 1) ˆp(ob = 1) ˆp(oc = 1) ... ⎞ ⎟ ⎟ ⎟ ⎠ (11) The correlations between the contract dimensions affect the uncertainty in the expected utility. To see this, consider the covariance 1072 The Sixth Intl. Joint Conf. on Autonomous Agents and Multi-Agent Systems (AAMAS 07) matrix that describes this uncertainty, Cov(p(X)), is now given by: Cov(p(X)) = ⎛ ⎜ ⎜ ⎜ ⎝ Va Cab Cac . . . Cab Vb Cbc . . . Cac Cbc Vc . . . ... ... ... ⎞ ⎟ ⎟ ⎟ ⎠ (12) In this matrix, the diagonal terms, Va, Vb and Vc, represent the uncertainties in p(oa = 1), p(ob = 1) and p(oc = 1) within p(X). The off-diagonal terms, Cab, Cac and Cbc, represent the correlations between these probabilities. In the next section, we use the Dirichlet distribution to calculate both ˆp(X) and Cov(p(X)) from an agent"s direct experience of previous contract outcomes. We first illustrate why this is necessary by considering an alternative approach to modelling multi-dimensional contracts whereby an agent na¨ıvely assumes that the dimensions are independent, and thus, it models each individually by separate beta distributions (as in the single dimensional case we presented in section 3). This is actually equivalent to setting the off-diagonal terms within the covariance matrix, Cov(p(X)), to zero. However, doing so can lead an agent to assume that its estimate of the expected utility of the contract is more accurate than it actually is. To illustrate this, consider a specific scenario with the following values: u(oa = 1) = u(ob = 1) = 1 and Va = Vb = 0.2. In this case, Var(E[U]) = 0.4(1 + Cab), and thus, if the correlation Cab is ignored then the variance in the expected utility is 0.4. However, if the contract outcomes are completely correlated then Cab = 1 and Var(E[U]) is actually 0.8. Thus, in order to have an accurate estimate of the variance of the expected contract utility, and to make a rational decision, it is essential that the agent is able to represent and calculate these correlation terms. In the next section, we describe how an agent may do so using the Dirichlet distribution. 4.2 The Dirichlet Distribution In this section, we describe how the agent may use its direct experience of previous contracts, and the standard results of the Dirichlet distribution, to determine an estimate of the probability that each contract dimension will be successful fulfilled, ˆp(X), and a measure of the uncertainties in these probabilities that expresses the correlations between the contract dimensions, Cov(p(X)). We first consider the calculation of ˆp(X) and the diagonal terms of the covariance matrix Cov(p(X)). As described above, the derivation of these results is identical to the case of the single dimensional beta distribution, where out of N contract outcomes, n are successfully fulfilled. In the multi-dimensional case, however, we have a vector {na, nb, nc, . . .} that represents the number of outcomes for which each of the individual contract dimensions were successfully fulfilled. Thus, in terms of the standard Dirichlet parameters where αa = na + 1 and α0 = N + 2, the agent can estimate the probability of this contract dimension being successfully fulfilled: ˆp(oa = 1) = αa α0 = na + 1 N + 2 and can also calculate the variance in any contract dimension: Va = αa(α0 − αa) α2 0(1 + α0) = (na + 1)(N − na + 1) (N + 2)2(N + 3) However, calculating the off-diagonal terms within Cov(p(X)) is more complex since it is necessary to consider the correlations between the contract dimensions. Thus, for each pair of dimensions (i.e. a and b), we must consider all possible combinations of contract outcomes, and thus we define nab ij as the number of contract outcomes for which both oa = i and ob = j. For example, nab 10 represents the number of contracts for which oa = 1 and ob = 0. Now, using the standard Dirichlet notation, we can define αab ij nab ij + 1 for all i and j taking values 0 and 1, and then, to calculate the cross-correlations between contract pairs a and b, we note that the Dirichlet distribution over pair-wise joint probabilities is: Prob(pab) = Kab i∈{0,1} j∈{0,1} p(oa = i, ob = j)αab ij −1 where: i∈{0,1} j∈{0,1} p(oa = i, ob = j) = 1 and Kab is a normalising constant [3]. From this we can derive pair-wise probability estimates and variances: E[p(oa = i, ob = j)] = αab ij α0 (13) V [p(oa = i, ob = j)] = αab ij (α0 − αab ij ) α2 0(1 + α0) (14) where: α0 = i∈{0,1} j∈{0,1} αab ij (15) and in fact, α0 = N + 2, where N is the total number of contracts observed. Likewise, we can express the covariance in these pairwise probabilities in similar terms: C[p(oa = i, ob = j), p(oa = m, ob = n)] = −αab ij αab mn α2 0(1 + α0) Finally, we can use the expression: p(oa = 1) = j∈{0,1} p(oa = 1, ob = j) to determine the covariance Cab. To do so, we first simplify the notation by defining V ab ij V [p(oa = i, ob = j)] and Cab ijmn C[p(oa = i, ob = j), p(oa = m, ob = n)]. The covariance for the probability of positive contract outcomes is then the covariance between j∈{0,1} p(oa = 1, ob = j) and i∈{0,1} p(oa = i, ob = 1), and thus: Cab = Cab 1001 + Cab 1101 + Cab 1011 + V ab 11 . Thus, given a set of contract outcomes that represent the agent"s previous interactions with a supplier, we may use the Dirichlet distribution to calculate the mean and variance of the probability of any contract dimension being successfully fulfilled (i.e. ˆp(oa = 1) and Va). In addition, by a somewhat more complex procedure we can also calculate the correlations between these probabilities (i.e. Cab). This allows us to calculate an estimate of the probability that any contract dimension will be successfully fulfilled, ˆp(X), and also represent the uncertainty and correlations in these probabilities by the covariance matrix, Cov(p(X)). In turn, these results may be used to calculate the estimate and uncertainty in the expected utility of the contract. In the next section we present empirical results that show that in practise this formalism yields significant improvements in these estimates compared to the na¨ıve approximation using multiple independent beta distributions. 4.3 Empirical Comparison In order to evaluate the effectiveness of our formalism, and show the importance of the off-diagonal terms in Cov(p(X)), we compare two approaches: The Sixth Intl. Joint Conf. on Autonomous Agents and Multi-Agent Systems (AAMAS 07) 1073 −1 −0.5 0 0.5 1 0.2 0.4 0.6 0.8 Correlation (ρ) Var(E[U]) Dirichlet Distribution Indepedent Beta Distributions −1 −0.5 0 0.5 1 0.5 1 1.5 2 2.5 x 10 4 Correlation (ρ) Information(I) Dirichlet Distribution Indepedent Beta Distributions Figure 1: Plots showing (i) the variance of the expected contract utility and (ii) the information content of the estimates computed using the Dirichlet distribution and multiple independent beta distributions. Results are averaged over 106 runs, and the error bars show the standard error in the mean. • Dirichlet Distribution: We use the full Dirichlet distribution, as described above, to calculate ˆp(X) and Cov(p(X)) including all its off-diagonal terms that represent the correlations between the contract dimensions. • Independent Beta Distributions: We use independent beta distributions to represent each contract dimension, in order to calculate ˆp(X), and then, as described earlier, we approximate Cov(p(X)) and ignore the correlations by setting all the off-diagonal terms to zero. We consider a two-dimensional case where u(oa = 1) = 6 and u(ob = 1) = 2, since this allows us to plot ˆp(X) and Cov(p(X)) as ellipses in a two-dimensional plane, and thus explain the differences between the two approaches. Specifically, we initially allocate the agent some previous contract outcomes that represents its direct experience with a supplier. The number of contracts is drawn uniformly between 10 and 20, and the actual contract outcomes are drawn from an arbitrary joint distribution intended to induce correlations between the contract dimensions. For each set of contracts, we use the approaches described above to calculate ˆp(X) and Cov(p(X)), and hence, the variance in the expected contract utility, Var(E[U]). In addition, we calculate a scalar measure of the information content, I, of the covariance matrix Cov(p(X)), which is a standard way of measuring the uncertainty encoded within the covariance matrix [1]. More specifically, we calculate the determinant of the inverse of the covariance matrix: I = det(Cov(p(X))−1 ) (16) and note that the larger the information content, the more precise ˆp(X) will be, and thus, the better the estimate of the expected utility that the agent is able to calculate. Finally, we use the results 0.3 0.4 0.5 0.6 0.7 0.8 0.3 0.4 0.5 0.6 0.7 0.8 p(o =1) p(o=1) a b Dirichlet Distribution Indepedent Beta Distributions Figure 2: Examples of ˆp(X) and Cov(p(X)) plotted as second standard error ellipses. presented in section 4.2 to calculate the actual correlation, ρ, associated with this particular set of contract outcomes: ρ = Cab √ VaVb (17) where Cab, Va and Vb are calculated as described in section 4.2. The results of this analysis are shown in figure 1. Here we show the values of I and Var(E[U]) calculated by the agents, plotted against the correlation of the contract outcomes, ρ, that constituted their direct experience. The results are averaged over 106 simulation runs. Note that as expected, when the dimensions of the contract outcomes are uncorrelated (i.e. ρ = 0), then both approaches give the same results. However, the value of using our formalism with the full Dirichlet distribution is shown when the correlation between the dimensions increases (either negatively or positively). As can be seen, if we approximate the Dirichlet distribution with multiple independent beta distributions, all of the correlation information contained within the covariance matrix, Cov(p(X)), is lost, and thus, the information content of the matrix is much lower. The loss of this correlation information leads the variance of the expected utility of the contract to be incorrect (either over or under estimated depending on the correlation)5 , with the exact amount of mis-estimation depending on the actual utility function chosen (i.e. the values of u(oa = 1) and u(ob = 1)). In addition, in figure 2 we illustrate an example of the estimates calculated through both methods, for a single exemplar set of contract outcomes. We represent the probability estimates, ˆp(X), and the covariance matrix, Cov(p(X)), in the standard way as an ellipse [1]. That is, ˆp(X) determines the position of the center of the ellipse, Cov(p(X)) defines its size and shape. Note that whilst the ellipse resulting from the full Dirichlet formalism accurately reflects the true distribution (samples of which are plotted as points), that calculated by using multiple independent Beta distributions (and thus ignoring the correlations) results in a much larger ellipse that does not reflect the true distribution. The larger size of this ellipse is a result of the off-diagonal terms of the covariance matrix being set to zero, and corresponds to the agent miscalculating the uncertainty in the probability of each contract dimension being fulfilled. This, in turn, leads it to miscalculate the uncertainty in the expected utility of a contract (shown in figure 1 as Var(E[U]). 5. COMMUNICATING REPUTATION Having described how an individual agent can use its own direct experience of contract outcomes in order to estimate the probabil5 Note that the plots are not smooth due to the fact that given a limited number of contract outcomes, then the mean of Va and Vb do not vary smoothly with ρ. 1074 The Sixth Intl. Joint Conf. on Autonomous Agents and Multi-Agent Systems (AAMAS 07) ity that a multi-dimensional contract will be successfully fulfilled, we now go on to consider how agents within an open multi-agent system can communicate these estimates to one another. This is commonly referred to as reputation and allows agents with limited direct experience of a supplier to make rational decisions. Both Jøsang and Ismail, and Teacy et al. present models whereby reputation is communicated between agents using the sufficient statistics of the beta distribution [6, 11]. This approach is attractive since these sufficient statistics are simple aggregations of contract outcomes (more precisely, they are simply the total number of contracts observed, N, and the number of these that were successfully fulfilled, n). Under the probabilistic framework of the beta distribution, reputation reports in this form may simply be aggregated with an agent"s own direct experience, in order to gain a more precise estimate based on a larger set of contract outcomes. We can immediately extend this approach to the multi-dimensional case considered here, by requiring that the agents exchange the sufficient statistics of the Dirichlet distribution instead of the beta distribution. In this case, for each pair of dimensions (i.e. a and b), the agents must communicate a vector of contract outcomes, N, which are the sufficient statistics of the Dirichlet distribution, given by: N =< nab ij > ∀a, b, i ∈ {0, 1}, j ∈ {0, 1} (18) Thus, an agent is able to communicate the sufficient statistics of its own Dirichlet distribution in terms of just 2d(d − 1) numbers (where d is the number of contract dimensions). For instance, in the case of three dimensions, N, is given by: N =< nab 00, nab 01, nab 10, nab 11, nac 00, nac 01, nac 10, nac 11, nbc 00, nbc 01, nbc 10, nbc 11 > and, hence, large sets of contract outcomes may be communicated within a relatively small message size, with no loss of information. Again, agents receiving these sufficient statistics may simply aggregate them with their own direct experience in order to gain a more precise estimate of the trustworthiness of a supplier. Finally, we note that whilst it is not the focus of our work here, by adopting the same principled approach as Jøsang and Ismail, and Teacy et al., many of the techniques that they have developed (such as discounting reports from unreliable agents, and filtering inconsistent reports from selfish agents) may be directly applied within this multi-dimensional model. However, we now go on to consider a new issue that arises in both the single and multi-dimensional models, namely the problems that arise when such aggregated sufficient statistics are propagated within decentralised agent networks. 6. RUMOUR PROPAGATION WITHIN REPUTATION SYSTEMS In the previous section, we described the use of sufficient statistics to communicate reputation, and we showed that by aggregating contract outcomes together into these sufficient statistics, a large number of contract outcomes can be represented and communicated in a compact form. Whilst, this is an attractive property, it can be problematic in practise, since the individual provenance of each contract outcome is lost in the aggregation. Thus, to ensure an accurate estimate, the reputation system must ensure that each observation of a contract outcome is included within the aggregated statistics no more than once. Within a centralised reputation system, where all agents report their direct experience to a trusted center, such double counting of contract outcomes is easy to avoid. However, in a decentralised reputation system, where agents communicate reputation to one another, and aggregate their direct experience with these reputation reports on-the-fly, avoiding double counting is much more difficult. a1 a2 a3 ¨ ¨¨ ¨¨ ¨¨B E T N1 N1 N1 + N2 Figure 3: Example of rumour propagation in a decentralised reputation system. For example, consider the case shown in figure 3 where three agents (a1 . . . a3), each with some direct experience of a supplier, share reputation reports regarding this supplier. If agent a1 were to provide its estimate to agents a2 and a3 in the form of the sufficient statistics of its Dirichlet distribution, then these agents can aggregate these contract outcomes with their own, and thus obtain more precise estimates. If at a later stage, agent a2 were to send its aggregate vector of contract outcomes to agent a3, then agent a3 being unaware of the full history of exchanges, may attempt to combine these contract outcomes with its own aggregated vector. However, since both vectors contain a contribution from agent a1, these will be counted twice in the final aggregated vector, and will result in a biased and overconfident estimate. This is termed rumour propagation or data incest in the data fusion literature [9]. One possible solution would be to uniquely identify the source of each contract outcome, and then propagate each vector, along with its label, through the network. Agents can thus identify identical observations that have arrived through different routes, and after removing the duplicates, can aggregate these together to form their estimates. Whilst this appears to be attractive in principle, for a number of reasons, it is not always a viable solution in practise [12]. Firstly, agents may not actually wish to have their uniquely labelled contract outcomes passed around an open system, since such information may have commercial or practical significance that could be used to their disadvantage. As such, agents may only be willing to exchange identifiable contract outcomes with a small number of other agents (perhaps those that they have some sort of reciprocal relationship with). Secondly, the fact that there is no aggregation of the contract outcomes as they pass around the network means that the message size increases over time, and the ultimate size of these messages is bounded only by the number of agents within the system (possibly an extremely large number for a global system). Finally, it may actually be difficult to assign globally agreeable, consistent, and unique labels for each agent within an open system. In the next section, we develop a novel solution to the problem of rumour propagation within decentralised reputation systems. Our solution is based on an approach developed within the area of target tracking and data fusion [9]. It avoids the need to uniquely identify an agent, it allows agents to restrict the number of other agents who they reveal their private estimates to, and yet it still allows information to propagate throughout the network. 6.1 Private and Shared Information Our solution to rumour propagation within decentralised reputation systems introduces the notion of private information that an agent knows it has not communicated to any other agent, and shared information that has been communicated to, or received from, another agent. Thus, the agent can decompose its contract outcome vector, N, into two vectors, a private one, Np, that has not been communicated to another agent, and a shared one, Ns, that has been shared with, or received from, another agent: N = Np + Ns (19) Now, whenever an agent communicates reputation, it communicates both its private and shared vectors separately. Both the origThe Sixth Intl. Joint Conf. on Autonomous Agents and Multi-Agent Systems (AAMAS 07) 1075 inating and receiving agents then update their two vectors appropriately. To understand this, consider the case that agent aα sends its private and shared contract outcome vectors, Nα p and Nα s , to agent aβ that itself has private and shared contract outcomes Nβ p and Nβ s . Each agent updates its vectors of contract outcomes according to the following procedure: • Originating Agent: Once the originating agent has sent both its shared and private contract outcome vectors to another agent, its private information is no longer private. Thus, it must remove the contract outcomes that were in its private vector, and add them into its shared vector: Nα s ← Nα s + Nα p Nα p ← ∅. • Receiving Agent: The goal of the receiving agent is to accumulate the largest number contract outcomes (since this will result in the most precise estimate) without including shared information from both itself and the other agent (since this may result in double counting of contract outcomes). It has two choices depending on the total number of contract outcomes6 within its own shared vector, Nβ s , and within that of the originating agent, Nα s . Thus, it updates its vector according to the procedure below: - Nβ s > Nα s : If the receiving agent"s shared vector represents a greater number of contract outcomes than that of the shared vector of the originating agent, then the agent combines its shared vector with the private vector of the originating agent: Nβ s ← Nβ s + Nα p Nβ p unchanged. - Nβ s < Nα s : Alternatively if the receiving agent"s shared vector represents a smaller number contract outcomes than that of the shared vector of the originating agent, then the receiving agent discards its own shared vector and forms a new one from both the private and shared vectors of the originating agent: Nβ s ← Nα s + Nα p Nβ p unchanged. In the case that Nβ s = Nα s then either option is appropriate. Once the receiving agent has updated its sets, it uses the contract outcomes within both to form its trust estimate. If agents receive several vectors simultaneously, this approach generalises to the receiving agent using the largest shared vector, and the private vectors of itself and all the originating agents to form its new shared vector. This procedure has a number of attractive properties. Firstly, since contract outcomes in an agent"s shared vector are never combined with those in the shared vector of another agent, outcomes that originated from the same agent are never combined together, and thus, rumour propagation is completely avoided. However, since the receiving agent may discard its own shared vector, and adopt the shared vector of the originating agent, information is still propagated around the network. Moreover, since contract outcomes are aggregated together within the private and shared vectors, the message size is constant and does not increase as the number of interactions increases. Finally, an agent only communicates its own private contract outcomes to its immediate neighbours. If this agent 6 Note that this may be calculated from N = nab 00 +nab 01 +nab 10 +nab 11. subsequently passes it on, it does so as unidentifiable aggregated information within its shared information. Thus, an agent may limit the number of agents with which it is willing to reveal identifiable contract outcomes, and yet these contract outcomes can still propagate within the network, and thus, improve estimates of other agents. Next, we demonstrate empirically that these properties can indeed be realised in practise. 6.2 Empirical Comparison In order to evaluate the effectiveness of this procedure we simulated random networks consisting of ten agents. Each agent has some direct experience of interacting with a supplier (as described in section 4.3). At each iteration of the simulation, it interacts with its immediate neighbours and exchanges reputation reports through the sufficient statistics of their Dirichlet distributions. We compare our solution to two of the most obvious decentralised alternatives: • Private and Shared Information: The agents follow the procedure described in the previous section. That is, they maintain separate private and shared vectors of contract outcomes, and at each iteration they communicate both these vectors to their immediate neighbours. • Rumour Propagation: The agents do not differentiate between private and shared contract outcomes. At the first iteration they communicate all of the contract outcomes that constitute their direct experience. In subsequent iterations, they propagate contract outcomes that they receive from any of the neighbours, to all their other immediate neighbours. • Private Information Only: The agents only communicate the contract outcomes that constitute their direct experience. In all cases, at each iteration, the agents use the Dirichlet distribution in order to calculate their trust estimates. We compare these three decentralised approaches to a centralised reputation system: • Centralised Reputation: All the agents pass their direct experience to a centralised reputation system that aggregates them together, and passes this estimate back to each agent. This centralised solution makes the most effective use of information available in the network. However, most real world problems demand decentralised solutions due to scalability, modularity and communication concerns. Thus, this centralised solution is included since it represents the optimal case, and allows us to benchmark our decentralised solution. The results of these comparisons are shown in figure 4. Here we show the sum of the information content of each agent"s covariance matrix (calculated as discussed earlier in section 4.3), for each of these four different approaches. We first note that where private information only is communicated, there is no change in information after the first iteration. Once each agent has received the direct experience of its immediate neighbours, no further increase in information can be achieved. This represents the minimum communication, and it exhibits the lowest total information of the four cases. Next, we note that in the case of rumour propagation, the information content increases continually, and rapidly exceeds the centralised reputation result. The fact that the rumour propagation case incorrectly exceeds this limit, indicates that it is continuously counting the same contract outcomes as they cycle around the network, in the belief that they are independent events. Finally, we note that using private and shared information represents a compromise between the private information only case and the centralised reputation case. Information is still allowed to propagate around the network, however rumours are eliminated. As before, we also plot a single instance of the trust estimates from one agent (i.e. ˆp(X) and Cov(p(X))) as a set of ellipses on a 1076 The Sixth Intl. Joint Conf. on Autonomous Agents and Multi-Agent Systems (AAMAS 07) 1 2 3 4 5 10 4 10 6 10 8 10 10 Iteration Information(I) Private & Shared Information Rumour Propagation Private Information Only Centralised Reputation Figure 4: Sum of information over all agents as a function of the communication iteration. two-dimensional plane (along with samples from the true distribution). As expected, the centralised reputation system achieves the best estimate of the true distribution, since it uses the direct experience of all agents. The private information only case shows the largest ellipse since it propagates the least information around the network. The rumour propagation case shows the smallest ellipse, but it is inconsistent with the actual distribution p(X). Thus, propagating rumours around the network and double counting contract outcomes in the belief that they are independent events, results in an overconfident estimate. However, we note that our solution, using separate vectors of private and shared information, allows us to propagate more information than the private information only case, but we completely avoid the problems of rumour propagation. Finally, we consider the effect that this has on the agents" calculation of the expected utility of the contract. We assume the same utility function as used in section 4.3 (i.e. u(oa = 1) = 6 and u(ob = 1) = 2), and in table 1 we present the estimate of the expected utility, and its standard deviation calculated for all four cases by a single agent at iteration five (after communication has ceased to have any further effect for all methods other than rumour propagation). We note that the rumour propagation case is clearly inconsistent with the centralised reputation system, since its standard deviation is too small and does not reflect the true uncertainty in the expected utility, given the contract outcomes. However, we observe that our solution represents the closest case to the centralised reputation system, and thus succeeds in propagating information throughout the network, whilst also avoiding bias and overconfidence. The exact difference between it and the centralised reputation system depends upon the topology of the network, and the history of exchanges that take place within it. 7. CONCLUSIONS In this paper we addressed the need for a principled probabilistic model of computational trust that deals with contracts that have multiple correlated dimensions. Our starting point was an agent estimating the expected utility of a contract, and we showed that this leads to a model of computational trust that uses the Dirichlet distribution to calculate a trust estimate from the direct experience of an agent. We then showed how agents may use the sufficient statistics of this Dirichlet distribution to represent and communicate reputation within a decentralised reputation system, and we presented a solution to rumour propagation within these systems. Our future work in this area is to extend the exchange of reputation to the case where contracts are not homogeneous. That is, not all agents observe the same contract dimensions. This is a challenging extension, since in this case, the sufficient statistics of the Dirichlet distribution can not be used directly. However, by 0.2 0.3 0.4 0.5 0.6 0.7 0.1 0.2 0.3 0.4 0.5 0.6 0.7 p(o =1) p(o=1) a b Private & Shared Information Rumour Propagation Private Information Only Centralised Reputation Figure 5: Instances of ˆp(X) and Cov(p(X)) plotted as second standard error ellipses after 5 communication iterations. Method E[E[U]] ± Var(E[U]) Private and Shared Information 3.18 ± 0.54 Rumour Propagation 3.33 ± 0.07 Private Information Only 3.20 ± 0.65 Centralised Reputation 3.17 ± 0.42 Table 1: Estimated expected utility and its standard error as calculated by a single agent after 5 communication iterations. addressing this challenge, we hope to be able to apply these techniques to a setting in which a suppliers provides a range of services whose failures are correlated, and agents only have direct experiences of different subsets of these services. 8. ACKNOWLEDGEMENTS This research was undertaken as part of the ALADDIN (Autonomous Learning Agents for Decentralised Data and Information Networks) project and is jointly funded by a BAE Systems and EPSRC strategic partnership (EP/C548051/1). 9. REFERENCES [1] Y. Bar-Shalom, X. R. Li, and T. Kirubarajan. Estimation with Applications to Tracking and Navigation. Wiley Interscience, 2001. [2] C. Boutilier. The foundations of expected expected utility. In Proc. of the 4th Int. Joint Conf. on on Artificial Intelligence, pages 285-290, Acapulco, Mexico, 2003. [3] M. Evans, N. Hastings, and B. Peacock. Statistical Distributions. John Wiley & Sons, Inc., 1993. [4] N. Griffiths. Task delegation using experience-based multi-dimensional trust. In Proc. of the 4th Int. Joint Conf. on Autonomous Agents and Multiagent Systems, pages 489-496, New York, USA, 2005. [5] N. Gukrai, D. DeAngelis, K. K. Fullam, and K. S. Barber. Modelling multi-dimensional trust. In Proc. of the 9th Int. Workshop on Trust in Agent Systems, Hakodate, Japan, 2006. [6] A. Jøsang and R. Ismail. The beta reputation system. In Proc. of the 15th Bled Electronic Commerce Conf., pages 324-337, Bled, Slovenia, 2002. [7] E. M. Maximilien and M. P. Singh. Agent-based trust model involving multiple qualities. In Proc. of the 4th Int. Joint Conf. on Autonomous Agents and Multiagent Systems, pages 519-526, Utrecht, The Netherlands, 2005. [8] S. D. Ramchurn, D. Hunyh, and N. R. Jennings. Trust in multi-agent systems. Knowledge Engineering Review, 19(1):1-25, 2004. [9] S. Reece and S. Roberts. Robust, low-bandwidth, multi-vehicle mapping. In Proc. of the 8th Int. Conf. on Information Fusion, Philadelphia, USA, 2005. [10] J. Sabater and C. Sierra. REGRET: A reputation model for gregarious societies. In Proc. of the 4th Workshop on Deception, Fraud and Trust in Agent Societies, pages 61-69, Montreal, Canada, 2001. [11] W. T. L. Teacy, J. Patel, N. R. Jennings, and M. Luck. TRAVOS: Trust and reputation in the context of inaccurate information sources. Autonomous Agents and Multi-Agent Systems, 12(2):183-198, 2006. [12] S. Utete. Network Management in Decentralised Sensing Systems. PhD thesis, University of Oxford, UK, 1994. The Sixth Intl. Joint Conf. on Autonomous Agents and Multi-Agent Systems (AAMAS 07) 1077

dirichlet distribution;rumour propogation;double counting;datum fusion;overconfidence;multi-dimensional trust;trust model;probability theory;reputation system;correlation;rumour propagation;heuristic;anonymity

train_I-58

An Efficient Heuristic Approach for Security Against Multiple Adversaries

In adversarial multiagent domains, security, commonly defined as the ability to deal with intentional threats from other agents, is a critical issue. This paper focuses on domains where these threats come from unknown adversaries. These domains can be modeled as Bayesian games; much work has been done on finding equilibria for such games. However, it is often the case in multiagent security domains that one agent can commit to a mixed strategy which its adversaries observe before choosing their own strategies. In this case, the agent can maximize reward by finding an optimal strategy, without requiring equilibrium. Previous work has shown this problem of optimal strategy selection to be NP-hard. Therefore, we present a heuristic called ASAP, with three key advantages to address the problem. First, ASAP searches for the highest-reward strategy, rather than a Bayes-Nash equilibrium, allowing it to find feasible strategies that exploit the natural first-mover advantage of the game. Second, it provides strategies which are simple to understand, represent, and implement. Third, it operates directly on the compact, Bayesian game representation, without requiring conversion to normal form. We provide an efficient Mixed Integer Linear Program (MILP) implementation for ASAP, along with experimental results illustrating significant speedups and higher rewards over other approaches.

1. INTRODUCTION In many multiagent domains, agents must act in order to provide security against attacks by adversaries. A common issue that agents face in such security domains is uncertainty about the adversaries they may be facing. For example, a security robot may need to make a choice about which areas to patrol, and how often [16]. However, it will not know in advance exactly where a robber will choose to strike. A team of unmanned aerial vehicles (UAVs) [1] monitoring a region undergoing a humanitarian crisis may also need to choose a patrolling policy. They must make this decision without knowing in advance whether terrorists or other adversaries may be waiting to disrupt the mission at a given location. It may indeed be possible to model the motivations of types of adversaries the agent or agent team is likely to face in order to target these adversaries more closely. However, in both cases, the security robot or UAV team will not know exactly which kinds of adversaries may be active on any given day. A common approach for choosing a policy for agents in such scenarios is to model the scenarios as Bayesian games. A Bayesian game is a game in which agents may belong to one or more types; the type of an agent determines its possible actions and payoffs. The distribution of adversary types that an agent will face may be known or inferred from historical data. Usually, these games are analyzed according to the solution concept of a Bayes-Nash equilibrium, an extension of the Nash equilibrium for Bayesian games. However, in many settings, a Nash or Bayes-Nash equilibrium is not an appropriate solution concept, since it assumes that the agents" strategies are chosen simultaneously [5]. In some settings, one player can (or must) commit to a strategy before the other players choose their strategies. These scenarios are known as Stackelberg games [6]. In a Stackelberg game, a leader commits to a strategy first, and then a follower (or group of followers) selfishly optimize their own rewards, considering the action chosen by the leader. For example, the security agent (leader) must first commit to a strategy for patrolling various areas. This strategy could be a mixed strategy in order to be unpredictable to the robbers (followers). The robbers, after observing the pattern of patrols over time, can then choose their strategy (which location to rob). Often, the leader in a Stackelberg game can attain a higher reward than if the strategies were chosen simultaneously. To see the advantage of being the leader in a Stackelberg game, consider a simple game with the payoff table as shown in Table 1. The leader is the row player and the follower is the column player. Here, the leader"s payoff is listed first. 1 2 3 1 5,5 0,0 3,10 2 0,0 2,2 5,0 Table 1: Payoff table for example normal form game. The only Nash equilibrium for this game is when the leader plays 2 and the follower plays 2 which gives the leader a payoff of 2. 311 978-81-904262-7-5 (RPS) c 2007 IFAAMAS However, if the leader commits to a uniform mixed strategy of playing 1 and 2 with equal (0.5) probability, the follower"s best response is to play 3 to get an expected payoff of 5 (10 and 0 with equal probability). The leader"s payoff would then be 4 (3 and 5 with equal probability). In this case, the leader now has an incentive to deviate and choose a pure strategy of 2 (to get a payoff of 5). However, this would cause the follower to deviate to strategy 2 as well, resulting in the Nash equilibrium. Thus, by committing to a strategy that is observed by the follower, and by avoiding the temptation to deviate, the leader manages to obtain a reward higher than that of the best Nash equilibrium. The problem of choosing an optimal strategy for the leader to commit to in a Stackelberg game is analyzed in [5] and found to be NP-hard in the case of a Bayesian game with multiple types of followers. Thus, efficient heuristic techniques for choosing highreward strategies in these games is an important open issue. Methods for finding optimal leader strategies for non-Bayesian games [5] can be applied to this problem by converting the Bayesian game into a normal-form game by the Harsanyi transformation [8]. If, on the other hand, we wish to compute the highest-reward Nash equilibrium, new methods using mixed-integer linear programs (MILPs) [17] may be used, since the highest-reward Bayes-Nash equilibrium is equivalent to the corresponding Nash equilibrium in the transformed game. However, by transforming the game, the compact structure of the Bayesian game is lost. In addition, since the Nash equilibrium assumes a simultaneous choice of strategies, the advantages of being the leader are not considered. This paper introduces an efficient heuristic method for approximating the optimal leader strategy for security domains, known as ASAP (Agent Security via Approximate Policies). This method has three key advantages. First, it directly searches for an optimal strategy, rather than a Nash (or Bayes-Nash) equilibrium, thus allowing it to find high-reward non-equilibrium strategies like the one in the above example. Second, it generates policies with a support which can be expressed as a uniform distribution over a multiset of fixed size as proposed in [12]. This allows for policies that are simple to understand and represent [12], as well as a tunable parameter (the size of the multiset) that controls the simplicity of the policy. Third, the method allows for a Bayes-Nash game to be expressed compactly without conversion to a normal-form game, allowing for large speedups over existing Nash methods such as [17] and [11]. The rest of the paper is organized as follows. In Section 2 we fully describe the patrolling domain and its properties. Section 3 introduces the Bayesian game, the Harsanyi transformation, and existing methods for finding an optimal leader"s strategy in a Stackelberg game. Then, in Section 4 the ASAP algorithm is presented for normal-form games, and in Section 5 we show how it can be adapted to the structure of Bayesian games with uncertain adversaries. Experimental results showing higher reward and faster policy computation over existing Nash methods are shown in Section 6, and we conclude with a discussion of related work in Section 7. 2. THE PATROLLING DOMAIN In most security patrolling domains, the security agents (like UAVs [1] or security robots [16]) cannot feasibly patrol all areas all the time. Instead, they must choose a policy by which they patrol various routes at different times, taking into account factors such as the likelihood of crime in different areas, possible targets for crime, and the security agents" own resources (number of security agents, amount of available time, fuel, etc.). It is usually beneficial for this policy to be nondeterministic so that robbers cannot safely rob certain locations, knowing that they will be safe from the security agents [14]. To demonstrate the utility of our algorithm, we use a simplified version of such a domain, expressed as a game. The most basic version of our game consists of two players: the security agent (the leader) and the robber (the follower) in a world consisting of m houses, 1 . . . m. The security agent"s set of pure strategies consists of possible routes of d houses to patrol (in an order). The security agent can choose a mixed strategy so that the robber will be unsure of exactly where the security agent may patrol, but the robber will know the mixed strategy the security agent has chosen. For example, the robber can observe over time how often the security agent patrols each area. With this knowledge, the robber must choose a single house to rob. We assume that the robber generally takes a long time to rob a house. If the house chosen by the robber is not on the security agent"s route, then the robber successfully robs it. Otherwise, if it is on the security agent"s route, then the earlier the house is on the route, the easier it is for the security agent to catch the robber before he finishes robbing it. We model the payoffs for this game with the following variables: • vl,x: value of the goods in house l to the security agent. • vl,q: value of the goods in house l to the robber. • cx: reward to the security agent of catching the robber. • cq: cost to the robber of getting caught. • pl: probability that the security agent can catch the robber at the lth house in the patrol (pl < pl ⇐⇒ l < l). The security agent"s set of possible pure strategies (patrol routes) is denoted by X and includes all d-tuples i =< w1, w2, ..., wd > with w1 . . . wd = 1 . . . m where no two elements are equal (the agent is not allowed to return to the same house). The robber"s set of possible pure strategies (houses to rob) is denoted by Q and includes all integers j = 1 . . . m. The payoffs (security agent, robber) for pure strategies i, j are: • −vl,x, vl,q, for j = l /∈ i. • plcx +(1−pl)(−vl,x), −plcq +(1−pl)(vl,q), for j = l ∈ i. With this structure it is possible to model many different types of robbers who have differing motivations; for example, one robber may have a lower cost of getting caught than another, or may value the goods in the various houses differently. If the distribution of different robber types is known or inferred from historical data, then the game can be modeled as a Bayesian game [6]. 3. BAYESIAN GAMES A Bayesian game contains a set of N agents, and each agent n must be one of a given set of types θn. For our patrolling domain, we have two agents, the security agent and the robber. θ1 is the set of security agent types and θ2 is the set of robber types. Since there is only one type of security agent, θ1 contains only one element. During the game, the robber knows its type but the security agent does not know the robber"s type. For each agent (the security agent or the robber) n, there is a set of strategies σn and a utility function un : θ1 × θ2 × σ1 × σ2 → . A Bayesian game can be transformed into a normal-form game using the Harsanyi transformation [8]. Once this is done, new, linear-program (LP)-based methods for finding high-reward strategies for normal-form games [5] can be used to find a strategy in the transformed game; this strategy can then be used for the Bayesian game. While methods exist for finding Bayes-Nash equilibria directly, without the Harsanyi transformation [10], they find only a 312 The Sixth Intl. Joint Conf. on Autonomous Agents and Multi-Agent Systems (AAMAS 07) single equilibrium in the general case, which may not be of high reward. Recent work [17] has led to efficient mixed-integer linear program techniques to find the best Nash equilibrium for a given agent. However, these techniques do require a normal-form game, and so to compare the policies given by ASAP against the optimal policy, as well as against the highest-reward Nash equilibrium, we must apply these techniques to the Harsanyi-transformed matrix. The next two subsections elaborate on how this is done. 3.1 Harsanyi Transformation The first step in solving Bayesian games is to apply the Harsanyi transformation [8] that converts the Bayesian game into a normal form game. Given that the Harsanyi transformation is a standard concept in game theory, we explain it briefly through a simple example in our patrolling domain without introducing the mathematical formulations. Let us assume there are two robber types a and b in the Bayesian game. Robber a will be active with probability α, and robber b will be active with probability 1 − α. The rules described in Section 2 allow us to construct simple payoff tables. Assume that there are two houses in the world (1 and 2) and hence there are two patrol routes (pure strategies) for the agent: {1,2} and {2,1}. The robber can rob either house 1 or house 2 and hence he has two strategies (denoted as 1l, 2l for robber type l). Since there are two types assumed (denoted as a and b), we construct two payoff tables (shown in Table 2) corresponding to the security agent playing a separate game with each of the two robber types with probabilities α and 1 − α. First, consider robber type a. Borrowing the notation from the domain section, we assign the following values to the variables: v1,x = v1,q = 3/4, v2,x = v2,q = 1/4, cx = 1/2, cq = 1, p1 = 1, p2 = 1/2. Using these values we construct a base payoff table as the payoff for the game against robber type a. For example, if the security agent chooses route {1,2} when robber a is active, and robber a chooses house 1, the robber receives a reward of -1 (for being caught) and the agent receives a reward of 0.5 for catching the robber. The payoffs for the game against robber type b are constructed using different values. Security agent: {1,2} {2,1} Robber a 1a -1, .5 -.375, .125 2a -.125, -.125 -1, .5 Robber b 1b -.9, .6 -.275, .225 2b -.025, -.025 -.9, .6 Table 2: Payoff tables: Security Agent vs Robbers a and b Using the Harsanyi technique involves introducing a chance node, that determines the robber"s type, thus transforming the security agent"s incomplete information regarding the robber into imperfect information [3]. The Bayesian equilibrium of the game is then precisely the Nash equilibrium of the imperfect information game. The transformed, normal-form game is shown in Table 3. In the transformed game, the security agent is the column player, and the set of all robber types together is the row player. Suppose that robber type a robs house 1 and robber type b robs house 2, while the security agent chooses patrol {1,2}. Then, the security agent and the robber receive an expected payoff corresponding to their payoffs from the agent encountering robber a at house 1 with probability α and robber b at house 2 with probability 1 − α. 3.2 Finding an Optimal Strategy Although a Nash equilibrium is the standard solution concept for games in which agents choose strategies simultaneously, in our security domain, the security agent (the leader) can gain an advantage by committing to a mixed strategy in advance. Since the followers (the robbers) will know the leader"s strategy, the optimal response for the followers will be a pure strategy. Given the common assumption, taken in [5], in the case where followers are indifferent, they will choose the strategy that benefits the leader, there must exist a guaranteed optimal strategy for the leader [5]. From the Bayesian game in Table 2, we constructed the Harsanyi transformed bimatrix in Table 3. The strategies for each player (security agent or robber) in the transformed game correspond to all combinations of possible strategies taken by each of that player"s types. Therefore, we denote X = σθ1 1 = σ1 and Q = σθ2 2 as the index sets of the security agent and robbers" pure strategies respectively, with R and C as the corresponding payoff matrices. Rij is the reward of the security agent and Cij is the reward of the robbers when the security agent takes pure strategy i and the robbers take pure strategy j. A mixed strategy for the security agent is a probability distribution over its set of pure strategies and will be represented by a vector x = (px1, px2, . . . , px|X|), where pxi ≥ 0 and P pxi = 1. Here, pxi is the probability that the security agent will choose its ith pure strategy. The optimal mixed strategy for the security agent can be found in time polynomial in the number of rows in the normal form game using the following linear program formulation from [5]. For every possible pure strategy j by the follower (the set of all robber types), max P i∈X pxiRij s.t. ∀j ∈ Q, P i∈σ1 pxiCij ≥ P i∈σ1 pxiCij P i∈X pxi = 1 ∀i∈X , pxi >= 0 (1) Then, for all feasible follower strategies j, choose the one that maximizes P i∈X pxiRij, the reward for the security agent (leader). The pxi variables give the optimal strategy for the security agent. Note that while this method is polynomial in the number of rows in the transformed, normal-form game, the number of rows increases exponentially with the number of robber types. Using this method for a Bayesian game thus requires running |σ2||θ2| separate linear programs. This is no surprise, since finding the leader"s optimal strategy in a Bayesian Stackelberg game is NP-hard [5]. 4. HEURISTIC APPROACHES Given that finding the optimal strategy for the leader is NP-hard, we provide a heuristic approach. In this heuristic we limit the possible mixed strategies of the leader to select actions with probabilities that are integer multiples of 1/k for a predetermined integer k. Previous work [14] has shown that strategies with high entropy are beneficial for security applications when opponents" utilities are completely unknown. In our domain, if utilities are not considered, this method will result in uniform-distribution strategies. One advantage of such strategies is that they are compact to represent (as fractions) and simple to understand; therefore they can be efficiently implemented by real organizations. We aim to maintain the advantage provided by simple strategies for our security application problem, incorporating the effect of the robbers" rewards on the security agent"s rewards. Thus, the ASAP heuristic will produce strategies which are k-uniform. A mixed strategy is denoted k-uniform if it is a uniform distribution on a multiset S of pure strategies with |S| = k. A multiset is a set whose elements may be repeated multiple times; thus, for example, the mixed strategy corresponding to the multiset {1, 1, 2} would take strategy 1 The Sixth Intl. Joint Conf. on Autonomous Agents and Multi-Agent Systems (AAMAS 07) 313 {1,2} {2,1} {1a, 1b} −1α − .9(1 − α), .5α + .6(1 − α) −.375α − .275(1 − α), .125α + .225(1 − α) {1a, 2b} −1α − .025(1 − α), .5α − .025(1 − α) −.375α − .9(1 − α), .125α + .6(1 − α) {2a, 1b} −.125α − .9(1 − α), −.125α + .6(1 − α) −1α − .275(1 − α), .5α + .225(1 − α) {2a, 2b} −.125α − .025(1 − α), −.125α − .025(1 − α) −1α − .9(1 − α), .5α + .6(1 − α) Table 3: Harsanyi Transformed Payoff Table with probability 2/3 and strategy 2 with probability 1/3. ASAP allows the size of the multiset to be chosen in order to balance the complexity of the strategy reached with the goal that the identified strategy will yield a high reward. Another advantage of the ASAP heuristic is that it operates directly on the compact Bayesian representation, without requiring the Harsanyi transformation. This is because the different follower (robber) types are independent of each other. Hence, evaluating the leader strategy against a Harsanyi-transformed game matrix is equivalent to evaluating against each of the game matrices for the individual follower types. This independence property is exploited in ASAP to yield a decomposition scheme. Note that the LP method introduced by [5] to compute optimal Stackelberg policies is unlikely to be decomposable into a small number of games as it was shown to be NP-hard for Bayes-Nash problems. Finally, note that ASAP requires the solution of only one optimization problem, rather than solving a series of problems as in the LP method of [5]. For a single follower type, the algorithm works the following way. Given a particular k, for each possible mixed strategy x for the leader that corresponds to a multiset of size k, evaluate the leader"s payoff from x when the follower plays a reward-maximizing pure strategy. We then take the mixed strategy with the highest payoff. We need only to consider the reward-maximizing pure strategies of the followers (robbers), since for a given fixed strategy x of the security agent, each robber type faces a problem with fixed linear rewards. If a mixed strategy is optimal for the robber, then so are all the pure strategies in the support of that mixed strategy. Note also that because we limit the leader"s strategies to take on discrete values, the assumption from Section 3.2 that the followers will break ties in the leader"s favor is not significant, since ties will be unlikely to arise. This is because, in domains where rewards are drawn from any random distribution, the probability of a follower having more than one pure optimal response to a given leader strategy approaches zero, and the leader will have only a finite number of possible mixed strategies. Our approach to characterize the optimal strategy for the security agent makes use of properties of linear programming. We briefly outline these results here for completeness, for detailed discussion and proofs see one of many references on the topic, such as [2]. Every linear programming problem, such as: max cT x | Ax = b, x ≥ 0, has an associated dual linear program, in this case: min bT y | AT y ≥ c. These primal/dual pairs of problems satisfy weak duality: For any x and y primal and dual feasible solutions respectively, cT x ≤ bT y. Thus a pair of feasible solutions is optimal if cT x = bT y, and the problems are said to satisfy strong duality. In fact if a linear program is feasible and has a bounded optimal solution, then the dual is also feasible and there is a pair x∗ , y∗ that satisfies cT x∗ = bT y∗ . These optimal solutions are characterized with the following optimality conditions (as defined in [2]): • primal feasibility: Ax = b, x ≥ 0 • dual feasibility: AT y ≥ c • complementary slackness: xi(AT y − c)i = 0 for all i. Note that this last condition implies that cT x = xT AT y = bT y, which proves optimality for primal dual feasible solutions x and y. In the following subsections, we first define the problem in its most intuititive form as a mixed-integer quadratic program (MIQP), and then show how this problem can be converted into a mixedinteger linear program (MILP). 4.1 Mixed-Integer Quadratic Program We begin with the case of a single type of follower. Let the leader be the row player and the follower the column player. We denote by x the vector of strategies of the leader and q the vector of strategies of the follower. We also denote X and Q the index sets of the leader and follower"s pure strategies, respectively. The payoff matrices R and C correspond to: Rij is the reward of the leader and Cij is the reward of the follower when the leader takes pure strategy i and the follower takes pure strategy j. Let k be the size of the multiset. We first fix the policy of the leader to some k-uniform policy x. The value xi is the number of times pure strategy i is used in the k-uniform policy, which is selected with probability xi/k. We formulate the optimization problem the follower solves to find its optimal response to x as the following linear program: max X j∈Q X i∈X 1 k Cijxi qj s.t. P j∈Q qj = 1 q ≥ 0. (2) The objective function maximizes the follower"s expected reward given x, while the constraints make feasible any mixed strategy q for the follower. The dual to this linear programming problem is the following: min a s.t. a ≥ X i∈X 1 k Cijxi j ∈ Q. (3) From strong duality and complementary slackness we obtain that the follower"s maximum reward value, a, is the value of every pure strategy with qj > 0, that is in the support of the optimal mixed strategy. Therefore each of these pure strategies is optimal. Optimal solutions to the follower"s problem are characterized by linear programming optimality conditions: primal feasibility constraints in (2), dual feasibility constraints in (3), and complementary slackness qj a − X i∈X 1 k Cijxi ! = 0 j ∈ Q. These conditions must be included in the problem solved by the leader in order to consider only best responses by the follower to the k-uniform policy x. 314 The Sixth Intl. Joint Conf. on Autonomous Agents and Multi-Agent Systems (AAMAS 07) The leader seeks the k-uniform solution x that maximizes its own payoff, given that the follower uses an optimal response q(x). Therefore the leader solves the following integer problem: max X i∈X X j∈Q 1 k Rijq(x)j xi s.t. P i∈X xi = k xi ∈ {0, 1, . . . , k}. (4) Problem (4) maximizes the leader"s reward with the follower"s best response (qj for fixed leader"s policy x and hence denoted q(x)j) by selecting a uniform policy from a multiset of constant size k. We complete this problem by including the characterization of q(x) through linear programming optimality conditions. To simplify writing the complementary slackness conditions, we will constrain q(x) to be only optimal pure strategies by just considering integer solutions of q(x). The leader"s problem becomes: maxx,q X i∈X X j∈Q 1 k Rijxiqj s.t. P i xi = kP j∈Q qj = 1 0 ≤ (a − P i∈X 1 k Cijxi) ≤ (1 − qj)M xi ∈ {0, 1, ...., k} qj ∈ {0, 1}. (5) Here, the constant M is some large number. The first and fourth constraints enforce a k-uniform policy for the leader, and the second and fifth constraints enforce a feasible pure strategy for the follower. The third constraint enforces dual feasibility of the follower"s problem (leftmost inequality) and the complementary slackness constraint for an optimal pure strategy q for the follower (rightmost inequality). In fact, since only one pure strategy can be selected by the follower, say qh = 1, this last constraint enforces that a = P i∈X 1 k Cihxi imposing no additional constraint for all other pure strategies which have qj = 0. We conclude this subsection noting that Problem (5) is an integer program with a non-convex quadratic objective in general, as the matrix R need not be positive-semi-definite. Efficient solution methods for non-linear, non-convex integer problems remains a challenging research question. In the next section we show a reformulation of this problem as a linear integer programming problem, for which a number of efficient commercial solvers exist. 4.2 Mixed-Integer Linear Program We can linearize the quadratic program of Problem 5 through the change of variables zij = xiqj, obtaining the following problem maxq,z P i∈X P j∈Q 1 k Rijzij s.t. P i∈X P j∈Q zij = k P j∈Q zij ≤ k kqj ≤ P i∈X zij ≤ k P j∈Q qj = 1 0 ≤ (a − P i∈X 1 k Cij( P h∈Q zih)) ≤ (1 − qj)M zij ∈ {0, 1, ...., k} qj ∈ {0, 1} (6) PROPOSITION 1. Problems (5) and (6) are equivalent. Proof: Consider x, q a feasible solution of (5). We will show that q, zij = xiqj is a feasible solution of (6) of same objective function value. The equivalence of the objective functions, and constraints 4, 6 and 7 of (6) are satisfied by construction. The fact that P j∈Q zij = xi as P j∈Q qj = 1 explains constraints 1, 2, and 5 of (6). Constraint 3 of (6) is satisfied because P i∈X zij = kqj. Let us now consider q, z feasible for (6). We will show that q and xi = P j∈Q zij are feasible for (5) with the same objective value. In fact all constraints of (5) are readily satisfied by construction. To see that the objectives match, notice that if qh = 1 then the third constraint in (6) implies that P i∈X zih = k, which means that zij = 0 for all i ∈ X and all j = h. Therefore, xiqj = X l∈Q zilqj = zihqj = zij. This last equality is because both are 0 when j = h. This shows that the transformation preserves the objective function value, completing the proof. Given this transformation to a mixed-integer linear program (MILP), we now show how we can apply our decomposition technique on the MILP to obtain significant speedups for Bayesian games with multiple follower types. 5. DECOMPOSITION FOR MULTIPLE ADVERSARIES The MILP developed in the previous section handles only one follower. Since our security scenario contains multiple follower (robber) types, we change the response function for the follower from a pure strategy into a weighted combination over various pure follower strategies where the weights are probabilities of occurrence of each of the follower types. 5.1 Decomposed MIQP To admit multiple adversaries in our framework, we modify the notation defined in the previous section to reason about multiple follower types. We denote by x the vector of strategies of the leader and ql the vector of strategies of follower l, with L denoting the index set of follower types. We also denote by X and Q the index sets of leader and follower l"s pure strategies, respectively. We also index the payoff matrices on each follower l, considering the matrices Rl and Cl . Using this modified notation, we characterize the optimal solution of follower l"s problem given the leaders k-uniform policy x, with the following optimality conditions: X j∈Q ql j = 1 al − X i∈X 1 k Cl ijxi ≥ 0 ql j(al − X i∈X 1 k Cl ijxi) = 0 ql j ≥ 0 Again, considering only optimal pure strategies for follower l"s problem we can linearize the complementarity constraint above. We incorporate these constraints on the leader"s problem that selects the optimal k-uniform policy. Therefore, given a priori probabilities pl , with l ∈ L of facing each follower, the leader solves the following problem: The Sixth Intl. Joint Conf. on Autonomous Agents and Multi-Agent Systems (AAMAS 07) 315 maxx,q X i∈X X l∈L X j∈Q pl k Rl ijxiql j s.t. P i xi = kP j∈Q ql j = 1 0 ≤ (al − P i∈X 1 k Cl ijxi) ≤ (1 − ql j)M xi ∈ {0, 1, ...., k} ql j ∈ {0, 1}. (7) Problem (7) for a Bayesian game with multiple follower types is indeed equivalent to Problem (5) on the payoff matrix obtained from the Harsanyi transformation of the game. In fact, every pure strategy j in Problem (5) corresponds to a sequence of pure strategies jl, one for each follower l ∈ L. This means that qj = 1 if and only if ql jl = 1 for all l ∈ L. In addition, given the a priori probabilities pl of facing player l, the reward in the Harsanyi transformation payoff table is Rij = P l∈L pl Rl ijl . The same relation holds between C and Cl . These relations between a pure strategy in the equivalent normal form game and pure strategies in the individual games with each followers are key in showing these problems are equivalent. 5.2 Decomposed MILP We can linearize the quadratic programming problem 7 through the change of variables zl ij = xiql j, obtaining the following problem maxq,z P i∈X P l∈L P j∈Q pl k Rl ijzl ij s.t. P i∈X P j∈Q zl ij = k P j∈Q zl ij ≤ k kql j ≤ P i∈X zl ij ≤ k P j∈Q ql j = 1 0 ≤ (al − P i∈X 1 k Cl ij( P h∈Q zl ih)) ≤ (1 − ql j)M P j∈Q zl ij = P j∈Q z1 ij zl ij ∈ {0, 1, ...., k} ql j ∈ {0, 1} (8) PROPOSITION 2. Problems (7) and (8) are equivalent. Proof: Consider x, ql , al with l ∈ L a feasible solution of (7). We will show that ql , al , zl ij = xiql j is a feasible solution of (8) of same objective function value. The equivalence of the objective functions, and constraints 4, 7 and 8 of (8) are satisfied by construction. The fact that P j∈Q zl ij = xi as P j∈Q ql j = 1 explains constraints 1, 2, 5 and 6 of (8). Constraint 3 of (8) is satisfied because P i∈X zl ij = kql j. Lets now consider ql , zl , al feasible for (8). We will show that ql , al and xi = P j∈Q z1 ij are feasible for (7) with the same objective value. In fact all constraints of (7) are readily satisfied by construction. To see that the objectives match, notice for each l one ql j must equal 1 and the rest equal 0. Let us say that ql jl = 1, then the third constraint in (8) implies that P i∈X zl ijl = k, which means that zl ij = 0 for all i ∈ X and all j = jl. In particular this implies that xi = X j∈Q z1 ij = z1 ij1 = zl ijl , the last equality from constraint 6 of (8). Therefore xiql j = zl ijl ql j = zl ij. This last equality is because both are 0 when j = jl. Effectively, constraint 6 ensures that all the adversaries are calculating their best responses against a particular fixed policy of the agent. This shows that the transformation preserves the objective function value, completing the proof. We can therefore solve this equivalent linear integer program with efficient integer programming packages which can handle problems with thousands of integer variables. We implemented the decomposed MILP and the results are shown in the following section. 6. EXPERIMENTAL RESULTS The patrolling domain and the payoffs for the associated game are detailed in Sections 2 and 3. We performed experiments for this game in worlds of three and four houses with patrols consisting of two houses. The description given in Section 2 is used to generate a base case for both the security agent and robber payoff functions. The payoff tables for additional robber types are constructed and added to the game by adding a random distribution of varying size to the payoffs in the base case. All games are normalized so that, for each robber type, the minimum and maximum payoffs to the security agent and robber are 0 and 1, respectively. Using the data generated, we performed the experiments using four methods for generating the security agent"s strategy: • uniform randomization • ASAP • the multiple linear programs method from [5] (to find the true optimal strategy) • the highest reward Bayes-Nash equilibrium, found using the MIP-Nash algorithm [17] The last three methods were applied using CPLEX 8.1. Because the last two methods are designed for normal-form games rather than Bayesian games, the games were first converted using the Harsanyi transformation [8]. The uniform randomization method is simply choosing a uniform random policy over all possible patrol routes. We use this method as a simple baseline to measure the performance of our heuristics. We anticipated that the uniform policy would perform reasonably well since maximum-entropy policies have been shown to be effective in multiagent security domains [14]. The highest-reward Bayes-Nash equilibria were used in order to demonstrate the higher reward gained by looking for an optimal policy rather than an equilibria in Stackelberg games such as our security domain. Based on our experiments we present three sets of graphs to demonstrate (1) the runtime of ASAP compared to other common methods for finding a strategy, (2) the reward guaranteed by ASAP compared to other methods, and (3) the effect of varying the parameter k, the size of the multiset, on the performance of ASAP. In the first two sets of graphs, ASAP is run using a multiset of 80 elements; in the third set this number is varied. The first set of graphs, shown in Figure 1 shows the runtime graphs for three-house (left column) and four-house (right column) domains. Each of the three rows of graphs corresponds to a different randomly-generated scenario. The x-axis shows the number of robber types the security agent faces and the y-axis of the graph shows the runtime in seconds. All experiments that were not concluded in 30 minutes (1800 seconds) were cut off. The runtime for the uniform policy is always negligible irrespective of the number of adversaries and hence is not shown. The ASAP algorithm clearly outperforms the optimal, multipleLP method as well as the MIP-Nash algorithm for finding the highestreward Bayes-Nash equilibrium with respect to runtime. For a 316 The Sixth Intl. Joint Conf. on Autonomous Agents and Multi-Agent Systems (AAMAS 07) Figure 1: Runtimes for various algorithms on problems of 3 and 4 houses. domain of three houses, the optimal method cannot reach a solution for more than seven robber types, and for four houses it cannot solve for more than six types within the cutoff time in any of the three scenarios. MIP-Nash solves for even fewer robber types within the cutoff time. On the other hand, ASAP runs much faster, and is able to solve for at least 20 adversaries for the three-house scenarios and for at least 12 adversaries in the four-house scenarios within the cutoff time. The runtime of ASAP does not increase strictly with the number of robber types for each scenario, but in general, the addition of more types increases the runtime required. The second set of graphs, Figure 2, shows the reward to the patrol agent given by each method for three scenarios in the three-house (left column) and four-house (right column) domains. This reward is the utility received by the security agent in the patrolling game, and not as a percentage of the optimal reward, since it was not possible to obtain the optimal reward as the number of robber types increased. The uniform policy consistently provides the lowest reward in both domains; while the optimal method of course produces the optimal reward. The ASAP method remains consistently close to the optimal, even as the number of robber types increases. The highest-reward Bayes-Nash equilibria, provided by the MIPNash method, produced rewards higher than the uniform method, but lower than ASAP. This difference clearly illustrates the gains in the patrolling domain from committing to a strategy as the leader in a Stackelberg game, rather than playing a standard Bayes-Nash strategy. The third set of graphs, shown in Figure 3 shows the effect of the multiset size on runtime in seconds (left column) and reward (right column), again expressed as the reward received by the security agent in the patrolling game, and not a percentage of the optimal Figure 2: Reward for various algorithms on problems of 3 and 4 houses. reward. Results here are for the three-house domain. The trend is that as as the multiset size is increased, the runtime and reward level both increase. Not surprisingly, the reward increases monotonically as the multiset size increases, but what is interesting is that there is relatively little benefit to using a large multiset in this domain. In all cases, the reward given by a multiset of 10 elements was within at least 96% of the reward given by an 80-element multiset. The runtime does not always increase strictly with the multiset size; indeed in one example (scenario 2 with 20 robber types), using a multiset of 10 elements took 1228 seconds, while using 80 elements only took 617 seconds. In general, runtime should increase since a larger multiset means a larger domain for the variables in the MILP, and thus a larger search space. However, an increase in the number of variables can sometimes allow for a policy to be constructed more quickly due to more flexibility in the problem. 7. SUMMARY AND RELATED WORK This paper focuses on security for agents patrolling in hostile environments. In these environments, intentional threats are caused by adversaries about whom the security patrolling agents have incomplete information. Specifically, we deal with situations where the adversaries" actions and payoffs are known but the exact adversary type is unknown to the security agent. Agents acting in the real world quite frequently have such incomplete information about other agents. Bayesian games have been a popular choice to model such incomplete information games [3]. The Gala toolkit is one method for defining such games [9] without requiring the game to be represented in normal form via the Harsanyi transformation [8]; Gala"s guarantees are focused on fully competitive games. Much work has been done on finding optimal Bayes-Nash equilbria for The Sixth Intl. Joint Conf. on Autonomous Agents and Multi-Agent Systems (AAMAS 07) 317 Figure 3: Reward for ASAP using multisets of 10, 30, and 80 elements subclasses of Bayesian games, finding single Bayes-Nash equilibria for general Bayesian games [10] or approximate Bayes-Nash equilibria [18]. Less attention has been paid to finding the optimal strategy to commit to in a Bayesian game (the Stackelberg scenario [15]). However, the complexity of this problem was shown to be NP-hard in the general case [5], which also provides algorithms for this problem in the non-Bayesian case. Therefore, we present a heuristic called ASAP, with three key advantages towards addressing this problem. First, ASAP searches for the highest reward strategy, rather than a Bayes-Nash equilibrium, allowing it to find feasible strategies that exploit the natural first-mover advantage of the game. Second, it provides strategies which are simple to understand, represent, and implement. Third, it operates directly on the compact, Bayesian game representation, without requiring conversion to normal form. We provide an efficient Mixed Integer Linear Program (MILP) implementation for ASAP, along with experimental results illustrating significant speedups and higher rewards over other approaches. Our k-uniform strategies are similar to the k-uniform strategies of [12]. While that work provides epsilon error-bounds based on the k-uniform strategies, their solution concept is still that of a Nash equilibrium, and they do not provide efficient algorithms for obtaining such k-uniform strategies. This contrasts with ASAP, where our emphasis is on a highly efficient heuristic approach that is not focused on equilibrium solutions. Finally the patrolling problem which motivated our work has recently received growing attention from the multiagent community due to its wide range of applications [4, 13]. However most of this work is focused on either limiting energy consumption involved in patrolling [7] or optimizing on criteria like the length of the path traveled [4, 13], without reasoning about any explicit model of an adversary[14]. Acknowledgments : This research is supported by the United States Department of Homeland Security through Center for Risk and Economic Analysis of Terrorism Events (CREATE). It is also supported by the Defense Advanced Research Projects Agency (DARPA), through the Department of the Interior, NBC, Acquisition Services Division, under Contract No. NBCHD030010. Sarit Kraus is also affiliated with UMIACS. 8. REFERENCES [1] R. W. Beard and T. McLain. Multiple UAV cooperative search under collision avoidance and limited range communication constraints. In IEEE CDC, 2003. [2] D. Bertsimas and J. Tsitsiklis. Introduction to Linear Optimization. Athena Scientific, 1997. [3] J. Brynielsson and S. Arnborg. Bayesian games for threat prediction and situation analysis. In FUSION, 2004. [4] Y. Chevaleyre. Theoretical analysis of multi-agent patrolling problem. In AAMAS, 2004. [5] V. Conitzer and T. Sandholm. Choosing the best strategy to commit to. In ACM Conference on Electronic Commerce, 2006. [6] D. Fudenberg and J. Tirole. Game Theory. MIT Press, 1991. [7] C. Gui and P. Mohapatra. Virtual patrol: A new power conservation design for surveillance using sensor networks. In IPSN, 2005. [8] J. C. Harsanyi and R. Selten. A generalized Nash solution for two-person bargaining games with incomplete information. Management Science, 18(5):80-106, 1972. [9] D. Koller and A. Pfeffer. Generating and solving imperfect information games. In IJCAI, pages 1185-1193, 1995. [10] D. Koller and A. Pfeffer. Representations and solutions for game-theoretic problems. Artificial Intelligence, 94(1):167-215, 1997. [11] C. Lemke and J. Howson. Equilibrium points of bimatrix games. Journal of the Society for Industrial and Applied Mathematics, 12:413-423, 1964. [12] R. J. Lipton, E. Markakis, and A. Mehta. Playing large games using simple strategies. In ACM Conference on Electronic Commerce, 2003. [13] A. Machado, G. Ramalho, J. D. Zucker, and A. Drougoul. Multi-agent patrolling: an empirical analysis on alternative architectures. In MABS, 2002. [14] P. Paruchuri, M. Tambe, F. Ordonez, and S. Kraus. Security in multiagent systems by policy randomization. In AAMAS, 2006. [15] T. Roughgarden. Stackelberg scheduling strategies. In ACM Symposium on TOC, 2001. [16] S. Ruan, C. Meirina, F. Yu, K. R. Pattipati, and R. L. Popp. Patrolling in a stochastic environment. In 10th Intl. Command and Control Research Symp., 2005. [17] T. Sandholm, A. Gilpin, and V. Conitzer. Mixed-integer programming methods for finding nash equilibria. In AAAI, 2005. [18] S. Singh, V. Soni, and M. Wellman. Computing approximate Bayes-Nash equilibria with tree-games of incomplete information. In ACM Conference on Electronic Commerce, 2004. 318 The Sixth Intl. Joint Conf. on Autonomous Agents and Multi-Agent Systems (AAMAS 07)

heuristic approach;game theory;security of agent system;decomposition for multiple adversary;bayesian game;agent system security;patrolling domain;bayesian and stackelberg game;bayes-nash equilibrium;adversarial multiagent domain;mixed-integer linear program;np-hard;agent security via approximate policy

train_I-59

An Agent-Based Approach for Privacy-Preserving Recommender Systems

Recommender Systems are used in various domains to generate personalized information based on personal user data. The ability to preserve the privacy of all participants is an essential requirement of the underlying Information Filtering architectures, because the deployed Recommender Systems have to be accepted by privacy-aware users as well as information and service providers. Existing approaches neglect to address privacy in this multilateral way. We have developed an approach for privacy-preserving Recommender Systems based on Multi-Agent System technology which enables applications to generate recommendations via various filtering techniques while preserving the privacy of all participants. We describe the main modules of our solution as well as an application we have implemented based on this approach.

1. INTRODUCTION Information Filtering (IF) systems aim at countering information overload by extracting information that is relevant for a given user out of a large body of information available via an information provider. In contrast to Information Retrieval (IR) systems, where relevant information is extracted based on search queries, IF architectures generate personalized information based on user profiles containing, for each given user, personal data, preferences, and rated items. The provided body of information is usually structured and collected in provider profiles. Filtering techniques operate on these profiles in order to generate recommendations of items that are probably relevant for a given user, or in order to determine users with similar interests, or both. Depending on the respective goal, the resulting systems constitute Recommender Systems [5], Matchmaker Systems [10], or a combination thereof. The aspect of privacy is an essential issue in all IF systems: Generating personalized information obviously requires the use of personal data. According to surveys indicating major privacy concerns of users in the context of Recommender Systems and e-commerce in general [23], users can be expected to be less reluctant to provide personal information if they trust the system to be privacy-preserving with regard to personal data. Similar considerations also apply to the information provider, who may want to control the dissemination of the provided information, and to the provider of the filtering techniques, who may not want the details of the utilized filtering algorithms to become common knowledge. A privacy-preserving IF system should therefore balance these requirements and protect the privacy of all parties involved in a multilateral way, while addressing general requirements regarding performance, security and quality of the recommendations as well. As described in the following section, there are several approaches with similar goals, but none of these provide a generic approach in which the privacy of all parties is preserved. We have developed an agent-based approach for privacypreserving IF which has been utilized for realizing a combined Recommender/Matchmaker System as part of an application supporting users in planning entertainment-related activities. In this paper, we focus on the Recommender System functionality. Our approach is based on Multi-Agent System (MAS) technology because fundamental features of agents such as autonomy, adaptability and the ability to communicate are essential requirements of our approach. In other words, the realized approach does not merely constitute a solution for privacy-preserving IF within a MAS context, but rather utilizes a MAS architecture in order to realize a solution for privacy-preserving IF, which could not be realized easily otherwise. The paper is structured as follows: Section 2 describes related work. Section 3 describes the general ideas of our approach. In Section 4, we describe essential details of the 319 978-81-904262-7-5 (RPS) c 2007 IFAAMAS modules of our approach and their implementation. In Section 5, we evaluate the approach, mainly via the realized application. Section 6 concludes the paper with an outlook and outlines further work. 2. RELATED WORK There is a large amount of work in related areas, such as Private Information Retrieval [7], Privacy-Preserving Data Mining [2], and other privacy-preserving protocols [4, 16], most of which is based on Secure Multi-Party Computation [27]. We have ruled out Secure Multi-Party Computation approaches mainly because of their complexity, and because the algorithm that is computed securely is not considered to be private in these approaches. Various enforcement mechanisms have been suggested that are applicable in the context of privacy-preserving Information Filtering, such as enterprise privacy policies [17] or hippocratic databases [1], both of which annotate user data with additional meta-information specifying how the data is to be handled on the provider side. These approaches ultimately assume that the provider actually intends to protect the privacy of the user data, and offer support for this task, but they are not intended to prevent the provider from acting in a malicious manner. Trusted computing, as specified by the Trusted Computing Group, aims at realizing trusted systems by increasing the security of open systems to a level comparable with the level of security that is achievable in closed systems. It is based on a combination of tamper-proof hardware and various software components. Some example applications, including peer-to-peer networks, distributed firewalls, and distributed computing in general, are listed in [13]. There are some approaches for privacy-preserving Recommender Systems based on distributed collaborative filtering, in which recommendations are generated via a public model aggregating the distributed user profiles without containing explicit information about user profiles themselves. This is achieved via Secure Multi-Party Computation [6], or via random perturbation of the user data [20]. In [19], various approaches are integrated within a single architecture. In [10], an agent-based approach is described in which user agents representing similar users are discovered via a transitive traversal of user agents. Privacy is preserved through pseudonymous interaction between the agents and through adding obfuscating data to personal information. More recent related approaches are described in [18]. In [3], an agent-based architecture for privacy-preserving demographic filtering is described which may be generalized in order to support other kinds of filtering techniques. While in some aspects similar to our approach, this architecture addresses at least two aspects inadequately, namely the protection of the filter against manipulation attempts, and the prevention of collusions between the filter and the provider. 3. PRIVACY-PRESERVING INFORMATION FILTERING We identify three main abstract entities participating in an information filtering process within a distributed system: A user entity, a provider entity, and a filter entity. Whereas in some applications the provider and filter entities explicitly trust each other, because they are deployed by the same party, our solution is applicable more generically because it does not require any trust between the main abstract entities. In this paper, we focus on aspects related to the information filtering process itself, and omit all aspects related to information collection and processing, i.e. the stages in which profiles are generated and maintained, mainly because these stages are less critical with regard to privacy, as they involve fewer different entities. 3.1 Requirements Our solution aims at meeting the following requirements with regard to privacy: • User Privacy: No linkable information about user profiles should be acquired permanently by any other entity or external party, including other user entities. Single user profile items, however, may be acquired permanently if they are unlinkable, i.e. if they cannot be attributed to a specific user or linked to other user profile items. Temporary acquisition of private information is permitted as well. Sets of recommendations may be acquired permanently by the provider, but they should not be linkable to a specific user. These concessions simplify the resulting protocol and allow the provider to obtain recommendations and single unlinkable user profile items, and thus to determine frequently requested information and optimize the offered information accordingly. • Provider Privacy: No information about provider profiles, with the exception of the recommendations, should be acquired permanently by other entities or external parties. Again, temporary acquisition of private information is permitted. Additionally, the propagation of provider information is entirely under the control of the provider. Thus, the provider is enabled to prevent misuse such as the automatic large-scale extraction of information. • Filter Privacy: Details of the algorithms applied by the filtering techniques should not be acquired permanently by any other entity or external party. General information about the algorithm may be provided by the filter entity in order to help other entities to reach a decision on whether to apply the respective filtering technique. In addition, general requirements regarding the quality of the recommendations as well as security aspects, performance and broadness of the resulting system have to be addressed as well. While minor trade-offs may be acceptable, the resulting system should reach a level similar to regular Recommender Systems with regard to these requirements. 3.2 Outline of the Solution The basic idea for realizing a protocol fulfilling these privacy-related requirements in Recommender Systems is implied by allowing the temporary acquisition of private information (see [8] for the original approach): User and provider entity both propagate the respective profile data to the filter entity. The filter entity provides the recommendations, and subsequently deletes all private information, thus fulfilling the requirement regarding permanent acquisition of private information. 320 The Sixth Intl. Joint Conf. on Autonomous Agents and Multi-Agent Systems (AAMAS 07) The entities whose private information is propagated have to be certain that the respective information is actually acquired temporarily only. Trust in this regard may be established in two main ways: • Trusted Software: The respective entity itself is trusted to remove the respective information as specified. • Trusted Environment: The respective entity operates in an environment that is trusted to control the communication and life cycle of the entity to an extent that the removal of the respective information may be achieved regardless of the attempted actions of the entity itself. Additionally, the environment itself is trusted not to act in a malicious manner (e.g. it is trusted not to acquire and propagate the respective information itself). In both cases, trust may be established in various ways. Reputation-based mechanisms, additional trusted third parties certifying entities or environments, or trusted computing mechanisms may be used. Our approach is based on a trusted environment realized via trusted computing mechanisms, because we see this solution as the most generic and realistic approach. This decision is discussed briefly in Section 5. We are now able to specify the abstract information filtering protocol as shown in Figure 1: The filter entity deploys a Temporary Filter Entity (TFE) operating in a trusted environment. The user entity deploys an additional relay entity operating in the same environment. Through mechanisms provided by this environment, the relay entity is able to control the communication of the TFE, and the provider entity is able to control the communication of both relay entity and the TFE. Thus, it is possible to ensure that the controlled entities are only able to propagate recommendations, but no other private information. In the first stage (steps 1.1 to 1.3 of Figure 1), the relay entity establishes control of the TFE, and thus prevents it from propagating user profile information. User profile data is propagated without participation of the provider entity from the user entity to the TFE via the relay entity. In the second stage (steps 2.1 to 2.3 of Figure 1), the provider entity establishes control of both relay and TFE, and thus prevents them from propagating provider profile information. Provider profile data is propagated from the provider entity to the TFE via the relay entity. In the third stage (steps 3.1 to 3.5 of Figure 1), the TFE returns the recommendations via the relay entity, and the controlled entities are terminated. Taken together, these steps ensure that all private information is acquired temporarily only by the other main entities. The problems of determining acceptable queries on the provider profile and ensuring unlinkability of the recommendations are discussed in the following section. Our approach requires each entity in the distributed architecture to have the following five main abilities: The ability to perform certain well-defined tasks (such as carrying out a filtering process) with a high degree of autonomy, i.e. largely independent of other entities (e.g. because the respective entity is not able to communicate in an unrestricted manner), the ability to be deployable dynamically in a well-defined environment, the ability to communicate with other entities, the ability to achieve protection against external manipulation attempts, and the ability to control and restrict the communication of other entities. Figure 1: The abstract privacy-preserving information filtering protocol. All communication across the environments indicated by dashed lines is prevented with the exception of communication with the controlling entity. MAS architectures are an ideal solution for realizing a distributed system characterized by these features, because they provide agents constituting entities that are actually characterized by autonomy, mobility and the ability to communicate [26], as well as agent platforms as environments providing means to realize the security of agents. In this context, the issue of malicious hosts, i.e. hosts attacking agents, has to be addressed explicitly. Furthermore, existing MAS architectures generally do not allow agents to control the communication of other agents. It is possible, however, to expand a MAS architecture and to provide designated agents with this ability. For these reasons, our solution is based on a FIPA[11]-compliant MAS architecture. The entities introduced above are mapped directly to agents, and the trusted environment in which they exist is realized in the form of agent platforms. In addition to the MAS architecture itself, which is assumed as given, our solution consists of the following five main modules: • The Controller Module described in Section 4.1 provides functionality for controlling the communication capabilities of agents. • The Transparent Persistence Module facilitates the use of different data storage mechanisms, and provides a uniform interface for accessing persistent information, which may be utilized for monitoring critical interactions involving potentially private information e.g. as part of queries. Its description is outside the scope of this paper. • The Recommender Module, details of which are described in Section 4.2, provides Recommender System functionality. • The Matchmaker Module provides Matchmaker System functionality. It additionally utilizes social aspects of MAS technology. Its description is outside the scope of this paper. The Sixth Intl. Joint Conf. on Autonomous Agents and Multi-Agent Systems (AAMAS 07) 321 • Finally, a separate module described in Section 4.4 provides Exemplary Filtering Techniques in order to show that various restrictions imposed on filtering techniques by our approach may actually be fulfilled. The trusted environment introduced above encompasses the MAS architecture itself and the Controller Module, which have to be trusted to act in a non-malicious manner in order to rule out the possibility of malicious hosts. 4. MAIN MODULES AND IMPLEMENTATION In this section, we describe the main modules of our approach, and outline the implementation. While we have chosen a specific architecture for the implementation, the specification of the module is applicable to any FIPA-compliant MAS architecture. A module basically encompasses ontologies, functionality provided by agents via agent services, and internal functionality. Throughout this paper, {m}KX denotes a message m encrypted via a non-specified symmetric encryption scheme with a secret key KX used for encryption and decryption which is initially known only to participant X. A key KXY is a key shared by participants X and Y . A cryptographic hash function is used at various points of the protocol, i.e. a function returning a hash value h(x) for given data x that is both preimage-resistant and collision-resistant1 . We denote a set of hash values for a data set X = {x1, .., xn} as H(X) = {h(x1), .., h(xn)}, whereas h(X) denotes a single hash value of a data set. 4.1 Controller Module As noted above, the ability to control the communication of agents is generally not a feature of existing MAS architectures2 but at the same time a central requirement of our approach for privacy-preserving Information Filtering. The required functionality cannot be realized based on regular agent services or components, because an agent on a platform is usually not allowed to interfere with the actions of other agents in any way. Therefore, we add additional infrastructure providing the required functionality to the MAS architecture itself, resulting in an agent environment with extended functionality and responsibilities. Controlling the communication capabilities of an agent is realized by restricting via rules, in a manner similar to a firewall, but with the consent of the respective agent, its incoming and outgoing communication to specific platforms or agents on external platforms as well as other possible communication channels, such as the file system. Consent is required because otherwise the overall security would be compromised, as attackers could arbitrarily block various communication channels. Our approach does not require controlling the communication between agents on the same platform, and therefore this aspect is not addressed. Consequently, all rules addressing communication capabilities have to be enforced across entire platforms, because otherwise a controlled agent could just use a non-controlled agent 1 In the implementation, we have used the Advanced Encryption Standard (AES) as the symmetric encryption scheme and SHA-1 as the cryptographic hash function. 2 A recent survey on agent environments [24] concludes that aspects related to agent environments are often neglected, and does not indicate any existing work in this particular area. on the same platform as a relay for communicating with agents residing on external platforms. Various agent services provide functionality for adding and revoking control of platforms, including functionality required in complex scenarios where controlled agents in turn control further platforms. The implementation of the actual control mechanism depends on the actual MAS architecture. In our implementation, we have utilized methods provided via the Java Security Manager as part of the Java security model. Thus, the supervisor agent is enabled to define custom security policies, thereby granting or denying other agents access to resources required for communication with other agents as well as communication in general, such as files or sockets for TCP/IP-based communication. 4.2 Recommender Module The Recommender Module is mainly responsible for carrying out information filtering processes, according to the protocol described in Table 1. The participating entities are realized as agents, and the interactions as agent services. We assume that mechanisms for secure agent communication are available within the respective MAS architecture. Two issues have to be addressed in this module: The relevant parts of the provider profile have to be retrieved without compromising the user"s privacy, and the recommendations have to be propagated in a privacy-preserving way. Our solution is based on a threat model in which no main abstract entity may safely assume any other abstract entity to act in an honest manner: Each entity has to assume that other entities may attempt to obtain private information, either while following the specified protocol or even by deviating from the protocol. According to [15], we classify the former case as honest-but-curious behavior (as an example, the TFE may propagate recommendations as specified, but may additionally attempt to propagate private information), and the latter case as malicious behavior (as an example, the filter may attempt to propagate private information instead of the recommendations). 4.2.1 Retrieving the Provider Profile As outlined above, the relay agent relays data between the TFE agent and the provider agent. These agents are not allowed to communicate directly, because the TFE agent cannot be assumed to act in an honest manner. Unlike the user profile, which is usually rather small, the provider profile is often too voluminous to be propagated as a whole efficiently. A typical example is a user profile containing ratings of about 100 movies, while the provider profile contains some 10,000 movies. Retrieving only the relevant part of the provider profile, however, is problematic because it has to be done without leaking sensitive information about the user profile. Therefore, the relay agent has to analyze all queries on the provider profile, and reject potentially critical queries, such as queries containing a set of user profile items. Because the propagation of single unlinkable user profile items is assumed to be uncritical, we extend the information filtering protocol as follows: The relevant parts of the provider profile are retrieved based on single anonymous interactions between the relay and the provider. If the MAS architecture used for the implementation does not provide an infrastructure for anonymous agent communication, this feature has to be provided explicitly: The most straightforward way is to use additional relay agents deployed via 322 The Sixth Intl. Joint Conf. on Autonomous Agents and Multi-Agent Systems (AAMAS 07) Table 1: The basic information filtering protocol with participants U = user agent, P = provider agent, F = TFE agent, R = relay agent, based on the abstract protocol shown in Figure 1. UP denotes the user profile with UP = {up1, .., upn}, PP denotes the provider profile, and REC denotes the set of recommendations with REC = {rec1, .., recm}. Phase. Sender → Message or Action Step Receiver 1.1 R → F establish control 1.2 U → R UP 1.3 R → F UP 2.1 P → R, F establish control 2.2 P → R PP 2.3 R → F PP 3.1 F → R REC 3.2 R → P REC 3.3 P → U REC 3.4 R → F terminate F 3.5 P → R terminate R the main relay agent and used once for a single anonymous interaction. Obviously, unlinkability is only achieved if multiple instances of the protocol are executed simultaneously between the provider and different users. Because agents on controlled platforms are unable to communicate anonymously with the respective controlling agent, control has to be established after the anonymous interactions have been completed. To prevent the uncontrolled relay agents from propagating provider profile data, the respective data is encrypted and the key is provided only after control has been established. Therefore, the second phase of the protocol described in Table 1 is replaced as described in Table 2. Additionally, the relay agent may allow other interactions as long as no user profile items are used within the queries. In this case, the relay agent has to ensure that the provider does not obtain any information exceeding the information deducible via the recommendations themselves. The clusterbased filtering technique described in Section 4.3 is an example for a filtering technique operating in this manner. 4.2.2 Recommendation Propagation The propagation of the recommendations is even more problematic mainly because more participants are involved: Recommendations have to be propagated from the TFE agent via the relay and provider agent to the user agent. No participant should be able to alter the recommendations or use them for the propagation of private information. Therefore, every participant in this chain has to obtain and verify the recommendations in unencrypted form prior to the next agent in the chain, i.e. the relay agent has to verify the recommendations before the provider obtains them, and so on. Therefore, the final phase of the protocol described in Table 1 is replaced as described in Table 3. It basically consists of two parts (Step 3.1 to 3.4, and Step 3.5 to Step 3.8), each of which provide a solution for a problem related to the prisoners" problem [22], in which two participants (the prisoners) intend to exchange a message via a third, untrusted participant (the warden) who may read the message but must not be able to alter it in an undetectable manner. There are various solutions for protocols addressing the prisoners" probTable 2: The updated second stage of the information filtering protocol with definitions as above. PPq is the part of the provider profile PP returned as the result of the query q. Phase. Sender → Message or Action Step Receiver repeat 2.1 to 2.3 ∀ up ∈ UP: 2.1 F → R q(up) (a query based on up) 2.2 R anon → P q(up) (R remains anonymous) 2.3 P → R anon {PPq(up)}KP 2.4 P → R, F establish control 2.5 P → R KP 2.6 R → F PPq(UP ) Table 3: The updated final stage of the information filtering protocol with definitions as above. Phase. Sender → Message or Action Step Receiver 3.1 F → R REC, {H(REC)}KPF 3.2 R → P h(KR), {{H(REC)}KPF }KR 3.3 P → R KP F 3.4 R → P KR repeat 3.5 ∀ rec ∈ REC: 3.5 R → P {rec}KURrec repeat 3.6 ∀ rec ∈ REC: 3.6 P → U h(KPrec ), {{rec}KURrec }KPrec repeat 3.7 to 3.8 ∀ rec ∈ REC: 3.7 U → P KURrec 3.8 P → U KPrec 3.9 U → F terminate F 3.10 P → U terminate U lem. The more obvious of these, however, such as protocols based on the use of digital signatures, introduce additional threats e.g. via the possibility of additional subliminal channels [22]. In order to minimize the risk of possible threats, we have decided to use a protocol that only requires a symmetric encryption scheme. The first part of the final phase is carried out as follows: In order to prevent the relay from altering recommendations, they are propagated by the filter together with an encrypted hash in Step 3.1. Thus, the relay is able to verify the recommendations before they are propagated further. The relay, however, may suspect the data propagated as the encrypted hash to contain private information instead of the actual hash value. Therefore, the encrypted hash is encrypted again and propagated together with a hash on the respective key in Step 3.2. In Step 3.3, the key KP F is revealed to the relay, allowing the relay to validate the encrypted hash. In Step 3.4, the key KR is revealed to the provider, allowing the provider to decrypt the data received in Step 3.2 and thus to obtain H(REC). Propagating the hash of the key KR prevents the relay from altering the recommendations to REC after Step 3.3, which would be undetectable otherwise because the relay could choose a key KR so that {{H(REC)}KPF }KR = {{H(REC )}KPF }KR . The encryption scheme used for encrypting the hash has to be secure against known-plaintext attacks, because otherwise the relay may be able to obtain KP F after Step 3.1 and subsequently alter the recommendations in an undetectable The Sixth Intl. Joint Conf. on Autonomous Agents and Multi-Agent Systems (AAMAS 07) 323 way. Additionally, the encryption scheme must not be commutative for similar reasons. The remaining protocol steps are interactions between relay, provider and user agent. The interactions of Step 3.5 to Step 3.8 ensure, via the same mechanisms used in Step 3.1 to 3.4, that the provider is able to analyze the recommendations before the user obtains them, but at the same time prevent the provider from altering the recommendations. Additionally, the recommendations are not processed at once, but rather one at a time, to prevent the provider from withholding all recommendations. Upon completion of the protocol, both user and provider have obtained a set of recommendations. If the user wants these recommendations to be unlinkable to himself, the user agent has to carry out the entire protocol anonymously. Again, the most straightforward way to achieve this is to use additional relay agents deployed via the user agent which are used once for a single information filtering process. 4.3 Exemplary Filtering Techniques The filtering technique applied by the TFE agent cannot be chosen freely: All collaboration-based approaches, such as collaborative filtering techniques based on the profiles of a set of users, are not applicable because the provider profile does not contain user profile data (unless this data has been collected externally). Instead, these approaches are realized via the Matchmaker Module, which is outside the scope of this paper. Learning-based approaches are not applicable because the TFE agent cannot propagate any acquired data to the filter, which effectively means that the filter is incapable of learning. Filtering techniques that are actually applicable are feature-based approaches, such as content-based filtering (in which profile items are compared via their attributes) and knowledge-based filtering (in which domain-specific knowledge is applied in order to match user and provider profile items). An overview of different classes and hybrid combinations of filtering techniques is given in [5]. We have implemented two generic content-based filtering approaches that are applicable within our approach: A direct content-based filtering technique based on the class of item-based top-N recommendation algorithms [9] is used in cases where the user profile contains items that are also contained in the provider profile. In a preprocessing stage, i.e. prior to the actual information filtering processes, a model is generated containing the k most similar items for each provider profile item. While computationally rather complex, this approach is feasible because it has to be done only once, and it is carried out in a privacy-preserving way via interactions between the provider agent and a TFE agent. The resulting model is stored by the provider agent and can be seen as an additional part of the provider profile. In the actual information filtering process, the k most similar items are retrieved for each single user profile item via queries on the model (as described in Section 4.2.1, this is possible in a privacy-preserving way via anonymous communication). Recommendations are generated by selecting the n most frequent items from the result sets that are not already contained within the user profile. As an alternative approach applicable when the user profile contains information in addition to provider profile items, we provide a cluster-based approach in which provider profile items are clustered in a preprocessing stage via an agglomerative hierarchical clustering approach. Each cluster is represented by a centroid item, and the cluster elements are either sub-clusters or, on the lowest level, the items themselves. In the information filtering stage, the relevant items are retrieved by descending through the cluster hierarchy in the following manner: The cluster items of the highest level are retrieved independent of the user profile. By comparing these items with the user profile data, the most relevant sub-clusters are determined and retrieved in a subsequent iteration. This process is repeated until the lowest level is reached, which contains the items themselves as recommendations. Throughout the process, user profile items are never propagated to the provider as such. The information deducible about the user profile does not exceed the information deducible via the recommendations themselves (because essentially only a chain of cluster centroids leading to the recommendations is retrieved), and therefore it is not regarded as privacy-critical. 4.4 Implementation We have implemented the approach for privacy-preserving IF based on JIAC IV [12], a FIPA-compliant MAS architecture. JIAC IV integrates fundamental aspects of autonomous agents regarding pro-activeness, intelligence, communication capabilities and mobility by providing a scalable component-based architecture. Additionally, JIAC IV offers components realizing management and security functionality, and provides a methodology for Agent-Oriented Software Engineering. JIAC IV stands out among MAS architectures as the only security-certified architecture, since it has been certified by the German Federal Office for Information Security according to the EAL3 of the Common Criteria for Information Technology Security standard [14]. JIAC IV offers several security features in the areas of access control for agent services, secure communication between agents, and low-level security based on Java security policies [21], and thus provides all security-related functionality required for our approach. We have extended the JIAC IV architecture by adding the mechanisms for communication control described in Section 4.1. Regarding the issue of malicious hosts, we currently assume all providers of agent platforms to be trusted. We are additionally developing a solution that is actually based on a trusted computing infrastructure. 5. EVALUATION For the evaluation of our approach, we have examined whether and to which extent the requirements (mainly regarding privacy, performance, and quality) are actually met. Privacy aspects are directly addressed by the modules and protocols described above and therefore not evaluated further here. Performance is a critical issue, mainly because of the overhead caused by creating additional agents and agent platforms for controlling communication, and by the additional interactions within the Recommender Module. Overall, a single information filtering process takes about ten times longer than a non-privacy-preserving information filtering process leading to the same results, which is a considerable overhead but still acceptable under certain conditions, as described in the following section. 5.1 The Smart Event Assistant As a proof of concept, and in order to evaluate performance and quality under real-life conditions, we have ap324 The Sixth Intl. Joint Conf. on Autonomous Agents and Multi-Agent Systems (AAMAS 07) Figure 2: The Smart Event Assistant, a privacypreserving Recommender System supporting users in planning entertainment-related activities. plied our approach within the Smart Event Assistant, a MAS-based Recommender System which integrates various personalized services for entertainment planning in different German cities, such as a restaurant finder and a movie finder [25]. Additional services, such as a calendar, a routing service and news services complement the information services. An intelligent day planner integrates all functionality by providing personalized recommendations for the various information services, based on the user"s preferences and taking into account the location of the user as well as the potential venues. All services are accessible via mobile devices as well3 . Figure 2 shows a screenshot of the intelligent day planner"s result dialog. The Smart Event Assistant is entirely realized as a MAS system providing, among other functionality, various filter agents and different service provider agents, which together with the user agents utilize the functionality provided by our approach. Recommendations are generated in two ways: A push service delivers new recommendations to the user in regular intervals (e.g. once per day) via email or SMS. Because the user is not online during these interactions, they are less critical with regard to performance and the protracted duration of the information filtering process is acceptable in this case. Recommendations generated for the intelligent day planner, however, have to be delivered with very little latency because the process is triggered by the user, who expects to receive results promptly. In this scenario, the overall performance is substantially improved by setting up the relay agent and the TFE agent offline, i.e. prior to the user"s request, and by an efficient retrieval of the relevant 3 The Smart Event Assistant may be accessed online via http://www.smarteventassistant.de. Table 4: Complexity of typical privacy-preserving (PP) vs. non-privacy-preserving (NPP) filtering processes in the realized application. In the nonprivacy-preserving version, an agent retrieves the profiles directly and propagates the result to a provider agent. scenario push day planning version NPP PP NPP PP profile size (retrieved/total amount of items) user 25/25 25/25 provider 125/10,000 500/10,000 elapsed time in filtering process (in seconds) setup n/a 2.2 n/a offline database access 0.2 0.5 0.4 0.4 profile propagation n/a 0.8 n/a 0.3 filtering algorithm 0.2 0.2 0.2 0.2 result propagation 0.1 1.1 0.1 1.1 complete time 0.5 4.8 0.7 2.0 part of the provider profile: Because the user is only interested in items, such as movies, available within a certain time period and related to specific locations, such as screenings at cinemas in a specific city, the relevant part of the provider profile is usually small enough to be propagated entirely. Because these additional parameters are not seen as privacy-critical (as they are not based on the user profile, but rather constitute a short-term information need), the relevant part of the provider profile may be propagated as a whole, with no need for complex interactions. Taken together, these improvements result in a filtering process that takes about three times as long as the respective nonprivacy-preserving filtering process, which we regard as an acceptable trade-off for the increased level of privacy. Table 4 shows the results of the performance evaluation in more detail. In these scenarios, a direct content-based filtering technique similar to the one described in Section 4.3 is applied. Because equivalent filtering techniques have been applied successfully in regular Recommender Systems [9], there are no negative consequences with regard to the quality of the recommendations. 5.2 Alternative Approaches As described in Section 3.2, our solution is based on trusted computing. There are more straightforward ways to realize privacy-preserving IF, e.g. by utilizing a centralized architecture in which the privacy-preserving providerside functionality is realized as trusted software based on trusted computing. However, we consider these approaches to be unsuitable because they are far less generic: Whenever some part of the respective software is patched, upgraded or replaced, the entire system has to be analyzed again in order to determine its trustworthiness, a process that is problematic in itself due to its complexity. In our solution, only a comparatively small part of the overall system is based on trusted computing. Because agent platforms can be utilized for a large variety of tasks, and because we see trusted computing as the most promising approach to realize secure and trusted agent environments, it seems reasonable to assume that these respective mechanisms will be generally available in the future, independent of specific solutions such as the one described here. The Sixth Intl. Joint Conf. on Autonomous Agents and Multi-Agent Systems (AAMAS 07) 325 6. CONCLUSION & FURTHER WORK We have developed an agent-based approach for privacypreserving Recommender Systems. By utilizing fundamental features of agents such as autonomy, adaptability and the ability to communicate, by extending the capabilities of agent platform managers regarding control of agent communication, by providing a privacy-preserving protocol for information filtering processes, and by utilizing suitable filtering techniques we have been able to realize an approach which actually preserves privacy in Information Filtering architectures in a multilateral way. As a proof of concept, we have used the approach within an application supporting users in planning entertainment-related activities. We envision various areas of future work: To achieve complete user privacy, the protocol should be extended in order to keep the recommendations themselves private as well. Generally, the feedback we have obtained from users of the Smart Event Assistant indicates that most users are indifferent to privacy in the context of entertainment-related personal information. Therefore, we intend to utilize the approach to realize a Recommender System in a more privacysensitive domain, such as health or finance, which would enable us to better evaluate user acceptance. 7. ACKNOWLEDGMENTS We would like to thank our colleagues Andreas Rieger and Nicolas Braun, who co-developed the Smart Event Assistant. The Smart Event Assistant is based on a project funded by the German Federal Ministry of Education and Research under Grant No. 01AK037, and a project funded by the German Federal Ministry of Economics and Labour under Grant No. 01MD506. 8. REFERENCES [1] R. Agrawal, J. Kiernan, R. Srikant, and Y. Xu. Hippocratic databases. In 28th Int"l Conf. on Very Large Databases (VLDB), Hong Kong, 2002. [2] R. Agrawal and R. Srikant. Privacy-preserving data mining. In Proc. of the ACM SIGMOD Conference on Management of Data, pages 439-450. ACM Press, May 2000. [3] E. A¨ımeur, G. Brassard, J. M. Fernandez, and F. S. Mani Onana. Privacy-preserving demographic filtering. In SAC "06: Proceedings of the 2006 ACM symposium on Applied computing, pages 872-878, New York, NY, USA, 2006. ACM Press. [4] M. Bawa, R. Bayardo, Jr., and R. Agrawal. Privacy-preserving indexing of documents on the network. In Proc. of the 2003 VLDB, 2003. [5] R. Burke. Hybrid recommender systems: Survey and experiments. User Modeling and User-Adapted Interaction, 12(4):331-370, 2002. [6] J. Canny. Collaborative filtering with privacy. In IEEE Symposium on Security and Privacy, pages 45-57, 2002. [7] B. Chor, O. Goldreich, E. Kushilevitz, and M. Sudan. Private information retrieval. In IEEE Symposium on Foundations of Computer Science, pages 41-50, 1995. [8] R. Ciss´ee. An architecture for agent-based privacy-preserving information filtering. In Proceedings of the 6th International Workshop on Trust, Privacy, Deception and Fraud in Agent Systems, 2003. [9] M. Deshpande and G. Karypis. Item-based top-N recommendation algorithms. ACM Trans. Inf. Syst., 22(1):143-177, 2004. [10] L. Foner. Political artifacts and personal privacy: The yenta multi-agent distributed matchmaking system. PhD thesis, MIT, 1999. [11] Foundation for Intelligent Physical Agents. FIPA Abstract Architecture Specification, Version L, 2002. [12] S. Fricke, K. Bsufka, J. Keiser, T. Schmidt, R. Sesseler, and S. Albayrak. Agent-based telematic services and telecom applications. Communications of the ACM, 44(4), April 2001. [13] T. Garfinkel, M. Rosenblum, and D. Boneh. Flexible OS support and applications for trusted computing. In Proceedings of HotOS-IX, May 2003. [14] T. Geissler and O. Kroll-Peters. Applying security standards to multi agent systems. In AAMAS Workshop: Safety & Security in Multiagent Systems, 2004. [15] O. Goldreich, S. Micali, and A. Wigderson. How to play any mental game. In Proc. of STOC "87, pages 218-229, New York, NY, USA, 1987. ACM Press. [16] S. Jha, L. Kruger, and P. McDaniel. Privacy preserving clustering. In ESORICS 2005, volume 3679 of LNCS. Springer, 2005. [17] G. Karjoth, M. Schunter, and M. Waidner. The platform for enterprise privacy practices: Privacy-enabled management of customer data. In PET 2002, volume 2482 of LNCS. Springer, 2003. [18] H. Link, J. Saia, T. Lane, and R. A. LaViolette. The impact of social networks on multi-agent recommender systems. In Proc. of the Workshop on Cooperative Multi-Agent Learning (ECML/PKDD "05), 2005. [19] B. N. Miller, J. A. Konstan, and J. Riedl. PocketLens: Toward a personal recommender system. ACM Trans. Inf. Syst., 22(3):437-476, 2004. [20] H. Polat and W. Du. SVD-based collaborative filtering with privacy. In Proc. of SAC "05, pages 791-795, New York, NY, USA, 2005. ACM Press. [21] T. Schmidt. Advanced Security Infrastructure for Multi-Agent-Applications in the Telematic Area. PhD thesis, Technische Universit¨at Berlin, 2002. [22] G. J. Simmons. The prisoners" problem and the subliminal channel. In D. Chaum, editor, Proc. of Crypto "83, pages 51-67. Plenum Press, 1984. [23] M. Teltzrow and A. Kobsa. Impacts of user privacy preferences on personalized systems: a comparative study. In Designing personalized user experiences in eCommerce, pages 315-332. 2004. [24] D. Weyns, H. Parunak, F. Michel, T. Holvoet, and J. Ferber. Environments for multiagent systems: State-of-the-art and research challenges. In Environments for Multiagent Systems, volume 3477 of LNCS. Springer, 2005. [25] J. Wohltorf, R. Ciss´ee, and A. Rieger. Berlintainment: An agent-based context-aware entertainment planning system. IEEE Communications Magazine, 43(6), 2005. [26] M. Wooldridge and N. R. Jennings. Intelligent agents: Theory and practice. Knowledge Engineering Review, 10(2):115-152, 1995. [27] A. Yao. Protocols for secure computation. In Proc. of IEEE FOGS "82, pages 160-164, 1982. 326 The Sixth Intl. Joint Conf. on Autonomous Agents and Multi-Agent Systems (AAMAS 07)

privacy;secure multi-party computation;privacy-preserving recommender system;information filter;recommender system;multi-agent system;distributed artificial intelligence-multiagent system;trusted software;java security model;feature-based approach;learning-based approach;trust;information search;multi-agent system technology;retrieval-information filtering

train_I-60

On the Benefits of Cheating by Self-Interested Agents in Vehicular Networks

As more and more cars are equipped with GPS and Wi-Fi transmitters, it becomes easier to design systems that will allow cars to interact autonomously with each other, e.g., regarding traffic on the roads. Indeed, car manufacturers are already equipping their cars with such devices. Though, currently these systems are a proprietary, we envision a natural evolution where agent applications will be developed for vehicular systems, e.g., to improve car routing in dense urban areas. Nonetheless, this new technology and agent applications may lead to the emergence of self-interested car owners, who will care more about their own welfare than the social welfare of their peers. These car owners will try to manipulate their agents such that they transmit false data to their peers. Using a simulation environment, which models a real transportation network in a large city, we demonstrate the benefits achieved by self-interested agents if no counter-measures are implemented.

1. INTRODUCTION As technology advances, more and more cars are being equipped with devices, which enable them to act as autonomous agents. An important advancement in this respect is the introduction of ad-hoc communication networks (such as Wi-Fi), which enable the exchange of information between cars, e.g., for locating road congestions [1] and optimal routes [15] or improving traffic safety [2]. Vehicle-To-Vehicle (V2V) communication is already onboard by some car manufactures, enabling the collaboration between different cars on the road. For example, GM"s proprietary algorithm [6], called the threat assessment algorithm, constantly calculates, in real time, other vehicles" positions and speeds, and enables messaging other cars when a collision is imminent; Also, Honda has began testing its system in which vehicles talk with each other and with the highway system itself [7]. In this paper, we investigate the attraction of being a selfish agent in vehicular networks. That is, we investigate the benefits achieved by car owners, who tamper with on-board devices and incorporate their own self-interested agents in them, which act for their benefit. We build on the notion of Gossip Networks, introduced by Shavitt and Shay [15], in which the agents can obtain road congestion information by gossiping with peer agents using ad-hoc communication. We recognize two typical behaviors that the self-interested agents could embark upon, in the context of vehicular networks. In the first behavior, described in Section 4, the objective of the self-interested agents is to maximize their own utility, expressed by their average journey duration on the road. This situation can be modeled in real life by car owners, whose aim is to reach their destination as fast as possible, and would like to have their way free of other cars. To this end they will let their agents cheat the other agents, by injecting false information into the network. This is achieved by reporting heavy traffic values for the roads on their route to other agents in the network in the hope of making the other agents believe that the route is jammed, and causing them to choose a different route. The second type of behavior, described in Section 5, is modeled by the self-interested agents" objective to cause disorder in the network, more than they are interested in maximizing their own utility. This kind of behavior could be generated, for example, by vandalism or terrorists, who aim to cause as much mayhem in the network as possible. We note that the introduction of self-interested agents to the network, would most probably motivate other agents to try and detect these agents in order to minimize their effect. This is similar, though in a different context, to the problem introduced by Lamport et al. [8] as the Byzantine Generals Problem. However, the introduction of mechanisms to deal with self-interested agents is costly and time consuming. In this paper we focus mainly on the attractiveness of selfish behavior by these agents, while we also provide some insights 327 978-81-904262-7-5 (RPS) c 2007 IFAAMAS into the possibility of detecting self-interested agents and minimizing their effect. To demonstrate the benefits achieved by self-interested agents, we have used a simulation environment, which models the transportation network in a central part of a large real city. The simulation environment is further described in Section 3. Our simulations provide insights to the benefits of self-interested agents cheating. Our findings can motivate future research in this field in order to minimize the effect of selfish-agents. The rest of this paper is organized as follows. In Section 2 we review related work in the field of self-interested agents and V2V communications. We continue and formally describe our environment and simulation settings in Section 3. Sections 4 and 5 describe the different behaviors of the selfinterested agents and our findings. Finally, we conclude the paper with open questions and future research directions. 2. RELATED WORK In their seminal paper, Lamport et al. [8] describe the Byzantine Generals problem, in which processors need to handle malfunctioning components that give conflicting information to different parts of the system. They also present a model in which not all agents are connected, and thus an agent cannot send a message to all the other agents. Dolev et al. [5] has built on this problem and has analyzed the number of faulty agents that can be tolerated in order to eventually reach the right conclusion about true data. Similar work is presented by Minsky et al. [11], who discuss techniques for constructing gossip protocols that are resilient to up to t malicious host failures. As opposed to the above works, our work focuses on vehicular networks, in which the agents are constantly roaming the network and exchanging data. Also, the domain of transportation networks introduces dynamic data, as the load of the roads is subject to change. In addition, the system in transportation networks has a feedback mechanism, since the load in the roads depends on the reports and the movement of the agents themselves. Malkhi et al. [10] present a gossip algorithm for propagating information in a network of processors, in the presence of malicious parties. Their algorithm prevents the spreading of spurious gossip and diffuses genuine data. This is done in time, which is logarithmic in the number of processes and linear in the number of corrupt parties. Nevertheless, their work assumes that the network is static and also that the agents are static (they discuss a network of processors). This is not true for transportation networks. For example, in our model, agents might gossip about heavy traffic load of a specific road, which is currently jammed, yet this information might be false several minutes later, leaving the agents to speculate whether the spreading agents are indeed malicious or not. In addition, as the agents are constantly moving, each agent cannot choose with whom he interacts and exchanges data. In the context of analyzing the data and deciding whether the data is true or not, researchers have focused on distributed reputation systems or decision mechanisms to decide whether or not to share data. Yu and Singh [18] build a social network of agents" reputations. Every agent keeps a list of its neighbors, which can be changed over time, and computes the trustworthiness of other agents by updating the current values of testimonies obtained from reliable referral chains. After a bad experience with another agent every agent decreases the rating of the "bad" agent and propagates this bad experience throughout the network so that other agents can update their ratings accordingly. This approach might be implemented in our domain to allow gossip agents to identify self-interested agents and thus minimize their effect. However, the implementation of such a mechanism is an expensive addition to the infrastructure of autonomous agents in transportation networks. This is mainly due to the dynamic nature of the list of neighbors in transportation networks. Thus, not only does it require maintaining the neighbors" list, since the neighbors change frequently, but it is also harder to build a good reputation system. Leckie et al. [9] focus on the issue of when to share information between the agents in the network. Their domain involves monitoring distributed sensors. Each agent monitors a subset of the sensors and evaluates a hypothesis based on the local measurements of its sensors. If the agent believes that a hypothesis is sufficient likely he exchanges this information with the other agents. In their domain, the goal of all the agents is to reach a global consensus about the likelihood of the hypothesis. In our domain, however, as the agents constantly move, they have many samples, which they exchange with each other. Also, the data might also vary (e.g., a road might be reported as jammed, but a few minutes later it could be free), thus making it harder to decide whether to trust the agent, who sent the data. Moreover, the agent might lie only about a subset of its samples, thus making it even harder to detect his cheating. Some work has been done in the context of gossip networks or transportation networks regarding the spreading of data and its dissemination. Datta et al. [4] focus on information dissemination in mobile ad-hoc networks (MANET). They propose an autonomous gossiping algorithm for an infrastructure-less mobile ad-hoc networking environment. Their autonomous gossiping algorithm uses a greedy mechanism to spread data items in the network. The data items are spread to immediate neighbors that are interested in the information, and avoid ones that are not interested. The decision which node is interested in the information is made by the data item itself, using heuristics. However, their work concentrates on the movement of the data itself, and not on the agents who propagate the data. This is different from our scenario in which each agent maintains the data it has gathered, while the agent itself roams the road and is responsible (and has the capabilities) for spreading the data to other agents in the network. Das et al. [3] propose a cooperative strategy for content delivery in vehicular networks. In their domain, peers download a file from a mesh and exchange pieces of the file among themselves. We, on the other hand, are interested in vehicular networks in which there is no rule forcing the agents to cooperate among themselves. Shibata et al. [16] propose a method for cars to cooperatively and autonomously collect traffic jam statistics to estimate arrival time to destinations for each car. The communication is based on IEEE 802.11, without using a fixed infrastructure on the ground. While we use the same domain, we focus on a different problem. Shibata et al. [16] mainly focus on efficiently broadcasting the data between agents (e.g., avoid duplicates and communication overhead), as we focus on the case where agents are not cooperative in 328 The Sixth Intl. Joint Conf. on Autonomous Agents and Multi-Agent Systems (AAMAS 07) nature, and on how selfish agents affect other agents and the network load. Wang et al. [17] also assert, in the context of wireless networks, that individual agents are likely to do what is most beneficial for their owners, and will act selfishly. They design a protocol for communication in networks in which all agents are selfish. Their protocol motivates every agent to maximize its profit only when it behaves truthfully (a mechanism of incentive compatibility). However, the domain of wireless networks is quite different from the domain of transportation networks. In the wireless network, the wireless terminal is required to contribute its local resources to transmit data. Thus, Wang et al. [17] use a payment mechanism, which attaches costs to terminals when transmitting data, and thus enables them to maximize their utility when transmitting data, instead of acting selfishly. Unlike this, in the context of transportation networks, constructing such a mechanism is not quite a straightforward task, as self-interested agents and regular gossip agents might incur the same cost when transmitting data. The difference between the two types of agents only exists regarding the credibility of the data they exchange. In the next section, we will describe our transportation network model and gossiping between the agents. We will also describe the different agents in our system. 3. MODEL AND SIMULATIONS We first describe the formal transportation network model, and then we describe the simulations designs. 3.1 Formal Model Following Shavitt and Shay [15] and Parshani [13], the transportation network is represented by a directed graph G(V, E), where V is the set of vertices representing junctions, and E is the set of edges, representing roads. An edge e ∈ E is associated with a weight w > 0, which specifies the time it takes to traverse the road associated with that edge. The roads" weights vary in time according to the network (traffic) load. Each car, which is associated with an autonomous agent, is given a pair of origin and destination points (vertices). A journey is defined as the (not necessarily simple) path taken by an agent between the origin vertex and the destination vertex. We assume that there is always a path between a source and a destination. A journey length is defined as the sum of all weights of the edges constituting this path. Every agent has to travel between its origin and destination points and aims to minimize its journey length. Initially, agents are ignorant about the state of the roads. Regular agents are only capable of gathering information about the roads as they traverse them. However, we assume that some agents have means of inter-vehicle communication (e.g., IEEE 802.11) with a given communication range, which enables them to communicate with other agents with the same device. Those agents are referred to as gossip agents. Since the communication range is limited, the exchange of information using gossiping is done in one of two ways: (a) between gossip agents passing one another, or (b) between gossip agents located at the same junction. We assume that each agent stores the most recent information it has received or gathered around the edges in the network. A subset of the gossip agents are those agents who are selfinterested and manipulate the devices for their own benefit. We will refer to these agents as self-interested agents. A detailed description of their behavior is given in Sections 4 and 5. 3.2 Simulation Design Building on [13], the network in our simulations replicates a central part of a large city, and consists of 50 junctions and 150 roads, which are approximately the number of main streets in the city. Each simulation consists of 6 iterations. The basic time unit of the iteration is a step, which equivalents to about 30 seconds. Each iteration simulates six hours of movements. The average number of cars passing through the network during the iteration is about 70,000 and the average number of cars in the network at a specific time unit is about 3,500 cars. In each iteration the same agents are used with the same origin and destination points, whereas the data collected in earlier iterations is preserved in the future iterations (referred to as the history of the agent). This allows us to simulate somewhat a daily routine in the transportation network (e.g., a working week). Each of the experiments that we describe below is run with 5 different traffic scenarios. Each such traffic scenario differs from one another by the initial load of the roads and the designated routes of the agents (cars) in the network. For each such scenario 5 simulations are run, creating a total of 25 simulations for each experiment. It has been shown by Parshani et al. [13, 14] that the information propagation in the network is very efficient when the percentage of gossiping agents is 10% or more. Yet, due to congestion caused by too many cars rushing to what is reported as the less congested part of the network 20-30% of gossiping agents leads to the most efficient routing results in their experiments. Thus, in our simulation, we focus only on simulations in which the percentage of gossip agents is 20%. The simulations were done with different percentages of self-interested agents. To gain statistical significance we ran each simulation with changes in the set of the gossip agents, and the set of the self-interested agents. In order to gain a similar ordinal scale, the results were normalized. The normalized values were calculated by comparing each agent"s result to his results when the same scenario was run with no self-interested agents. This was done for all of the iterations. Using the normalized values enabled us to see how worse (or better) each agent would perform compared to the basic setting. For example, if an average journey length of a certain agent in iteration 1 with no selfinterested agent was 50, and the length was 60 in the same scenario and iteration in which self-interested agents were involved, then the normalized value for that agent would be 60/50 = 1.2. More details regarding the simulations are described in Sections 4 and 5. 4. SPREADING LIES, MAXIMIZING UTILITY In the first set of experiments we investigated the benefits achieved by the self-interested agents, whose aim was to minimize their own journey length. The self-interested agents adopted a cheating approach, in which they sent false data to their peers. In this section we first describe the simulations with the self-interested agents. Then, we model the scenario as a The Sixth Intl. Joint Conf. on Autonomous Agents and Multi-Agent Systems (AAMAS 07) 329 game with two types of agents, and prove that the equilibrium result can only be achieved when there is no efficient exchange of gossiping information in the network. 4.1 Modeling the Self-Interested Agents" Behavior While the gossip agents gather data and send it to other agents, the self-interested agents" behavior is modeled as follows: 1. Calculate the shortest path from origin to destination. 2. Communicate the following data to other agents: (a) If the road is not in the agent"s route - send the true data about it (e.g., data about roads it has received from other agents) (b) For all roads in the agent"s route, which the agent has not yet traversed, send a random high weight. Basically, the self-interested agent acts the same as the gossip agent. It collects data regarding the weight of the roads (either by traversing the road or by getting the data from other agents) and sends the data it has collected to other agents. However, the self-interested agent acts differently when the road is in its route. Since the agent"s goal is to reach its destination as fast as possible, the agent will falsely report that all the roads in its route are heavily congested. This is in order to free the path for itself, by making other agents recalculate their paths, this time without including roads on the self-interested agent"s route. To this end, for all the roads in its route, which the agent has not yet passed, the agent generates a random weight, which is above the average weight of the roads in the network. It then associates these new weights with the roads in its route and sends them to the other agents. While an agent can also divert cars from its route by falsely reporting congested roads in parallel to its route as free, this behavior is not very likely since other agents, attempting to use the roads, will find the mistake within a short time and spread the true congestion on the road. On the other hand, if an agent manages to persuade other agents not to use a road, it will be harder for them to detect that the said roads are not congested. In addition, to avoid being influenced by its own lies and other lies spreading in the network, all self-interested agents will ignore data received about roads with heavy traffic (note that data about roads that are not heavily traffic will not be ignored)1 . In the next subsection we describe the simulation results, involving the self-interested agents. 4.2 Simulation Results To test the benefits of cheating by the self-interested agents we ran several experiments. In the first set of experiments, we created a scenario, in which a small group of self-interested agents spread lies on the same route, and tested its effect on the journey length of all the agents in the network. 1 In other simulations we have run, in which there had been several real congestions in the network, we indeed saw that even when the roads are jammed, the self-interested agents were less affected if they ignored all reported heavy traffic, since by such they also discarded all lies roaming the network Table 1: Normalized journey length values, selfinterested agents with the same route Iteration Self-Interested Gossip - Gossip - Regular Number Agents SR Others Agents 1 1.38 1.27 1.06 1.06 2 0.95 1.56 1.18 1.14 3 1.00 1.86 1.28 1.17 4 1.06 2.93 1.35 1.16 5 1.13 2.00 1.40 1.17 6 1.08 2.02 1.43 1.18 Thus, several cars, which had the same origin and destination points, were designated as self-interested agents. In this simulation, we selected only 6 agents to be part of the group of the self-interested agents, as we wanted to investigate the effect achieved by only a small number of agents. In each simulation in this experiment, 6 different agents were randomly chosen to be part of the group of self-interested agents, as described above. In addition, one road, on the route of these agents, was randomly selected to be partially blocked, letting only one car go through that road at each time step. About 8,000 agents were randomly selected as regular gossip agents, and the other 32,000 agents were designated as regular agents. We analyzed the average journey length of the self-interested agents as opposed to the average journey length of other regular gossip agents traveling along the same route. Table 1 summarizes the normalized results for the self-interested agents, the gossip agents (those having the same origin and destination points as the self-interested agents, denoted Gossip - SR, and all other gossip agents, denoted Gossip - Others) and the regular agents, as a function of the iteration number. We can see from the results that the first time the selfinterested agents traveled the route while spreading the false data about the roads did not help them (using the paired t-test we show that those agents had significantly lower journey lengths in the scenario in which they did not spread any lies, with p < 0.01). This is mainly due to the fact that the lies do not bypass the self-interested agent and reach other cars that are ahead of the self-interested car on the same route. Thus, spreading the lies in the first iteration does not help the self-interested agent to free the route he is about to travel, in the first iteration. Only when the self-interested agents had repeated their journey in the next iteration (iteration 2) did it help them significantly (p = 0.04). The reason for this is that other gossip agents received this data and used it to recalculate their shortest path, thus avoiding entrance to the roads, for which the self-interested agents had spread false information about congestion. It is also interesting to note the large value attained by the self-interested agents in the first iteration. This is mainly due to several self-interested agents, who entered the jammed road. This situation occurred since the self-interested agents ignored all heavy traffic data, and thus ignored the fact that the road was jammed. As they started spreading lies about this road, more cars shifted from this route, thus making the road free for the future iterations. However, we also recall that the self-interested agents ignore all information about the heavy traffic roads. Thus, 330 The Sixth Intl. Joint Conf. on Autonomous Agents and Multi-Agent Systems (AAMAS 07) Table 2: Normalized journey length values, spreading lies for a beneficiary agent Iteration Beneficiary Gossip - Gossip - Regular Number Agent SR Others Agents 1 1.10 1.05 0.94 1.11 2 1.09 1.14 0.99 1.14 3 1.04 1.19 1.02 1.14 4 1.03 1.26 1.03 1.14 5 1.05 1.32 1.05 1.12 6 0.92 1.40 1.06 1.11 when the network becomes congested, more self-interested cars are affected, since they might enter jammed roads, which they would otherwise not have entered. This can be seen, for example, in iterations 4-6, in which the normalized value of the self-interested agents increased above 1.00. Using the paired t-test to compare these values with the values achieved by these agents when no lies are used, we see that there is no significant difference between the two scenarios. As opposed to the gossip agents, we can see how little effect the self-interested agents have on the regular agents. As compared to the gossip agents on the same route that have traveled as much as 193% more, when self-interested agents are introduced, the average journey length for the regular agents has only increased by about 15%. This result is even lower than the effect on other gossip agents in the entire network. Since we noticed that cheating by the self-interested agents does not benefit them in the first iteration, we devised another set of experiments. In the second set of experiments, the self-interested agents have the objective to help another agent, who is supposed to enter the network some time after the self-interested agent entered. We refer to the latter agent as the beneficiary agent. Just like a self-interested agent, the beneficiary agent also ignores all data regarding heavy traffic. In real-life this can be modeled, for example, by a husband, who would like to help his wife find a faster route to her destination. Table 2 summarizes the normalized values for the different agents. As in the first set of experiments, 5 simulations were run for each scenario, with a total of 25 simulations. In each of these simulation one agent was randomly selected as a self-interested agent, and then another agent, with the same origin as the selfinterested agent, was randomly selected as the beneficiary agent. The other 8,000 and 32,000 agents were designated as regular gossip agents and regular agents, respectively. We can see that as the number of iterations advances, the lower the normalized value for the beneficiary agent. In this scenario, just like the previous one, in the first iterations, not only does the beneficiary agent not avoid the jammed roads, since he ignores all heavy traffic, he also does not benefit from the lies spread by the self-interested agent. This is due to the fact that the lies are not yet incorporated by other gossip agents. Thus, if we compare the average journey length in the first iteration when lies are spread and when there are no lies, the average is significantly lower when there are no lies (p < 0.03). On the other hand, if we compare the average journey length in all of the iterations, there is no significant difference between the two settings. Still, in most of the iterations, the average journey length of the beneficiary agent is longer than in the case when no lies are spread. We can also see the impact on the other agents in the system. While the gossip agents, which are not on the route of the beneficiary agent, virtually are not affected by the self-interested agent, those on the route and the regular agents are affected and have higher normalized values. That is, even with just one self-interested car, we can see that both the gossip agents that follow the same route as the lies spread by the self-interested agents, and other regular agents, increase their journey length by more than 14%. In our third set of experiments we examined a setting in which there was an increasing number of agents, and the agents did not necessarily have the same origin and destination points. To model this we randomly selected selfinterested agents, whose objective was to minimize their average journey length, assuming the cars were repeating their journeys (that is, more than one iteration was made). As opposed to the first set of experiments, in this set the self-interested agents were selected randomly, and we did not enforce the constraint that they will all have the same origin and destination points. As in the previous sets of experiments we ran 5 different simulations per scenario. In each simulation 11 runs were made, each run with different numbers of self-interested agents: 0 (no self-interested agents), 1, 2, 4, 8, and 16. Each agent adopted the behavior modeled in Section 4.1. Figure 1 shows the normalized value achieved by the self-interested agents as a function of their number. The figure shows these values for iterations 2-6. The first iteration is not shown intentionally, as we assume repeated journeys. Also, we have seen in the previous set of experiments and we have provided explanations as to why the self-interested agents do not gain much from their behavior in the first iteration. 0.955 0.96 0.965 0.97 0.975 0.98 0.985 0.99 0.995 1 1.005 1.01 1.015 1.02 1.025 1.03 0 1 2 3 4 5 6 7 8 9 10111213141516 Self-Interested Agents Number NormalizedValue Iteration 2 Iteration 3 Iteration 4 Iteration 5 Iteration 6 Figure 1: Self-interested agents normalized values as a function of the number of self-interested agents. Using these simulations we examined what the threshold could be for the number of randomly selected self-interested agents in order to allow themselves to benefit from their selfish behavior. We can see that up to 8 self-interested agents, the average normalized value is below 1. That is, they benefit from their malicious behavior. In the case of one self-interested agent there is a significant difference between the average journey length of when the agent spread lies and when no lies are spread (p < 0.001), while when The Sixth Intl. Joint Conf. on Autonomous Agents and Multi-Agent Systems (AAMAS 07) 331 there are 2, 4, 8 and 16 self-interested agents there is no significance difference. Yet, as the number of self-interested agents increases, the normalized value also increases. In such cases, the normalized value is larger than 1, and the self-interested agents journey length becomes significantly higher than their journey length, in cases where there are no self-interested agents in the system. In the next subsection we analyze the scenario as a game and show that when in equilibrium the exchange of gossiping between the agents becomes inefficient. 4.3 When Gossiping is Inefficient We continued and modeled our scenario as a game, in order to find the equilibrium. There are two possible types for the agents: (a) regular gossip agents, and (b) self-interested agents. Each of these agents is a representative of its group, and thus all agents in the same group have similar behavior. We note that the advantage of using gossiping in transportation networks is to allow the agents to detect anomalies in the network (e.g., traffic jams) and to quickly adapt to them by recalculating their routes [14]. We also assume that the objective of the self-interested agents is to minimize their own journey length, thus they spread lies on their routes, as described in Section 4.1. We also assume that sophisticated methods for identifying the self-interested agents or managing reputation are not used. This is mainly due to the complexity of incorporating and maintaining such mechanisms, as well as due to the dynamics of the network, in which interactions between different agents are frequent, agents may leave the network, and data about the road might change as time progresses (e.g., a road might be reported by a regular gossip agent as free at a given time, yet it may currently be jammed due to heavy traffic on the road). Let Tavg be the average time it takes to traverse an edge in the transportation network (that is, the average load of an edge). Let Tmax be the maximum time it takes to traverse an edge. We will investigate the game, in which the self-interested and the regular gossip agents can choose the following actions. The self-interested agents can choose how much to lie, that is, they can choose to spread how long (not necessarily the true duration) it takes to traverse certain roads. Since the objective of the self-interested agents is to spread messages as though some roads are jammed, the traversal time they report is obviously larger than the average time. We denote the time the self-interested agents spread as Ts, such that Tavg ≤ Ts ≤ Tmax. Motivated by the results of the simulations we have described above, we saw that the agents are less affected if they discard the heavy traffic values. Thus, the regular gossip cars, attempting to mitigate the effect of the liars, can choose a strategy to ignore abnormal congestion values above a certain threshold, Tg. Obviously, Tavg ≤ Tg ≤ Tmax. In order to prevent the gossip agents from detecting the lies and just discarding those values, the self-interested agents send lies in a given range, [Ts, Tmax], with an inverse geometric distribution, that is, the higher the T value, the higher its frequency. Now we construct the utility functions for each type of agents, which is defined by the values of Ts and Tg. If the self-interested agents spread traversal times higher than or equal to the regular gossip cars" threshold, they will not benefit from those lies. Thus, the utility value of the selfinterested agents in this case is 0. On the other hand, if the self-interested agents spread traversal time which is lower than the threshold, they will gain a positive utility value. From the regular gossip agents point-of-view, if they accept messages from the self-interested agents, then they incorporate the lies in their calculation, thus they will lose utility points. On the other hand, if they discard the false values the self-interested agents send, that is, they do not incorporate the lies, they will gain utility values. Formally, we use us to denote the utility of the self-interested agents and ug to denote the utility of the regular gossip agents. We also denote the strategy profile in the game as {Ts, Tg}. The utility functions are defined as: us = 0 if Ts ≥ Tg Ts − Tavg + 1 if Ts < Tg (1) ug = Tg − Tavg if Ts ≥ Tg Ts − Tg if Ts < Tg (2) We are interested in finding the Nash equilibrium. We recall from [12], that the Nash equilibrium is a strategy profile, where no player has anything to gain by deviating from his strategy, given that the other agent follows his strategy profile. Formally, let (S, u) denote the game, where S is the set of strategy profiles and u is the set of utility functions. When each agent i ∈ {regular gossip, self-interested} chooses a strategy Ti resulting in a strategy profile T = (Ts, Tg) then agent i obtains a utility of ui (T). A strategy profile T∗ ∈ S is a Nash equilibrium if no deviation in the strategy by any single agent is profitable, that is, if for all i, ui (T∗ ) ≥ ui (Ti, T∗ −i). That is, (Ts, Tg) is a Nash equilibrium if the self-interested agents have no other value Ts such that us (Ts, Tg) > us (Ts, Tg), and similarly for the gossip agents. We now have the following theorem. Theorem 4.1. (Tavg, Tavg) is the only Nash equilibrium. Proof. First we will show that (Tavg, Tavg) is a Nash equilibrium. Assume, by contradiction, that the gossip agents choose another value Tg > Tavg. Thus, ug (Tavg, Tg ) = Tavg − Tg < 0. On the other hand, ug (Tavg, Tavg) = 0. Thus, the regular gossip agents have no incentive to deviate from this strategy. The self-interested agents also have no incentive to deviate from this strategy. By contradiction, again assume that the self-interested agents choose another value Ts > Tavg. Thus, us (Ts , Tavg) = 0, while us (Tavg, Tavg) = 0. We will now show that the above solution is unique. We will show that any other tuple (Ts, Tg), such that Tavg < Tg ≤ Tmax and Tavg < Ts ≤ Tmax is not a Nash equilibrium. We have three cases. In the first Tavg < Tg < Ts ≤ Tmax. Thus, us (Ts, Tg) = 0 and ug (Ts, Tg) = Tg − Tavg. In this case, the regular gossip agents have an incentive to deviate and choose another strategy Tg + 1, since by doing so they increase their own utility: ug (Ts, Tg + 1) = Tg + 1 − Tavg. In the second case we have Tavg < Ts < Tg ≤ Tmax. Thus, ug (Ts, Tg) = Ts − Tg < 0. Also, the regular gossip agents have an incentive to deviate and choose another strategy Tg −1, in which their utility value is higher: ug (Ts, Tg −1) = Ts − Tg + 1. In the last case we have Tavg < Ts = Tg ≤ Tmax. Thus, us (Ts, Tg) = Ts − Tg = 0. In this case, the self-interested agents have an incentive to deviate and choose another strategy Tg − 1, in which their utility value is higher: us (Tg − 1, Tg) = Tg − 1 − Tavg + 1 = Tg − Tavg > 0. 332 The Sixth Intl. Joint Conf. on Autonomous Agents and Multi-Agent Systems (AAMAS 07) Table 3: Normalized journey length values for the first iteration Self-Interested Self-Interested Gossip Regular Agents Number Agents Agents Agents 1 0.98 1.01 1.05 2 1.09 1.02 1.05 4 1.07 1.02 1.05 8 1.06 1.04 1.05 16 1.03 1.08 1.06 32 1.07 1.17 1.08 50 1.12 1.28 1.1 64 1.14 1.4 1.13 80 1.15 1.5 1.14 100 1.17 1.63 1.16 Table 4: Normalized journey length values for all iterations Self-Interested Self-Interested Gossip Regular Agents Number Agents Agents Agents 1 0.98 1.02 1.06 2 1.0 1.04 1.07 4 1.0 1.08 1.07 8 1.01 1.33 1.11 16 1.02 1.89 1.17 32 1.06 2.46 1.25 50 1.13 2.24 1.29 64 1.21 2.2 1.32 80 1.21 2.13 1.27 100 1.26 2.11 1.27 The above theorem proves that the equilibrium point is reached only when the self-interested agents send the time to traverse certain edges equals the average time, and on the other hand the regular gossip agents discard all data regarding roads that are associated with an average time or higher. Thus, for this equilibrium point the exchange of gossiping information between agents is inefficient, as the gossip agents are unable to detect any anomalies in the network. In the next section we describe another scenario for the self-interested agents, in which they are not concerned with their own utility, but rather interested in maximizing the average journey length of other gossip agents. 5. SPREADING LIES, CAUSING CHAOS Another possible behavior that can be adopted by selfinterested agents is characterized by their goal to cause disorder in the network. This can be achieved, for example, by maximizing the average journey length of all agents, even at the cost of maximizing their own journey length. To understand the vulnerability of the gossip based transportation support system, we ran 5 different simulations for each scenario. In each simulation different agents were randomly chosen (using a uniform distribution) to act as gossip agents, among them self-interested agents were chosen. Each self-interested agent behaved in the same manner as described in Section 4.1. Every simulation consisted of 11 runs with each run comprising different numbers of self-interested agents: 0 (no selfinterested agents), 1, 2, 4, 8, 16, 32, 50, 64, 80 and 100. Also, in each run the number of self-interested agents was increased incrementally. For example: the run with 50 selfinterested agents consisted of all the self-interested agents that were used in the run with 32 self-interested agents, but with an additional 18 self-interested agents. Tables 3 and 4 summarize the normalized journey length for the self-interested agents, the regular gossip agents and the regular (non-gossip) agents. Table 3 summarizes the data for the first iteration and Table 4 summarizes the data for the average of all iterations. Figure 2 demonstrates the changes in the normalized values for the regular gossip agents and the regular agents, as a function of the iteration number. Similar to the results in our first set of experiments, described in Section 4.2, we can see that randomly selected self-interested agents who follow different randomly selected routes do not benefit from their malicious behavior (that is, their average journey length does not decrease). However, when only one self-interested agent is involved, it does benefit from the malicious behavior, even in the first iteration. The results also indicate that the regular gossip agents are more sensitive to malicious behavior than regular agentsthe average journey length for the gossip agents increases significantly (e.g., with 32 self-interested agents the average journey length for the gossip agents was 146% higher than in the setting with no self-interested agents at all, as opposed to an increase of only 25% for the regular agents). In contrast, these results also indicate that the self-interested agents do not succeed in causing a significant load in the network by their malicious behavior. 1 1.1 1.2 1.3 1.4 1.5 1.6 1.7 1.8 1.9 2 2.1 2.2 2.3 2.4 2.5 2.6 2.7 2.8 2.9 1 2 3 4 5 6 Iteration Number NormalizedValue 32 self-interested agents, gossip agents normalized value 100 self-interested agents, gossip agents normalized value 32 self-interested agents, regular agents normalized value 100 self-interested agents, regular agents normalized value Figure 2: Gossip and regular agents normalized values, as a function of the iteration. Since the goal of the self-interested agents in this case is to cause disorder in the network rather than use the lies for their own benefits, the question arises as to why would the behavior of the self-interested agents be to send lies about their routes only. Furthermore, we hypothesize that if they all send lies about the same major roads the damage they might inflict on the entire network would be larger that had each of them sent lies about its own route. To examine this hypothesis, we designed another set of experiments. In this set of experiments, all the self-interested agents spread lies about the same 13 main roads in the network. However, the results show quite a smaller impact on other gossip and reguThe Sixth Intl. Joint Conf. on Autonomous Agents and Multi-Agent Systems (AAMAS 07) 333 Table 5: Normalized journey length values for all iterations. Network with congestions. Self-Interested Self-Interested Gossip Regular Agents Number Agents Agents Agents 1 1.04 1.02 1.22 2 1.06 1.04 1.22 4 1.04 1.06 1.23 8 1.07 1.15 1.26 16 1.09 1.55 1.39 32 1.12 2.25 1.56 50 1.24 2.25 1.60 64 1.28 2.47 1.63 80 1.50 2.41 1.64 100 1.69 2.61 1.75 lar agents in the network. The average normalized value for the gossip agents in these simulations was only about 1.07, as opposed to 1.7 in the original scenario. When analyzing the results we saw that although the false data was spread, it did not cause other gossip cars to change their route. The main reason was that the lies were spread on roads that were not on the route of the self-interested agents. Thus, it took the data longer to reach agents on the main roads, and when the agents reached the relevant roads this data was too old to be incorporated in the other agents calculations. We also examined the impact of sending lies in order to cause chaos when there are already congestions in the network. To this end, we simulated a network in which 13 main roads are jammed. The behavior of the self-interested agents is as described in Section 4.1, and the self-interested agents spread lies about their own route. The simulation results, detailed in Table 5, show that there is a greater incentive for the self-interested agents to cheat when the network is already congested, as their cheating causes more damage to the other agents in the network. For example, whereas the average journey length of the regular agents increased only by about 15% in the original scenario, in which the network was not congested, in this scenario the average journey length of the agents had increased by about 60%. 6. CONCLUSIONS In this paper we investigated the benefits achieved by self-interested agents in vehicular networks. Using simulations we investigated two behaviors that might be taken by self-interested agents: (a) trying to minimize their journey length, and (b) trying to cause chaos in the network. Our simulations indicate that in both behaviors the selfinterested agents have only limited success achieving their goal, even if no counter-measures are taken. This is in contrast to the greater impact inflicted by self-interested agents in other domains (e.g., E-Commerce). Some reasons for this are the special characteristics of vehicular networks and their dynamic nature. While the self-interested agents spread lies, they cannot choose which agents with whom they will interact. Also, by the time their lies reach other agents, they might become irrelevant, as more recent data has reached the same agents. Motivated by the simulation results, future research in this field will focus on modeling different behaviors of the self-interested agents, which might cause more damage to the network. Another research direction would be to find ways of minimizing the effect of selfish-agents by using distributed reputation or other measures. 7. REFERENCES [1] A. Bejan and R. Lawrence. Peer-to-peer cooperative driving. In Proceedings of ISCIS, pages 259-264, Orlando, USA, October 2002. [2] I. Chisalita and N. Shahmehri. A novel architecture for supporting vehicular communication. In Proceedings of VTC, pages 1002-1006, Canada, September 2002. [3] S. Das, A. Nandan, and G. Pau. Spawn: A swarming protocol for vehicular ad-hoc wireless networks. In Proceedings of VANET, pages 93-94, 2004. [4] A. Datta, S. Quarteroni, and K. Aberer. Autonomous gossiping: A self-organizing epidemic algorithm for selective information dissemination in mobile ad-hoc networks. In Proceedings of IC-SNW, pages 126-143, Maison des Polytechniciens, Paris, France, June 2004. [5] D. Dolev, R. Reischuk, and H. R. Strong. Early stopping in byzantine agreement. JACM, 37(4):720-741, 1990. [6] GM. Threat assessment algorithm. http://www.nhtsa.dot.gov/people/injury/research/pub/ acas/acas-fieldtest/, 2000. [7] Honda. http://world.honda.com/news/2005/c050902.html. [8] Lamport, Shostak, and Pease. The byzantine generals problem. In Advances in Ultra-Dependable Distributed Systems, N. Suri, C. J. Walter, and M. M. Hugue (Eds.). IEEE Computer Society Press, 1982. [9] C. Leckie and R. Kotagiri. Policies for sharing distributed probabilistic beliefs. In Proceedings of ACSC, pages 285-290, Adelaide, Australia, 2003. [10] D. Malkhi, E. Pavlov, and Y. Sella. Gossip with malicious parties. Technical report: 2003-9, School of Computer Science and Engineering - The Hebrew University of Jerusalem, Israel, March 2003. [11] Y. M. Minsky and F. B. Schneider. Tolerating malicious gossip. Distributed Computing, 16(1):49-68, February 2003. [12] M. J. Osborne and A. Rubinstein. A Course In Game Theory. MIT Press, Cambridge MA, 1994. [13] R. Parshani. Routing in gossip networks. Master"s thesis, Department of Computer Science, Bar-Ilan University, Ramat-Gan, Israel, October 2004. [14] R. Parshani, S. Kraus, and Y. Shavitt. A study of gossiping in transportation networks. Submitted for publication, 2006. [15] Y. Shavitt and A. Shay. Optimal routing in gossip networks. IEEE Transactions on Vehicular Technology, 54(4):1473-1487, July 2005. [16] N. Shibata, T. Terauchi, T. Kitani, K. Yasumoto, M. Ito, and T. Higashino. A method for sharing traffic jam information using inter-vehicle communication. In Proceedings of V2VCOM, USA, 2006. [17] W. Wang, X.-Y. Li, and Y. Wang. Truthful multicast routing in selfish wireless networks. In Proceedings of MobiCom, pages 245-259, USA, 2004. [18] B. Yu and M. P. Singh. A social mechanism of reputation management in electronic communities. In Proceedings of CIA, 2000. 334 The Sixth Intl. Joint Conf. on Autonomous Agents and Multi-Agent Systems (AAMAS 07)

artificial social system;chaos;vehicular network;intelligent agent;journey length;social network;self-interested agent;selfinterested agent;agent-base deploy application

train_I-61

Distributed Agent-Based Air Traffic Flow Management

Air traffic flow management is one of the fundamental challenges facing the Federal Aviation Administration (FAA) today. The FAA estimates that in 2005 alone, there were over 322,000 hours of delays at a cost to the industry in excess of three billion dollars. Finding reliable and adaptive solutions to the flow management problem is of paramount importance if the Next Generation Air Transportation Systems are to achieve the stated goal of accommodating three times the current traffic volume. This problem is particularly complex as it requires the integration and/or coordination of many factors including: new data (e.g., changing weather info), potentially conflicting priorities (e.g., different airlines), limited resources (e.g., air traffic controllers) and very heavy traffic volume (e.g., over 40,000 flights over the US airspace). In this paper we use FACET - an air traffic flow simulator developed at NASA and used extensively by the FAA and industry - to test a multi-agent algorithm for traffic flow management. An agent is associated with a fix (a specific location in 2D space) and its action consists of setting the separation required among the airplanes going though that fix. Agents use reinforcement learning to set this separation and their actions speed up or slow down traffic to manage congestion. Our FACET based results show that agents receiving personalized rewards reduce congestion by up to 45% over agents receiving a global reward and by up to 67% over a current industry approach (Monte Carlo estimation).

1. INTRODUCTION The efficient, safe and reliable management of our ever increasing air traffic is one of the fundamental challenges facing the aerospace industry today. On a typical day, more than 40,000 commercial flights operate within the US airspace [14]. In order to efficiently and safely route this air traffic, current traffic flow control relies on a centralized, hierarchical routing strategy that performs flow projections ranging from one to six hours. As a consequence, the system is slow to respond to developing weather or airport conditions leading potentially minor local delays to cascade into large regional congestions. In 2005, weather, routing decisions and airport conditions caused 437,667 delays, accounting for 322,272 hours of delays. The total cost of these delays was estimated to exceed three billion dollars by industry [7]. Furthermore, as the traffic flow increases, the current procedures increase the load on the system, the airports, and the air traffic controllers (more aircraft per region) without providing any of them with means to shape the traffic patterns beyond minor reroutes. The Next Generation Air Transportation Systems (NGATS) initiative aims to address this issues and, not only account for a threefold increase in traffic, but also for the increasing heterogeneity of aircraft and decreasing restrictions on flight paths. Unlike many other flow problems where the increasing traffic is to some extent absorbed by improved hardware (e.g., more servers with larger memories and faster CPUs for internet routing) the air traffic domain needs to find mainly algorithmic solutions, as the infrastructure (e.g., number of the airports) will not change significantly to impact the flow problem. There is therefore a strong need to explore new, distributed and adaptive solutions to the air flow control problem. An adaptive, multi-agent approach is an ideal fit to this naturally distributed problem where the complex interaction among the aircraft, airports and traffic controllers renders a pre-determined centralized solution severely suboptimal at the first deviation from the expected plan. Though a truly distributed and adaptive solution (e.g., free flight where aircraft can choose almost any path) offers the most potential in terms of optimizing flow, it also provides the most radical departure from the current system. As a consequence, a shift to such a system presents tremendous difficulties both in terms of implementation (e.g., scheduling and airport capacity) and political fallout (e.g., impact on air traffic controllers). In this paper, we focus on agent based system that can be implemented readily. In this approach, we assign an 342 978-81-904262-7-5 (RPS) c 2007 IFAAMAS agent to a fix, a specific location in 2D. Because aircraft flight plans consist of a sequence of fixes, this representation allows localized fixes (or agents) to have direct impact on the flow of air traffic1 . In this approach, the agents" actions are to set the separation that approaching aircraft are required to keep. This simple agent-action pair allows the agents to slow down or speed up local traffic and allows agents to a have significant impact on the overall air traffic flow. Agents learn the most appropriate separation for their location using a reinforcement learning (RL) algorithm [15]. In a reinforcement learning approach, the selection of the agent reward has a large impact on the performance of the system. In this work, we explore four different agent reward functions, and compare them to simulating various changes to the system and selecting the best solution (e.g, equivalent to a Monte-Carlo search). The first explored reward consisted of the system reward. The second reward was a personalized agent reward based on collectives [3, 17, 18]. The last two rewards were personalized rewards based on estimations to lower the computational burden of the reward computation. All three personalized rewards aim to align agent rewards with the system reward and ensure that the rewards remain sensitive to the agents" actions. Previous work in this domain fell into one of two distinct categories: The first principles based modeling approaches used by domain experts [5, 8, 10, 13] and the algorithmic approaches explored by the learning and/or agents community [6, 9, 12]. Though our approach comes from the second category, we aim to bridge the gap by using FACET to test our algorithms, a simulator introduced and widely used (i.e., over 40 organizations and 5000 users) by work in the first category [4, 11]. The main contribution of this paper is to present a distributed adaptive air traffic flow management algorithm that can be readily implemented and test that algorithm using FACET. In Section 2, we describe the air traffic flow problem and the simulation tool, FACET. In Section 3, we present the agent-based approach, focusing on the selection of the agents and their action space along with the agents" learning algorithms and reward structures. In Section 4 we present results in domains with one and two congestions, explore different trade-offs of the system objective function, discuss the scaling properties of the different agent rewards and discuss the computational cost of achieving certain levels of performance. Finally, in Section 5, we discuss the implications of these results and provide and map the required work to enable the FAA to reach its stated goal of increasing the traffic volume by threefold. 2. AIR TRAFFIC FLOW MANAGEMENT With over 40,000 flights operating within the United States airspace on an average day, the management of traffic flow is a complex and demanding problem. Not only are there concerns for the efficiency of the system, but also for fairness (e.g., different airlines), adaptability (e.g., developing weather patterns), reliability and safety (e.g., airport management). In order to address such issues, the management of this traffic flow occurs over four hierarchical levels: 1. Separation assurance (2-30 minute decisions); 1 We discuss how flight plans with few fixes can be handled in more detail in Section 2. 2. Regional flow (20 minutes to 2 hours); 3. National flow (1-8 hours); and 4. Dynamic airspace configuration (6 hours to 1 year). Because of the strict guidelines and safety concerns surrounding aircraft separation, we will not address that control level in this paper. Similarly, because of the business and political impact of dynamic airspace configuration, we will not address the outermost flow control level either. Instead, we will focus on the regional and national flow management problems, restricting our impact to decisions with time horizons between twenty minutes and eight hours. The proposed algorithm will fit between long term planning by the FAA and the very short term decisions by air traffic controllers. The continental US airspace consists of 20 regional centers (handling 200-300 flights on a given day) and 830 sectors (handling 10-40 flights). The flow control problem has to address the integration of policies across these sectors and centers, account for the complexity of the system (e.g., over 5200 public use airports and 16,000 air traffic controllers) and handle changes to the policies caused by weather patterns. Two of the fundamental problems in addressing the flow problem are: (i) modeling and simulating such a large complex system as the fidelity required to provide reliable results is difficult to achieve; and (ii) establishing the method by which the flow management is evaluated, as directly minimizing the total delay may lead to inequities towards particular regions or commercial entities. Below, we discuss how we addressed both issues, namely, we present FACET a widely used simulation tool and discuss our system evaluation function. Figure 1: FACET screenshot displaying traffic routes and air flow statistics. 2.1 FACET FACET (Future ATM Concepts Evaluation Tool), a physics based model of the US airspace was developed to accurately model the complex air traffic flow problem [4]. It is based on propagating the trajectories of proposed flights forward in time. FACET can be used to either simulate and display air traffic (a 24 hour slice with 60,000 flights takes 15 minutes to simulate on a 3 GHz, 1 GB RAM computer) or provide rapid statistics on recorded data (4D trajectories for 10,000 flights including sectors, airports, and fix statistics in 10 seconds on the same computer) [11]. FACET is extensively used by The Sixth Intl. Joint Conf. on Autonomous Agents and Multi-Agent Systems (AAMAS 07) 343 the FAA, NASA and industry (over 40 organizations and 5000 users) [11]. FACET simulates air traffic based on flight plans and through a graphical user interface allows the user to analyze congestion patterns of different sectors and centers (Figure 1). FACET also allows the user to change the flow patterns of the aircraft through a number of mechanisms, including metering aircraft through fixes. The user can then observe the effects of these changes to congestion. In this paper, agents use FACET directly through batch mode, where agents send scripts to FACET asking it to simulate air traffic based on metering orders imposed by the agents. The agents then produce their rewards based on receive feedback from FACET about the impact of these meterings. 2.2 System Evaluation The system performance evaluation function we select focuses on delay and congestion but does not account for fairness impact on different commercial entities. Instead it focuses on the amount of congestion in a particular sector and on the amount of measured air traffic delay. The linear combination of these two terms gives the full system evaluation function, G(z) as a function of the full system state z. More precisely, we have: G(z) = −((1 − α)B(z) + αC(z)) , (1) where B(z) is the total delay penalty for all aircraft in the system, and C(z) is the total congestion penalty. The relative importance of these two penalties is determined by the value of α, and we explore various trade-offs based on α in Section 4. The total delay, B, is a sum of delays over a set of sectors S and is given by: B(z) = X s∈S Bs(z) (2) where Bs(z) = X t Θ(t − τs)kt,s(t − τs) , (3) where ks,t is the number of aircraft in sector s at a particular time, τs is a predetermined time, and Θ(·) is the step function that equals 1 when its argument is greater or equal to zero, and has a value of zero otherwise. Intuitively, Bs(z) provides the total number of aircraft that remain in a sector s past a predetermined time τs, and scales their contribution to count by the amount by which they are late. In this manner Bs(z) provides a delay factor that not only accounts for all aircraft that are late, but also provides a scale to measure their lateness. This definition is based on the assumption that most aircraft should have reached the sector by time τs and that aircraft arriving after this time are late. In this paper the value of τs is determined by assessing aircraft counts in the sector in the absence of any intervention or any deviation from predicted paths. Similarly, the total congestion penalty is a sum over the congestion penalties over the sectors of observation, S: C(z) = X s∈S Cs(z) (4) where Cs(z) = a X t Θ(ks,t − cs)eb(ks,t−cs) , (5) where a and b are normalizing constants, and cs is the capacity of sector s as defined by the FAA. Intuitively, Cs(z) penalizes a system state where the number of aircraft in a sector exceeds the FAAs official sector capacity. Each sector capacity is computed using various metrics which include the number of air traffic controllers available. The exponential penalty is intended to provide strong feedback to return the number of aircraft in a sector to below the FAA mandated capacities. 3. AGENT BASED AIR TRAFFIC FLOW The multi agent approach to air traffic flow management we present is predicated on adaptive agents taking independent actions that maximize the system evaluation function discussed above. To that end, there are four critical decisions that need to be made: agent selection, agent action set selection, agent learning algorithm selection and agent reward structure selection. 3.1 Agent Selection Selecting the aircraft as agents is perhaps the most obvious choice for defining an agent. That selection has the advantage that agent actions can be intuitive (e.g., change of flight plan, increase or decrease speed and altitude) and offer a high level of granularity, in that each agent can have its own policy. However, there are several problems with that approach. First, there are in excess of 40,000 aircraft in a given day, leading to a massively large multi-agent system. Second, as the agents would not be able to sample their state space sufficiently, learning would be prohibitively slow. As an alternative, we assign agents to individual ground locations throughout the airspace called fixes. Each agent is then responsible for any aircraft going through its fix. Fixes offer many advantages as agents: 1. Their number can vary depending on need. The system can have as many agents as required for a given situation (e.g., agents coming live around an area with developing weather conditions). 2. Because fixes are stationary, collecting data and matching behavior to reward is easier. 3. Because aircraft flight plans consist of fixes, agent will have the ability to affect traffic flow patterns. 4. They can be deployed within the current air traffic routing procedures, and can be used as tools to help air traffic controllers rather than compete with or replace them. Figure 2 shows a schematic of this agent based system. Agents surrounding a congestion or weather condition affect the flow of traffic to reduce the burden on particular regions. 3.2 Agent Actions The second issue that needs to be addressed, is determining the action set of the agents. Again, an obvious choice may be for fixes to bid on aircraft, affecting their flight plans. Though appealing from a free flight perspective, that approach makes the flight plans too unreliable and significantly complicates the scheduling problem (e.g., arrival at airports and the subsequent gate assignment process). Instead, we set the actions of an agent to determining the separation (distance between aircraft) that aircraft have 344 The Sixth Intl. Joint Conf. on Autonomous Agents and Multi-Agent Systems (AAMAS 07) to maintain, when going through the agent"s fix. This is known as setting the Miles in Trail or MIT. When an agent sets the MIT value to d, aircraft going towards its fix are instructed to line up and keep d miles of separation (though aircraft will always keep a safe distance from each other regardless of the value of d). When there are many aircraft going through a fix, the effect of issuing higher MIT values is to slow down the rate of aircraft that go through the fix. By increasing the value of d, an agent can limit the amount of air traffic downstream of its fix, reducing congestion at the expense of increasing the delays upstream. Figure 2: Schematic of agent architecture. The agents corresponding to fixes surrounding a possible congestion become live and start setting new separation times. 3.3 Agent Learning The objective of each agent is to learn the best values of d that will lead to the best system performance, G. In this paper we assume that each agent will have a reward function and will aim to maximize its reward using its own reinforcement learner [15] (though alternatives such as evolving neuro-controllers are also effective [1]). For complex delayed-reward problems, relatively sophisticated reinforcement learning systems such as temporal difference may have to be used. However, due to our agent selection and agent action set, the air traffic congestion domain modeled in this paper only needs to utilize immediate rewards. As a consequence, simple table-based immediate reward reinforcement learning is used. Our reinforcement learner is equivalent to an -greedy Q-learner with a discount rate of 0 [15]. At every episode an agent takes an action and then receives a reward evaluating that action. After taking action a and receiving reward R an agent updates its Q table (which contains its estimate of the value for taking that action [15]) as follows: Q (a) = (1 − l)Q(a) + l(R), (6) where l is the learning rate. At every time step the agent chooses the action with the highest table value with probability 1 − and chooses a random action with probability . In the experiments described in this paper, α is equal to 0.5 and is equal to 0.25. The parameters were chosen experimentally, though system performance was not overly sensitive to these parameters. 3.4 Agent Reward Structure The final issue that needs to be addressed is selecting the reward structure for the learning agents. The first and most direct approach is to let each agent receive the system performance as its reward. However, in many domains such a reward structure leads to slow learning. We will therefore also set up a second set of reward structures based on agent-specific rewards. Given that agents aim to maximize their own rewards, a critical task is to create good agent rewards, or rewards that when pursued by the agents lead to good overall system performance. In this work we focus on difference rewards which aim to provide a reward that is both sensitive to that agent"s actions and aligned with the overall system reward [2, 17, 18]. 3.4.1 Difference Rewards Consider difference rewards of the form [2, 17, 18]: Di ≡ G(z) − G(z − zi + ci) , (7) where zi is the action of agent i. All the components of z that are affected by agent i are replaced with the fixed constant ci 2 . In many situations it is possible to use a ci that is equivalent to taking agent i out of the system. Intuitively this causes the second term of the difference reward to evaluate the performance of the system without i and therefore D evaluates the agent"s contribution to the system performance. There are two advantages to using D: First, because the second term removes a significant portion of the impact of other agents in the system, it provides an agent with a cleaner signal than G. This benefit has been dubbed learnability (agents have an easier time learning) in previous work [2, 17]. Second, because the second term does not depend on the actions of agent i, any action by agent i that improves D, also improves G. This term which measures the amount of alignment between two rewards has been dubbed factoredness in previous work [2, 17]. 3.4.2 Estimates of Difference Rewards Though providing a good compromise between aiming for system performance and removing the impact of other agents from an agent"s reward, one issue that may plague D is computational cost. Because it relies on the computation of the counterfactual term G(z − zi + ci) (i.e., the system performance without agent i) it may be difficult or impossible to compute, particularly when the exact mathematical form of G is not known. Let us focus on G functions in the following form: G(z) = Gf (f(z)), (8) where Gf () is non-linear with a known functional form and, f(z) = X i fi(zi) , (9) where each fi is an unknown non-linear function. We assume that we can sample values from f(z), enabling us to compute G, but that we cannot sample from each fi(zi). 2 This notation uses zero padding and vector addition rather than concatenation to form full state vectors from partial state vectors. The vector zi in our notation would be ziei in standard vector notation, where ei is a vector with a value of 1 in the ith component and is zero everywhere else. The Sixth Intl. Joint Conf. on Autonomous Agents and Multi-Agent Systems (AAMAS 07) 345 In addition, we assume that Gf is much easier to compute than f(z), or that we may not be able to even compute f(z) directly and must sample it from a black box computation. This form of G matches our system evaluation in the air traffic domain. When we arrange agents so that each aircraft is typically only affected by a single agent, each agent"s impact of the counts of the number of aircraft in a sector, kt,s, will be mostly independent of the other agents. These values of kt,s are the f(z)s in our formulation and the penalty functions form Gf . Note that given aircraft counts, the penalty functions (Gf ) can be easily computed in microseconds, while aircraft counts (f) can only be computed by running FACET taking on the order of seconds. To compute our counterfactual G(z − zi + ci) we need to compute: Gf (f(z − zi + ci)) = Gf 0 @ X j=i fj(zj) + fi(ci) 1 A (10) = Gf (f(z) − fi(zi) + fi(ci)) .(11) Unfortunately, we cannot compute this directly as the values of fi(zi) are unknown. However, if agents take actions independently (it does not observe how other agents act before taking its own action) we can take advantage of the linear form of f(z) in the fis with the following equality: E(f−i(z−i)|zi) = E(f−i(z−i)|ci) (12) where E(f−i(z−i)|zi) is the expected value of all of the fs other than fi given the value of zi and E(f−i(z−i)|ci) is the expected value of all of the fs other than fi given the value of zi is changed to ci. We can then estimate f(z − zi + ci): f(z) − fi(zi) + fi(ci) = f(z) − fi(zi) + fi(ci) + E(f−i(z−i)|ci) − E(f−i(z−i)|zi) = f(z) − E(fi(zi)|zi) + E(fi(ci)|ci) + E(f−i(z−i)|ci) − E(f−i(z−i)|zi) = f(z) − E(f(z)|zi) + E(f(z)|ci) . Therefore we can evaluate Di = G(z) − G(z − zi + ci) as: Dest1 i = Gf (f(z)) − Gf (f(z) − E(f(z)|zi) + E(f(z)|ci)) , (13) leaving us with the task of estimating the values of E(f(z)|zi) and E(f(z)|ci)). These estimates can be computed by keeping a table of averages where we average the values of the observed f(z) for each value of zi that we have seen. This estimate should improve as the number of samples increases. To improve our estimates, we can set ci = E(z) and if we make the mean squared approximation of f(E(z)) ≈ E(f(z)) then we can estimate G(z) − G(z − zi + ci) as: Dest2 i = Gf (f(z)) − Gf (f(z) − E(f(z)|zi) + E(f(z))) . (14) This formulation has the advantage in that we have more samples at our disposal to estimate E(f(z)) than we do to estimate E(f(z)|ci)). 4. SIMULATION RESULTS In this paper we test the performance of our agent based air traffic optimization method on a series of simulations using the FACET air traffic simulator. In all experiments we test the performance of five different methods. The first method is Monte Carlo estimation, where random policies are created, with the best policy being chosen. The other four methods are agent based methods where the agents are maximizing one of the following rewards: 1. The system reward, G(z), as define in Equation 1. 2. The difference reward, Di(z), assuming that agents can calculate counterfactuals. 3. Estimation to the difference reward, Dest1 i (z), where agents estimate the counterfactual using E(f(z)|zi) and E(f(z)|ci). 4. Estimation to the difference reward, Dest2 i (z), where agents estimate the counterfactual using E(f(z)|zi) and E(f(z)). These methods are first tested on an air traffic domain with 300 aircraft, where 200 of the aircraft are going through a single point of congestion over a four hour simulation. Agents are responsible for reducing congestion at this single point, while trying to minimize delay. The methods are then tested on a more difficult problem, where a second point of congestion is added with the 100 remaining aircraft going through this second point of congestion. In all experiments the goal of the system is to maximize the system performance given by G(z) with the parameters, a = 50, b = 0.3, τs1 equal to 200 minutes and τs1 equal to 175 minutes. These values of τ are obtained by examining the time at which most of the aircraft leave the sectors, when no congestion control is being performed. Except where noted, the trade-off between congestion and lateness, α is set to 0.5. In all experiments to make the agent results comparable to the Monte Carlo estimation, the best policies chosen by the agents are used in the results. All results are an average of thirty independent trials with the differences in the mean (σ/ √ n) shown as error bars, though in most cases the error bars are too small to see. Figure 3: Performance on single congestion problem, with 300 Aircraft , 20 Agents and α = .5. 4.1 Single Congestion In the first experiment we test the performance of the five methods when there is a single point of congestion, with twenty agents. This point of congestion is created by setting up a series of flight plans that cause the number of aircraft in 346 The Sixth Intl. Joint Conf. on Autonomous Agents and Multi-Agent Systems (AAMAS 07) the sector of interest to be significantly more than the number allowed by the FAA. The results displayed in Figures 3 and 4 show the performance of all five algorithms on two different system evaluations. In both cases, the agent based methods significantly outperform the Monte Carlo method. This result is not surprising since the agent based methods intelligently explore their space, where as the Monte Carlo method explores the space randomly. Figure 4: Performance on single congestion problem, with 300 Aircraft , 20 Agents and α = .75. Among the agent based methods, agents using difference rewards perform better than agents using the system reward. Again this is not surprising, since with twenty agents, an agent directly trying to maximize the system reward has difficulty determining the effect of its actions on its own reward. Even if an agent takes an action that reduces congestion and lateness, other agents at the same time may take actions that increase congestion and lateness, causing the agent to wrongly believe that its action was poor. In contrast agents using the difference reward have more influence over the value of their own reward, therefore when an agent takes a good action, the value of this action is more likely to be reflected in its reward. This experiment also shows that estimating the difference reward is not only possible, but also quite effective, when the true value of the difference reward cannot be computed. While agents using the estimates do not achieve as high of results as agents using the true difference reward, they still perform significantly better than agents using the system reward. Note, however, that the benefit of the estimated difference rewards are only present later in learning. Earlier in learning, the estimates are poor, and agents using the estimated difference rewards perform no better then agents using the system reward. 4.2 Two Congestions In the second experiment we test the performance of the five methods on a more difficult problem with two points of congestion. On this problem the first region of congestion is the same as in the previous problem, and the second region of congestion is added in a different part of the country. The second congestion is less severe than the first one, so agents have to form different policies depending which point of congestion they are influencing. Figure 5: Performance on two congestion problem, with 300 Aircraft, 20 Agents and α = .5. Figure 6: Performance on two congestion problem, with 300 Aircraft, 50 Agents and α = .5. The results displayed in Figure 5 show that the relative performance of the five methods is similar to the single congestion case. Again agent based methods perform better than the Monte Carlo method and the agents using difference rewards perform better than agents using the system reward. To verify that the performance improvement of our methods is maintained when there are a different number of agents, we perform additional experiments with 50 agents. The results displayed in Figure 6 show that indeed the relative performances of the methods are comparable when the number of agents is increased to 50. Figure 7 shows scaling results and demonstrates that the conclusions hold over a wide range of number of agents. Agents using Dest2 perform slightly better than agents using Dest1 in all cases but for 50 agents. This slight advantage stems from Dest2 providing the agents with a cleaner signal, since its estimate uses more data points. 4.3 Penalty Tradeoffs The system evaluation function used in the experiments is G(z) = −((1−α)D(z)+αC(z)), which comprises of penalties for both congestion and lateness. This evaluation function The Sixth Intl. Joint Conf. on Autonomous Agents and Multi-Agent Systems (AAMAS 07) 347 Figure 7: Impact of number of agents on system performance. Two congestion problem, with 300 Aircraft and α = .5. forces the agents to tradeoff these relative penalties depending on the value of α. With high α the optimization focuses on reducing congestion, while with low α the system focuses on reducing lateness. To verify that the results obtained above are not specific to a particular value of α, we repeat the experiment with 20 agents for α = .75. Figure 8 shows that qualitatively the relative performance of the algorithms remain the same. Next, we perform a series of experiments where α ranges from 0.0 to 1.0 . Figure 9 shows the results which lead to three interesting observations: • First, there is a zero congestion penalty solution. This solution has agents enforce large MIT values to block all air traffic, which appears viable when the system evaluation does not account for delays. All algorithms find this solution, though it is of little interest in practice due to the large delays it would cause. • Second, if the two penalties were independent, an optimal solution would be a line from the two end points. Therefore, unless D is far from being optimal, the two penalties are not independent. Note that for α = 0.5 the difference between D and this hypothetical line is as large as it is anywhere else, making α = 0.5 a reasonable choice for testing the algorithms in a difficult setting. • Third, Monte Carlo and G are particularly poor at handling multiple objectives. For both algorithms, the performance degrades significantly for mid-ranges of α. 4.4 Computational Cost The results in the previous section show the performance of the different algorithms after a specific number of episodes. Those results show that D is significantly superior to the other algorithms. One question that arises, though, is what computational overhead D puts on the system, and what results would be obtained if the additional computational expense of D is made available to the other algorithms. The computation cost of the system evaluation, G (Equation 1) is almost entirely dependent on the computation of Figure 8: Performance on two congestion problem, with 300 Aircraft, 20 Agents and α = .75. Figure 9: Tradeoff Between Objectives on two congestion problem, with 300 Aircraft and 20 Agents. Note that Monte Carlo and G are particularly bad at handling multiple objectives. the airplane counts for the sectors kt,s, which need to be computed using FACET. Except when D is used, the values of k are computed once per episode. However, to compute the counterfactual term in D, if FACET is treated as a black box, each agent would have to compute their own values of k for their counterfactual resulting in n + 1 computations of k per episode. While it may be possible to streamline the computation of D with some knowledge of the internals of FACET, given the complexity of the FACET simulation, it is not unreasonable in this case to treat it as a black box. Table 1 shows the performance of the algorithms after 2100 G computations for each of the algorithms for the simulations presented in Figure 5 where there were 20 agents, 2 congestions and α = .5. All the algorithms except the fully computed D reach 2100 k computations at time step 2100. D however computes k once for the system, and then once for each agent, leading to 21 computations per time step. It therefore reaches 2100 computations at time step 100. We also show the results of the full D computation at t=2100, which needs 44100 computations of k as D44K . 348 The Sixth Intl. Joint Conf. on Autonomous Agents and Multi-Agent Systems (AAMAS 07) Table 1: System Performance for 20 Agents, 2 congestions and α = .5, after 2100 G evaluations (except for D44K which has 44100 G evaluations at t=2100). Reward G σ/ √ n time Dest2 -232.5 7.55 2100 Dest1 -234.4 6.83 2100 D -277.0 7.8 100 D44K -219.9 4.48 2100 G -412.6 13.6 2100 MC -639.0 16.4 2100 Although D44K provides the best result by a slight margin, it is achieved at a considerable computational cost. Indeed, the performance of the two D estimates is remarkable in this case as they were obtained with about twenty times fewer computations of k. Furthermore, the two D estimates, significantly outperform the full D computation for a given number of computations of k and validate the assumptions made in Section 3.4.2. This shows that for this domain, in practice it is more fruitful to perform more learning steps and approximate D, than few learning steps with full D computation when we treat FACET as a black box. 5. DISCUSSION The efficient, safe and reliable management of air traffic flow is a complex problem, requiring solutions that integrate control policies with time horizons ranging from minutes up to a year. The main contribution of this paper is to present a distributed adaptive air traffic flow management algorithm that can be readily implemented and to test that algorithm using FACET, a simulation tool widely used by the FAA, NASA and the industry. Our method is based on agents representing fixes and having each agent determine the separation between aircraft approaching its fix. It offers the significant benefit of not requiring radical changes to the current air flow management structure and is therefore readily deployable. The agents use reinforcement learning to learn control policies and we explore different agent reward functions and different ways of estimating those functions. We are currently extending this work in three directions. First, we are exploring new methods of estimating agent rewards, to further speed up the simulations. Second we are investigating deployment strategies and looking for modifications that would have larger impact. One such modification is to extend the definition of agents from fixes to sectors, giving agents more opportunity to control the traffic flow, and allow them to be more efficient in eliminating congestion. Finally, in cooperation with domain experts, we are investigating different system evaluation functions, above and beyond the delay and congestion dependent G presented in this paper. Acknowledgments: The authors thank Banavar Sridhar for his invaluable help in describing both current air traffic flow management and NGATS, and Shon Grabbe for his detailed tutorials on FACET. 6. REFERENCES [1] A. Agogino and K. Tumer. Efficient evaluation functions for multi-rover systems. In The Genetic and Evolutionary Computation Conference, pages 1-12, Seatle, WA, June 2004. [2] A. Agogino and K. Tumer. Multi agent reward analysis for learning in noisy domains. In Proceedings of the Fourth International Joint Conference on Autonomous Agents and Multi-Agent Systems, Utrecht, Netherlands, July 2005. [3] A. K. Agogino and K. Tumer. Handling communiction restrictions and team formation in congestion games. Journal of Autonous Agents and Multi Agent Systems, 13(1):97-115, 2006. [4] K. D. Bilimoria, B. Sridhar, G. B. Chatterji, K. S. Shethand, and S. R. Grabbe. FACET: Future ATM concepts evaluation tool. Air Traffic Control Quarterly, 9(1), 2001. [5] Karl D. Bilimoria. A geometric optimization approach to aircraft conflict resolution. In AIAA Guidance, Navigation, and Control Conf, Denver, CO, 2000. [6] Martin S. Eby and Wallace E. Kelly III. Free flight separation assurance using distributed algorithms. In Proc of Aerospace Conf, 1999, Aspen, CO, 1999. [7] FAA OPSNET data Jan-Dec 2005. US Department of Transportation website. [8] S. Grabbe and B. Sridhar. Central east pacific flight routing. In AIAA Guidance, Navigation, and Control Conference and Exhibit, Keystone, CO, 2006. [9] Jared C. Hill, F. Ryan Johnson, James K. Archibald, Richard L. Frost, and Wynn C. Stirling. A cooperative multi-agent approach to free flight. In AAMAS "05: Proceedings of the fourth international joint conference on Autonomous agents and multiagent systems, pages 1083-1090, New York, NY, USA, 2005. ACM Press. [10] P. K. Menon, G. D. Sweriduk, and B. Sridhar. Optimal strategies for free flight air traffic conflict resolution. Journal of Guidance, Control, and Dynamics, 22(2):202-211, 1999. [11] 2006 NASA Software of the Year Award Nomination. FACET: Future ATM concepts evaluation tool. Case no. ARC-14653-1, 2006. [12] M. Pechoucek, D. Sislak, D. Pavlicek, and M. Uller. Autonomous agents for air-traffic deconfliction. In Proc of the Fifth Int Jt Conf on Autonomous Agents and Multi-Agent Systems, Hakodate, Japan, May 2006. [13] B. Sridhar and S. Grabbe. Benefits of direct-to in national airspace system. In AIAA Guidance, Navigation, and Control Conf, Denver, CO, 2000. [14] B. Sridhar, T. Soni, K. Sheth, and G. B. Chatterji. Aggregate flow model for air-traffic management. Journal of Guidance, Control, and Dynamics, 29(4):992-997, 2006. [15] R. S. Sutton and A. G. Barto. Reinforcement Learning: An Introduction. MIT Press, Cambridge, MA, 1998. [16] C. Tomlin, G. Pappas, and S. Sastry. Conflict resolution for air traffic management. IEEE Tran on Automatic Control, 43(4):509-521, 1998. [17] K. Tumer and D. Wolpert, editors. Collectives and the Design of Complex Systems. Springer, New York, 2004. [18] D. H. Wolpert and K. Tumer. Optimal payoff functions for members of collectives. Advances in Complex Systems, 4(2/3):265-279, 2001. The Sixth Intl. Joint Conf. on Autonomous Agents and Multi-Agent Systems (AAMAS 07) 349

reinforcement learning;air traffic control;optimization;new method of estimating agent reward;reinforcement learn;traffic flow;congestion;deployment strategy;multiagent system;future atm concept evaluation tool

train_I-62

A Q-decomposition and Bounded RTDP Approach to Resource Allocation

This paper contributes to solve effectively stochastic resource allocation problems known to be NP-Complete. To address this complex resource management problem, a Qdecomposition approach is proposed when the resources which are already shared among the agents, but the actions made by an agent may influence the reward obtained by at least another agent. The Q-decomposition allows to coordinate these reward separated agents and thus permits to reduce the set of states and actions to consider. On the other hand, when the resources are available to all agents, no Qdecomposition is possible and we use heuristic search. In particular, the bounded Real-time Dynamic Programming (bounded rtdp) is used. Bounded rtdp concentrates the planning on significant states only and prunes the action space. The pruning is accomplished by proposing tight upper and lower bounds on the value function. Categories and Subject Descriptors

1. INTRODUCTION This paper aims to contribute to solve complex stochastic resource allocation problems. In general, resource allocation problems are known to be NP-Complete [12]. In such problems, a scheduling process suggests the action (i.e. resources to allocate) to undertake to accomplish certain tasks, according to the perfectly observable state of the environment. When executing an action to realize a set of tasks, the stochastic nature of the problem induces probabilities on the next visited state. In general, the number of states is the combination of all possible specific states of each task and available resources. In this case, the number of possible actions in a state is the combination of each individual possible resource assignment to the tasks. The very high number of states and actions in this type of problem makes it very complex. There can be many types of resource allocation problems. Firstly, if the resources are already shared among the agents, and the actions made by an agent does not influence the state of another agent, the globally optimal policy can be computed by planning separately for each agent. A second type of resource allocation problem is where the resources are already shared among the agents, but the actions made by an agent may influence the reward obtained by at least another agent. To solve this problem efficiently, we adapt Qdecomposition proposed by Russell and Zimdars [9]. In our Q-decomposition approach, a planning agent manages each task and all agents have to share the limited resources. The planning process starts with the initial state s0. In s0, each agent computes their respective Q-value. Then, the planning agents are coordinated through an arbitrator to find the highest global Q-value by adding the respective possible Q-values of each agents. When implemented with heuristic search, since the number of states and actions to consider when computing the optimal policy is exponentially reduced compared to other known approaches, Q-decomposition allows to formulate the first optimal decomposed heuristic search algorithm in a stochastic environments. On the other hand, when the resources are available to all agents, no Q-decomposition is possible. A common way of addressing this large stochastic problem is by using Markov Decision Processes (mdps), and in particular real-time search where many algorithms have been developed recently. For instance Real-Time Dynamic Programming (rtdp) [1], lrtdp [4], hdp [3], and lao [5] are all state-of-the-art heuristic search approaches in a stochastic environment. Because of its anytime quality, an interesting approach is rtdp introduced by Barto et al. [1] which updates states in trajectories from an initial state s0 to a goal state sg. rtdp is used in this paper to solve efficiently a constrained resource allocation problem. rtdp is much more effective if the action space can be pruned of sub-optimal actions. To do this, McMahan et 1212 978-81-904262-7-5 (RPS) c 2007 IFAAMAS al. [6], Smith and Simmons [11], and Singh and Cohn [10] proposed solving a stochastic problem using a rtdp type heuristic search with upper and lower bounds on the value of states. McMahan et al. [6] and Smith and Simmons [11] suggested, in particular, an efficient trajectory of state updates to further speed up the convergence, when given upper and lower bounds. This efficient trajectory of state updates can be combined to the approach proposed here since this paper focusses on the definition of tight bounds, and efficient state update for a constrained resource allocation problem. On the other hand, the approach by Singh and Cohn is suitable to our case, and extended in this paper using, in particular, the concept of marginal revenue [7] to elaborate tight bounds. This paper proposes new algorithms to define upper and lower bounds in the context of a rtdp heuristic search approach. Our marginal revenue bounds are compared theoretically and empirically to the bounds proposed by Singh and Cohn. Also, even if the algorithm used to obtain the optimal policy is rtdp, our bounds can be used with any other algorithm to solve an mdp. The only condition on the use of our bounds is to be in the context of stochastic constrained resource allocation. The problem is now modelled. 2. PROBLEM FORMULATION A simple resource allocation problem is one where there are the following two tasks to realize: ta1 = {wash the dishes}, and ta2 = {clean the floor}. These two tasks are either in the realized state, or not realized state. To realize the tasks, two type of resources are assumed: res1 = {brush}, and res2 = {detergent}. A computer has to compute the optimal allocation of these resources to cleaner robots to realize their tasks. In this problem, a state represents a conjunction of the particular state of each task, and the available resources. The resources may be constrained by the amount that may be used simultaneously (local constraint), and in total (global constraint). Furthermore, the higher is the number of resources allocated to realize a task, the higher is the expectation of realizing the task. For this reason, when the specific states of the tasks change, or when the number of available resources changes, the value of this state may change. When executing an action a in state s, the specific states of the tasks change stochastically, and the remaining resource are determined with the resource available in s, subtracted from the resources used by action a, if the resource is consumable. Indeed, our model may consider consumable and non-consumable resource types. A consumable resource type is one where the amount of available resource is decreased when it is used. On the other hand, a nonconsumable resource type is one where the amount of available resource is unchanged when it is used. For example, a brush is a non-consumable resource, while the detergent is a consumable resource. 2.1 Resource Allocation as a MDPs In our problem, the transition function and the reward function are both known. A Markov Decision Process (mdp) framework is used to model our stochastic resource allocation problem. mdps have been widely adopted by researchers today to model a stochastic process. This is due to the fact that mdps provide a well-studied and simple, yet very expressive model of the world. An mdp in the context of a resource allocation problem with limited resources is defined as a tuple Res, T a, S, A, P, W, R, , where: • Res = res1, ..., res|Res| is a finite set of resource types available for a planning process. Each resource type may have a local resource constraint Lres on the number that may be used in a single step, and a global resource constraint Gres on the number that may be used in total. The global constraint only applies for consumable resource types (Resc) and the local constraints always apply to consumable and nonconsumable resource types. • T a is a finite set of tasks with ta ∈ T a to be accomplished. • S is a finite set of states with s ∈ S. A state s is a tuple T a, res1, ..., res|Resc| , which is the characteristic of each unaccomplished task ta ∈ T a in the environment, and the available consumable resources. sta is the specific state of task ta. Also, S contains a non empty set sg ⊆ S of goal states. A goal state is a sink state where an agent stays forever. • A is a finite set of actions (or assignments). The actions a ∈ A(s) applicable in a state are the combination of all resource assignments that may be executed, according to the state s. In particular, a is simply an allocation of resources to the current tasks, and ata is the resource allocation to task ta. The possible actions are limited by Lres and Gres. • Transition probabilities Pa(s |s) for s ∈ S and a ∈ A(s). • W = [wta] is the relative weight (criticality) of each task. • State rewards R = [rs] : ta∈T a rsta ← sta × wta. The relative reward of the state of a task rsta is the product of a real number sta by the weight factor wta. For our problem, a reward of 1 × wta is given when the state of a task (sta) is in an achieved state, and 0 in all other cases. • A discount (preference) factor γ, which is a real number between 0 and 1. A solution of an mdp is a policy π mapping states s into actions a ∈ A(s). In particular, πta(s) is the action (i.e. resources to allocate) that should be executed on task ta, considering the global state s. In this case, an optimal policy is one that maximizes the expected total reward for accomplishing all tasks. The optimal value of a state, V (s), is given by: V (s) = R(s) + max a∈A(s) γ s ∈S Pa(s |s)V (s ) (1) where the remaining consumable resources in state s are Resc \ res(a), where res(a) are the consumable resources used by action a. Indeed, since an action a is a resource assignment, Resc \ res(a) is the new set of available resources after the execution of action a. Furthermore, one may compute the Q-Values Q(a, s) of each state action pair using the The Sixth Intl. Joint Conf. on Autonomous Agents and Multi-Agent Systems (AAMAS 07) 1213 following equation: Q(a, s) = R(s) + γ s ∈S Pa(s |s) max a ∈A(s ) Q(a , s ) (2) where the optimal value of a state is V (s) = max a∈A(s) Q(a, s). The policy is subjected to the local resource constraints res(π(s)) ≤ Lres∀ s ∈ S , and ∀ res ∈ Res. The global constraint is defined according to all system trajectories tra ∈ T RA. A system trajectory tra is a possible sequence of state-action pairs, until a goal state is reached under the optimal policy π. For example, state s is entered, which may transit to s or to s , according to action a. The two possible system trajectories are (s, a), (s ) and (s, a), (s ) . The global resource constraint is res(tra) ≤ Gres∀ tra ∈ T RA ,and ∀ res ∈ Resc where res(tra) is a function which returns the resources used by trajectory tra. Since the available consumable resources are represented in the state space, this condition is verified by itself. In other words, the model is Markovian as the history has not to be considered in the state space. Furthermore, the time is not considered in the model description, but it may also include a time horizon by using a finite horizon mdp. Since resource allocation in a stochastic environment is NP-Complete, heuristics should be employed. Q-decomposition which decomposes a planning problem to many agents to reduce the computational complexity associated to the state and/or action spaces is now introduced. 2.2 Q-decomposition for Resource Allocation There can be many types of resource allocation problems. Firstly, if the resources are already shared among the agents, and the actions made by an agent does not influence the state of another agent, the globally optimal policy can be computed by planning separately for each agent. A second type of resource allocation problem is where the resources are already shared among the agents, but the actions made by an agent may influence the reward obtained by at least another agent. For instance, a group of agents which manages the oil consummated by a country falls in this group. These agents desire to maximize their specific reward by consuming the right amount of oil. However, all the agents are penalized when an agent consumes oil because of the pollution it generates. Another example of this type comes from our problem of interest, explained in Section 3, which is a naval platform which must counter incoming missiles (i.e. tasks) by using its resources (i.e. weapons, movements). In some scenarios, it may happens that the missiles can be classified in two types: Those requiring a set of resources Res1 and those requiring a set of resources Res2. This can happen depending on the type of missiles, their range, and so on. In this case, two agents can plan for both set of tasks to determine the policy. However, there are interaction between the resource of Res1 and Res2, so that certain combination of resource cannot be assigned. IN particular, if an agent i allocate resources Resi to the first set of tasks T ai, and agent i allocate resources Resi to second set of tasks T ai , the resulting policy may include actions which cannot be executed together. To result these conflicts, we use Q-decomposition proposed by Russell and Zimdars [9] in the context of reinforcement learning. The primary assumption underlying Qdecomposition is that the overall reward function R can be additively decomposed into separate rewards Ri for each distinct agent i ∈ Ag, where |Ag| is the number of agents. That is, R = i∈Ag Ri. It requires each agent to compute a value, from its perspective, for every action. To coordinate with each other, each agent i reports its action values Qi(ai, si) for each state si ∈ Si to an arbitrator at each learning iteration. The arbitrator then chooses an action maximizing the sum of the agent Q-values for each global state s ∈ S. The next time state s is updated, an agent i considers the value as its respective contribution, or Q-value, to the global maximal Q-value. That is, Qi(ai, si) is the value of a state such that it maximizes maxa∈A(s) i∈Ag Qi(ai, si). The fact that the agents use a determined Q-value as the value of a state is an extension of the Sarsa on-policy algorithm [8] to Q-decomposition. Russell and Zimdars called this approach local Sarsa. In this way, an ideal compromise can be found for the agents to reach a global optimum. Indeed, rather than allowing each agent to choose the successor action, each agent i uses the action ai executed by the arbitrator in the successor state si: Qi(ai, si) = Ri(si) + γ si∈Si Pai (si|si)Qi(ai, si) (3) where the remaining consumable resources in state si are Resci \ resi(ai) for a resource allocation problem. Russell and Zimdars [9] demonstrated that local Sarsa converges to the optimum. Also, in some cases, this form of agent decomposition allows the local Q-functions to be expressed by a much reduced state and action space. For our resource allocation problem described briefly in this section, Q-decomposition can be applied to generate an optimal solution. Indeed, an optimal Bellman backup can be applied in a state as in Algorithm 1. In Line 5 of the Qdec-backup function, each agent managing a task computes its respective Q-value. Here, Qi (ai, s ) determines the optimal Q-value of agent i in state s . An agent i uses as the value of a possible state transition s the Q-value for this agent which determines the maximal global Q-value for state s as in the original Q-decomposition approach. In brief, for each visited states s ∈ S, each agent computes its respective Q-values with respect to the global state s. So the state space is the joint state space of all agents. Some of the gain in complexity to use Q-decomposition resides in the si∈Si Pai (si|s) part of the equation. An agent considers as a possible state transition only the possible states of the set of tasks it manages. Since the number of states is exponential with the number of tasks, using Q-decomposition should reduce the planning time significantly. Furthermore, the action space of the agents takes into account only their available resources which is much less complex than a standard action space, which is the combination of all possible resource allocation in a state for all agents. Then, the arbitrator functionalities are in Lines 8 to 20. The global Q-value is the sum of the Q-values produced by each agent managing each task as shown in Line 11, considering the global action a. In this case, when an action of an agent i cannot be executed simultaneously with an action of another agent i , the global action is simply discarded from the action space A(s). Line 14 simply allocate the current value with respect to the highest global Q-value, as in a standard Bellman backup. Then, the optimal policy and Q-value of each agent is updated in Lines 16 and 17 to the sub-actions ai and specific Q-values Qi(ai, s) of each agent 1214 The Sixth Intl. Joint Conf. on Autonomous Agents and Multi-Agent Systems (AAMAS 07) for action a. Algorithm 1 The Q-decomposition Bellman Backup. 1: Function Qdec-backup(s) 2: V (s) ← 0 3: for all i ∈ Ag do 4: for all ai ∈ Ai(s) do 5: Qi(ai, s) ← Ri(s) + γ si ∈Si Pai (si|s)Qi (ai, s ) {where Qi (ai, s ) = hi(s ) when s is not yet visited, and s has Resci \ resi(ai) remaining consumable resources for each agent i} 6: end for 7: end for 8: for all a ∈ A(s) do 9: Q(a, s) ← 0 10: for all i ∈ Ag do 11: Q(a, s) ← Q(a, s) + Qi(ai, s) 12: end for 13: if Q(a, s) > V (s) then 14: V (s) ← Q(a, s) 15: for all i ∈ Ag do 16: πi(s) ← ai 17: Qi (ai, s) ← Qi(ai, s) 18: end for 19: end if 20: end for A standard Bellman backup has a complexity of O(|A| × |SAg|), where |SAg| is the number of joint states for all agents excluding the resources, and |A| is the number of joint actions. On the other hand, the Q-decomposition Bellman backup has a complexity of O((|Ag| × |Ai| × |Si)|) + (|A| × |Ag|)), where |Si| is the number of states for an agent i, excluding the resources and |Ai| is the number of actions for an agent i. Since |SAg| is combinatorial with the number of tasks, so |Si| |S|. Also, |A| is combinatorial with the number of resource types. If the resources are already shared among the agents, the number of resource type for each agent will usually be lower than the set of all available resource types for all agents. In these circumstances, |Ai| |A|. In a standard Bellman backup, |A| is multiplied by |SAg|, which is much more complex than multiplying |A| by |Ag| with the Q-decomposition Bellman backup. Thus, the Q-decomposition Bellman backup is much less complex than a standard Bellman backup. Furthermore, the communication cost between the agents and the arbitrator is null since this approach does not consider a geographically separated problem. However, when the resources are available to all agents, no Q-decomposition is possible. In this case, Bounded RealTime Dynamic Programming (bounded-rtdp) permits to focuss the search on relevant states, and to prune the action space A by using lower and higher bound on the value of states. bounded-rtdp is now introduced. 2.3 Bounded-RTDP Bonet and Geffner [4] proposed lrtdp as an improvement to rtdp [1]. lrtdp is a simple dynamic programming algorithm that involves a sequence of trial runs, each starting in the initial state s0 and ending in a goal or a solved state. Each lrtdp trial is the result of simulating the policy π while updating the values V (s) using a Bellman backup (Equation 1) over the states s that are visited. h(s) is a heuristic which define an initial value for state s. This heuristic has to be admissible - The value given by the heuristic has to overestimate (or underestimate) the optimal value V (s) when the objective function is maximized (or minimized). For example, an admissible heuristic for a stochastic shortest path problem is the solution of a deterministic shortest path problem. Indeed, since the problem is stochastic, the optimal value is lower than for the deterministic version. It has been proven that lrtdp, given an admissible initial heuristic on the value of states cannot be trapped in loops, and eventually yields optimal values [4]. The convergence is accomplished by means of a labeling procedure called checkSolved(s, ). This procedure tries to label as solved each traversed state in the current trajectory. When the initial state is labelled as solved, the algorithm has converged. In this section, a bounded version of rtdp (boundedrtdp) is presented in Algorithm 2 to prune the action space of sub-optimal actions. This pruning enables to speed up the convergence of lrtdp. bounded-rtdp is similar to rtdp except there are two distinct initial heuristics for unvisited states s ∈ S; hL(s) and hU (s). Also, the checkSolved(s, ) procedure can be omitted because the bounds can provide the labeling of a state as solved. On the one hand, hL(s) defines a lower bound on the value of s such that the optimal value of s is higher than hL(s). For its part, hU (s) defines an upper bound on the value of s such that the optimal value of s is lower than hU (s). The values of the bounds are computed in Lines 3 and 4 of the bounded-backup function. Computing these two Q-values is made simultaneously as the state transitions are the same for both Q-values. Only the values of the state transitions change. Thus, having to compute two Q-values instead of one does not augment the complexity of the approach. In fact, Smith and Simmons [11] state that the additional time to compute a Bellman backup for two bounds, instead of one, is no more than 10%, which is also what we obtained. In particular, L(s) is the lower bound of state s, while U(s) is the upper bound of state s. Similarly, QL(a, s) is the Q-value of the lower bound of action a in state s, while QU (a, s) is the Q-value of the upper bound of action a in state s. Using these two bounds allow significantly reducing the action space A. Indeed, in Lines 5 and 6 of the bounded-backup function, if QU (a, s) ≤ L(s) then action a may be pruned from the action space of s. In Line 13 of this function, a state can be labeled as solved if the difference between the lower and upper bounds is lower than . When the execution goes back to the bounded-rtdp function, the next state in Line 10 has a fixed number of consumable resources available Resc, determined in Line 9. In brief, pickNextState(res) selects a none-solved state s reachable under the current policy which has the highest Bellman error (|U(s) − L(s)|). Finally, in Lines 12 to 15, a backup is made in a backward fashion on all visited state of a trajectory, when this trajectory has been made. This strategy has been proven as efficient [11] [6]. As discussed by Singh and Cohn [10], this type of algorithm has a number of desirable anytime characteristics: if an action has to be picked in state s before the algorithm has converged (while multiple competitive actions remains), the action with the highest lower bound is picked. Since the upper bound for state s is known, it may be estimated The Sixth Intl. Joint Conf. on Autonomous Agents and Multi-Agent Systems (AAMAS 07) 1215 Algorithm 2 The bounded-rtdp algorithm. Adapted from [4] and [10]. 1: Function bounded-rtdp(S) 2: returns a value function V 3: repeat 4: s ← s0 5: visited ← null 6: repeat 7: visited.push(s) 8: bounded-backup(s) 9: Resc ← Resc \ {π(s)} 10: s ← s.pickNextState(Resc) 11: until s is a goal 12: while visited = null do 13: s ← visited.pop() 14: bounded-backup(s) 15: end while 16: until s0 is solved or |A(s)| = 1 ∀ s ∈ S reachable from s0 17: return V Algorithm 3 The bounded Bellman backup. 1: Function bounded-backup(s) 2: for all a ∈ A(s) do 3: QU (a, s) ← R(s) + γ s ∈S Pa(s |s)U(s ) 4: QL(a, s) ← R(s) + γ s ∈S Pa(s |s)L(s ) {where L(s ) ← hL(s ) and U(s ) ← hU (s ) when s is not yet visited and s has Resc \ res(a) remaining consumable resources} 5: if QU (a, s) ≤ L(s) then 6: A(s) ← A(s) \ res(a) 7: end if 8: end for 9: L(s) ← max a∈A(s) QL(a, s) 10: U(s) ← max a∈A(s) QU (a, s) 11: π(s) ← arg max a∈A(s) QL(a, s) 12: if |U(s) − L(s)| < then 13: s ← solved 14: end if how far the lower bound is from the optimal. If the difference between the lower and upper bound is too high, one can choose to use another greedy algorithm of one"s choice, which outputs a fast and near optimal solution. Furthermore, if a new task dynamically arrives in the environment, it can be accommodated by redefining the lower and upper bounds which exist at the time of its arrival. Singh and Cohn [10] proved that an algorithm that uses admissible lower and upper bounds to prune the action space is assured of converging to an optimal solution. The next sections describe two separate methods to define hL(s) and hU (s). First of all, the method of Singh and Cohn [10] is briefly described. Then, our own method proposes tighter bounds, thus allowing a more effective pruning of the action space. 2.4 Singh and Cohn"s Bounds Singh and Cohn [10] defined lower and upper bounds to prune the action space. Their approach is pretty straightforward. First of all, a value function is computed for all tasks to realize, using a standard rtdp approach. Then, using these task-value functions, a lower bound hL, and upper bound hU can be defined. In particular, hL(s) = max ta∈T a Vta(sta), and hU (s) = ta∈T a Vta(sta). For readability, the upper bound by Singh and Cohn is named SinghU, and the lower bound is named SinghL. The admissibility of these bounds has been proven by Singh and Cohn, such that, the upper bound always overestimates the optimal value of each state, while the lower bound always underestimates the optimal value of each state. To determine the optimal policy π, Singh and Cohn implemented an algorithm very similar to bounded-rtdp, which uses the bounds to initialize L(s) and U(s). The only difference between bounded-rtdp, and the rtdp version of Singh and Cohn is in the stopping criteria. Singh and Cohn proposed that the algorithm terminates when only one competitive action remains for each state, or when the range of all competitive actions for any state are bounded by an indifference parameter . bounded-rtdp labels states for which |U(s) − L(s)| < as solved and the convergence is reached when s0 is solved or when only one competitive action remains for each state. This stopping criteria is more effective since it is similar to the one used by Smith and Simmons [11] and McMahan et al. brtdp [6]. In this paper, the bounds defined by Singh and Cohn and implemented using bounded-rtdp define the Singh-rtdp approach. The next sections propose to tighten the bounds of Singh-rtdp to permit a more effective pruning of the action space. 2.5 Reducing the Upper Bound SinghU includes actions which may not be possible to execute because of resource constraints, which overestimates the upper bound. To consider only possible actions, our upper bound, named maxU is introduced: hU (s) = max a∈A(s) ta∈T a Qta(ata, sta) (4) where Qta(ata, sta) is the Q-value of task ta for state sta, and action ata computed using a standard lrtdp approach. Theorem 2.1. The upper bound defined by Equation 4 is admissible. Proof: The local resource constraints are satisfied because the upper bound is computed using all global possible actions a. However, hU (s) still overestimates V (s) because the global resource constraint is not enforced. Indeed, each task may use all consumable resources for its own purpose. Doing this produces a higher value for each task, than the one obtained when planning for all tasks globally with the shared limited resources. Computing the maxU bound in a state has a complexity of O(|A| × |T a|), and O(|T a|) for SinghU. A standard Bellman backup has a complexity of O(|A| × |S|). Since |A|×|T a| |A|×|S|, the computation time to determine the upper bound of a state, which is done one time for each visited state, is much less than the computation time required to compute a standard Bellman backup for a state, which is usually done many times for each visited state. Thus, the computation time of the upper bound is negligible. 1216 The Sixth Intl. Joint Conf. on Autonomous Agents and Multi-Agent Systems (AAMAS 07) 2.6 Increasing the Lower Bound The idea to increase SinghL is to allocate the resources a priori among the tasks. When each task has its own set of resources, each task may be solved independently. The lower bound of state s is hL(s) = ta∈T a Lowta(sta), where Lowta(sta) is a value function for each task ta ∈ T a, such that the resources have been allocated a priori. The allocation a priori of the resources is made using marginal revenue, which is a highly used concept in microeconomics [7], and has recently been used for coordination of a Decentralized mdp [2]. In brief, marginal revenue is the extra revenue that an additional unit of product will bring to a firm. Thus, for a stochastic resource allocation problem, the marginal revenue of a resource is the additional expected value it involves. The marginal revenue of a resource res for a task ta in a state sta is defined as following: mrta(sta) = max ata∈A(sta) Qta(ata, sta)− max ata∈A(sta) Qta(ata|res /∈ ata, sta) (5) The concept of marginal revenue of a resource is used in Algorithm 4 to allocate the resources a priori among the tasks which enables to define the lower bound value of a state. In Line 4 of the algorithm, a value function is computed for all tasks in the environment using a standard lrtdp [4] approach. These value functions, which are also used for the upper bound, are computed considering that each task may use all available resources. The Line 5 initializes the valueta variable. This variable is the estimated value of each task ta ∈ T a. In the beginning of the algorithm, no resources are allocated to a specific task, thus the valueta variable is initialized to 0 for all ta ∈ T a. Then, in Line 9, a resource type res (consumable or non-consumable) is selected to be allocated. Here, a domain expert may separate all available resources in many types or parts to be allocated. The resources are allocated in the order of its specialization. In other words, the more a resource is efficient on a small group of tasks, the more it is allocated early. Allocating the resources in this order improves the quality of the resulting lower bound. The Line 12 computes the marginal revenue of a consumable resource res for each task ta ∈ T a. For a non-consumable resource, since the resource is not considered in the state space, all other reachable states from sta consider that the resource res is still usable. The approach here is to sum the difference between the real value of a state to the maximal Q-value of this state if resource res cannot be used for all states in a trajectory given by the policy of task ta. This heuristic proved to obtain good results, but other ones may be tried, for example Monte-Carlo simulation. In Line 21, the marginal revenue is updated in function of the resources already allocated to each task. R(sgta ) is the reward to realize task ta. Thus, Vta(sta)−valueta R(sgta ) is the residual expected value that remains to be achieved, knowing current allocation to task ta, and normalized by the reward of realizing the tasks. The marginal revenue is multiplied by this term to indicate that, the more a task has a high residual value, the more its marginal revenue is going to be high. Then, a task ta is selected in Line 23 with the highest marginal revenue, adjusted with residual value. In Line 24, the resource type res is allocated to the group of resources Resta of task ta. Afterwards, Line 29 recomAlgorithm 4 The marginal revenue lower bound algorithm. 1: Function revenue-bound(S) 2: returns a lower bound LowT a 3: for all ta ∈ T a do 4: Vta ←lrtdp(Sta) 5: valueta ← 0 6: end for 7: s ← s0 8: repeat 9: res ← Select a resource type res ∈ Res 10: for all ta ∈ T a do 11: if res is consumable then 12: mrta(sta) ← Vta(sta) − Vta(sta(Res \ res)) 13: else 14: mrta(sta) ← 0 15: repeat 16: mrta(sta) ← mrta(sta) + Vta(sta)max (ata∈A(sta)|res/∈ata) Qta(ata, sta) 17: sta ← sta.pickNextState(Resc) 18: until sta is a goal 19: s ← s0 20: end if 21: mrrvta(sta) ← mrta(sta) × Vta(sta)−valueta R(sgta ) 22: end for 23: ta ← Task ta ∈ T a which maximize mrrvta(sta) 24: Resta ← Resta {res} 25: temp ← ∅ 26: if res is consumable then 27: temp ← res 28: end if 29: valueta ← valueta + ((Vta(sta) − valueta)× max ata∈A(sta,res) Qta(ata,sta(temp)) Vta(sta) ) 30: until all resource types res ∈ Res are assigned 31: for all ta ∈ T a do 32: Lowta ←lrtdp(Sta, Resta) 33: end for 34: return LowT a putes valueta. The first part of the equation to compute valueta represents the expected residual value for task ta. This term is multiplied by max ata∈A(sta) Qta(ata,sta(res)) Vta(sta) , which is the ratio of the efficiency of resource type res. In other words, valueta is assigned to valueta + (the residual value × the value ratio of resource type res). For a consumable resource, the Q-value consider only resource res in the state space, while for a non-consumable resource, no resources are available. All resource types are allocated in this manner until Res is empty. All consumable and non-consumable resource types are allocated to each task. When all resources are allocated, the lower bound components Lowta of each task are computed in Line 32. When the global solution is computed, the lower bound is as follow: hL(s) = max(SinghL, max a∈A(s) ta∈T a Lowta(sta)) (6) We use the maximum of the SinghL bound and the sum of the lower bound components Lowta, thus marginalrevenue ≥ SinghL. In particular, the SinghL bound may The Sixth Intl. Joint Conf. on Autonomous Agents and Multi-Agent Systems (AAMAS 07) 1217 be higher when a little number of tasks remain. As the components Lowta are computed considering s0; for example, if in a subsequent state only one task remains, the bound of SinghL will be higher than any of the Lowta components. The main difference of complexity between SinghL and revenue-bound is in Line 32 where a value for each task has to be computed with the shared resource. However, since the resource are shared, the state space and action space is greatly reduced for each task, reducing greatly the calculus compared to the value functions computed in Line 4 which is done for both SinghL and revenue-bound. Theorem 2.2. The lower bound of Equation 6 is admissible. Proof: Lowta(sta) is computed with the resource being shared. Summing the Lowta(sta) value functions for each ta ∈ T a does not violates the local and global resource constraints. Indeed, as the resources are shared, the tasks cannot overuse them. Thus, hL(s) is a realizable policy, and an admissible lower bound. 3. DISCUSSION AND EXPERIMENTS The domain of the experiments is a naval platform which must counter incoming missiles (i.e. tasks) by using its resources (i.e. weapons, movements). For the experiments, 100 randomly resource allocation problems were generated for each approach, and possible number of tasks. In our problem, |Sta| = 4, thus each task can be in four distinct states. There are two types of states; firstly, states where actions modify the transition probabilities; and then, there are goal states. The state transitions are all stochastic because when a missile is in a given state, it may always transit in many possible states. In particular, each resource type has a probability to counter a missile between 45% and 65% depending on the state of the task. When a missile is not countered, it transits to another state, which may be preferred or not to the current state, where the most preferred state for a task is when it is countered. The effectiveness of each resource is modified randomly by ±15% at the start of a scenario. There are also local and global resource constraints on the amount that may be used. For the local constraints, at most 1 resource of each type can be allocated to execute tasks in a specific state. This constraint is also present on a real naval platform because of sensor and launcher constraints and engagement policies. Furthermore, for consumable resources, the total amount of available consumable resource is between 1 and 2 for each type. The global constraint is generated randomly at the start of a scenario for each consumable resource type. The number of resource type has been fixed to 5, where there are 3 consumable resource types and 2 non-consumable resources types. For this problem a standard lrtdp approach has been implemented. A simple heuristic has been used where the value of an unvisited state is assigned as the value of a goal state such that all tasks are achieved. This way, the value of each unvisited state is assured to overestimate its real value since the value of achieving a task ta is the highest the planner may get for ta. Since this heuristic is pretty straightforward, the advantages of using better heuristics are more evident. Nevertheless, even if the lrtdp approach uses a simple heuristic, still a huge part of the state space is not visited when computing the optimal policy. The approaches described in this paper are compared in Figures 1 and 2. Lets summarize these approaches here: • Qdec-lrtdp: The backups are computed using the Qdec-backup function (Algorithm 1), but in a lrtdp context. In particular the updates made in the checkSolved function are also made using the the Qdecbackup function. • lrtdp-up: The upper bound of maxU is used for lrtdp. • Singh-rtdp: The SinghL and SinghU bounds are used for bounded-rtdp. • mr-rtdp: The revenue-bound and maxU bounds are used for bounded-rtdp. To implement Qdec-lrtdp, we divided the set of tasks in two equal parts. The set of task T ai, managed by agent i, can be accomplished with the set of resources Resi, while the second set of task T ai , managed by agent Agi , can be accomplished with the set of resources Resi . Resi had one consumable resource type and one non-consumable resource type, while Resi had two consumable resource types and one non-consumable resource type. When the number of tasks is odd, one more task was assigned to T ai . There are constraint between the group of resource Resi and Resi such that some assignments are not possible. These constraints are managed by the arbitrator as described in Section 2.2. Q-decomposition permits to diminish the planning time significantly in our problem settings, and seems a very efficient approach when a group of agents have to allocate resources which are only available to themselves, but the actions made by an agent may influence the reward obtained by at least another agent. To compute the lower bound of revenue-bound, all available resources have to be separated in many types or parts to be allocated. For our problem, we allocated each resource of each type in the order of of its specialization like we said when describing the revenue-bound function. In terms of experiments, notice that the lrtdp lrtdp-up and approaches for resource allocation, which doe not prune the action space, are much more complex. For instance, it took an average of 1512 seconds to plan for the lrtdp-up approach with six tasks (see Figure 1). The Singh-rtdp approach diminished the planning time by using a lower and upper bound to prune the action space. mr-rtdp further reduce the planning time by providing very tight initial bounds. In particular, Singh-rtdp needed 231 seconds in average to solve problem with six tasks and mr-rtdp required 76 seconds. Indeed, the time reduction is quite significant compared to lrtdp-up, which demonstrates the efficiency of using bounds to prune the action space. Furthermore, we implemented mr-rtdp with the SinghU bound, and this was slightly less efficient than with the maxU bound. We also implemented mr-rtdp with the SinghL bound, and this was slightly more efficient than Singh-rtdp. From these results, we conclude that the difference of efficiency between mr-rtdp and Singh-rtdp is more attributable to the marginal-revenue lower bound that to the maxU upper bound. Indeed, when the number of task to execute is high, the lower bounds by Singh-rtdp takes the values of a single task. On the other hand, the lower bound of mr-rtdp takes into account the value of all 1218 The Sixth Intl. Joint Conf. on Autonomous Agents and Multi-Agent Systems (AAMAS 07) 0.01 0.1 1 10 100 1000 10000 100000 1 2 3 4 5 6 7 8 9 10 11 12 13 Timeinseconds Number of tasks LRTDP QDEC-LRTDP Figure 1: Efficiency of Q-decomposition LRTDP and LRTDP. 0.01 0.1 1 10 100 1000 10000 1 2 3 4 5 6 7 8 Timeinseconds Number of tasks LRTDP LRTDP-up Singh-RTDP MR-RTDP Figure 2: Efficiency of MR-RTDP compared to SINGH-RTDP. task by using a heuristic to distribute the resources. Indeed, an optimal allocation is one where the resources are distributed in the best way to all tasks, and our lower bound heuristically does that. 4. CONCLUSION The experiments have shown that Q-decomposition seems a very efficient approach when a group of agents have to allocate resources which are only available to themselves, but the actions made by an agent may influence the reward obtained by at least another agent. On the other hand, when the available resource are shared, no Q-decomposition is possible and we proposed tight bounds for heuristic search. In this case, the planning time of bounded-rtdp, which prunes the action space, is significantly lower than for lrtdp. Furthermore, The marginal revenue bound proposed in this paper compares favorably to the Singh and Cohn [10] approach. boundedrtdp with our proposed bounds may apply to a wide range of stochastic environments. The only condition for the use our bounds is that each task possesses consumable and/or non-consumable limited resources. An interesting research avenue would be to experiment our bounds with other heuristic search algorithms. For instance, frtdp [11], and brtdp [6] are both efficient heuristic search algorithms. In particular, both these approaches proposed an efficient state trajectory updates, when given upper and lower bounds. Our tight bounds would enable, for both frtdp and brtdp, to reduce the number of backup to perform before convergence. Finally, the bounded-rtdp function prunes the action space when QU (a, s) ≤ L(s), as Singh and Cohn [10] suggested. frtdp and brtdp could also prune the action space in these circumstances to further reduce their planning time. 5. REFERENCES [1] A. Barto, S. Bradtke, and S. Singh. Learning to act using real-time dynamic programming. Artificial Intelligence, 72(1):81-138, 1995. [2] A. Beynier and A. I. Mouaddib. An iterative algorithm for solving constrained decentralized markov decision processes. In Proceeding of the Twenty-First National Conference on Artificial Intelligence (AAAI-06), 2006. [3] B. Bonet and H. Geffner. Faster heuristic search algorithms for planning with uncertainty and full feedback. In Proceedings of the Eighteenth International Joint Conference on Artificial Intelligence (IJCAI-03), August 2003. [4] B. Bonet and H. Geffner. Labeled lrtdp approach: Improving the convergence of real-time dynamic programming. In Proceeding of the Thirteenth International Conference on Automated Planning & Scheduling (ICAPS-03), pages 12-21, Trento, Italy, 2003. [5] E. A. Hansen and S. Zilberstein. lao : A heuristic search algorithm that finds solutions with loops. Artificial Intelligence, 129(1-2):35-62, 2001. [6] H. B. McMahan, M. Likhachev, and G. J. Gordon. Bounded real-time dynamic programming: rtdp with monotone upper bounds and performance guarantees. In ICML "05: Proceedings of the Twenty-Second International Conference on Machine learning, pages 569-576, New York, NY, USA, 2005. ACM Press. [7] R. S. Pindyck and D. L. Rubinfeld. Microeconomics. Prentice Hall, 2000. [8] G. A. Rummery and M. Niranjan. On-line Q-learning using connectionist systems. Technical report CUED/FINFENG/TR 166, Cambridge University Engineering Department, 1994. [9] S. J. Russell and A. Zimdars. Q-decomposition for reinforcement learning agents. In ICML, pages 656-663, 2003. [10] S. Singh and D. Cohn. How to dynamically merge markov decision processes. In Advances in Neural Information Processing Systems, volume 10, pages 1057-1063, Cambridge, MA, USA, 1998. MIT Press. [11] T. Smith and R. Simmons. Focused real-time dynamic programming for mdps: Squeezing more out of a heuristic. In Proceedings of the Twenty-First National Conference on Artificial Intelligence (AAAI), Boston, USA, 2006. [12] W. Zhang. Modeling and solving a resource allocation problem with soft constraint techniques. Technical report: wucs-2002-13, Washington University, Saint-Louis, Missouri, 2002. The Sixth Intl. Joint Conf. on Autonomous Agents and Multi-Agent Systems (AAMAS 07) 1219

planning agent;real-time dynamic programming;heuristic search;reward separated agent;stochastic environment;marginal revenue;markov decision process;complex stochastic resource allocation problem;marginal revenue bound;resource management;resource allocation;q-decomposition

train_I-63

Combinatorial Resource Scheduling for Multiagent MDPs

Optimal resource scheduling in multiagent systems is a computationally challenging task, particularly when the values of resources are not additive. We consider the combinatorial problem of scheduling the usage of multiple resources among agents that operate in stochastic environments, modeled as Markov decision processes (MDPs). In recent years, efficient resource-allocation algorithms have been developed for agents with resource values induced by MDPs. However, this prior work has focused on static resource-allocation problems where resources are distributed once and then utilized in infinite-horizon MDPs. We extend those existing models to the problem of combinatorial resource scheduling, where agents persist only for finite periods between their (predefined) arrival and departure times, requiring resources only for those time periods. We provide a computationally efficient procedure for computing globally optimal resource assignments to agents over time. We illustrate and empirically analyze the method in the context of a stochastic jobscheduling domain.

1. INTRODUCTION The tasks of optimal resource allocation and scheduling are ubiquitous in multiagent systems, but solving such optimization problems can be computationally difficult, due to a number of factors. In particular, when the value of a set of resources to an agent is not additive (as is often the case with resources that are substitutes or complements), the utility function might have to be defined on an exponentially large space of resource bundles, which very quickly becomes computationally intractable. Further, even when each agent has a utility function that is nonzero only on a small subset of the possible resource bundles, obtaining optimal allocation is still computationally prohibitive, as the problem becomes NP-complete [14]. Such computational issues have recently spawned several threads of work in using compact models of agents" preferences. One idea is to use any structure present in utility functions to represent them compactly, via, for example, logical formulas [15, 10, 4, 3]. An alternative is to directly model the mechanisms that define the agents" utility functions and perform resource allocation directly with these models [9]. A way of accomplishing this is to model the processes by which an agent might utilize the resources and define the utility function as the payoff of these processes. In particular, if an agent uses resources to act in a stochastic environment, its utility function can be naturally modeled with a Markov decision process, whose action set is parameterized by the available resources. This representation can then be used to construct very efficient resource-allocation algorithms that lead to an exponential speedup over a straightforward optimization problem with flat representations of combinatorial preferences [6, 7, 8]. However, this existing work on resource allocation with preferences induced by resource-parameterized MDPs makes an assumption that the resources are only allocated once and are then utilized by the agents independently within their infinite-horizon MDPs. This assumption that no reallocation of resources is possible can be limiting in domains where agents arrive and depart dynamically. In this paper, we extend the work on resource allocation under MDP-induced preferences to discrete-time scheduling problems, where agents are present in the system for finite time intervals and can only use resources within these intervals. In particular, agents arrive and depart at arbitrary (predefined) times and within these intervals use resources to execute tasks in finite-horizon MDPs. We address the problem of globally optimal resource scheduling, where the objective is to find an allocation of resources to the agents across time that maximizes the sum of the expected rewards that they obtain. In this context, our main contribution is a mixed-integerprogramming formulation of the scheduling problem that chooses globally optimal resource assignments, starting times, and execution horizons for all agents (within their arrival1220 978-81-904262-7-5 (RPS) c 2007 IFAAMAS departure intervals). We analyze and empirically compare two flavors of the scheduling problem: one, where agents have static resource assignments within their finite-horizon MDPs, and another, where resources can be dynamically reallocated between agents at every time step. In the rest of the paper, we first lay down the necessary groundwork in Section 2 and then introduce our model and formal problem statement in Section 3. In Section 4.2, we describe our main result, the optimization program for globally optimal resource scheduling. Following the discussion of our experimental results on a job-scheduling problem in Section 5, we conclude in Section 6 with a discussion of possible extensions and generalizations of our method. 2. BACKGROUND Similarly to the model used in previous work on resourceallocation with MDP-induced preferences [6, 7], we define the value of a set of resources to an agent as the value of the best MDP policy that is realizable, given those resources. However, since the focus of our work is on scheduling problems, and a large part of the optimization problem is to decide how resources are allocated in time among agents with finite arrival and departure times, we model the agents" planning problems as finite-horizon MDPs, in contrast to previous work that used infinite-horizon discounted MDPs. In the rest of this section, we first introduce some necessary background on finite-horizon MDPs and present a linear-programming formulation that serves as the basis for our solution algorithm developed in Section 4. We also outline the standard methods for combinatorial resource scheduling with flat resource values, which serve as a comparison benchmark for the new model developed here. 2.1 Markov Decision Processes A stationary, finite-domain, discrete-time MDP (see, for example, [13] for a thorough and detailed development) can be described as S, A, p, r , where: S is a finite set of system states; A is a finite set of actions that are available to the agent; p is a stationary stochastic transition function, where p(σ|s, a) is the probability of transitioning to state σ upon executing action a in state s; r is a stationary reward function, where r(s, a) specifies the reward obtained upon executing action a in state s. Given such an MDP, a decision problem under a finite horizon T is to choose an optimal action at every time step to maximize the expected value of the total reward accrued during the agent"s (finite) lifetime. The agent"s optimal policy is then a function of current state s and the time until the horizon. An optimal policy for such a problem is to act greedily with respect to the optimal value function, defined recursively by the following system of finite-time Bellman equations [2]: v(s, t) = max a r(s, a) + X σ p(σ|s, a)v(σ, t + 1), ∀s ∈ S, t ∈ [1, T − 1]; v(s, T) = 0, ∀s ∈ S; where v(s, t) is the optimal value of being in state s at time t ∈ [1, T]. This optimal value function can be easily computed using dynamic programming, leading to the following optimal policy π, where π(s, a, t) is the probability of executing action a in state s at time t: π(s, a, t) = ( 1, a = argmaxa r(s, a) + P σ p(σ|s, a)v(σ, t + 1), 0, otherwise. The above is the most common way of computing the optimal value function (and therefore an optimal policy) for a finite-horizon MDP. However, we can also formulate the problem as the following linear program (similarly to the dual LP for infinite-horizon discounted MDPs [13, 6, 7]): max X s X a r(s, a) X t x(s, a, t) subject to: X a x(σ, a, t + 1) = X s,a p(σ|s, a)x(s, a, t) ∀σ, t ∈ [1, T − 1]; X a x(s, a, 1) = α(s), ∀s ∈ S; (1) where α(s) is the initial distribution over the state space, and x is the (non-stationary) occupation measure (x(s, a, t) ∈ [0, 1] is the total expected number of times action a is executed in state s at time t). An optimal (non-stationary) policy is obtained from the occupation measure as follows: π(s, a, t) = x(s, a, t)/ X a x(s, a, t) ∀s ∈ S, t ∈ [1, T]. (2) Note that the standard unconstrained finite-horizon MDP, as described above, always has a uniformly-optimal solution (optimal for any initial distribution α(s)). Therefore, an optimal policy can be obtained by using an arbitrary constant α(s) > 0 (in particular, α(s) = 1 will result in x(s, a, t) = π(s, a, t)). However, for MDPs with resource constraints (as defined below in Section 3), uniformly-optimal policies do not in general exist. In such cases, α becomes a part of the problem input, and a resulting policy is only optimal for that particular α. This result is well known for infinite-horizon MDPs with various types of constraints [1, 6], and it also holds for our finite-horizon model, which can be easily established via a line of reasoning completely analogous to the arguments in [6]. 2.2 Combinatorial Resource Scheduling A straightforward approach to resource scheduling for a set of agents M, whose values for the resources are induced by stochastic planning problems (in our case, finite-horizon MDPs) would be to have each agent enumerate all possible resource assignments over time and, for each one, compute its value by solving the corresponding MDP. Then, each agent would provide valuations for each possible resource bundle over time to a centralized coordinator, who would compute the optimal resource assignments across time based on these valuations. When resources can be allocated at different times to different agents, each agent must submit valuations for every combination of possible time horizons. Let each agent m ∈ M execute its MDP within the arrival-departure time interval τ ∈ [τa m, τd m]. Hence, agent m will execute an MDP with time horizon no greater than Tm = τd m−τa m+1. Let bτ be the global time horizon for the problem, before which all of the agents" MDPs must finish. We assume τd m < bτ, ∀m ∈ M. The Sixth Intl. Joint Conf. on Autonomous Agents and Multi-Agent Systems (AAMAS 07) 1221 For the scheduling problem where agents have static resource requirements within their finite-horizon MDPs, the agents provide a valuation for each resource bundle for each possible time horizon (from [1, Tm]) that they may use. Let Ω be the set of resources to be allocated among the agents. An agent will get at most one resource bundle for one of the time horizons. Let the variable ψ ∈ Ψm enumerate all possible pairs of resource bundles and time horizons for agent m, so there are 2|Ω| × Tm values for ψ (the space of bundles is exponential in the number of resource types |Ω|). The agent m must provide a value vψ m for each ψ, and the coordinator will allocate at most one ψ (resource, time horizon) pair to each agent. This allocation is expressed as an indicator variable zψ m ∈ {0, 1} that shows whether ψ is assigned to agent m. For time τ and resource ω, the function nm(ψ, τ, ω) ∈ {0, 1} indicates whether the bundle in ψ uses resource ω at time τ (we make the assumption that agents have binary resource requirements). This allocation problem is NP-complete, even when considering only a single time step, and its difficulty increases significantly with multiple time steps because of the increasing number of values of ψ. The problem of finding an optimal allocation that satisfies the global constraint that the amount of each resource ω allocated to all agents does not exceed the available amount bϕ(ω) can be expressed as the following integer program: max X m∈M X ψ∈Ψm zψ mvψ m subject to: X ψ∈Ψm zψ m ≤ 1, ∀m ∈ M; X m∈M X ψ∈Ψm zψ mnm(ψ, τ, ω) ≤ bϕ(ω), ∀τ ∈ [1, bτ], ∀ω ∈ Ω; (3) The first constraint in equation 3 says that no agent can receive more than one bundle, and the second constraint ensures that the total assignment of resource ω does not, at any time, exceed the resource bound. For the scheduling problem where the agents are able to dynamically reallocate resources, each agent must specify a value for every combination of bundles and time steps within its time horizon. Let the variable ψ ∈ Ψm in this case enumerate all possible resource bundles for which at most one bundle may be assigned to agent m at each time step. Therefore, in this case there are P t∈[1,Tm](2|Ω| )t ∼ 2|Ω|Tm possibilities of resource bundles assigned to different time slots, for the Tm different time horizons. The same set of equations (3) can be used to solve this dynamic scheduling problem, but the integer program is different because of the difference in how ψ is defined. In this case, the number of ψ values is exponential in each agent"s planning horizon Tm, resulting in a much larger program. This straightforward approach to solving both of these scheduling problems requires an enumeration and solution of either 2|Ω| Tm (static allocation) or P t∈[1,Tm] 2|Ω|t (dynamic reallocation) MDPs for each agent, which very quickly becomes intractable with the growth of the number of resources |Ω| or the time horizon Tm. 3. MODEL AND PROBLEM STATEMENT We now formally introduce our model of the resourcescheduling problem. The problem input consists of the following components: • M, Ω, bϕ, τa m, τd m, bτ are as defined above in Section 2.2. • {Θm} = {S, A, pm, rm, αm} are the MDPs of all agents m ∈ M. Without loss of generality, we assume that state and action spaces of all agents are the same, but each has its own transition function pm, reward function rm, and initial conditions αm. • ϕm : A×Ω → {0, 1} is the mapping of actions to resources for agent m. ϕm(a, ω) indicates whether action a of agent m needs resource ω. An agent m that receives a set of resources that does not include resource ω cannot execute in its MDP policy any action a for which ϕm(a, ω) = 0. We assume all resource requirements are binary; as discussed below in Section 6, this assumption is not limiting. Given the above input, the optimization problem we consider is to find the globally optimal-maximizing the sum of expected rewards-mapping of resources to agents for all time steps: Δ : τ × M × Ω → {0, 1}. A solution is feasible if the corresponding assignment of resources to the agents does not violate the global resource constraint: X m Δm(τ, ω) ≤ bϕ(ω), ∀ω ∈ Ω, τ ∈ [1, bτ]. (4) We consider two flavors of the resource-scheduling problem. The first formulation restricts resource assignments to the space where the allocation of resources to each agent is static during the agent"s lifetime. The second formulation allows reassignment of resources between agents at every time step within their lifetimes. Figure 1 depicts a resource-scheduling problem with three agents M = {m1, m2, m3}, three resources Ω = {ω1, ω2, ω3}, and a global problem horizon of bτ = 11. The agents" arrival and departure times are shown as gray boxes and are {1, 6}, {3, 7}, and {2, 11}, respectively. A solution to this problem is shown via horizontal bars within each agents" box, where the bars correspond to the allocation of the three resource types. Figure 1a shows a solution to a static scheduling problem. According to the shown solution, agent m1 begins the execution of its MDP at time τ = 1 and has a lock on all three resources until it finishes execution at time τ = 3. Note that agent m1 relinquishes its hold on the resources before its announced departure time of τd m1 = 6, ostensibly because other agents can utilize the resources more effectively. Thus, at time τ = 4, resources ω1 and ω3 are allocated to agent m2, who then uses them to execute its MDP (using only actions supported by resources ω1 and ω3) until time τ = 7. Agent m3 holds resource ω3 during the interval τ ∈ [4, 10]. Figure 1b shows a possible solution to the dynamic version of the same problem. There, resources can be reallocated between agents at every time step. For example, agent m1 gives up its use of resource ω2 at time τ = 2, although it continues the execution of its MDP until time τ = 6. Notice that an agent is not allowed to stop and restart its MDP, so agent m1 is only able to continue executing in the interval τ ∈ [3, 4] if it has actions that do not require any resources (ϕm(a, ω) = 0). Clearly, the model and problem statement described above make a number of assumptions about the problem and the desired solution properties. We discuss some of those assumptions and their implications in Section 6. 1222 The Sixth Intl. Joint Conf. on Autonomous Agents and Multi-Agent Systems (AAMAS 07) (a) (b) Figure 1: Illustration of a solution to a resource-scheduling problem with three agents and three resources: a) static resource assignments (resource assignments are constant within agents" lifetimes; b) dynamic assignment (resource assignments are allowed to change at every time step). 4. RESOURCE SCHEDULING Our resource-scheduling algorithm proceeds in two stages. First, we perform a preprocessing step that augments the agent MDPs; this process is described in Section 4.1. Second, using these augmented MDPs we construct a global optimization problem, which is described in Section 4.2. 4.1 Augmenting Agents" MDPs In the model described in the previous section, we assume that if an agent does not possess the necessary resources to perform actions in its MDP, its execution is halted and the agent leaves the system. In other words, the MDPs cannot be paused and resumed. For example, in the problem shown in Figure 1a, agent m1 releases all resources after time τ = 3, at which point the execution of its MDP is halted. Similarly, agents m2 and m3 only execute their MDPs in the intervals τ ∈ [4, 6] and τ ∈ [4, 10], respectively. Therefore, an important part of the global decision-making problem is to decide the window of time during which each of the agents is active (i.e., executing its MDP). To accomplish this, we augment each agent"s MDP with two new states (start and finish states sb , sf , respectively) and a new start/stop action a∗ , as illustrated in Figure 2. The idea is that an agent stays in the start state sb until it is ready to execute its MDP, at which point it performs the start/stop action a∗ and transitions into the state space of the original MDP with the transition probability that corresponds to the original initial distribution α(s). For example, in Figure 1a, for agent m2 this would happen at time τ = 4. Once the agent gets to the end of its activity window (time τ = 6 for agent m2 in Figure 1a), it performs the start/stop action, which takes it into the sink finish state sf at time τ = 7. More precisely, given an MDP S, A, pm, rm, αm , we define an augmented MDP S , A , pm, rm, αm as follows: S = S ∪ sb ∪ sf ; A = A ∪ a∗ ; p (s|sb , a∗ ) = α(s), ∀s ∈ S; p (sb |sb , a) = 1.0, ∀a ∈ A; p (sf |s, a∗ ) = 1.0, ∀s ∈ S; p (σ|s, a) = p(σ|s, a), ∀s, σ ∈ S, a ∈ A; r (sb , a) = r (sf , a) = 0, ∀a ∈ A ; r (s, a) = r(s, a), ∀s ∈ S, a ∈ A; α (sb ) = 1; α (s) = 0, ∀s ∈ S; where all non-specified transition probabilities are assumed to be zero. Further, in order to account for the new starting state, we begin the MDP one time-step earlier, setting τa m ← τa m − 1. This will not affect the resource allocation due to the resource constraints only being enforced for the original MDP states, as will be discussed in the next section. For example, the augmented MDPs shown in Figure 2b (which starts in state sb at time τ = 2) would be constructed from an MDP with original arrival time τ = 3. Figure 2b also shows a sample trajectory through the state space: the agent starts in state sb , transitions into the state space S of the original MDP, and finally exists into the sink state sf . Note that if we wanted to model a problem where agents could pause their MDPs at arbitrary time steps (which might be useful for domains where dynamic reallocation is possible), we could easily accomplish this by including an extra action that transitions from each state to itself with zero reward. 4.2 MILP for Resource Scheduling Given a set of augmented MDPs, as defined above, the goal of this section is to formulate a global optimization program that solves the resource-scheduling problem. In this section and below, all MDPs are assumed to be the augmented MDPs as defined in Section 4.1. Our approach is similar to the idea used in [6]: we begin with the linear-program formulation of agents" MDPs (1) and augment it with constraints that ensure that the corresponding resource allocation across agents and time is valid. The resulting optimization problem then simultaneously solves the agents" MDPs and resource-scheduling problems. In the rest of this section, we incrementally develop a mixed integer program (MILP) that achieves this. In the absence of resource constraints, the agents" finitehorizon MDPs are completely independent, and the globally optimal solution can be trivially obtained via the following LP, which is simply an aggregation of single-agent finitehorizon LPs: max X m X s X a rm(s, a) X t xm(s, a, t) subject to: X a xm(σ, a, t + 1) = X s,a pm(σ|s, a)xm(s, a, t), ∀m ∈ M, σ ∈ S, t ∈ [1, Tm − 1]; X a xm(s, a, 1) = αm(s), ∀m ∈ M, s ∈ S; (12) where xm(s, a, t) is the occupation measure of agent m, and The Sixth Intl. Joint Conf. on Autonomous Agents and Multi-Agent Systems (AAMAS 07) 1223 (a) (b) Figure 2: Illustration of augmenting an MDP to allow for variable starting and stopping times: a) (left) the original two-state MDP with a single action; (right) the augmented MDP with new states sb and sf and the new action a∗ (note that the origianl transitions are not changed in the augmentation process); b) the augmented MDP displayed as a trajectory through time (grey lines indicate all transitions, while black lines indicate a given trajectory. Objective Function (sum of expected rewards over all agents) max X m X s X a rm(s, a) X t xm(s, a, t) (5) Meaning Implication Linear Constraints Tie x to θ. Agent is only active when occupation measure is nonzero in original MDP states. θm(τ) = 0 =⇒ xm(s, a, τ −τa m+1) = 0 ∀s /∈ {sb , sf }, a ∈ A X s/∈{sb,sf } X a xm(s, a, t) ≤ θm(τa m + t − 1) ∀m ∈ M, ∀t ∈ [1, Tm] (6) Agent can only be active in τ ∈ (τa m, τd m) θm(τ) = 0 ∀m ∈ M, τ /∈ (τa m, τd m) (7) Cannot use resources when not active θm(τ) = 0 =⇒ Δm(τ, ω) = 0 ∀τ ∈ [0, bτ], ω ∈ Ω Δm(τ, ω) ≤ θm(τ) ∀m ∈ M, τ ∈ [0, bτ], ω ∈ Ω (8) Tie x to Δ (nonzero x forces corresponding Δ to be nonzero.) Δm(τ, ω) = 0, ϕm(a, ω) = 1 =⇒ xm(s, a, τ − τa m + 1) = 0 ∀s /∈ {sb , sf } 1/|A| X a ϕm(a, ω) X s/∈{sb,sf } xm(s, a, t) ≤ Δm(t + τa m − 1, ω) ∀m ∈ M, ω ∈ Ω, t ∈ [1, Tm] (9) Resource bounds X m Δm(τ, ω) ≤ bϕ(ω) ∀ω ∈ Ω, τ ∈ [0, bτ] (10) Agent cannot change resources while active. Only enabled for scheduling with static assignments. θm(τ) = 1 and θm(τ + 1) = 1 =⇒ Δm(τ, ω) = Δm(τ + 1, ω) Δm(τ, ω) − Z(1 − θm(τ + 1)) ≤ Δm(τ + 1, ω) + Z(1 − θm(τ)) Δm(τ, ω) + Z(1 − θm(τ + 1)) ≥ Δm(τ + 1, ω) − Z(1 − θm(τ)) ∀m ∈ M, ω ∈ Ω, τ ∈ [0, bτ] (11) Table 1: MILP for globally optimal resource scheduling. Tm = τd m − τa m + 1 is the time horizon for the agent"s MDP. Using this LP as a basis, we augment it with constraints that ensure that the resource usage implied by the agents" occupation measures {xm} does not violate the global resource requirements bϕ at any time step τ ∈ [0, bτ]. To formulate these resource constraints, we use the following binary variables: • Δm(τ, ω) = {0, 1}, ∀m ∈ M, τ ∈ [0, bτ], ω ∈ Ω, which serve as indicator variables that define whether agent m possesses resource ω at time τ. These are analogous to the static indicator variables used in the one-shot static resource-allocation problem in [6]. • θm = {0, 1}, ∀m ∈ M, τ ∈ [0, bτ] are indicator variables that specify whether agent m is active (i.e., executing its MDP) at time τ. The meaning of resource-usage variables Δ is illustrated in Figure 1: Δm(τ, ω) = 1 only if resource ω is allocated to agent m at time τ. The meaning of the activity indicators θ is illustrated in Figure 2b: when agent m is in either the start state sb or the finish state sf , the corresponding θm = 0, but once the agent becomes active and enters one of the other states, we set θm = 1 . This meaning of θ can be enforced with a linear constraint that synchronizes the values of the agents" occupation measures xm and the activity 1224 The Sixth Intl. Joint Conf. on Autonomous Agents and Multi-Agent Systems (AAMAS 07) indicators θ, as shown in (6) in Table 1. Another constraint we have to add-because the activity indicators θ are defined on the global timeline τ-is to enforce the fact that the agent is inactive outside of its arrivaldeparture window. This is accomplished by constraint (7) in Table 1. Furthermore, agents should not be using resources while they are inactive. This constraint can also be enforced via a linear inequality on θ and Δ, as shown in (8). Constraint (6) sets the value of θ to match the policy defined by the occupation measure xm. In a similar fashion, we have to make sure that the resource-usage variables Δ are also synchronized with the occupation measure xm. This is done via constraint (9) in Table 1, which is nearly identical to the analogous constraint from [6]. After implementing the above constraint, which enforces the meaning of Δ, we add a constraint that ensures that the agents" resource usage never exceeds the amounts of available resources. This condition is also trivially expressed as a linear inequality (10) in Table 1. Finally, for the problem formulation where resource assignments are static during a lifetime of an agent, we add a constraint that ensures that the resource-usage variables Δ do not change their value while the agent is active (θ = 1). This is accomplished via the linear constraint (11), where Z ≥ 2 is a constant that is used to turn off the constraints when θm(τ) = 0 or θm(τ + 1) = 0. This constraint is not used for the dynamic problem formulation, where resources can be reallocated between agents at every time step. To summarize, Table 1 together with the conservationof-flow constraints from (12) defines the MILP that simultaneously computes an optimal resource assignment for all agents across time as well as optimal finite-horizon MDP policies that are valid under that resource assignment. As a rough measure of the complexity of this MILP, let us consider the number of optimization variables and constraints. Let TM = P Tm = P m(τa m − τd m + 1) be the sum of the lengths of the arrival-departure windows across all agents. Then, the number of optimization variables is: TM + bτ|M||Ω| + bτ|M|, TM of which are continuous (xm), and bτ|M||Ω| + bτ|M| are binary (Δ and θ). However, notice that all but TM|M| of the θ are set to zero by constraint (7), which also immediately forces all but TM|M||Ω| of the Δ to be zero via the constraints (8). The number of constraints (not including the degenerate constraints in (7)) in the MILP is: TM + TM|Ω| + bτ|Ω| + bτ|M||Ω|. Despite the fact that the complexity of the MILP is, in the worst case, exponential1 in the number of binary variables, the complexity of this MILP is significantly (exponentially) lower than that of the MILP with flat utility functions, described in Section 2.2. This result echos the efficiency gains reported in [6] for single-shot resource-allocation problems, but is much more pronounced, because of the explosion of the flat utility representation due to the temporal aspect of the problem (recall the prohibitive complexity of the combinatorial optimization in Section 2.2). We empirically analyze the performance of this method in Section 5. 1 Strictly speaking, solving MILPs to optimality is NPcomplete in the number of integer variables. 5. EXPERIMENTAL RESULTS Although the complexity of solving MILPs is in the worst case exponential in the number of integer variables, there are many efficient methods for solving MILPs that allow our algorithm to scale well for parameters common to resource allocation and scheduling problems. In particular, this section introduces a problem domain-the repairshop problem-used to empirically evaluate our algorithm"s scalability in terms of the number of agents |M|, the number of shared resources |Ω|, and the varied lengths of global time bτ during which agents may enter and exit the system. The repairshop problem is a simple parameterized MDP adopting the metaphor of a vehicular repair shop. Agents in the repair shop are mechanics with a number of independent tasks that yield reward only when completed. In our MDP model of this system, actions taken to advance through the state space are only allowed if the agent holds certain resources that are publicly available to the shop. These resources are in finite supply, and optimal policies for the shop will determine when each agent may hold the limited resources to take actions and earn individual rewards. Each task to be completed is associated with a single action, although the agent is required to repeat the action numerous times before completing the task and earning a reward. This model was parameterized in terms of the number of agents in the system, the number of different types of resources that could be linked to necessary actions, a global time during which agents are allowed to arrive and depart, and a maximum length for the number of time steps an agent may remain in the system. All datapoints in our experiments were obtained with 20 evaluations using CPLEX to solve the MILPs on a Pentium4 computer with 2Gb of RAM. Trials were conducted on both the static and the dynamic version of the resourcescheduling problem, as defined earlier. Figure 3 shows the runtime and policy value for independent modifications to the parameter set. The top row shows how the solution time for the MILP scales as we increase the number of agents |M|, the global time horizon bτ, and the number of resources |Ω|. Increasing the number of agents leads to exponential complexity scaling, which is to be expected for an NP-complete problem. However, increasing the global time limit bτ or the total number of resource types |Ω|-while holding the number of agents constantdoes not lead to decreased performance. This occurs because the problems get easier as they become under-constrained, which is also a common phenomenon for NP-complete problems. We also observe that the solution to the dynamic version of the problem can often be computed much faster than the static version. The bottom row of Figure 3 shows the joint policy value of the policies that correspond to the computed optimal resource-allocation schedules. We can observe that the dynamic version yields higher reward (as expected, since the reward for the dynamic version is always no less than the reward of the static version). We should point out that these graphs should not be viewed as a measure of performance of two different algorithms (both algorithms produce optimal solutions but to different problems), but rather as observations about how the quality of optimal solutions change as more flexibility is allowed in the reallocation of resources. Figure 4 shows runtime and policy value for trials in which common input variables are scaled together. This allows The Sixth Intl. Joint Conf. on Autonomous Agents and Multi-Agent Systems (AAMAS 07) 1225 2 4 6 8 10 10 −3 10 −2 10 −1 10 0 10 1 10 2 10 3 10 4 Number of Agents |M| CPUTime,sec |Ω| = 5, τ = 50 static dynamic 50 100 150 200 10 −2 10 −1 10 0 10 1 10 2 10 3 Global Time Boundary τ CPUTime,sec |M| = 5, |Ω| = 5 static dynamic 10 20 30 40 50 10 −2 10 −1 10 0 10 1 10 2 Number of Resources |Ω| CPUTime,sec |M| = 5, τ = 50 static dynamic 2 4 6 8 10 200 400 600 800 1000 1200 1400 1600 Number of Agents |M| Value |Ω| = 5, τ = 50 static dynamic 50 100 150 200 400 500 600 700 800 900 1000 1100 1200 1300 1400 Global Time Boundary τ Value |M| = 5, |Ω| = 5 static dynamic 10 20 30 40 50 500 600 700 800 900 1000 1100 1200 1300 1400 Number of Resources |Ω| Value |M| = 5, τ = 50 static dynamic Figure 3: Evaluation of our MILP for variable numbers of agents (column 1), lengths of global-time window (column 2), and numbers of resource types (column 3). Top row shows CPU time, and bottom row shows the joint reward of agents" MDP policies. Error bars show the 1st and 3rd quartiles (25% and 75%). 2 4 6 8 10 10 −3 10 −2 10 −1 10 0 10 1 10 2 10 3 Number of Agents |M| CPUTime,sec τ = 10|M| static dynamic 2 4 6 8 10 10 −3 10 −2 10 −1 10 0 10 1 10 2 10 3 10 4 Number of Agents |M| CPUTime,sec |Ω| = 2|M| static dynamic 2 4 6 8 10 10 −3 10 −2 10 −1 10 0 10 1 10 2 10 3 10 4 Number of Agents |M| CPUTime,sec |Ω| = 5|M| static dynamic 2 4 6 8 10 200 400 600 800 1000 1200 1400 1600 1800 2000 2200 Number of Agents |M| Value τ = 10|M| static dynamic 2 4 6 8 10 200 400 600 800 1000 1200 1400 1600 1800 2000 Number of Agents |M| Value |Ω| = 2|M| static dynamic 2 4 6 8 10 0 500 1000 1500 2000 2500 Number of Agents |M| Value |Ω| = 5|M| static dynamic Figure 4: Evaluation of our MILP using correlated input variables. The left column tracks the performance and CPU time as the number of agents and global-time window increase together (bτ = 10|M|). The middle and the right column track the performance and CPU time as the number of resources and the number of agents increase together as |Ω| = 2|M| and |Ω| = 5|M|, respectively. Error bars show the 1st and 3rd quartiles (25% and 75%). 1226 The Sixth Intl. Joint Conf. on Autonomous Agents and Multi-Agent Systems (AAMAS 07) us to explore domains where the total number of agents scales proportionally to the total number of resource types or the global time horizon, while keeping constant the average agent density (per unit of global time) or the average number of resources per agent (which commonly occurs in real-life applications). Overall, we believe that these experimental results indicate that our MILP formulation can be used to effectively solve resource-scheduling problems of nontrivial size. 6. DISCUSSION AND CONCLUSIONS Throughout the paper, we have made a number of assumptions in our model and solution algorithm; we discuss their implications below. • Continual execution. We assume that once an agent stops executing its MDP (transitions into state sf ), it exits the system and cannot return. It is easy to relax this assumption for domains where agents" MDPs can be paused and restarted. All that is required is to include an additional pause action which transitions from a given state back to itself, and has zero reward. • Indifference to start time. We used a reward model where agents" rewards depend only on the time horizon of their MDPs and not the global start time. This is a consequence of our MDP-augmentation procedure from Section 4.1. It is easy to extend the model so that the agents incur an explicit penalty for idling by assigning a non-zero negative reward to the start state sb . • Binary resource requirements. For simplicity, we have assumed that resource costs are binary: ϕm(a, ω) = {0, 1}, but our results generalize in a straightforward manner to non-binary resource mappings, analogously to the procedure used in [5]. • Cooperative agents. The optimization procedure discussed in this paper was developed in the context of cooperative agents, but it can also be used to design a mechanism for scheduling resources among self-interested agents. This optimization procedure can be embedded in a VickreyClarke-Groves auction, completely analogously to the way it was done in [7]. In fact, all the results of [7] about the properties of the auction and information privacy directly carry over to the scheduling domain discussed in this paper, requiring only slight modifications to deal with finitehorizon MDPs. • Known, deterministic arrival and departure times. Finally, we have assumed that agents" arrival and departure times (τa m and τd m) are deterministic and known a priori. This assumption is fundamental to our solution method. While there are many domains where this assumption is valid, in many cases agents arrive and depart dynamically and their arrival and departure times can only be predicted probabilistically, leading to online resource-allocation problems. In particular, in the case of self-interested agents, this becomes an interesting version of an online-mechanism-design problem [11, 12]. In summary, we have presented an MILP formulation for the combinatorial resource-scheduling problem where agents" values for possible resource assignments are defined by finitehorizon MDPs. This result extends previous work ([6, 7]) on static one-shot resource allocation under MDP-induced preferences to resource-scheduling problems with a temporal aspect. As such, this work takes a step in the direction of designing an online mechanism for agents with combinatorial resource preferences induced by stochastic planning problems. Relaxing the assumption about deterministic arrival and departure times of the agents is a focus of our future work. We would like to thank the anonymous reviewers for their insightful comments and suggestions. 7. REFERENCES [1] E. Altman and A. Shwartz. Adaptive control of constrained Markov chains: Criteria and policies. Annals of Operations Research, special issue on Markov Decision Processes, 28:101-134, 1991. [2] R. Bellman. Dynamic Programming. Princeton University Press, 1957. [3] C. Boutilier. Solving concisely expressed combinatorial auction problems. In Proc. of AAAI-02, pages 359-366, 2002. [4] C. Boutilier and H. H. Hoos. Bidding languages for combinatorial auctions. In Proc. of IJCAI-01, pages 1211-1217, 2001. [5] D. Dolgov. Integrated Resource Allocation and Planning in Stochastic Multiagent Environments. PhD thesis, Computer Science Department, University of Michigan, February 2006. [6] D. A. Dolgov and E. H. Durfee. Optimal resource allocation and policy formulation in loosely-coupled Markov decision processes. In Proc. of ICAPS-04, pages 315-324, June 2004. [7] D. A. Dolgov and E. H. Durfee. Computationally efficient combinatorial auctions for resource allocation in weakly-coupled MDPs. In Proc. of AAMAS-05, New York, NY, USA, 2005. ACM Press. [8] D. A. Dolgov and E. H. Durfee. Resource allocation among agents with preferences induced by factored MDPs. In Proc. of AAMAS-06, 2006. [9] K. Larson and T. Sandholm. Mechanism design and deliberative agents. In Proc. of AAMAS-05, pages 650-656, New York, NY, USA, 2005. ACM Press. [10] N. Nisan. Bidding and allocation in combinatorial auctions. In Electronic Commerce, 2000. [11] D. C. Parkes and S. Singh. An MDP-based approach to Online Mechanism Design. In Proc. of the Seventeenths Annual Conference on Neural Information Processing Systems (NIPS-03), 2003. [12] D. C. Parkes, S. Singh, and D. Yanovsky. Approximately efficient online mechanism design. In Proc. of the Eighteenths Annual Conference on Neural Information Processing Systems (NIPS-04), 2004. [13] M. L. Puterman. Markov Decision Processes. John Wiley & Sons, New York, 1994. [14] M. H. Rothkopf, A. Pekec, and R. M. Harstad. Computationally manageable combinational auctions. Management Science, 44(8):1131-1147, 1998. [15] T. Sandholm. An algorithm for optimal winner determination in combinatorial auctions. In Proc. of IJCAI-99, pages 542-547, San Francisco, CA, USA, 1999. Morgan Kaufmann Publishers Inc. The Sixth Intl. Joint Conf. on Autonomous Agents and Multi-Agent Systems (AAMAS 07) 1227

multiagent system;multiagent plan;resource-scheduling;optimal allocation;markov decision process;utility function;resource;resource allocation;scheduling;discrete-time scheduling problem;resource-scheduling algorithm;task and resource allocation in agent system;optimal resource scheduling;optimization problem;combinatorial resource scheduling

train_I-64

Organizational Self-Design in Semi-dynamic Environments

Organizations are an important basis for coordination in multiagent systems. However, there is no best way to organize and all ways of organizing are not equally effective. Attempting to optimize an organizational structure depends strongly on environmental features including problem characteristics, available resources, and agent capabilities. If the environment is dynamic, the environmental conditions or the problem task structure may change over time. This precludes the use of static, design-time generated, organizational structures in such systems. On the other hand, for many real environments, the problems are not totally unique either: certain characteristics and conditions change slowly, if at all, and these can have an important effect in creating stable organizational structures. Organizational-Self Design (OSD) has been proposed as an approach for constructing suitable organizational structures at runtime. We extend the existing OSD approach to include worthoriented domains, model other resources in addition to only processor resources and build in robustness into the organization. We then evaluate our approach against the contract-net approach and show that our OSD agents perform better, are more efficient, and more flexible to changes in the environment.

1. INTRODUCTION In this paper, we are primarily interested in the organizational design of a multiagent system - the roles enacted by the agents, ∗Primary author is a student the coordination between the roles and the number and assignment of roles and resources to the individual agents. The organizational design is complicated by the fact that there is no best way to organize and all ways of organizing are not equally effective [2]. Instead, the optimal organizational structure depends both on the problem at hand and the environmental conditions under which the problem needs to be solved. The environmental conditions may not be known a priori, or may change over time, which would preclude the use of a static organizational structure. On the other hand, all problem instances and environmental conditions are not always unique, which would render inefficient the use of a new, bespoke organizational structure for every problem instance. Organizational Self-Design (OSD) [4, 10] has been proposed as an approach to designing organizations at run-time in which the agents are responsible for generating their own organizational structures. We believe that OSD is especially suited to the above scenario in which the environment is semi-dynamic as the agents can adapt to changes in the task structures and environmental conditions, while still being able to generate relatively stable organizational structures that exploit the common characteristics across problem instances. In our approach (as in [10]), we define two operators for OSD - agent spawning and composition - when an agent becomes overloaded, it spawns off a new agent to handle part of its task load/responsibility; when an agent lies idle for an extended period of time, it may decide to compose with another agent. We use TÆMS as the underlying representation for our problem solving requests. TÆMS [11] (Task Analysis, Environment Modeling and Simulation) is a computational framework for representing and reasoning about complex task environments in which tasks (problems) are represented using extended hierarchical task structures [3]. The root node of the task structure represents the high-level goal that the agent is trying to achieve. The sub-nodes of a node represent the subtasks and methods that make up the highlevel task. The leaf nodes are at the lowest level of abstraction and represent executable methods - the primitive actions that the agents can perform. The executable methods, themselves, may have multiple outcomes, with different probabilities and different characteristics such as quality, cost and duration. TÆMS also allows various mechanisms for specifying subtask variations and alternatives, i.e. each node in TÆMS is labeled with a characteristic accumulation function that describes how many or which subgoals or sets of subgoals need to be achieved in order to achieve a particular higherlevel goal. TÆMS has been used to model many different problemsolving environments including distributed sensor networks, information gathering, hospital scheduling, EMS, and military planning. [5, 6, 3, 15]. The main contributions of this paper are as follows: 1. We extend existing OSD approaches to use TÆMS as the underlying problem representation, which allows us to model and use OSD for worth-oriented domains. This in turn allows us to reason about (1) alternative task and role assignments that make different quality/cost tradeoffs and generate different organizational structures and (2) uncertainties in the execution of tasks. 2. We model the use of resources other than only processor resources. 3. We incorporate robustness into the organizational structures. 2. RELATED WORK The concept of OSD is not new and has been around since the work of Corkill and Lesser on the DVMT system[4], even though the concept was not fully developed by them. More recently Dignum et. al.[8] have described OSD in the context of the reorganization of agent societies and attempt to classify the various kinds of reorganization possible according to the the reason for reorganization, the type of reorganization and who is responsible for the reorganization decision. According to their scheme, the type of reorganization done by our agents falls into the category of structural changes and the reorganization decision can be described as shared command. Our research primarily builds on the work done by Gasser and Ishida [10], in which they use OSD in the context of a production system in order to perform adaptive work allocation and load balancing. In their approach, they define two organizational primitives - composition and decomposition, which are similar to our organizational primitives for agent spawning and composition. The main difference between their work and our work is that we use TÆMS as the underlying representation for our problems, which allows, firstly, the representation of a larger, more general class of problems and, secondly, quantitative reasoning over task structures. The latter also allows us to incorporate different design-to-criteria schedulers [16]. Horling and Lesser [9] present a different, top-down approach to OSD that also uses TÆMS as the underlying representation. However, their approach assumes a fixed number of agents with designated (and fixed) roles. OSD is used in their work to change the interaction patterns between the agents and results in the agents using different subtasks or different resources to achieve their goals. We also extend on the work done by Sycara et. al.,[13] on Agent Cloning, which is another approach to resource allocation and load balancing. In this approach, the authors present agent cloning as a possible response to agent overload - if an agent detects that it is overloaded and that there are spare (unused) resources in the system, the agent clones itself and gives its clone some part of its task load. Hence, agent cloning can be thought of as akin to agent spawning in our approach. However, the two approaches are different in that there is no specialization of the agents in the formerthe cloned agents are perfect replicas of the original agents and fulfill the same roles and responsibilities as the original agents. In our approach, on the other hand, the spawned agents are specialized on a subpart of the spawning agent"s task structure, which is no longer the responsibility of the spawning agent. Hence, our approach also deals with explicit organization formation and the coordination of the agents" tasks which are not handled by their approach. Other approaches to OSD include the work of So and Durfee [14], who describe a top-down model of OSD in the context of Cooperative Distributive Problem Solving (CDPS) and Barber and Martin [1], who describe an adaptive decision making framework in which agents are able to reorganize decision-making groups by dynamically changing (1) who makes the decisions for a particular goal and (2) who must carry out these decisions.The latter work is primarily concerned with coordination decisions and can be used to complement our OSD work, which primarily deals with task and resource allocation. 3. TASK AND RESOURCE MODEL To ground our discussion of OSD, we now formally describe our task and resource model. In our model, the primary input to the multi-agent system (MAS) is an ordered set of problem solving requests or task instances, < P1, P2, P3, ..., Pn >, where each problem solving request, Pi, can be represented using the tuple < ti, ai, di >. In this scheme, ti is the underlying TÆMS task structure, ai ∈ N+ is the arrival time and di ∈ N+ is the deadline of the ith task instance1 . The MAS has no prior knowledge about the task ti before the arrival time, ai. In order for the MAS to accrue quality, the task ti must be completed before the deadline, di. Furthermore, every underlying task structure, ti, can be represented using the tuple < T, τ, M, Q, E, R, ρ, C >, where: • T is the set of tasks. The tasks are non-leaf nodes in a TÆMS task structure and are used to denote goals that the agents must achieve. Tasks have a characteristic accumulation function (see below) and are themselves composed of other subtasks and/or methods that need to be achieved in order to achieve the goal represented by that task. Formally, each task Tj can be represented using the pair (qj, sj), where qj ∈ Q and sj ⊂ (T ∪ M). For our convenience, we define two functions SUBTASKS(Task) : T → P(T ∪ M) and SUPERTASKS(TÆMS node) : T ∪ M → P(T), that return the subtasks and supertasks of a TÆMS node respectively2 . • τ ∈ T, is the root of the task structure, i.e. the highest level goal that the organization is trying to achieve. The quality accrued on a problem is equal to the quality of task τ. • M is the set executable methods, i.e., M = {m1, m2, ..., mn}, where each method, mk, is represented using the outcome distribution, {(o1, p1), (o2, p2), ..., (om, pm)}. In the pair (ol, pl), ol is an outcome and pl is the probability that executing mk will result in the outcome ol. Furthermore, each outcome, ol is represented using the triple (ql, cl, dl), where ql is the quality distribution, cl is the cost distribution and dl is the duration distribution of outcome ol. Each discrete distribution is itself a set of pairs, {(n1, p1), (n2, p2), ..., (nn, pn)}, where pi ∈ + is the probability that the outcome will have a quality/cost/duration of nl ∈ N depending on the type of distribution and Pm i=1 pl = 1. • Q is the set of quality/characteristic accumulation functions (CAFs). The CAFs determine how a task group accrues quality given the quality accrued by its subtasks/methods. For our research, we use four CAFs: MIN, MAX, SUM and EXACTLY ONE. See [5] for formal definitions. • E is the set of (non-local) effects. Again, see [5] for formal definitions. • R is the set of resources. • ρ is a mapping from an executable method and resource to the quantity of that resource needed (by an agent) to schedule/execute that method. That is ρ(method, resource) : M × R → N. 1 N is the set of natural numbers including zero and N+ is the set of positive natural numbers excluding zero. 2 P is the power set of set, i.e., the set of all subsets of a set The Sixth Intl. Joint Conf. on Autonomous Agents and Multi-Agent Systems (AAMAS 07) 1229 • C is a mapping from a resource to the cost of that resource, that is C(resource) : R → N+ We also make the following set of assumptions in our research: 1. The agents in the MAS are drawn from the infinite set A = {a1, a2, a3, ...}. That is, we do not assume a fixed set of agents - instead agents are created (spawned) and destroyed (combined) as needed. 2. All problem solving requests have the same underlying task structure, i.e. ∃t∀iti = t, where t is the task structure of the problem that the MAS is trying to solve. We believe that this assumption holds for many of the practical problems that we have in mind because TÆMS task structures are basically high-level plans for achieving some goal in which the steps required for achieving the goal-as well as the possible contingency situations-have been pre-computed offline and represented in the task structure. Because it represents many contingencies, alternatives, uncertain characteristics and runtime flexible choices, the same underlying task structure can play out very differently across specific instances. 3. All resources are exclusive, i.e., only one agent may use a resource at any given time. Furthermore, we assume that each agent has to own the set of resources that it needseven though the resource ownership can change during the evolution of the organization. 4. All resources are non-consumable. 4. ORGANIZATIONAL SELF DESIGN 4.1 Agent Roles and Relationships The organizational structure is primarily composed of roles and the relationships between the roles. One or more agents may enact a particular role and one or more roles must be enacted by every agent. The roles may be thought of as the parts played by the agents enacting the roles in the solution to the problem and reflect the long-term commitments made by the agents in question to a certain course of action (that includes task responsibility, authority, and mechanisms for coordination). The relationships between the roles are the coordination relationships that exist between the subparts of a problem. In our approach, the organizational design is directly contingent on the task structure and the environmental conditions under which the problems need to be solved. We define a role as a TÆMS subtree rooted at a particular node. Hence, the set (T ∪ M) encompasses the space of all possible roles. Note, by definition, a role may consist of one or more other (sub-) roles as a particular TÆMS node may itself be made up of one or more subtrees. Hence, we will use the terms role, task node and task interchangeably. We, also, differentiate between local and managed (non-local) roles. Local roles are roles that are the sole responsibility of a single agent, that is, the agent concerned is responsible for solving all the subproblems of the tree rooted at that node. For such roles, the agent concerned can do one or more subtasks, solely at its discretion and without consultation with any other agent. Managed roles, on the other hand, must be coordinated between two or more agents as such roles will have two or more descendent local roles that are the responsibility of two or more separate agents. Any of the existing coordination mechanisms (such as GPGP [11]) can be used to achieve this coordination. Formally, if the function TYPE(Agent, TÆMS Node) : A×(T ∪ M) → {Local, Managed, Unassigned}, returns the type of the responsibility of the agent towards the specified role, then TYPE(a, r) = Local ⇐⇒ ∀ri∈SUBTASKS(r)TYPE(a, ri) = Local TYPE(a, r) = Managed ⇐⇒ [∃a1∃r1(r1 ∈ SUBTASKS(r)) ∧ (TYPE(a1, r1) = Managed)] ∨ [∃a2∃a3∃r2∃r3(a2 = a3) ∧ (r2 = r3) ∧ (r2 ∈ SUBTASKS(r)) ∧ (r3 ∈ SUBTASKS(r)) ∧ (TYPE(a2, r2) = Local) ∧ (TYPE(a3, r3) = Local)] 4.2 Organization Formation and Adaptation To form or adapt their organizational structure, the agents use two organizational primitives: agent spawning and composition. These two primitives result in a change in the assignment of roles to the agents. Agent spawning is the generation of a new agent to handle a subset of the roles of the spawning agent. Agent composition, on the other hand, is orthogonal to agent spawning and involves the merging of two or more agents together - the combined agent is responsible for enacting all the roles of the agents being merged. In order to participate in the formation and adaption of an organization, the agents need to explicitly represent and reason about the role assignments. Hence, as a part of its organizational knowledge, each agent keeps a list of the local roles that it is enacting and the non-local roles that it is managing. Note that each agent only has limited organizational knowledge and is individually responsible for spawning off or combining with another agent, as needed, based on its estimate of its performance so far. To see how the organizational primitives work, we first describe four rules that can be thought of as the organizational invariants which will always hold before and after any organizational change: 1. For a local role, all the descendent nodes of that role will be local. TYPE(a, r) = Local =⇒ ∀ri∈SUBTASKS(r)TYPE(a, ri) = Local 2. Similarly, for a managed (non-local) role, all the ascendent nodes of that role will be managed. TYPE(a, r) = Managed =⇒ ∀ri∈SUPERTASKS(r)∃ai(ai ∈ A) ∧ (TYPE(ai, ri) = Managed) 3. If two local roles that are enacted by two different agents share a common ancestor, that ancestor will be a managed role. (TYPE(a1, r1) = Local) ∧ (TYPE(a2, r2) = Local)∧ (a1 = a2) ∧ (r1 = r2) =⇒ ∀ri∈(SUPERTASKS(r1)∩SUPERTASKS(r2))∃ai(ai ∈ A)∧ (TYPE(ai, ri) = Managed) 4. If all the direct descendants of a role are local and the sole responsibility of a single agent, that role will be a local role. ∃a∃r∀ri∈SUBTASKS(r)(a ∈ A) ∧ (r ∈ (T ∪ M))∧ (TYPE(a, ri) = Local) =⇒ (TYPE(a, r) = Local) When a new agent is spawned, the agent doing the spawning will assign one or more of its local roles to the newly spawned agent (Algorithm 1). To preserve invariant rules 2 and 3, the spawning agent will change the type of all the ascendent roles of the nodes assigned to the newly spawned agent from local to managed. Note that the spawning agent is only changing its local organizational knowledge and not the global organizational knowledge. At the 1230 The Sixth Intl. Joint Conf. on Autonomous Agents and Multi-Agent Systems (AAMAS 07) same time, the spawning agent is taking on the task of managing the previously local roles. Similarly, the newly spawned agent will only know of its just assigned local roles. When an agent (the composing agent) decides to compose with another agent (the composed agent), the organizational knowledge of the composing agent is merged with the organizational knowledge of the composed agent. To do this, the composed agent takes on the roles of all the local and managed tasks of the composing agent. Care is taken to preserve the organizational invariant rules 1 and 4. Algorithm 1 SpawnAgent(SpawningAgent) : A → A 1: LocalRoles ← {r ⊆ (T ∪ M) | TYPE(SpawningAgent, r)= Local} 2: NewAgent ← CREATENEWAGENT() 3: NewAgentRoles ← FINDROLESFORSPAWNEDAGENT (LocalRoles) 4: for role in NewAgentRoles do 5: TYPE(NewAgent, role) ← Local 6: TYPE(SpawningAgent, role) ← Unassigned 7: PRESERVEORGANIZATIONALINVARIANTS() 8: return NewAgent Algorithm 2 FINDROLESFORSPAWNEDAGENT (SpawningAgentRoles) : (T ∪ M) → (T ∪ M) 1: R ← SpawningAgentRoles 2: selectedRoles ← nil 3: for roleSet in [P(R) − {φ, R}] do 4: if COST(R, roleSet) < COST(R, selectedRoles) then 5: selectedRoles ← roleSet 6: return selectedRoles Algorithm 3 GETRESOURCECOST(Roles) : (T ∪ M) → 1: M ← (Roles ∩ M) 2: cost ← 0 3: for resource in R do 4: maxResourceUsage ← 0 5: for method in M do 6: if ρ(method, resource) > maxResourceUsage then 7: max ← ρ(method, resource) 8: cost ← cost + [C(resource) × maxResourceUsage] 9: return cost 4.2.1 Role allocation during spawning One of the key questions that the agent doing the spawning needs to answer is - which of its local-roles should it assign to the newly spawned agent and which of its local roles should it keep to itself? The onus of answering this question falls on the FINDROLESFORSPAWNEDAGENT() function, shown in Algorithm 2 above. This function takes the set of local roles that are the responsibility of the spawning agent and returns a subset of those roles for allocation to the newly spawned agent. This subset is selected based on the results of a cost function as is evident from line 4 of the algorithm. Since the use of different cost functions will result in different organizational structures and since we have no a priori reason to believe that one cost function will out-perform the other, we evaluated the performance of three different cost functions based on the following three different heuristics: Algorithm 4 GETEXPECTEDDURATION(Roles) : (T ∪ M) → N+ 1: M ← (Roles ∩ M) 2: exptDuration ← 0 3: for [outcome =< (q, c, d), outcomeProb >] in M do 4: exptOutcomeDuration ← 0 5: for (n,p) in d do 6: exptOutcomeDuration ← n × p 7: exptDuration ← exptDuration + [exptOutcomeDuration × outcomeProb] 8: return exptDuration Allocating top-most roles first: This heuristic always breaks up at the top-most nodes first. That is, if the nodes of a task structure were numbered, starting from the root, in a breadth-first fashion, then this heuristic would select the local-role of the spawning agent that had the lowest number and breakup that node (by allocating one of its subtasks to the newly spawned agent). We selected this heuristic because (a) it is the simplest to implement, (b) fastest to run (the role allocation can be done in constant time without the need of a search through the task structure) and (c) it makes sense from a human-organizational perspective as this heuristic corresponds to dividing an organization along functional lines. Minimizing total resources: This heuristic attempts to minimize the total cost of the resources needed by the agents in the organization to execute their roles. If R be the local roles of the spawning agent and R be the subset of roles being evaluated for allocation to the newly spawned agent, the cost function for this heuristic is given by: COST(R, R ) ← GETRESOURCECOST(R − R )+GETRESOURCECOST(R ) Balancing execution time: This heuristic attempts to allocate roles in a way that tries to ensure that each agent has an equal amount of work to do. For each potential role allocation, this heuristic works by calculating the absolute value of the difference between the expected duration of its own roles after spawning and the expected duration of the roles of the newly spawned agent. If this difference is close to zero, then the both the agents have roughly the same amount of work to do. Formally, if R be the local roles of the spawning agent and R be the subset of roles being evaluated for allocation to the newly spawned agent, then the cost function for this heuristic is given by: COST(R, R ) ← |GETEXPECTEDDURATION(R−R )−GETEXPECTEDDURATION(R )| To evaluate these heuristics, we ran a series of experiments that tested the performance of the resultant organization on randomly generated task structures. The results are given in Section 6. 4.3 Reasons for Organizational Change As organizational change is expensive (requiring clock cycles, allocation/deallocation of resources, etc.) we want a stable organizational structure that is suited to the task and environmental conditions at hand. Hence, we wish to change the organizational structure only if the task structure and/or environmental conditions change. Also to allow temporary changes to the environmental conditions to be overlooked, we want the probability of an organizational change to be inversely proportional to the time since the last organizational change. If this time is relatively short, the agents are still adjusting to the changes in the environment - hence the probability of an agent initiating an organizational change should be high. Similarly, if the time since the last organizational change is relatively large, we wish to have a low probability of organizational change. To allow this variation in probability of organizational change, we use simulated annealing to determine the probability of keepThe Sixth Intl. Joint Conf. on Autonomous Agents and Multi-Agent Systems (AAMAS 07) 1231 ing an existing organizational structure. This probability is calculated using the annealing formula: p = e− ΔE kT where ΔE is the amount of overload/underload, T is the time since the last organizational change and k is a constant. The mechanism of computing ΔE is different for agent spawning than for agent composition and is described below. From this formula, if T is large, p, or the probability of keeping the existing organizational structure is large. Note that the value of p is capped at a certain threshold in order to prevent the organization from being too sluggish in its reaction to environmental change. To compute if agent spawning is necessary, we use the annealing equation with ΔE = 1 α∗Slack where α is a constant and Slack is the difference between the total time available for completion of the outstanding tasks and the sum of the expected time required for completion of each task on the task queue. Also, if the amount of Slack is negative, immediate agent spawning will occur without use of the annealing equation. To calculate if agent composition is necessary, we again use the simulated annealing equation. However, in this case, ΔE = β ∗ Idle Time, where β is a constant and Idle Time is the amount of time for which the agent was idle. If the agent has been sitting idle for a long period of time, ΔE is large, which implies that p, the probability of keeping the existing organizational structure, is low. 5. ORGANIZATION AND ROBUSTNESS There are two approaches commonly used to achieve robustness in multiagent systems: 1. the Survivalist Approach [12], which involves replicating domain agents in order to allow the replicas to take over should the original agents fail; and 2. the Citizen Approach [7], which involves the use of special monitoring agents (called Sentinel Agents) in order to detect agent failure and dynamically startup new agents in lieu of the failed ones. The advantage of the survivalist approach is that recovery is relatively fast, since the replicas are pre-existing in the organization and can take over as soon as a failure is detected. The advantages of the citizen approach are that it requires fewer resources, little modification to the existing organizational structure and coordination protocol and is simpler to implement. Both of these approaches can be applied to achieve robustness in our OSD agents and it is not clear which approach would be better. Rather a thorough empirical evaluation of both approaches would be required. In this paper, we present the citizen approach as it has been shown by [7], to have a better performance than the survivalist approach in the Contract Net protocol, and leave the presentation and evaluation of the survivalist approach to a future paper. To implement the citizen approach, we designed special monitoring agents, that periodically poll the domain agents by sending them are you alive messages that the agents must respond to. If an agent fails, it will not respond to such messages - the monitoring agents can then create a new agent and delegate the responsibilities of the dead agent to the new agent. This delegation of responsibilities is non-trivial as the monitoring agents do not have access to the internal state of the domain agents, which is itself composed of two components - the organizational knowledge and the task information. The former consists of the information about the local and managerial roles of the agent while the latter is composed of the methods that are being scheduled and executed and the tasks that have been delegated to other agents. This state information can only be deduced by monitoring and recording the messages being sent and received by the domain agents. For example, in order to deduce the organizational knowledge, the monitoring agents need to keep a track of the spawn and compose messages sent by the agents in order to trigger the spawning and composition operations respectively. The deduction process is particularly complicated in the case of the task information as the monitoring agents do not have access to the private schedules of the domain agents. The details are beyond the scope of this paper. 6. EVALUATION To evaluate our approach, we ran a series of experiments that simulated the operation of both the OSD agents and the Contract Net agents on various task structures with varied arrival rates and deadlines. At the start of each experiment, a random TÆMS task structure was generated with a specified depth and branching factor. During the course of the experiment, a series of task instances (problems) arrive at the organization and must be completed by the agents before their specified deadlines. To directly compare the OSD approach with the Contract Net approach, each experiment was repeated several times - using OSD agents on the first run and a different number of Contract Net agents on each subsequent run. We were careful to use the same task structure, task arrival times, task deadlines and random numbers for each of these trials. We divided the experiments into two groups: experiments in which the environment was static (fixed task arrival rates and deadlines) and experiments in which the environment was dynamic (varying arrival rates and/or deadlines). The two graphs in Figure 1, show the average performance of the OSD organization against the Contract Net organizations with 8, 10, 12 and 14 agents. The results shown are the averages of running 40 experiments. 20 of those experiments had a static environment with a fixed task arrival time of 15 cycles and a deadline window of 20 cycles. The remaining 20 experiments had a varying task arrival rate - the task arrival rate was changed from 15 cycles to 30 cycles and back to 15 cycles after every 20 tasks. In all the experiments, the task structures were randomly generated with a maximum depth of 4 and a maximum branching factor of 3. The runtime of all the experiments was 2500 cycles. We tested several hypotheses relating to the comparative performance of our OSD approach using the Wilcoxon Matched-Pair Signed-Rank tests. Matched-Pair signifies that we are comparing the performance of each system on precisely the same randomized task set within each separate experiment. The tested hypothesis are: The OSD organization requires fewer agents to complete an equal or larger number of tasks when compared to the Contract Net organization: To test this hypothesis, we tested the stronger null hypothesis that states that the contract net agents complete more tasks. This null hypothesis is rejected for all contract net organizations with less than 14 agents (static: p < 0.0003; dynamic: p < 0.03). For large contract net organizations, the number of tasks completed is statistically equivalent to the number completed by the OSD agents, however the number of agents used by the OSD organization is smaller: 9.59 agents (in the static case) and 7.38 agents (in the dynamic case) versus 14 contract net agents3 . Thus the original hypothesis, that OSD requires fewer agents to 3 These values should not be construed as an indication of the scalability of our approach. We have tested our approach on organizations with more than 300 agents, which is significantly greater than the number of agents needed for the kind of applications that we have in mind (i.e. web service choreography, efficient dynamic use of grid computing, distributed information gathering, etc.). 1232 The Sixth Intl. Joint Conf. on Autonomous Agents and Multi-Agent Systems (AAMAS 07) Figure 1: Graph comparing the average performance of the OSD organization with the Contract Net organizations (with 8, 10, 12 and 14 agents). The error bars show the standard deviations. complete an equal or larger number of tasks, is upheld. The OSD organizations achieve an equal or greater average quality than the Contract Net organizations: The null hypothesis is that the Contract Net agents achieve a greater average quality. We can reject the null hypothesis for contract net organizations with less than 12 agents (static: p < 0.01; dynamic: p < 0.05). For larger contract net organizations, the average quality is statistically equivalent to that achieved by OSD. The OSD agents have a lower average response time as compared to the Contract Net agents: The null hypothesis that OSD has the same or higher response time is rejected for all contract net organizations (static: p < 0.0002; dynamic: p < 0.0004). The OSD agents send less messages than the Contract Net Agents: The null hypothesis that OSD sends the same or more messages is rejected for all contract net organizations (p < .0003 in all cases except 8 contract net agents in a static environment where p < 0.02) Hence, as demonstrated by the above tests, our agents perform better than the contract net agents as they complete a larger number of tasks, achieve a greater quality and also have a lower response time and communication overhead. These results make intuitive sense given our goals for the OSD approach. We expected the OSD organizations to have a faster average response time and to send less messages because the agents in the OSD organization are not wasting time and messages sending bid requests and replying to bids. The quality gained on the tasks is directly dependent on the Criteria/Heuristic BET TF MR Rand Number of Agents 572 567 100 139 No-Org-Changes 641 51 5 177 Total-Messages-Sent 586 499 13 11 Resource-Cost 346 418 337 66 Tasks-Completed 427 560 154 166 Average-Quality 367 492 298 339 Average-Response-Time 356 321 370 283 Average-Runtime 543 323 74 116 Average-Turnaround-Time 560 314 74 126 Table 1: The number of times that each heuristic performed the best or statistically equivalent to the best for each of the performance criteria. Heuristic Key: BET is Balancing Execution Time, TF is Topmost First, MR is Minimizing Resources and Rand is a random allocation strategy, in which every TÆMS node has a uniform probability of being selected for allocation. number of tasks completed, hence the more the number of tasks completed, the greater average quality. The results of testing the first hypothesis were slightly more surprising. It appears that due to the inherent inefficiency of the contract net protocol in bidding for each and every task instance, a greater number of agents are needed to complete an equal number of tasks. Next, we evaluated the performance of the three heuristics for allocating tasks. Some preliminary experiments (that are not reported here due to space constraints) demonstrated the lack of a clear winner amongst the three heuristics for most of the performance criteria that we evaluated. We suspected this to be the case because different heuristics are better for different task structures and environmental conditions, and since each experiment starts with a different random task structure, we couldn"t find one allocation strategy that always dominated the other for all the performance criteria. To determine which heuristic performs the best, given a set of task structures, environmental conditions and performance criteria, we performed a series of experiments that were controlled using the following five variables: • The depth of the task structure was varied from 3 to 5. • The branching factor was varied from 3 to 5. • The probability of any given task node having a MIN CAF was varied from 0.0 to 1.0 in increments of 0.2. The probability of any node having a SUM CAF was in turn modified to ensure that the probabilities add up to 14 . • The arrival rate: from 10 to 40 cycles in increments of 10. • The deadline slack: from 5 to 15 in increments of 5. Each experiment was repeated 20 times, with a new task structure being generated each time - these 20 experiments formed an experimental set. Hence, all the experiments in an experimental set had the same values for the exogenous variables that were used to control the experiment. Note that a static environment was used in each of these experiments, as we wanted to see the performance of the arrival rate and deadline slack on each of the three heuristics. Also the results of any experiment in which the OSD organization consisted of a single agent ware culled from the results. Similarly, 4 Since our preliminary analysis led is to believe that the number of MAX and EXACTLY ONE CAFs in a task structure have a minimal effect on the performance of the allocation strategies being evaluated, we set the probabilities of the MAX and EXACTLY ONE CAFs to 0 in order to reduce the combinatorial explosion of the full factorial experimental design. The Sixth Intl. Joint Conf. on Autonomous Agents and Multi-Agent Systems (AAMAS 07) 1233 experiments in which the generated task structures were unsatisfiable (given the deadline constraints), were removed from the final results. If any experimental set had more than 15 experiments thus removed, the whole set was ignored for performing the evaluation. The final evaluation was done on 673 experimental sets. We tested the potential of these three heuristics on the following performance criteria: 1. The average number of agents used. 2. The total number of organizational changes. 3. The total messages sent by all the agents. 4. The total resource cost of the organization. 5. The number of tasks completed. 6. The average quality accrued. The average quality is defined as the total quality accrued during the experimental run divided by the sum of the number of tasks completed and the number of tasks failed. 7. The average response time of the organization. The response time of a task is defined as the difference between the time at which any agent in the organization starts working on the task (the start time) and the time at which the task was generated (the generation time). Hence, the response time is equivalent to the wait time. For tasks that are never attempted/started, the response time is set at final runtime minus the generation time. 8. The average runtime of the tasks attempted by the organization. This time is defined as the difference between the time at which the task completed or failed and the start time. For tasks that were never stated, this time is set to zero. 9. The turnaround time is defined as the sum of the response time and runtime of a task. Except for the number of tasks completed and the average quality accrued, lower values for the various performance criteria indicate better performance. Again we ran the Wilcoxon Matched-Pair Signed-Rank tests on the experiments in each of the experimental sets. The null hypothesis in each case was that there is no difference between the pair of heuristics for the performance criteria under consideration. We were interested in the cases in which we could reject the null hypothesis with 95% confidence (p < 0.05). We noted the number of times that a heuristic performed the best or was in a group that performed statistically better than the rest. These counts are given in Tables 1 and 2. The number of experimental sets in which each heuristic performed the best or statistically equivalent to the best is shown in Table 1. The breakup of these numbers into (1) the number of times that each heuristic performed better than all the other heuristics and (2) the number of times each heuristic was statistically equivalent to another group of heuristics, all of which performed the best, is shown in Table 2. Both of these tables allow us to glean important information about the performance of the three heuristics. Particularly interesting were the following results: • Whereas Balancing Execution Time (BET) used the lowest number of agents in largest number of experimental sets (572), in most of these cases (337 experimental sets) it was statistically equivalent to Topmost First (TF). When these two heuristics didn"t perform equally, there was an almost even split between the number of experimental sets in which one outperformed the other. We believe this was the case because BET always bifurcates the agents into two agents that have a more or less equal task load. This often results in organizations that have an even Figure 2: Graph demonstrating the robustness of the citizen approach. The baseline shows the number of tasks completed in the absence of any failure. number of agents - none of which are small5 enough to combine into a larger agent. With TF, on the other hand, a large agent can successively spawn off smaller agents until it and the spawned agents are small enough to complete their tasks before the deadlines - this often results in organizations with an odd number of agents that is less than those used by BET. • As expected, BET achieved the lowest number of organizational changes in the largest number of experimental sets. In fact, it was over ten times as good as its second best competitor (TF). This shows that if the agents are conscientious in their initial task allocation, there is a lesser need for organizational change later on, especially for static environments. • A particularly interesting, yet easily explainable, result was that of the average response time. We found that the Minimizing Resources (MR) heuristic performed the best when it came to minimizing the average response time! This can be explained by the fact the MR heuristic is extremely greedy and prefers to spawn off small agents that have a tiny resource footprint (so as to minimize the total increase in the resource cost to the organization at the time of spawning). Whereas most of these small agents might compose with other agents over time, the presence of a single small agent is sufficient to reduce the response time. In fact the MR heuristic is not the most effective heuristic when it comes to minimizing the resource-cost of the organization - in fact, it only outperforms a random task/resource allocation. We believe this is in part due to the greedy nature of this heuristic and in part because of the fact that all spawning and composition operations only use local information. We believe that using some non-local information about the resource allocation might help in making better decisions, something that we plan to look at in the future. Finally we evaluated the performance of the citizens approach to robustness as applied to our OSD mechanism (Figure 2). As expected, as the probability of failure increases, the number of agents failing during a run also increases. This results in a slight decrease in the number of tasks completed, which can be explained by the fact that whenever an agent fails, its looses whatever work it was doing at the time. The newly created agent that fills in for the failed 5 For this discussion small agents are agents that have a low expected duration for their local roles (as calculated by Algorithm 4). 1234 The Sixth Intl. Joint Conf. on Autonomous Agents and Multi-Agent Systems (AAMAS 07) Criteria/Heuristic BET TF MR Rand BET+TF BET+Rand MR+Rand TF+MR BET+TF+MR All Number of Agents 94 88 3 7 337 2 0 0 12 85 No-Org-Changes 480 0 0 29 16 113 0 0 0 5 Total-Messages-Sent 170 85 0 2 399 1 0 0 7 5 Resource-Cost 26 100 170 42 167 0 7 6 128 15 Tasks-Completed 77 197 4 28 184 1 3 9 36 99 Average-Quality 38 147 26 104 76 0 11 11 34 208 Average-Response-Time 104 74 162 43 31 20 16 8 7 169 Average-Runtime 322 110 0 12 121 13 1 1 1 69 Average-Turnaround-Time 318 94 1 11 125 26 1 0 7 64 Table 2: Table showing the number of times that each individual heuristic performed the best and the number of times that a certain group of statistically equivalent heuristics performed the best. Only the more interesting heuristic groupings are shown. All shows the number of experimental sets in which there was no statistical difference between the three heuristics and a random allocation strategy one must redo the work, thus wasting precious time which might not be available close to a deadline. As a part of our future research, we wish to, firstly, evaluate the survivalist approach to robustness. The survivalist approach might actually be better than the citizen approach for higher probabilities of agent failure, as the replicated agents may be processing the task structures in parallel and can take over the moment the original agents fail - thus saving time around tight deadlines. Also, we strongly believe that the optimal organizational structure may vary, depending on the probability of failure and the desired level of robustness. For example, one way of achieving a higher level of robustness in the survivalist approach, given a large numbers of agent failures, would be to relax the task deadlines. However, such a relaxation would result in the system using fewer agents in order to conserve resources, which in turn would have a detrimental effect on the robustness. Therefore, towards this end, we have begun exploring the robustness properties of task structures and the ways in which the organizational design can be modified to take such properties into account. 7. CONCLUSION In this paper, we have presented a run-time approach to organization in which the agents use Organizational Self-Design to come up with a suitable organizational structure. We have also evaluated the performance of the organizations generated by the agents following our approach with the bespoke organization formation that takes place in the Contract Net protocol and have demonstrated that our approach is better than the Contract Net approach as evident by the larger number of tasks completed, larger quality achieved and lower response time. Finally, we tested the performance of three different resource allocation heuristics on various performance metrics and also evaluated the robustness of our approach. 8. REFERENCES [1] K. S. Barber and C. E. Martin. Dynamic reorganization of decision-making groups. In AGENTS "01, pages 513-520, New York, NY, USA, 2001. [2] K. M. Carley and L. Gasser. Computational organization theory. In G. Wiess, editor, Multiagent Systems: A Modern Approach to Distributed Artificial Intelligence, pages 299-330, MIT Press, 1999. [3] W. Chen and K. S. Decker. The analysis of coordination in an information system application - emergency medical services. In Lecture Notes in Computer Science (LNCS), number 3508, pages 36-51. Springer-Verlag, May 2005. [4] D. Corkill and V. Lesser. The use of meta-level control for coordination in a distributed problem solving network. Proceedings of the Eighth International Joint Conference on Artificial Intelligence, pages 748-756, August 1983. [5] K. S. Decker. Environment centered analysis and design of coordination mechanisms. Ph.D. Thesis, Dept. of Comp. Science, University of Massachusetts, Amherst, May 1995. [6] K. S. Decker and J. Li. Coordinating mutually exclusive resources using GPGP. Autonomous Agents and Multi-Agent Systems, 3(2):133-157, 2000. [7] C. Dellarocas and M. Klein. An experimental evaluation of domain-independent fault handling services in open multi-agent systems. Proceedings of the International Conference on Multi-Agent Systems (ICMAS-2000), July 2000. [8] V. Dignum, F. Dignum, and L. Sonenberg. Towards Dynamic Reorganization of Agent Societies. In Proceedings of CEAS: Workshop on Coordination in Emergent Agent Societies at ECAI, pages 22-27, Valencia, Spain, September 2004. [9] B. Horling, B. Benyo, and V. Lesser. Using self-diagnosis to adapt organizational structures. In AGENTS "01, pages 529-536, New York, NY, USA, 2001. ACM Press. [10] T. Ishida, L. Gasser, and M. Yokoo. Organization self-design of distributed production systems. IEEE Transactions on Knowledge and Data Engineering, 4(2):123-134, 1992. [11] V. R. Lesser et. al. Evolution of the gpgp/tæms domain-independent coordination framework. Autonomous Agents and Multi-Agent Systems, 9(1-2):87-143, 2004. [12] O. Marin, P. Sens, J. Briot, and Z. Guessoum. Towards adaptive fault tolerance for distributed multi-agent systems. Proceedings of ERSADS 2001, May 2001. [13] O. Shehory, K. Sycara, et. al. Agent cloning: an approach to agent mobility and resource allocation. IEEE Communications Magazine, 36(7):58-67, 1998. [14] Y. So and E. Durfee. An organizational self-design model for organizational change. In AAAI-93 Workshop on AI and Theories of Groups and Organizations, pages 8-15, Washington, D.C., July 1993. [15] T. Wagner. Coordination decision support assistants (coordinators). Technical Report 04-29, BAA, 2004. [16] T. Wagner and V. Lesser. Design-to-criteria scheduling: Real-time agent control. Proc. of AAAI 2000 Spring Symposium on Real-Time Autonomous Systems, 89-96. The Sixth Intl. Joint Conf. on Autonomous Agents and Multi-Agent Systems (AAMAS 07) 1235

organization;organizational-self design;task and resource allocation;agent spawning;robustness;organizational structure;extended hierarchical task structure;environment modeling;composition;task analysis;coordination;simulation;multiagent system;organizational self-design

train_I-65

Graphical Models for Online Solutions to Interactive POMDPs

We develop a new graphical representation for interactive partially observable Markov decision processes (I-POMDPs) that is significantly more transparent and semantically clear than the previous representation. These graphical models called interactive dynamic influence diagrams (I-DIDs) seek to explicitly model the structure that is often present in real-world problems by decomposing the situation into chance and decision variables, and the dependencies between the variables. I-DIDs generalize DIDs, which may be viewed as graphical representations of POMDPs, to multiagent settings in the same way that I-POMDPs generalize POMDPs. I-DIDs may be used to compute the policy of an agent online as the agent acts and observes in a setting that is populated by other interacting agents. Using several examples, we show how I-DIDs may be applied and demonstrate their usefulness.

1. INTRODUCTION Interactive partially observable Markov decision processes (IPOMDPs) [9] provide a framework for sequential decision-making in partially observable multiagent environments. They generalize POMDPs [13] to multiagent settings by including the other agents" computable models in the state space along with the states of the physical environment. The models encompass all information influencing the agents" behaviors, including their preferences, capabilities, and beliefs, and are thus analogous to types in Bayesian games [11]. I-POMDPs adopt a subjective approach to understanding strategic behavior, rooted in a decision-theoretic framework that takes a decision-maker"s perspective in the interaction. In [15], Polich and Gmytrasiewicz introduced interactive dynamic influence diagrams (I-DIDs) as the computational representations of I-POMDPs. I-DIDs generalize DIDs [12], which may be viewed as computational counterparts of POMDPs, to multiagents settings in the same way that I-POMDPs generalize POMDPs. I-DIDs contribute to a growing line of work [19] that includes multi-agent influence diagrams (MAIDs) [14], and more recently, networks of influence diagrams (NIDs) [8]. These formalisms seek to explicitly model the structure that is often present in real-world problems by decomposing the situation into chance and decision variables, and the dependencies between the variables. MAIDs provide an alternative to normal and extensive game forms using a graphical formalism to represent games of imperfect information with a decision node for each agent"s actions and chance nodes capturing the agent"s private information. MAIDs objectively analyze the game, efficiently computing the Nash equilibrium profile by exploiting the independence structure. NIDs extend MAIDs to include agents" uncertainty over the game being played and over models of the other agents. Each model is a MAID and the network of MAIDs is collapsed, bottom up, into a single MAID for computing the equilibrium of the game keeping in mind the different models of each agent. Graphical formalisms such as MAIDs and NIDs open up a promising area of research that aims to represent multiagent interactions more transparently. However, MAIDs provide an analysis of the game from an external viewpoint and the applicability of both is limited to static single play games. Matters are more complex when we consider interactions that are extended over time, where predictions about others" future actions must be made using models that change as the agents act and observe. I-DIDs address this gap by allowing the representation of other agents" models as the values of a special model node. Both, other agents" models and the original agent"s beliefs over these models are updated over time using special-purpose implementations. In this paper, we improve on the previous preliminary representation of the I-DID shown in [15] by using the insight that the static I-ID is a type of NID. Thus, we may utilize NID-specific language constructs such as multiplexers to represent the model node, and subsequently the I-ID, more transparently. Furthermore, we clarify the semantics of the special purpose policy link introduced in the representation of I-DID by [15], and show that it could be replaced by traditional dependency links. In the previous representation of the I-DID, the update of the agent"s belief over the models of others as the agents act and receive observations was denoted using a special link called the model update link that connected the model nodes over time. We explicate the semantics of this link by showing how it can be implemented using the traditional dependency links between the chance nodes that constitute the model nodes. The net result is a representation of I-DID that is significantly more transparent, semantically clear, and capable of being implemented using the standard algorithms for solving DIDs. We show how IDIDs may be used to model an agent"s uncertainty over others" models, that may themselves be I-DIDs. Solution to the I-DID is a policy that prescribes what the agent should do over time, given its beliefs over the physical state and others" models. Analogous to DIDs, I-DIDs may be used to compute the policy of an agent online as the agent acts and observes in a setting that is populated by other interacting agents. 2. BACKGROUND: FINITELY NESTED IPOMDPS Interactive POMDPs generalize POMDPs to multiagent settings by including other agents" models as part of the state space [9]. Since other agents may also reason about others, the interactive state space is strategically nested; it contains beliefs about other agents" models and their beliefs about others. For simplicity of presentation we consider an agent, i, that is interacting with one other agent, j. A finitely nested I-POMDP of agent i with a strategy level l is defined as the tuple: I-POMDPi,l = ISi,l, A, Ti, Ωi, Oi, Ri where: • ISi,l denotes a set of interactive states defined as, ISi,l = S × Mj,l−1, where Mj,l−1 = {Θj,l−1 ∪ SMj}, for l ≥ 1, and ISi,0 = S, where S is the set of states of the physical environment. Θj,l−1 is the set of computable intentional models of agent j: θj,l−1 = bj,l−1, ˆθj where the frame, ˆθj = A, Ωj, Tj, Oj, Rj, OCj . Here, j is Bayes rational and OCj is j"s optimality criterion. SMj is the set of subintentional models of j. Simple examples of subintentional models include a no-information model [10] and a fictitious play model [6], both of which are history independent. We give a recursive bottom-up construction of the interactive state space below. ISi,0 = S, Θj,0 = { bj,0, ˆθj | bj,0 ∈ Δ(ISj,0)} ISi,1 = S × {Θj,0 ∪ SMj}, Θj,1 = { bj,1, ˆθj | bj,1 ∈ Δ(ISj,1)} . . . . . . ISi,l = S × {Θj,l−1 ∪ SMj}, Θj,l = { bj,l, ˆθj | bj,l ∈ Δ(ISj,l)} Similar formulations of nested spaces have appeared in [1, 3]. • A = Ai × Aj is the set of joint actions of all agents in the environment; • Ti : S ×A×S → [0, 1], describes the effect of the joint actions on the physical states of the environment; • Ωi is the set of observations of agent i; • Oi : S × A × Ωi → [0, 1] gives the likelihood of the observations given the physical state and joint action; • Ri : ISi × A → R describes agent i"s preferences over its interactive states. Usually only the physical states will matter. Agent i"s policy is the mapping, Ω∗ i → Δ(Ai), where Ω∗ i is the set of all observation histories of agent i. Since belief over the interactive states forms a sufficient statistic [9], the policy can also be represented as a mapping from the set of all beliefs of agent i to a distribution over its actions, Δ(ISi) → Δ(Ai). 2.1 Belief Update Analogous to POMDPs, an agent within the I-POMDP framework updates its belief as it acts and observes. However, there are two differences that complicate the belief update in multiagent settings when compared to single agent ones. First, since the state of the physical environment depends on the actions of both agents, i"s prediction of how the physical state changes has to be made based on its prediction of j"s actions. Second, changes in j"s models have to be included in i"s belief update. Specifically, if j is intentional then an update of j"s beliefs due to its action and observation has to be included. In other words, i has to update its belief based on its prediction of what j would observe and how j would update its belief. If j"s model is subintentional, then j"s probable observations are appended to the observation history contained in the model. Formally, we have: Pr(ist |at−1 i , bt−1 i,l ) = β ISt−1:mt−1 j =θt j bt−1 i,l (ist−1 ) × at−1 j Pr(at−1 j |θt−1 j,l−1)Oi(st , at−1 i , at−1 j , ot i) ×Ti(st−1 , at−1 i , at−1 j , st ) ot j Oj(st , at−1 i , at−1 j , ot j) ×τ(SEθt j (bt−1 j,l−1, at−1 j , ot j) − bt j,l−1) (1) where β is the normalizing constant, τ is 1 if its argument is 0 otherwise it is 0, Pr(at−1 j |θt−1 j,l−1) is the probability that at−1 j is Bayes rational for the agent described by model θt−1 j,l−1, and SE(·) is an abbreviation for the belief update. For a version of the belief update when j"s model is subintentional, see [9]. If agent j is also modeled as an I-POMDP, then i"s belief update invokes j"s belief update (via the term SEθt j ( bt−1 j,l−1 , at−1 j , ot j)), which in turn could invoke i"s belief update and so on. This recursion in belief nesting bottoms out at the 0th level. At this level, the belief update of the agent reduces to a POMDP belief update. 1 For illustrations of the belief update, additional details on I-POMDPs, and how they compare with other multiagent frameworks, see [9]. 2.2 Value Iteration Each belief state in a finitely nested I-POMDP has an associated value reflecting the maximum payoff the agent can expect in this belief state: Un ( bi,l, θi ) = max ai∈Ai is∈ISi,l ERi(is, ai)bi,l(is)+ γ oi∈Ωi Pr(oi|ai, bi,l)Un−1 ( SEθi (bi,l, ai, oi), θi ) (2) where, ERi(is, ai) = aj Ri(is, ai, aj)Pr(aj|mj,l−1) (since is = (s, mj,l−1)). Eq. 2 is a basis for value iteration in I-POMDPs. Agent i"s optimal action, a∗ i , for the case of finite horizon with discounting, is an element of the set of optimal actions for the belief state, OPT(θi), defined as: OPT( bi,l, θi ) = argmax ai∈Ai is∈ISi,l ERi(is, ai)bi,l(is) +γ oi∈Ωi Pr(oi|ai, bi,l)Un ( SEθi (bi,l, ai, oi), θi ) (3) 3. INTERACTIVEINFLUENCEDIAGRAMS A naive extension of influence diagrams (IDs) to settings populated by multiple agents is possible by treating other agents as automatons, represented using chance nodes. However, this approach assumes that the agents" actions are controlled using a probability distribution that does not change over time. Interactive influence diagrams (I-IDs) adopt a more sophisticated approach by generalizing IDs to make them applicable to settings shared with other agents who may act and observe, and update their beliefs. 3.1 Syntax In addition to the usual chance, decision, and utility nodes, IIDs include a new type of node called the model node. We show a general level l I-ID in Fig. 1(a), where the model node (Mj,l−1) is denoted using a hexagon. We note that the probability distribution over the chance node, S, and the model node together represents agent i"s belief over its interactive states. In addition to the model 1 The 0th level model is a POMDP: Other agent"s actions are treated as exogenous events and folded into the T, O, and R functions. The Sixth Intl. Joint Conf. on Autonomous Agents and Multi-Agent Systems (AAMAS 07) 815 Figure 1: (a) A generic level l I-ID for agent i situated with one other agent j. The hexagon is the model node (Mj,l−1) whose structure we show in (b). Members of the model node are I-IDs themselves (m1 j,l−1, m2 j,l−1; diagrams not shown here for simplicity) whose decision nodes are mapped to the corresponding chance nodes (A1 j , A2 j ). Depending on the value of the node, Mod[Mj], the distribution of each of the chance nodes is assigned to the node Aj. (c) The transformed I-ID with the model node replaced by the chance nodes and the relationships between them. node, I-IDs differ from IDs by having a dashed link (called the policy link in [15]) between the model node and a chance node, Aj, that represents the distribution over the other agent"s actions given its model. In the absence of other agents, the model node and the chance node, Aj, vanish and I-IDs collapse into traditional IDs. The model node contains the alternative computational models ascribed by i to the other agent from the set, Θj,l−1 ∪ SMj, where Θj,l−1 and SMj were defined previously in Section 2. Thus, a model in the model node may itself be an I-ID or ID, and the recursion terminates when a model is an ID or subintentional. Because the model node contains the alternative models of the other agent as its values, its representation is not trivial. In particular, some of the models within the node are I-IDs that when solved generate the agent"s optimal policy in their decision nodes. Each decision node is mapped to the corresponding chance node, say A1 j , in the following way: if OPT is the set of optimal actions obtained by solving the I-ID (or ID), then Pr(aj ∈ A1 j ) = 1 |OP T | if aj ∈ OPT, 0 otherwise. Borrowing insights from previous work [8], we observe that the model node and the dashed policy link that connects it to the chance node, Aj, could be represented as shown in Fig. 1(b). The decision node of each level l − 1 I-ID is transformed into a chance node, as we mentioned previously, so that the actions with the largest value in the decision node are assigned uniform probabilities in the chance node while the rest are assigned zero probability. The different chance nodes (A1 j , A2 j ), one for each model, and additionally, the chance node labeled Mod[Mj] form the parents of the chance node, Aj. Thus, there are as many action nodes (A1 j , A2 j ) in Mj,l−1 as the number of models in the support of agent i"s beliefs. The conditional probability table of the chance node, Aj, is a multiplexer that assumes the distribution of each of the action nodes (A1 j , A2 j ) depending on the value of Mod[Mj]. The values of Mod[Mj] denote the different models of j. In other words, when Mod[Mj] has the value m1 j,l−1, the chance node Aj assumes the distribution of the node A1 j , and Aj assumes the distribution of A2 j when Mod[Mj] has the value m2 j,l−1. The distribution over the node, Mod[Mj], is the agent i"s belief over the models of j given a physical state. For more agents, we will have as many model nodes as there are agents. Notice that Fig. 1(b) clarifies the semantics of the policy link, and shows how it can be represented using the traditional dependency links. In Fig. 1(c), we show the transformed I-ID when the model node is replaced by the chance nodes and relationships between them. In contrast to the representation in [15], there are no special-purpose policy links, rather the I-ID is composed of only those types of nodes that are found in traditional IDs and dependency relationships between the nodes. This allows I-IDs to be represented and implemented using conventional application tools that target IDs. Note that we may view the level l I-ID as a NID. Specifically, each of the level l − 1 models within the model node are blocks in the NID (see Fig. 2). If the level l = 1, each block is a traditional ID, otherwise if l > 1, each block within the NID may itself be a NID. Note that within the I-IDs (or IDs) at each level, there is only a single decision node. Thus, our NID does not contain any MAIDs. Figure 2: A level l I-ID represented as a NID. The probabilities assigned to the blocks of the NID are i"s beliefs over j"s models conditioned on a physical state. 3.2 Solution The solution of an I-ID proceeds in a bottom-up manner, and is implemented recursively. We start by solving the level 0 models, which, if intentional, are traditional IDs. Their solutions provide probability distributions over the other agents" actions, which are entered in the corresponding chance nodes found in the model node of the level 1 I-ID. The mapping from the level 0 models" decision nodes to the chance nodes is carried out so that actions with the largest value in the decision node are assigned uniform probabilities in the chance node while the rest are assigned zero probability. Given the distributions over the actions within the different chance nodes (one for each model of the other agent), the level 1 I-ID is transformed as shown in Fig. 1(c). During the transformation, the conditional probability table (CPT) of the node, Aj, is populated such that the node assumes the distribution of each of the chance nodes depending on the value of the node, Mod[Mj]. As we mentioned previously, the values of the node Mod[Mj] denote the different models of the other agent, and its distribution is the agent i"s belief over the models of j conditioned on the physical state. The transformed level 1 I-ID is a traditional ID that may be solved us816 The Sixth Intl. Joint Conf. on Autonomous Agents and Multi-Agent Systems (AAMAS 07) (a) (b) Figure 3: (a) A generic two time-slice level l I-DID for agent i in a setting with one other agent j. Notice the dotted model update link that denotes the update of the models of j and the distribution over the models over time. (b) The semantics of the model update link. ing the standard expected utility maximization method [18]. This procedure is carried out up to the level l I-ID whose solution gives the non-empty set of optimal actions that the agent should perform given its belief. Notice that analogous to IDs, I-IDs are suitable for online decision-making when the agent"s current belief is known. 4. INTERACTIVE DYNAMIC INFLUENCE DIAGRAMS Interactive dynamic influence diagrams (I-DIDs) extend I-IDs (and NIDs) to allow sequential decision-making over several time steps. Just as DIDs are structured graphical representations of POMDPs, I-DIDs are the graphical online analogs for finitely nested I-POMDPs. I-DIDs may be used to optimize over a finite look-ahead given initial beliefs while interacting with other, possibly similar, agents. 4.1 Syntax We depict a general two time-slice I-DID in Fig. 3(a). In addition to the model nodes and the dashed policy link, what differentiates an I-DID from a DID is the model update link shown as a dotted arrow in Fig. 3(a). We explained the semantics of the model node and the policy link in the previous section; we describe the model updates next. The update of the model node over time involves two steps: First, given the models at time t, we identify the updated set of models that reside in the model node at time t + 1. Recall from Section 2 that an agent"s intentional model includes its belief. Because the agents act and receive observations, their models are updated to reflect their changed beliefs. Since the set of optimal actions for a model could include all the actions, and the agent may receive any one of |Ωj| possible observations, the updated set at time step t + 1 will have at most |Mt j,l−1||Aj||Ωj| models. Here, |Mt j,l−1| is the number of models at time step t, |Aj| and |Ωj| are the largest spaces of actions and observations respectively, among all the models. Second, we compute the new distribution over the updated models given the original distribution and the probability of the agent performing the action and receiving the observation that led to the updated model. These steps are a part of agent i"s belief update formalized using Eq. 1. In Fig. 3(b), we show how the dotted model update link is implemented in the I-DID. If each of the two level l − 1 models ascribed to j at time step t results in one action, and j could make one of two possible observations, then the model node at time step t + 1 contains four updated models (mt+1,1 j,l−1 ,mt+1,2 j,l−1 , mt+1,3 j,l−1 , and mt+1,4 j,l−1 ). These models differ in their initial beliefs, each of which is the result of j updating its beliefs due to its action and a possible observation. The decision nodes in each of the I-DIDs or DIDs that represent the lower level models are mapped to the corresponding Figure 4: Transformed I-DID with the model nodes and model update link replaced with the chance nodes and the relationships (in bold). chance nodes, as mentioned previously. Next, we describe how the distribution over the updated set of models (the distribution over the chance node Mod[Mt+1 j ] in Mt+1 j,l−1) is computed. The probability that j"s updated model is, say mt+1,1 j,l−1 , depends on the probability of j performing the action and receiving the observation that led to this model, and the prior distribution over the models at time step t. Because the chance node At j assumes the distribution of each of the action nodes based on the value of Mod[Mt j ], the probability of the action is given by this chance node. In order to obtain the probability of j"s possible observation, we introduce the chance node Oj, which depending on the value of Mod[Mt j ] assumes the distribution of the observation node in the lower level model denoted by Mod[Mt j ]. Because the probability of j"s observations depends on the physical state and the joint actions of both agents, the node Oj is linked with St+1 , At j, and At i. 2 Analogous to At j, the conditional probability table of Oj is also a multiplexer modulated by Mod[Mt j ]. Finally, the distribution over the prior models at time t is obtained from the chance node, Mod[Mt j ] in Mt j,l−1. Consequently, the chance nodes, Mod[Mt j ], At j, and Oj, form the parents of Mod[Mt+1 j ] in Mt+1 j,l−1. Notice that the model update link may be replaced by the dependency links between the chance nodes that constitute the model nodes in the two time slices. In Fig. 4 we show the two time-slice I-DID with the model nodes replaced by the chance nodes and the relationships between them. Chance nodes and dependency links that not in bold are standard, usually found in DIDs. Expansion of the I-DID over more time steps requires the repetition of the two steps of updating the set of models that form the 2 Note that Oj represents j"s observation at time t + 1. The Sixth Intl. Joint Conf. on Autonomous Agents and Multi-Agent Systems (AAMAS 07) 817 values of the model node and adding the relationships between the chance nodes, as many times as there are model update links. We note that the possible set of models of the other agent j grows exponentially with the number of time steps. For example, after T steps, there may be at most |Mt=1 j,l−1|(|Aj||Ωj|)T −1 candidate models residing in the model node. 4.2 Solution Analogous to I-IDs, the solution to a level l I-DID for agent i expanded over T time steps may be carried out recursively. For the purpose of illustration, let l=1 and T=2. The solution method uses the standard look-ahead technique, projecting the agent"s action and observation sequences forward from the current belief state [17], and finding the possible beliefs that i could have in the next time step. Because agent i has a belief over j"s models as well, the lookahead includes finding out the possible models that j could have in the future. Consequently, each of j"s subintentional or level 0 models (represented using a standard DID) in the first time step must be solved to obtain its optimal set of actions. These actions are combined with the set of possible observations that j could make in that model, resulting in an updated set of candidate models (that include the updated beliefs) that could describe the behavior of j. Beliefs over this updated set of candidate models are calculated using the standard inference methods using the dependency relationships between the model nodes as shown in Fig. 3(b). We note the recursive nature of this solution: in solving agent i"s level 1 I-DID, j"s level 0 DIDs must be solved. If the nesting of models is deeper, all models at all levels starting from 0 are solved in a bottom-up manner. We briefly outline the recursive algorithm for solving agent i"s Algorithm for solving I-DID Input : level l ≥ 1 I-ID or level 0 ID, T Expansion Phase 1. For t from 1 to T − 1 do 2. If l ≥ 1 then Populate Mt+1 j,l−1 3. For each mt j in Range(Mt j,l−1) do 4. Recursively call algorithm with the l − 1 I-ID (or ID) that represents mt j and the horizon, T − t + 1 5. Map the decision node of the solved I-ID (or ID), OPT(mt j), to a chance node Aj 6. For each aj in OPT(mt j) do 7. For each oj in Oj (part of mt j) do 8. Update j"s belief, bt+1 j ← SE(bt j, aj, oj) 9. mt+1 j ← New I-ID (or ID) with bt+1 j as the initial belief 10. Range(Mt+1 j,l−1) ∪ ← {mt+1 j } 11. Add the model node, Mt+1 j,l−1, and the dependency links between Mt j,l−1 and Mt+1 j,l−1 (shown in Fig. 3(b)) 12. Add the chance, decision, and utility nodes for t + 1 time slice and the dependency links between them 13. Establish the CPTs for each chance node and utility node Look-Ahead Phase 14. Apply the standard look-ahead and backup method to solve the expanded I-DID Figure 5: Algorithm for solving a level l ≥ 0 I-DID. level l I-DID expanded over T time steps with one other agent j in Fig. 5. We adopt a two-phase approach: Given an I-ID of level l (described previously in Section 3) with all lower level models also represented as I-IDs or IDs (if level 0), the first step is to expand the level l I-ID over T time steps adding the dependency links and the conditional probability tables for each node. We particularly focus on establishing and populating the model nodes (lines 3-11). Note that Range(·) returns the values (lower level models) of the random variable given as input (model node). In the second phase, we use a standard look-ahead technique projecting the action and observation sequences over T time steps in the future, and backing up the utility values of the reachable beliefs. Similar to I-IDs, the I-DIDs reduce to DIDs in the absence of other agents. As we mentioned previously, the 0-th level models are the traditional DIDs. Their solutions provide probability distributions over actions of the agent modeled at that level to I-DIDs at level 1. Given probability distributions over other agent"s actions the level 1 IDIDs can themselves be solved as DIDs, and provide probability distributions to yet higher level models. Assume that the number of models considered at each level is bound by a number, M. Solving an I-DID of level l in then equivalent to solving O(Ml ) DIDs. 5. EXAMPLE APPLICATIONS To illustrate the usefulness of I-DIDs, we apply them to three problem domains. We describe, in particular, the formulation of the I-DID and the optimal prescriptions obtained on solving it. 5.1 Followership-Leadership in the Multiagent Tiger Problem We begin our illustrations of using I-IDs and I-DIDs with a slightly modified version of the multiagent tiger problem discussed in [9]. The problem has two agents, each of which can open the right door (OR), the left door (OL) or listen (L). In addition to hearing growls (from the left (GL) or from the right (GR)) when they listen, the agents also hear creaks (from the left (CL), from the right (CR), or no creaks (S)), which noisily indicate the other agent"s opening one of the doors. When any door is opened, the tiger persists in its original location with a probability of 95%. Agent i hears growls with a reliability of 65% and creaks with a reliability of 95%. Agent j, on the other hand, hears growls with a reliability of 95%. Thus, the setting is such that agent i hears agent j opening doors more reliably than the tiger"s growls. This suggests that i could use j"s actions as an indication of the location of the tiger, as we discuss below. Each agent"s preferences are as in the single agent game discussed in [13]. The transition, observation, and reward functions are shown in [16]. A good indicator of the usefulness of normative methods for decision-making like I-DIDs is the emergence of realistic social behaviors in their prescriptions. In settings of the persistent multiagent tiger problem that reflect real world situations, we demonstrate followership between the agents and, as shown in [15], deception among agents who believe that they are in a follower-leader type of relationship. In particular, we analyze the situational and epistemological conditions sufficient for their emergence. The followership behavior, for example, results from the agent knowing its own weaknesses, assessing the strengths, preferences, and possible behaviors of the other, and realizing that its best for it to follow the other"s actions in order to maximize its payoffs. Let us consider a particular setting of the tiger problem in which agent i believes that j"s preferences are aligned with its own - both of them just want to get the gold - and j"s hearing is more reliable in comparison to itself. As an example, suppose that j, on listening can discern the tiger"s location 95% of the times compared to i"s 65% accuracy. Additionally, agent i does not have any initial information about the tiger"s location. In other words, i"s single-level nested belief, bi,1, assigns 0.5 to each of the two locations of the tiger. In addition, i considers two models of j, which differ in j"s flat level 0 initial beliefs. This is represented in the level 1 I-ID shown in Fig. 6(a). According to one model, j assigns a probability of 0.9 that the tiger is behind the left door, while the other 818 The Sixth Intl. Joint Conf. on Autonomous Agents and Multi-Agent Systems (AAMAS 07) Figure 6: (a) Level 1 I-ID of agent i, (b) two level 0 IDs of agent j whose decision nodes are mapped to the chance nodes, A1 j , A2 j , in (a). model assigns 0.1 to that location (see Fig. 6(b)). Agent i is undecided on these two models of j. If we vary i"s hearing ability, and solve the corresponding level 1 I-ID expanded over three time steps, we obtain the normative behavioral policies shown in Fig 7 that exhibit followership behavior. If i"s probability of correctly hearing the growls is 0.65, then as shown in the policy in Fig. 7(a), i begins to conditionally follow j"s actions: i opens the same door that j opened previously iff i"s own assessment of the tiger"s location confirms j"s pick. If i loses the ability to correctly interpret the growls completely, it blindly follows j and opens the same door that j opened previously (Fig. 7(b)). Figure 7: Emergence of (a) conditional followership, and (b) blind followership in the tiger problem. Behaviors of interest are in bold. * is a wildcard, and denotes any one of the observations. We observed that a single level of belief nesting - beliefs about the other"s models - was sufficient for followership to emerge in the tiger problem. However, the epistemological requirements for the emergence of leadership are more complex. For an agent, say j, to emerge as a leader, followership must first emerge in the other agent i. As we mentioned previously, if i is certain that its preferences are identical to those of j, and believes that j has a better sense of hearing, i will follow j"s actions over time. Agent j emerges as a leader if it believes that i will follow it, which implies that j"s belief must be nested two levels deep to enable it to recognize its leadership role. Realizing that i will follow presents j with an opportunity to influence i"s actions in the benefit of the collective good or its self-interest alone. For example, in the tiger problem, let us consider a setting in which if both i and j open the correct door, then each gets a payoff of 20 that is double the original. If j alone selects the correct door, it gets the payoff of 10. On the other hand, if both agents pick the wrong door, their penalties are cut in half. In this setting, it is in both j"s best interest as well as the collective betterment for j to use its expertise in selecting the correct door, and thus be a good leader. However, consider a slightly different problem in which j gains from i"s loss and is penalized if i gains. Specifically, let i"s payoff be subtracted from j"s, indicating that j is antagonistic toward i - if j picks the correct door and i the wrong one, then i"s loss of 100 becomes j"s gain. Agent j believes that i incorrectly thinks that j"s preferences are those that promote the collective good and that it starts off by believing with 99% confidence where the tiger is. Because i believes that its preferences are similar to those of j, and that j starts by believing almost surely that one of the two is the correct location (two level 0 models of j), i will start by following j"s actions. We show i"s normative policy on solving its singly-nested I-DID over three time steps in Fig. 8(a). The policy demonstrates that i will blindly follow j"s actions. Since the tiger persists in its original location with a probability of 0.95, i will select the same door again. If j begins the game with a 99% probability that the tiger is on the right, solving j"s I-DID nested two levels deep, results in the policy shown in Fig. 8(b). Even though j is almost certain that OL is the correct action, it will start by selecting OR, followed by OL. Agent j"s intention is to deceive i who, it believes, will follow j"s actions, so as to gain $110 in the second time step, which is more than what j would gain if it were to be honest. Figure 8: Emergence of deception between agents in the tiger problem. Behaviors of interest are in bold. * denotes as before. (a) Agent i"s policy demonstrating that it will blindly follow j"s actions. (b) Even though j is almost certain that the tiger is on the right, it will start by selecting OR, followed by OL, in order to deceive i. 5.2 Altruism and Reciprocity in the Public Good Problem The public good (PG) problem [7], consists of a group of M agents, each of whom must either contribute some resource to a public pot or keep it for themselves. Since resources contributed to the public pot are shared among all the agents, they are less valuable to the agent when in the public pot. However, if all agents choose to contribute their resources, then the payoff to each agent is more than if no one contributes. Since an agent gets its share of the public pot irrespective of whether it has contributed or not, the dominating action is for each agent to not contribute, and instead free ride on others" contributions. However, behaviors of human players in empirical simulations of the PG problem differ from the normative predictions. The experiments reveal that many players initially contribute a large amount to the public pot, and continue to contribute when the PG problem is played repeatedly, though in decreasing amounts [4]. Many of these experiments [5] report that a small core group of players persistently contributes to the public pot even when all others are defecting. These experiments also reveal that players who persistently contribute have altruistic or reciprocal preferences matching expected cooperation of others. For simplicity, we assume that the game is played between M = 2 agents, i and j. Let each agent be initially endowed with XT amount of resources. While the classical PG game formulation permits each agent to contribute any quantity of resources (≤ XT ) to the public pot, we simplify the action space by allowing two possible actions. Each agent may choose to either contribute (C) a fixed amount of the resources, or not contribute. The latter action is deThe Sixth Intl. Joint Conf. on Autonomous Agents and Multi-Agent Systems (AAMAS 07) 819 noted as defect (D). We assume that the actions are not observable to others. The value of resources in the public pot is discounted by ci for each agent i, where ci is the marginal private return. We assume that ci < 1 so that the agent does not benefit enough that it contributes to the public pot for private gain. Simultaneously, ciM > 1, making collective contribution pareto optimal. i/j C D C 2ciXT , 2cjXT ciXT − cp, XT + cjXT − P D XT + ciXT − P, cjXT − cp XT , XT Table 1: The one-shot PG game with punishment. In order to encourage contributions, the contributing agents punish free riders but incur a small cost for administering the punishment. Let P be the punishment meted out to the defecting agent and cp the non-zero cost of punishing for the contributing agent. For simplicity, we assume that the cost of punishing is same for both the agents. The one-shot PG game with punishment is shown in Table. 1. Let ci = cj, cp > 0, and if P > XT − ciXT , then defection is no longer a dominating action. If P < XT − ciXT , then defection is the dominating action for both. If P = XT − ciXT , then the game is not dominance-solvable. Figure 9: (a) Level 1 I-ID of agent i, (b) level 0 IDs of agent j with decision nodes mapped to the chance nodes, A1 j and A2 j , in (a). We formulate a sequential version of the PG problem with punishment from the perspective of agent i. Though in the repeated PG game, the quantity in the public pot is revealed to all the agents after each round of actions, we assume in our formulation that it is hidden from the agents. Each agent may contribute a fixed amount, xc, or defect. An agent on performing an action receives an observation of plenty (PY) or meager (MR) symbolizing the state of the public pot. Notice that the observations are also indirectly indicative of agent j"s actions because the state of the public pot is influenced by them. The amount of resources in agent i"s private pot, is perfectly observable to i. The payoffs are analogous to Table. 1. Borrowing from the empirical investigations of the PG problem [5], we construct level 0 IDs for j that model altruistic and non-altruistic types (Fig. 9(b)). Specifically, our altruistic agent has a high marginal private return (cj is close to 1) and does not punish others who defect. Let xc = 1 and the level 0 agent be punished half the times it defects. With one action remaining, both types of agents choose to contribute to avoid being punished. With two actions to go, the altruistic type chooses to contribute, while the other defects. This is because cj for the altruistic type is close to 1, thus the expected punishment, 0.5P > (1 − cj), which the altruistic type avoids. Because cj for the non-altruistic type is less, it prefers not to contribute. With three steps to go, the altruistic agent contributes to avoid punishment (0.5P > 2(1 − cj)), and the non-altruistic type defects. For greater than three steps, while the altruistic agent continues to contribute to the public pot depending on how close its marginal private return is to 1, the non-altruistic type prescribes defection. We analyzed the decisions of an altruistic agent i modeled using a level 1 I-DID expanded over 3 time steps. i ascribes the two level 0 models, mentioned previously, to j (see Fig. 9). If i believes with a probability 1 that j is altruistic, i chooses to contribute for each of the three steps. This behavior persists when i is unaware of whether j is altruistic (Fig. 10(a)), and when i assigns a high probability to j being the non-altruistic type. However, when i believes with a probability 1 that j is non-altruistic and will thus surely defect, i chooses to defect to avoid being punished and because its marginal private return is less than 1. These results demonstrate that the behavior of our altruistic type resembles that found experimentally. The non-altruistic level 1 agent chooses to defect regardless of how likely it believes the other agent to be altruistic. We analyzed the behavior of a reciprocal agent type that matches expected cooperation or defection. The reciprocal type"s marginal private return is similar to that of the non-altruistic type, however, it obtains a greater payoff when its action is similar to that of the other. We consider the case when the reciprocal agent i is unsure of whether j is altruistic and believes that the public pot is likely to be half full. For this prior belief, i chooses to defect. On receiving an observation of plenty, i decides to contribute, while an observation of meager makes it defect (Fig. 10(b)). This is because an observation of plenty signals that the pot is likely to be greater than half full, which results from j"s action to contribute. Thus, among the two models ascribed to j, its type is likely to be altruistic making it likely that j will contribute again in the next time step. Agent i therefore chooses to contribute to reciprocate j"s action. An analogous reasoning leads i to defect when it observes a meager pot. With one action to go, i believing that j contributes, will choose to contribute too to avoid punishment regardless of its observations. Figure 10: (a) An altruistic level 1 agent always contributes. (b) A reciprocal agent i starts off by defecting followed by choosing to contribute or defect based on its observation of plenty (indicating that j is likely altruistic) or meager (j is non-altruistic). 5.3 Strategies in Two-Player Poker Poker is a popular zero sum card game that has received much attention among the AI research community as a testbed [2]. Poker is played among M ≥ 2 players in which each player receives a hand of cards from a deck. While several flavors of Poker with varying complexity exist, we consider a simple version in which each player has three plys during which the player may either exchange a card (E), keep the existing hand (K), fold (F) and withdraw from the game, or call (C), requiring all players to show their hands. To keep matters simple, let M = 2, and each player receive a hand consisting of a single card drawn from the same suit. Thus, during a showdown, the player who has the numerically larger card (2 is the lowest, ace is the highest) wins the pot. During an exchange of cards, the discarded card is placed either in the L pile, indicating to the other agent that it was a low numbered card less than 8, or in the 820 The Sixth Intl. Joint Conf. on Autonomous Agents and Multi-Agent Systems (AAMAS 07) H pile, indicating that the card had a rank greater than or equal to 8. Notice that, for example, if a lower numbered card is discarded, the probability of receiving a low card in exchange is now reduced. We show the level 1 I-ID for the simplified two-player Poker in Fig. 11. We considered two models (personality types) of agent j. The conservative type believes that it is likely that its opponent has a high numbered card in its hand. On the other hand, the aggressive agent j believes with a high probability that its opponent has a lower numbered card. Thus, the two types differ in their beliefs over their opponent"s hand. In both these level 0 models, the opponent is assumed to perform its actions following a fixed, uniform distribution. With three actions to go, regardless of its hand (unless it is an ace), the aggressive agent chooses to exchange its card, with the intent of improving on its current hand. This is because it believes the other to have a low card, which improves its chances of getting a high card during the exchange. The conservative agent chooses to keep its card, no matter its hand because its chances of getting a high card are slim as it believes that its opponent has one. Figure 11: (a) Level 1 I-ID of agent i. The observation reveals information about j"s hand of the previous time step, (b) level 0 IDs of agent j whose decision nodes are mapped to the chance nodes, A1 j , A2 j , in (a). The policy of a level 1 agent i who believes that each card except its own has an equal likelihood of being in j"s hand (neutral personality type) and j could be either an aggressive or conservative type, is shown in Fig. 12. i"s own hand contains the card numbered 8. The agent starts by keeping its card. On seeing that j did not exchange a card (N), i believes with probability 1 that j is conservative and hence will keep its cards. i responds by either keeping its card or exchanging it because j is equally likely to have a lower or higher card. If i observes that j discarded its card into the L or H pile, i believes that j is aggressive. On observing L, i realizes that j had a low card, and is likely to have a high card after its exchange. Because the probability of receiving a low card is high now, i chooses to keep its card. On observing H, believing that the probability of receiving a high numbered card is high, i chooses to exchange its card. In the final step, i chooses to call regardless of its observation history because its belief that j has a higher card is not sufficiently high to conclude that its better to fold and relinquish the payoff. This is partly due to the fact that an observation of, say, L resets the agent i"s previous time step beliefs over j"s hand to the low numbered cards only. 6. DISCUSSION We showed how DIDs may be extended to I-DIDs that enable online sequential decision-making in uncertain multiagent settings. Our graphical representation of I-DIDs improves on the previous Figure 12: A level 1 agent i"s three step policy in the Poker problem. i starts by believing that j is equally likely to be aggressive or conservative and could have any card in its hand with equal probability. work significantly by being more transparent, semantically clear, and capable of being solved using standard algorithms that target DIDs. I-DIDs extend NIDs to allow sequential decision-making over multiple time steps in the presence of other interacting agents. I-DIDs may be seen as concise graphical representations for IPOMDPs providing a way to exploit problem structure and carry out online decision-making as the agent acts and observes given its prior beliefs. We are currently investigating ways to solve I-DIDs approximately with provable bounds on the solution quality. Acknowledgment: We thank Piotr Gmytrasiewicz for some useful discussions related to this work. The first author would like to acknowledge the support of a UGARF grant. 7. REFERENCES [1] R. J. Aumann. Interactive epistemology i: Knowledge. International Journal of Game Theory, 28:263-300, 1999. [2] D. Billings, A. Davidson, J. Schaeffer, and D. Szafron. The challenge of poker. AIJ, 2001. [3] A. Brandenburger and E. Dekel. Hierarchies of beliefs and common knowledge. Journal of Economic Theory, 59:189-198, 1993. [4] C. Camerer. Behavioral Game Theory: Experiments in Strategic Interaction. Princeton University Press, 2003. [5] E. Fehr and S. Gachter. Cooperation and punishment in public goods experiments. American Economic Review, 90(4):980-994, 2000. [6] D. Fudenberg and D. K. Levine. The Theory of Learning in Games. MIT Press, 1998. [7] D. Fudenberg and J. Tirole. Game Theory. MIT Press, 1991. [8] Y. Gal and A. Pfeffer. A language for modeling agent"s decision-making processes in games. In AAMAS, 2003. [9] P. Gmytrasiewicz and P. Doshi. A framework for sequential planning in multiagent settings. JAIR, 24:49-79, 2005. [10] P. Gmytrasiewicz and E. Durfee. Rational coordination in multi-agent environments. JAAMAS, 3(4):319-350, 2000. [11] J. C. Harsanyi. Games with incomplete information played by bayesian players. Management Science, 14(3):159-182, 1967. [12] R. A. Howard and J. E. Matheson. Influence diagrams. In R. A. Howard and J. E. Matheson, editors, The Principles and Applications of Decision Analysis. Strategic Decisions Group, Menlo Park, CA 94025, 1984. [13] L. Kaelbling, M. Littman, and A. Cassandra. Planning and acting in partially observable stochastic domains. Artificial Intelligence Journal, 2, 1998. [14] D. Koller and B. Milch. Multi-agent influence diagrams for representing and solving games. In IJCAI, pages 1027-1034, 2001. [15] K. Polich and P. Gmytrasiewicz. Interactive dynamic influence diagrams. In GTDT Workshop, AAMAS, 2006. [16] B. Rathnas., P. Doshi, and P. J. Gmytrasiewicz. Exact solutions to interactive pomdps using behavioral equivalence. In Autonomous Agents and Multi-Agent Systems Conference (AAMAS), 2006. [17] S. Russell and P. Norvig. Artificial Intelligence: A Modern Approach (Second Edition). Prentice Hall, 2003. [18] R. D. Shachter. Evaluating influence diagrams. Operations Research, 34(6):871-882, 1986. [19] D. Suryadi and P. Gmytrasiewicz. Learning models of other agents using influence diagrams. In UM, 1999. The Sixth Intl. Joint Conf. on Autonomous Agents and Multi-Agent Systems (AAMAS 07) 821

influence diagram;multiagent environment;influence diagram network;dynamic influence diagram;multiplexer;independence structure;dependency link;decision-make;online sequential decision-making;interactive dynamic influence diagram;nash equilibrium profile;agent online;network of influence diagram;policy link;multi-agent influence diagram;partially observable multiagent environment;interactive influence diagram;sequential decision-making;agent model;interactive partially observable markov decision process;graphical model

train_I-66

Letting loose a SPIDER on a network of POMDPs: Generating quality guaranteed policies

Distributed Partially Observable Markov Decision Problems (Distributed POMDPs) are a popular approach for modeling multi-agent systems acting in uncertain domains. Given the significant complexity of solving distributed POMDPs, particularly as we scale up the numbers of agents, one popular approach has focused on approximate solutions. Though this approach is efficient, the algorithms within this approach do not provide any guarantees on solution quality. A second less popular approach focuses on global optimality, but typical results are available only for two agents, and also at considerable computational cost. This paper overcomes the limitations of both these approaches by providing SPIDER, a novel combination of three key features for policy generation in distributed POMDPs: (i) it exploits agent interaction structure given a network of agents (i.e. allowing easier scale-up to larger number of agents); (ii) it uses a combination of heuristics to speedup policy search; and (iii) it allows quality guaranteed approximations, allowing a systematic tradeoff of solution quality for time. Experimental results show orders of magnitude improvement in performance when compared with previous global optimal algorithms.

1. INTRODUCTION Distributed Partially Observable Markov Decision Problems (Distributed POMDPs) are emerging as a popular approach for modeling sequential decision making in teams operating under uncertainty [9, 4, 1, 2, 13]. The uncertainty arises on account of nondeterminism in the outcomes of actions and because the world state may only be partially (or incorrectly) observable. Unfortunately, as shown by Bernstein et al. [3], the problem of finding the optimal joint policy for general distributed POMDPs is NEXP-Complete. Researchers have attempted two different types of approaches towards solving these models. The first category consists of highly efficient approximate techniques, that may not reach globally optimal solutions [2, 9, 11]. The key problem with these techniques has been their inability to provide any guarantees on the quality of the solution. In contrast, the second less popular category of approaches has focused on a global optimal result [13, 5, 10]. Though these approaches obtain optimal solutions, they typically consider only two agents. Furthermore, they fail to exploit structure in the interactions of the agents and hence are severely hampered with respect to scalability when considering more than two agents. To address these problems with the existing approaches, we propose approximate techniques that provide guarantees on the quality of the solution while focussing on a network of more than two agents. We first propose the basic SPIDER (Search for Policies In Distributed EnviRonments) algorithm. There are two key novel features in SPIDER: (i) it is a branch and bound heuristic search technique that uses a MDP-based heuristic function to search for an optimal joint policy; (ii) it exploits network structure of agents by organizing agents into a Depth First Search (DFS) pseudo tree and takes advantage of the independence in the different branches of the DFS tree. We then provide three enhancements to improve the efficiency of the basic SPIDER algorithm while providing guarantees on the quality of the solution. The first enhancement uses abstractions for speedup, but does not sacrifice solution quality. In particular, it initially performs branch and bound search on abstract policies and then extends to complete policies. The second enhancement obtains speedups by sacrificing solution quality, but within an input parameter that provides the tolerable expected value difference from the optimal solution. The third enhancement is again based on bounding the search for efficiency, however with a tolerance parameter that is provided as a percentage of optimal. We experimented with the sensor network domain presented in Nair et al. [10], a domain representative of an important class of problems with networks of agents working in uncertain environments. In our experiments, we illustrate that SPIDER dominates an existing global optimal approach called GOA [10], the only known global optimal algorithm with demonstrated experimental results for more than two agents. Furthermore, we demonstrate that abstraction improves the performance of SPIDER significantly (while providing optimal solutions). We finally demonstrate a key feature of SPIDER: by utilizing the approximation enhancements it enables principled tradeoffs in run-time versus solution quality. 822 978-81-904262-7-5 (RPS) c 2007 IFAAMAS 2. DOMAIN: DISTRIBUTED SENSOR NETS Distributed sensor networks are a large, important class of domains that motivate our work. This paper focuses on a set of target tracking problems that arise in certain types of sensor networks [6] first introduced in [10]. Figure 1 shows a specific problem instance within this type consisting of three sensors. Here, each sensor node can scan in one of four directions: North, South, East or West (see Figure 1). To track a target and obtain associated reward, two sensors with overlapping scanning areas must coordinate by scanning the same area simultaneously. In Figure 1, to track a target in Loc11, sensor1 needs to scan ‘East" and sensor2 needs to scan ‘West" simultaneously. Thus, sensors have to act in a coordinated fashion. We assume that there are two independent targets and that each target"s movement is uncertain and unaffected by the sensor agents. Based on the area it is scanning, each sensor receives observations that can have false positives and false negatives. The sensors" observations and transitions are independent of each other"s actions e.g.the observations that sensor1 receives are independent of sensor2"s actions. Each agent incurs a cost for scanning whether the target is present or not, but no cost if it turns off. Given the sensors" observational uncertainty, the targets" uncertain transitions and the distributed nature of the sensor nodes, these sensor nets provide a useful domains for applying distributed POMDP models. Figure 1: A 3-chain sensor configuration 3. BACKGROUND 3.1 Model: Network Distributed POMDP The ND-POMDP model was introduced in [10], motivated by domains such as the sensor networks introduced in Section 2. It is defined as the tuple S, A, P, Ω, O, R, b , where S = ×1≤i≤nSi × Su is the set of world states. Si refers to the set of local states of agent i and Su is the set of unaffectable states. Unaffectable state refers to that part of the world state that cannot be affected by the agents" actions, e.g. environmental factors like target locations that no agent can control. A = ×1≤i≤nAi is the set of joint actions, where Ai is the set of action for agent i. ND-POMDP assumes transition independence, where the transition function is defined as P(s, a, s ) = Pu(su, su) · 1≤i≤n Pi(si, su, ai, si), where a = a1, . . . , an is the joint action performed in state s = s1, . . . , sn, su and s = s1, . . . , sn, su is the resulting state. Ω = ×1≤i≤nΩi is the set of joint observations where Ωi is the set of observations for agents i. Observational independence is assumed in ND-POMDPs i.e., the joint observation function is defined as O(s, a, ω) = 1≤i≤n Oi(si, su, ai, ωi), where s = s1, . . . , sn, su is the world state that results from the agents performing a = a1, . . . , an in the previous state, and ω = ω1, . . . , ωn ∈ Ω is the observation received in state s. This implies that each agent"s observation depends only on the unaffectable state, its local action and on its resulting local state. The reward function, R, is defined as R(s, a) = l Rl(sl1, . . . , slr, su, al1, . . . , alr ), where each l could refer to any sub-group of agents and r = |l|. Based on the reward function, an interaction hypergraph is constructed. A hyper-link, l, exists between a subset of agents for all Rl that comprise R. The interaction hypergraph is defined as G = (Ag, E), where the agents, Ag, are the vertices and E = {l|l ⊆ Ag ∧ Rl is a component of R} are the edges. The initial belief state (distribution over the initial state), b, is defined as b(s) = bu(su) · 1≤i≤n bi(si), where bu and bi refer to the distribution over initial unaffectable state and agent i"s initial belief state, respectively. The goal in ND-POMDP is to compute the joint policy π = π1, . . . , πn that maximizes team"s expected reward over a finite horizon T starting from the belief state b. An ND-POMDP is similar to an n-ary Distributed Constraint Optimization Problem (DCOP)[8, 12] where the variable at each node represents the policy selected by an individual agent, πi with the domain of the variable being the set of all local policies, Πi. The reward component Rl where |l| = 1 can be thought of as a local constraint while the reward component Rl where l > 1 corresponds to a non-local constraint in the constraint graph. 3.2 Algorithm: Global Optimal Algorithm (GOA) In previous work, GOA has been defined as a global optimal algorithm for ND-POMDPs [10]. We will use GOA in our experimental comparisons, since GOA is a state-of-the-art global optimal algorithm, and in fact the only one with experimental results available for networks of more than two agents. GOA borrows from a global optimal DCOP algorithm called DPOP[12]. GOA"s message passing follows that of DPOP. The first phase is the UTIL propagation, where the utility messages, in this case values of policies, are passed up from the leaves to the root. Value for a policy at an agent is defined as the sum of best response values from its children and the joint policy reward associated with the parent policy. Thus, given a policy for a parent node, GOA requires an agent to iterate through all its policies, finding the best response policy and returning the value to the parent - while at the parent node, to find the best policy, an agent requires its children to return their best responses to each of its policies. This UTIL propagation process is repeated at each level in the tree, until the root exhausts all its policies. In the second phase of VALUE propagation, where the optimal policies are passed down from the root till the leaves. GOA takes advantage of the local interactions in the interaction graph, by pruning out unnecessary joint policy evaluations (associated with nodes not connected directly in the tree). Since the interaction graph captures all the reward interactions among agents and as this algorithm iterates through all the relevant joint policy evaluations, this algorithm yields a globally optimal solution. 4. SPIDER As mentioned in Section 3.1, an ND-POMDP can be treated as a DCOP, where the goal is to compute a joint policy that maximizes the overall joint reward. The brute-force technique for computing an optimal policy would be to examine the expected values for all possible joint policies. The key idea in SPIDER is to avoid computation of expected values for the entire space of joint policies, by utilizing upper bounds on the expected values of policies and the interaction structure of the agents. Akin to some of the algorithms for DCOP [8, 12], SPIDER has a pre-processing step that constructs a DFS tree corresponding to the given interaction structure. Note that these DFS trees are pseudo trees [12] that allow links between ancestors and children. We employ the Maximum Constrained Node (MCN) heuristic used in the DCOP algorithm, ADOPT [8], however other heuristics (such as MLSP heuristic from [7]) can also be employed. MCN heuristic tries to place agents with more number of constraints at the top of the tree. This tree governs how the search for the optimal joint polThe Sixth Intl. Joint Conf. on Autonomous Agents and Multi-Agent Systems (AAMAS 07) 823 icy proceeds in SPIDER. The algorithms presented in this paper are easily extendable to hyper-trees, however for expository purposes, we assume binary trees. SPIDER is an algorithm for centralized planning and distributed execution in distributed POMDPs. In this paper, we employ the following notation to denote policies and expected values: Ancestors(i) ⇒ agents from i to the root (not including i). Tree(i) ⇒ agents in the sub-tree (not including i) for which i is the root. πroot+ ⇒ joint policy of all agents. πi+ ⇒ joint policy of all agents in Tree(i) ∪ i. πi− ⇒ joint policy of agents that are in Ancestors(i). πi ⇒ policy of the ith agent. ˆv[πi, πi− ] ⇒ upper bound on the expected value for πi+ given πi and policies of ancestor agents i.e. πi− . ˆvj[πi, πi− ] ⇒ upper bound on the expected value for πi+ from the jth child. v[πi, πi− ] ⇒ expected value for πi given policies of ancestor agents, πi− . v[πi+ , πi− ] ⇒ expected value for πi+ given policies of ancestor agents, πi− . vj[πi+ , πi− ] ⇒ expected value for πi+ from the jth child. Figure 2: Execution of SPIDER, an example 4.1 Outline of SPIDER SPIDER is based on the idea of branch and bound search, where the nodes in the search tree represent partial/complete joint policies. Figure 2 shows an example search tree for the SPIDER algorithm, using an example of the three agent chain. Before SPIDER begins its search we create a DFS tree (i.e. pseudo tree) from the three agent chain, with the middle agent as the root of this tree. SPIDER exploits the structure of this DFS tree while engaging in its search. Note that in our example figure, each agent is assigned a policy with T=2. Thus, each rounded rectange (search tree node) indicates a partial/complete joint policy, a rectangle indicates an agent and the ovals internal to an agent show its policy. Heuristic or actual expected value for a joint policy is indicated in the top right corner of the rounded rectangle. If the number is italicized and underlined, it implies that the actual expected value of the joint policy is provided. SPIDER begins with no policy assigned to any of the agents (shown in the level 1 of the search tree). Level 2 of the search tree indicates that the joint policies are sorted based on upper bounds computed for root agent"s policies. Level 3 shows one SPIDER search node with a complete joint policy (a policy assigned to each of the agents). The expected value for this joint policy is used to prune out the nodes in level 2 (the ones with upper bounds < 234) When creating policies for each non-leaf agent i, SPIDER potentially performs two steps: 1. Obtaining upper bounds and sorting: In this step, agent i computes upper bounds on the expected values, ˆv[πi, πi− ] of the joint policies πi+ corresponding to each of its policy πi and fixed ancestor policies. An MDP based heuristic is used to compute these upper bounds on the expected values. Detailed description about this MDP heuristic is provided in Section 4.2. All policies of agent i, Πi are then sorted based on these upper bounds (also referred to as heuristic values henceforth) in descending order. Exploration of these policies (in step 2 below) are performed in this descending order. As indicated in the level 2 of the search tree (of Figure 2), all the joint policies are sorted based on the heuristic values, indicated in the top right corner of each joint policy. The intuition behind sorting and then exploring policies in descending order of upper bounds, is that the policies with higher upper bounds could yield joint policies with higher expected values. 2. Exploration and Pruning: Exploration implies computing the best response joint policy πi+,∗ corresponding to fixed ancestor policies of agent i, πi− . This is performed by iterating through all policies of agent i i.e. Πi and summing two quantities for each policy: (i) the best response for all of i"s children (obtained by performing steps 1 and 2 at each of the child nodes); (ii) the expected value obtained by i for fixed policies of ancestors. Thus, exploration of a policy πi yields actual expected value of a joint policy, πi+ represented as v[πi+ , πi− ]. The policy with the highest expected value is the best response policy. Pruning refers to avoiding exploring all policies (or computing expected values) at agent i by using the current best expected value, vmax [πi+ , πi− ]. Henceforth, this vmax [πi+ , πi− ] will be referred to as threshold. A policy, πi need not be explored if the upper bound for that policy, ˆv[πi, πi− ] is less than the threshold. This is because the expected value for the best joint policy attainable for that policy will be less than the threshold. On the other hand, when considering a leaf agent, SPIDER computes the best response policy (and consequently its expected value) corresponding to fixed policies of its ancestors, πi− . This is accomplished by computing expected values for each of the policies (corresponding to fixed policies of ancestors) and selecting the highest expected value policy. In Figure 2, SPIDER assigns best response policies to leaf agents at level 3. The policy for the left leaf agent is to perform action East at each time step in the policy, while the policy for the right leaf agent is to perform Off at each time step. These best response policies from the leaf agents yield an actual expected value of 234 for the complete joint policy. Algorithm 1 provides the pseudo code for SPIDER. This algorithm outputs the best joint policy, πi+,∗ (with an expected value greater than threshold) for the agents in Tree(i). Lines 3-8 compute the best response policy of a leaf agent i, while lines 9-23 computes the best response joint policy for agents in Tree(i). This best response computation for a non-leaf agent i includes: (a) Sorting of policies (in descending order) based on heuristic values on line 11; (b) Computing best response policies at each of the children for fixed policies of agent i in lines 16-20; and (c) Maintaining 824 The Sixth Intl. Joint Conf. on Autonomous Agents and Multi-Agent Systems (AAMAS 07) Algorithm 1 SPIDER(i, πi− , threshold) 1: πi+,∗ ← null 2: Πi ← GET-ALL-POLICIES (horizon, Ai, Ωi) 3: if IS-LEAF(i) then 4: for all πi ∈ Πi do 5: v[πi, πi−] ← JOINT-REWARD (πi, πi−) 6: if v[πi, πi−] > threshold then 7: πi+,∗ ← πi 8: threshold ← v[πi, πi−] 9: else 10: children ← CHILDREN (i) 11: ˆΠi ← UPPER-BOUND-SORT(i, Πi, πi−) 12: for all πi ∈ ˆΠi do 13: ˜πi+ ← πi 14: if ˆv[πi, πi−] < threshold then 15: Go to line 12 16: for all j ∈ children do 17: jThres ← threshold − v[πi, πi−]− Σk∈children,k=j ˆvk[πi, πi−] 18: πj+,∗ ← SPIDER(j, πi πi−, jThres) 19: ˜πi+ ← ˜πi+ πj+,∗ 20: ˆvj[πi, πi−] ← v[πj+,∗, πi πi−] 21: if v[˜πi+, πi−] > threshold then 22: threshold ← v[˜πi+, πi−] 23: πi+,∗ ← ˜πi+ 24: return πi+,∗ Algorithm 2 UPPER-BOUND-SORT(i, Πi, πi− ) 1: children ← CHILDREN (i) 2: ˆΠi ← null /* Stores the sorted list */ 3: for all πi ∈ Πi do 4: ˆv[πi, πi−] ← JOINT-REWARD (πi, πi−) 5: for all j ∈ children do 6: ˆvj[πi, πi−] ← UPPER-BOUND(i, j, πi πi−) 7: ˆv[πi, πi−] + ← ˆvj[πi, πi−] 8: ˆΠi ← INSERT-INTO-SORTED (πi, ˆΠi) 9: return ˆΠi best expected value, joint policy in lines 21-23. Algorithm 2 provides the pseudo code for sorting policies based on the upper bounds on the expected values of joint policies. Expected value for an agent i consists of two parts: value obtained from ancestors and value obtained from its children. Line 4 computes the expected value obtained from ancestors of the agent (using JOINT-REWARD function), while lines 5-7 compute the heuristic value from the children. The sum of these two parts yields an upper bound on the expected value for agent i, and line 8 of the algorithm sorts the policies based on these upper bounds. 4.2 MDP based heuristic function The heuristic function quickly provides an upper bound on the expected value obtainable from the agents in Tree(i). The subtree of agents is a distributed POMDP in itself and the idea here is to construct a centralized MDP corresponding to the (sub-tree) distributed POMDP and obtain the expected value of the optimal policy for this centralized MDP. To reiterate this in terms of the agents in DFS tree interaction structure, we assume full observability for the agents in Tree(i) and for fixed policies of the agents in {Ancestors(i) ∪ i}, we compute the joint value ˆv[πi+ , πi− ] . We use the following notation for presenting the equations for computing upper bounds/heuristic values (for agents i and k): Let Ei− denote the set of links between agents in {Ancestors(i)∪ i} and Tree(i), Ei+ denote the set of links between agents in Tree(i). Also, if l ∈ Ei− , then l1 is the agent in {Ancestors(i)∪ i} and l2 is the agent in Tree(i), that l connects together. We first compact the standard notation: ot k =Ok(st+1 k , st+1 u , πk(ωt k), ωt+1 k ) (1) pt k =Pk(st k, st u, πk(ωt k), st+1 k ) · ot k pt u =P(st u, st+1 u ) st l = st l1 , st l2 , st u ; ωt l = ωt l1 , ωt l2 rt l =Rl(st l , πl1 (ωt l1 ), πl2 (ωt l2 )) vt l =V t πl (st l , st u, ωt l1 , ωt l2 ) Depending on the location of agent k in the agent tree we have the following cases: IF k ∈ {Ancestors(i) ∪ i}, ˆpt k = pt k, (2) IF k ∈ Tree(i), ˆpt k = Pk(st k, st u, πk(ωt k), st+1 k ) IF l ∈ Ei− , ˆrt l = max {al2 } Rl(st l , πl1 (ωt l1 ), al2 ) IF l ∈ Ei+ , ˆrt l = max {al1 ,al2 } Rl(st l , al1 , al2 ) The value function for an agent i executing the joint policy πi+ at time η − 1 is provided by the equation: V η−1 πi+ (sη−1 , ωη−1 ) = l∈Ei− vη−1 l + l∈Ei+ vη−1 l (3) where vη−1 l = rη−1 l + ω η l ,sη pη−1 l1 pη−1 l2 pη−1 u vη l Algorithm 3 UPPER-BOUND (i, j, πj− ) 1: val ← 0 2: for all l ∈ Ej− ∪ Ej+ do 3: if l ∈ Ej− then πl1 ← φ 4: for all s0 l do 5: val + ← startBel[s0 l ]· UPPER-BOUND-TIME (i, s0 l , j, πl1 , ) 6: return val Algorithm 4 UPPER-BOUND-TIME (i, st l , j, πl1 , ωt l1 ) 1: maxV al ← −∞ 2: for all al1 , al2 do 3: if l ∈ Ei− and l ∈ Ej− then al1 ← πl1 (ωt l1 ) 4: val ← GET-REWARD(st l , al1 , al2 ) 5: if t < πi.horizon − 1 then 6: for all st+1 l , ωt+1 l1 do 7: futV al←pt u ˆpt l1 ˆpt l2 8: futV al ∗ ← UPPER-BOUND-TIME(st+1 l , j, πl1 , ωt l1 ωt+1 l1 ) 9: val + ← futV al 10: if val > maxV al then maxV al ← val 11: return maxV al Upper bound on the expected value for a link is computed by modifying the equation 3 to reflect the full observability assumption. This involves removing the observational probability term for agents in Tree(i) and maximizing the future value ˆvη l over the actions of those agents (in Tree(i)). Thus, the equation for the The Sixth Intl. Joint Conf. on Autonomous Agents and Multi-Agent Systems (AAMAS 07) 825 computation of the upper bound on a link l, is as follows: IF l ∈ Ei− , ˆvη−1 l =ˆrη−1 l + max al2 ω η l1 ,s η l ˆpη−1 l1 ˆpη−1 l2 pη−1 u ˆvη l IF l ∈ Ei+ , ˆvη−1 l =ˆrη−1 l + max al1 ,al2 s η l ˆpη−1 l1 ˆpη−1 l2 pη−1 u ˆvη l Algorithm 3 and Algorithm 4 provide the algorithm for computing upper bound for child j of agent i, using the equations descirbed above. While Algorithm 4 computes the upper bound on a link given the starting state, Algorithm 3 sums the upper bound values computed over each of the links in Ei− ∪ Ei+ . 4.3 Abstraction Algorithm 5 SPIDER-ABS(i, πi− , threshold) 1: πi+,∗ ← null 2: Πi ← GET-POLICIES (<>, 1) 3: if IS-LEAF(i) then 4: for all πi ∈ Πi do 5: absHeuristic ← GET-ABS-HEURISTIC (πi, πi−) 6: absHeuristic ∗ ← (timeHorizon − πi.horizon) 7: if πi.horizon = timeHorizon and πi.absNodes = 0 then 8: v[πi, πi−] ← JOINT-REWARD (πi, πi−) 9: if v[πi, πi−] > threshold then 10: πi+,∗ ← πi; threshold ← v[πi, πi−] 11: else if v[πi, πi−] + absHeuristic > threshold then 12: ˆΠi ← EXTEND-POLICY (πi, πi.absNodes + 1) 13: Πi + ← INSERT-SORTED-POLICIES ( ˆΠi) 14: REMOVE(πi) 15: else 16: children ← CHILDREN (i) 17: Πi ← UPPER-BOUND-SORT(i, Πi, πi−) 18: for all πi ∈ Πi do 19: ˜πi+ ← πi 20: absHeuristic ← GET-ABS-HEURISTIC (πi, πi−) 21: absHeuristic ∗ ← (timeHorizon − πi.horizon) 22: if πi.horizon = timeHorizon and πi.absNodes = 0 then 23: if ˆv[πi, πi−] < threshold and πi.absNodes = 0 then 24: Go to line 19 25: for all j ∈ children do 26: jThres ← threshold − v[πi, πi−]− Σk∈children,k=j ˆvk[πi, πi−] 27: πj+,∗ ← SPIDER(j, πi πi−, jThres) 28: ˜πi+ ← ˜πi+ πj+,∗; ˆvj[πi, πi−] ← v[πj+,∗, πi πi−] 29: if v[˜πi+, πi−] > threshold then 30: threshold ← v[˜πi+, πi−]; πi+,∗ ← ˜πi+ 31: else if ˆv[πi+, πi−] + absHeuristic > threshold then 32: ˆΠi ← EXTEND-POLICY (πi, πi.absNodes + 1) 33: Πi + ← INSERT-SORTED-POLICIES (ˆΠi) 34: REMOVE(πi) 35: return πi+,∗ In SPIDER, the exploration/pruning phase can only begin after the heuristic (or upper bound) computation and sorting for the policies has ended. We provide an approach to possibly circumvent the exploration of a group of policies based on heuristic computation for one abstract policy, thus leading to an improvement in runtime performance (without loss in solution quality). The important steps in this technique are defining the abstract policy and how heuristic values are computated for the abstract policies. In this paper, we propose two types of abstraction: 1. Horizon Based Abstraction (HBA): Here, the abstract policy is defined as a shorter horizon policy. It represents a group of longer horizon policies that have the same actions as the abstract policy for times less than or equal to the horizon of the abstract policy. In Figure 3(a), a T=1 abstract policy that performs East action, represents a group of T=2 policies, that perform East in the first time step. For HBA, there are two parts to heuristic computation: (a) Computing the upper bound for the horizon of the abstract policy. This is same as the heuristic computation defined by the GETHEURISTIC() algorithm for SPIDER, however with a shorter time horizon (horizon of the abstract policy). (b) Computing the maximum possible reward that can be accumulated in one time step (using GET-ABS-HEURISTIC()) and multiplying it by the number of time steps to time horizon. This maximum possible reward (for one time step) is obtained by iterating through all the actions of all the agents in Tree(i) and computing the maximum joint reward for any joint action. Sum of (a) and (b) is the heuristic value for a HBA abstract policy. 2. Node Based Abstraction (NBA): Here an abstract policy is obtained by not associating actions to certain nodes of the policy tree. Unlike in HBA, this implies multiple levels of abstraction. This is illustrated in Figure 3(b), where there are T=2 policies that do not have an action for observation ‘TP". These incomplete T=2 policies are abstractions for T=2 complete policies. Increased levels of abstraction leads to faster computation of a complete joint policy, πroot+ and also to shorter heuristic computation and exploration, pruning phases. For NBA, the heuristic computation is similar to that of a normal policy, except in cases where there is no action associated with policy nodes. In such cases, the immediate reward is taken as Rmax (maximum reward for any action). We combine both the abstraction techniques mentioned above into one technique, SPIDER-ABS. Algorithm 5 provides the algorithm for this abstraction technique. For computing optimal joint policy with SPIDER-ABS, a non-leaf agent i initially examines all abstract T=1 policies (line 2) and sorts them based on abstract policy heuristic computations (line 17). The abstraction horizon is gradually increased and these abstract policies are then explored in descending order of heuristic values and ones that have heuristic values less than the threshold are pruned (lines 23-24). Exploration in SPIDER-ABS has the same definition as in SPIDER if the policy being explored has a horizon of policy computation which is equal to the actual time horizon and if all the nodes of the policy have an action associated with them (lines 25-30). However, if those conditions are not met, then it is substituted by a group of policies that it represents (using EXTEND-POLICY () function) (lines 31-32). EXTEND-POLICY() function is also responsible for initializing the horizon and absNodes of a policy. absNodes represents the number of nodes at the last level in the policy tree, that do not have an action assigned to them. If πi.absNodes = |Ωi|πi.horizon−1 (i.e. total number of policy nodes possible at πi.horizon) , then πi.absNodes is set to zero and πi.horizon is increased by 1. Otherwise, πi.absNodes is increased by 1. Thus, this function combines both HBA and NBA by using the policy variables, horizon and absNodes. Before substituting the abstract policy with a group of policies, those policies are sorted based on heuristic values (line 33). Similar type of abstraction based best response computation is adopted at leaf agents (lines 3-14). 4.4 Value ApproXimation (VAX) In this section, we present an approximate enhancement to SPIDER called VAX. The input to this technique is an approximation parameter , which determines the difference from the optimal solution quality. This approximation parameter is used at each agent for pruning out joint policies. The pruning mechanism in SPIDER and SPIDER-Abs dictates that a joint policy be pruned only if the threshold is exactly greater than the heuristic value. However, the 826 The Sixth Intl. Joint Conf. on Autonomous Agents and Multi-Agent Systems (AAMAS 07) Figure 3: Example of abstraction for (a) HBA (Horizon Based Abstraction) and (b) NBA (Node Based Abstraction) idea in this technique is to prune out joint a policy if the following condition is satisfied: threshold + > ˆv[πi , πi− ]. Apart from the pruning condition, VAX is the same as SPIDER/SPIDER-ABS. In the example of Figure 2, if the heuristic value for the second joint policy (or second search tree node) in level 2 were 238 instead of 232, then that policy could not be be pruned using SPIDER or SPIDER-Abs. However, in VAX with an approximation parameter of 5, the joint policy in consideration would also be pruned. This is because the threshold (234) at that juncture plus the approximation parameter (5), i.e. 239 would have been greater than the heuristic value for that joint policy (238). It can be noted from the example (just discussed) that this kind of pruning can lead to fewer explorations and hence lead to an improvement in the overall run-time performance. However, this can entail a sacrifice in the quality of the solution because this technique can prune out a candidate optimal solution. A bound on the error introduced by this approximate algorithm as a function of , is provided by Proposition 3. 4.5 Percentage ApproXimation (PAX) In this section, we present the second approximation enhancement over SPIDER called PAX. Input to this technique is a parameter, δ that represents the minimum percentage of the optimal solution quality that is desired. Output of this technique is a policy with an expected value that is at least δ% of the optimal solution quality. A policy is pruned if the following condition is satisfied: threshold > δ 100 ˆv[πi , πi− ]. Like in VAX, the only difference between PAX and SPIDER/SPIDER-ABS is this pruning condition. Again in Figure 2, if the heuristic value for the second search tree node in level 2 were 238 instead of 232, then PAX with an input parameter of 98% would be able to prune that search tree node (since 98 100 ∗238 < 234). This type of pruning leads to fewer explorations and hence an improvement in run-time performance, while potentially leading to a loss in quality of the solution. Proposition 4 provides the bound on quality loss. 4.6 Theoretical Results PROPOSITION 1. Heuristic provided using the centralized MDP heuristic is admissible. Proof. For the value provided by the heuristic to be admissible, it should be an over estimate of the expected value for a joint policy. Thus, we need to show that: For l ∈ Ei+ ∪ Ei− : ˆvt l ≥ vt l (refer to notation in Section 4.2) We use mathematical induction on t to prove this. Base case: t = T − 1. Irrespective of whether l ∈ Ei− or l ∈ Ei+ , ˆrt l is computed by maximizing over all actions of the agents in Tree(i), while rt l is computed for fixed policies of the same agents. Hence, ˆrt l ≥ rt l and also ˆvt l ≥ vt l . Assumption: Proposition holds for t = η, where 1 ≤ η < T − 1. We now have to prove that the proposition holds for t = η − 1. We show the proof for l ∈ Ei− and similar reasoning can be adopted to prove for l ∈ Ei+ . The heuristic value function for l ∈ Ei− is provided by the following equation: ˆvη−1 l =ˆrη−1 l + max al2 ω η l1 ,s η l ˆpη−1 l1 ˆpη−1 l2 pη−1 u ˆvη l Rewriting the RHS and using Eqn 2 (in Section 4.2) =ˆrη−1 l + max al2 ω η l1 ,s η l pη−1 u pη−1 l1 ˆpη−1 l2 ˆvη l =ˆrη−1 l + ω η l1 ,s η l pη−1 u pη−1 l1 max al2 ˆpη−1 l2 ˆvη l Since maxal2 ˆpη−1 l2 ˆvη l ≥ ωl2 oη−1 l2 ˆpη−1 l2 ˆvη l and pη−1 l2 = oη−1 l2 ˆpη−1 l2 ≥ˆrη−1 l + ω η l1 ,s η l pη−1 u pη−1 l1 ωl2 pη−1 l2 ˆvη l Since ˆvη l ≥ vη l (from the assumption) ≥ˆrη−1 l + ω η l1 ,s η l pη−1 u pη−1 l1 ωl2 pη−1 l2 vη l Since ˆrη−1 l ≥ rη−1 l (by definition) ≥rη−1 l + ω η l1 ,s η l pη−1 u pη−1 l1 ωl2 pη−1 l2 vη l =rη−1 l + (ω η l ,s η l ) pη−1 u pη−1 l1 pη−1 l2 vη l = vη−1 l Thus proved. PROPOSITION 2. SPIDER provides an optimal solution. Proof. SPIDER examines all possible joint policies given the interaction structure of the agents. The only exception being when a joint policy is pruned based on the heuristic value. Thus, as long as a candidate optimal policy is not pruned, SPIDER will return an optimal policy. As proved in Proposition 1, the expected value for a joint policy is always an upper bound. Hence when a joint policy is pruned, it cannot be an optimal solution. PROPOSITION 3. Error bound on the solution quality for VAX (implemented over SPIDER-ABS) with an approximation parameter of is ρ , where ρ is the number of leaf nodes in the DFS tree. The Sixth Intl. Joint Conf. on Autonomous Agents and Multi-Agent Systems (AAMAS 07) 827 Proof. We prove this proposition using mathematical induction on the depth of the DFS tree. Base case: depth = 1 (i.e. one node). Best response is computed by iterating through all policies, Πk. A policy,πk is pruned if ˆv[πk, πk− ] < threshold + . Thus the best response policy computed by VAX would be at most away from the optimal best response. Hence the proposition holds for the base case. Assumption: Proposition holds for d, where 1 ≤ depth ≤ d. We now have to prove that the proposition holds for d + 1. Without loss of generality, lets assume that the root node of this tree has k children. Each of this children is of depth ≤ d, and hence from the assumption, the error introduced in kth child is ρk , where ρk is the number of leaf nodes in kth child of the root. Therefore, ρ = k ρk, where ρ is the number of leaf nodes in the tree. In SPIDER-ABS, threshold at the root agent, thresspider = k v[πk+ , πk− ]. However, with VAX the threshold at the root agent will be (in the worst case), threshvax = k v[πk+ , πk− ]− k ρk . Hence, with VAX a joint policy is pruned at the root agent if ˆv[πroot, πroot− ] < threshvax + ⇒ ˆv[πroot, πroot− ] < threshspider − (( k ρk) − 1) ≤ threshspider − ( k ρk) ≤ threshspider − ρ . Hence proved. PROPOSITION 4. For PAX (implemented over SPIDER-ABS) with an input parameter of δ, the solution quality is at least δ 100 v[πroot+,∗ ], where v[πroot+,∗ ] denotes the optimal solution quality. Proof. We prove this proposition using mathematical induction on the depth of the DFS tree. Base case: depth = 1 (i.e. one node). Best response is computed by iterating through all policies, Πk. A policy,πk is pruned if δ 100 ˆv[πk, πk− ] < threshold. Thus the best response policy computed by PAX would be at least δ 100 times the optimal best response. Hence the proposition holds for the base case. Assumption: Proposition holds for d, where 1 ≤ depth ≤ d. We now have to prove that the proposition holds for d + 1. Without loss of generality, lets assume that the root node of this tree has k children. Each of this children is of depth ≤ d, and hence from the assumption, the solution quality in the kth child is at least δ 100 v[πk+,∗ , πk− ] for PAX. With SPIDER-ABS, a joint policy is pruned at the root agent if ˆv[πroot, πroot− ] < k v[πk+,∗ , πk− ]. However with PAX, a joint policy is pruned if δ 100 ˆv[πroot, πroot− ] < k δ 100 v[πk+,∗ , πk− ] ⇒ ˆv[πroot, πroot− ] < k v[πk+,∗ , πk− ]. Since the pruning condition at the root agent in PAX is the same as the one in SPIDER-ABS, there is no error introduced at the root agent and all the error is introduced in the children. Thus, overall solution quality is at least δ 100 of the optimal solution. Hence proved. 5. EXPERIMENTAL RESULTS All our experiments were conducted on the sensor network domain from Section 2. The five network configurations employed are shown in Figure 4. Algorithms that we experimented with are GOA, SPIDER, SPIDER-ABS, PAX and VAX. We compare against GOA because it is the only global optimal algorithm that considers more than two agents. We performed two sets of experiments: (i) firstly, we compared the run-time performance of the above algorithms and (ii) secondly, we experimented with PAX and VAX to study the tradeoff between run-time and solution quality. Experiments were terminated after 10000 seconds1 . Figure 5(a) provides run-time comparisons between the optimal algorithms GOA, SPIDER, SPIDER-Abs and the approximate algorithms, PAX ( of 30) and VAX(δ of 80). X-axis denotes the 1 Machine specs for all experiments: Intel Xeon 3.6 GHZ processor, 2GB RAM sensor network configuration used, while Y-axis indicates the runtime (on a log-scale). The time horizon of policy computation was 3. For each configuration (3-chain, 4-chain, 4-star and 5-star), there are five bars indicating the time taken by GOA, SPIDER, SPIDERAbs, PAX and VAX. GOA did not terminate within the time limit for 4-star and 5-star configurations. SPIDER-Abs dominated the SPIDER and GOA for all the configurations. For instance, in the 3chain configuration, SPIDER-ABS provides 230-fold speedup over GOA and 2-fold speedup over SPIDER and for the 4-chain configuration it provides 58-fold speedup over GOA and 2-fold speedup over SPIDER. The two approximation approaches, VAX and PAX provided further improvement in performance over SPIDER-Abs. For instance, in the 5-star configuration VAX provides a 15-fold speedup and PAX provides a 8-fold speedup over SPIDER-Abs. Figures 5(b) provides a comparison of the solution quality obtained using the different algorithms for the problems tested in Figure 5(a). X-axis denotes the sensor network configuration while Y-axis indicates the solution quality. Since GOA, SPIDER, and SPIDER-Abs are all global optimal algorithms, the solution quality is the same for all those algorithms. For 5-P configuration, the global optimal algorithms did not terminate within the limit of 10000 seconds, so the bar for optimal quality indicates an upper bound on the optimal solution quality. With both the approximations, we obtained a solution quality that was close to the optimal solution quality. In 3-chain and 4-star configurations, it is remarkable that both PAX and VAX obtained almost the same actual quality as the global optimal algorithms, despite the approximation parameter and δ. For other configurations as well, the loss in quality was less than 20% of the optimal solution quality. Figure 5(c) provides the time to solution with PAX (for varying epsilons). X-axis denotes the approximation parameter, δ (percentage to optimal) used, while Y-axis denotes the time taken to compute the solution (on a log-scale). The time horizon for all the configurations was 4. As δ was decreased from 70 to 30, the time to solution decreased drastically. For instance, in the 3-chain case there was a total speedup of 170-fold when the δ was changed from 70 to 30. Interestingly, even with a low δ of 30%, the actual solution quality remained equal to the one obtained at 70%. Figure 5(d) provides the time to solution for all the configurations with VAX (for varying epsilons). X-axis denotes the approximation parameter, used, while Y-axis denotes the time taken to compute the solution (on a log-scale). The time horizon for all the configurations was 4. As was increased, the time to solution decreased drastically. For instance, in the 4-star case there was a total speedup of 73-fold when the was changed from 60 to 140. Again, the actual solution quality did not change with varying epsilon. Figure 4: Sensor network configurations 828 The Sixth Intl. Joint Conf. on Autonomous Agents and Multi-Agent Systems (AAMAS 07) Figure 5: Comparison of GOA, SPIDER, SPIDER-Abs and VAX for T = 3 on (a) Runtime and (b) Solution quality; (c) Time to solution for PAX with varying percentage to optimal for T=4 (d) Time to solution for VAX with varying epsilon for T=4 6. SUMMARY AND RELATED WORK This paper presents four algorithms SPIDER, SPIDER-ABS, PAX and VAX that provide a novel combination of features for policy search in distributed POMDPs: (i) exploiting agent interaction structure given a network of agents (i.e. easier scale-up to larger number of agents); (ii) using branch and bound search with an MDP based heuristic function; (iii) utilizing abstraction to improve runtime performance without sacrificing solution quality; (iv) providing a priori percentage bounds on quality of solutions using PAX; and (v) providing expected value bounds on the quality of solutions using VAX. These features allow for systematic tradeoff of solution quality for run-time in networks of agents operating under uncertainty. Experimental results show orders of magnitude improvement in performance over previous global optimal algorithms. Researchers have typically employed two types of techniques for solving distributed POMDPs. The first set of techniques compute global optimal solutions. Hansen et al. [5] present an algorithm based on dynamic programming and iterated elimination of dominant policies, that provides optimal solutions for distributed POMDPs. Szer et al. [13] provide an optimal heuristic search method for solving Decentralized POMDPs. This algorithm is based on the combination of a classical heuristic search algorithm, A∗ and decentralized control theory. The key differences between SPIDER and MAA* are: (a) Enhancements to SPIDER (VAX and PAX) provide for quality guaranteed approximations, while MAA* is a global optimal algorithm and hence involves significant computational complexity; (b) Due to MAA*"s inability to exploit interaction structure, it was illustrated only with two agents. However, SPIDER has been illustrated for networks of agents; and (c) SPIDER explores the joint policy one agent at a time, while MAA* expands it one time step at a time (simultaneously for all the agents). The second set of techniques seek approximate policies. EmeryMontemerlo et al. [4] approximate POSGs as a series of one-step Bayesian games using heuristics to approximate future value, trading off limited lookahead for computational efficiency, resulting in locally optimal policies (with respect to the selected heuristic). Nair et al. [9]"s JESP algorithm uses dynamic programming to reach a local optimum solution for finite horizon decentralized POMDPs. Peshkin et al. [11] and Bernstein et al. [2] are examples of policy search techniques that search for locally optimal policies. Though all the above techniques improve the efficiency of policy computation considerably, they are unable to provide error bounds on the quality of the solution. This aspect of quality bounds differentiates SPIDER from all the above techniques. Acknowledgements. This material is based upon work supported by the Defense Advanced Research Projects Agency (DARPA), through the Department of the Interior, NBC, Acquisition Services Division under Contract No. NBCHD030010. The views and conclusions contained in this document are those of the authors, and should not be interpreted as representing the official policies, either expressed or implied, of the Defense Advanced Research Projects Agency or the U.S. Government. 7. REFERENCES [1] R. Becker, S. Zilberstein, V. Lesser, and C.V. Goldman. Solving transition independent decentralized Markov decision processes. JAIR, 22:423-455, 2004. [2] D. S. Bernstein, E.A. Hansen, and S. Zilberstein. Bounded policy iteration for decentralized POMDPs. In IJCAI, 2005. [3] D. S. Bernstein, S. Zilberstein, and N. Immerman. The complexity of decentralized control of MDPs. In UAI, 2000. [4] R. Emery-Montemerlo, G. Gordon, J. Schneider, and S. Thrun. Approximate solutions for partially observable stochastic games with common payoffs. In AAMAS, 2004. [5] E. Hansen, D. Bernstein, and S. Zilberstein. Dynamic programming for partially observable stochastic games. In AAAI, 2004. [6] V. Lesser, C. Ortiz, and M. Tambe. Distributed sensor nets: A multiagent perspective. Kluwer, 2003. [7] R. Maheswaran, M. Tambe, E. Bowring, J. Pearce, and P. Varakantham. Taking dcop to the real world : Efficient complete solutions for distributed event scheduling. In AAMAS, 2004. [8] P. J. Modi, W. Shen, M. Tambe, and M. Yokoo. An asynchronous complete method for distributed constraint optimization. In AAMAS, 2003. [9] R. Nair, D. Pynadath, M. Yokoo, M. Tambe, and S. Marsella. Taming decentralized POMDPs: Towards efficient policy computation for multiagent settings. In IJCAI, 2003. [10] R. Nair, P. Varakantham, M. Tambe, and M. Yokoo. Networked distributed POMDPs: A synthesis of distributed constraint optimization and POMDPs. In AAAI, 2005. [11] L. Peshkin, N. Meuleau, K.-E. Kim, and L. Kaelbling. Learning to cooperate via policy search. In UAI, 2000. [12] A. Petcu and B. Faltings. A scalable method for multiagent constraint optimization. In IJCAI, 2005. [13] D. Szer, F. Charpillet, and S. Zilberstein. MAA*: A heuristic search algorithm for solving decentralized POMDPs. In IJCAI, 2005. The Sixth Intl. Joint Conf. on Autonomous Agents and Multi-Agent Systems (AAMAS 07) 829

quality guaranteed approximation;policy search;branch and bound heuristic search technique;globally optimal solution;distributed partially observable markov decision problem;global optimality;network structure;heuristic;partially observable markov decision process;distributed sensor network;pomdp;approximate solution;multi-agent system;network of agent;maximum constrained node;distribute pomdp;agent network;depth first search;overall joint reward;quality guaranteed policy;optimal joint policy;heuristic function;network;uncertain domain;agent interaction structure

train_I-68

On Opportunistic Techniques for Solving Decentralized Markov Decision Processes with Temporal Constraints

Decentralized Markov Decision Processes (DEC-MDPs) are a popular model of agent-coordination problems in domains with uncertainty and time constraints but very difficult to solve. In this paper, we improve a state-of-the-art heuristic solution method for DEC-MDPs, called OC-DEC-MDP, that has recently been shown to scale up to larger DEC-MDPs. Our heuristic solution method, called Value Function Propagation (VFP), combines two orthogonal improvements of OC-DEC-MDP. First, it speeds up OC-DECMDP by an order of magnitude by maintaining and manipulating a value function for each state (as a function of time) rather than a separate value for each pair of sate and time interval. Furthermore, it achieves better solution qualities than OC-DEC-MDP because, as our analytical results show, it does not overestimate the expected total reward like OC-DEC- MDP. We test both improvements independently in a crisis-management domain as well as for other types of domains. Our experimental results demonstrate a significant speedup of VFP over OC-DEC-MDP as well as higher solution qualities in a variety of situations.

1. INTRODUCTION The development of algorithms for effective coordination of multiple agents acting as a team in uncertain and time critical domains has recently become a very active research field with potential applications ranging from coordination of agents during a hostage rescue mission [11] to the coordination of Autonomous Mars Exploration Rovers [2]. Because of the uncertain and dynamic characteristics of such domains, decision-theoretic models have received a lot of attention in recent years, mainly thanks to their expressiveness and the ability to reason about the utility of actions over time. Key decision-theoretic models that have become popular in the literature include Decentralized Markov Decision Processes (DECMDPs) and Decentralized, Partially Observable Markov Decision Processes (DEC-POMDPs). Unfortunately, solving these models optimally has been proven to be NEXP-complete [3], hence more tractable subclasses of these models have been the subject of intensive research. In particular, Network Distributed POMDP [13] which assume that not all the agents interact with each other, Transition Independent DEC-MDP [2] which assume that transition function is decomposable into local transition functions or DEC-MDP with Event Driven Interactions [1] which assume that interactions between agents happen at fixed time points constitute good examples of such subclasses. Although globally optimal algorithms for these subclasses have demonstrated promising results, domains on which these algorithms run are still small and time horizons are limited to only a few time ticks. To remedy that, locally optimal algorithms have been proposed [12] [4] [5]. In particular, Opportunity Cost DEC-MDP [4] [5], referred to as OC-DEC-MDP, is particularly notable, as it has been shown to scale up to domains with hundreds of tasks and double digit time horizons. Additionally, OC-DEC-MDP is unique in its ability to address both temporal constraints and uncertain method execution durations, which is an important factor for real-world domains. OC-DEC-MDP is able to scale up to such domains mainly because instead of searching for the globally optimal solution, it carries out a series of policy iterations; in each iteration it performs a value iteration that reuses the data computed during the previous policy iteration. However, OC-DEC-MDP is still slow, especially as the time horizon and the number of methods approach large values. The reason for high runtimes of OC-DEC-MDP for such domains is a consequence of its huge state space, i.e., OC-DEC-MDP introduces a separate state for each possible pair of method and method execution interval. Furthermore, OC-DEC-MDP overestimates the reward that a method expects to receive for enabling the execution of future methods. This reward, also referred to as the opportunity cost, plays a crucial role in agent decision making, and as we show later, its overestimation leads to highly suboptimal policies. In this context, we present VFP (= Value Function P ropagation), an efficient solution technique for the DEC-MDP model with temporal constraints and uncertain method execution durations, that builds on the success of OC-DEC-MDP. VFP introduces our two orthogonal ideas: First, similarly to [7] [9] and [10], we maintain 830 978-81-904262-7-5 (RPS) c 2007 IFAAMAS and manipulate a value function over time for each method rather than a separate value for each pair of method and time interval. Such representation allows us to group the time points for which the value function changes at the same rate (= its slope is constant), which results in fast, functional propagation of value functions. Second, we prove (both theoretically and empirically) that OC-DEC- MDP overestimates the opportunity cost, and to remedy that, we introduce a set of heuristics, that correct the opportunity cost overestimation problem. This paper is organized as follows: In section 2 we motivate this research by introducing a civilian rescue domain where a team of fire- brigades must coordinate in order to rescue civilians trapped in a burning building. In section 3 we provide a detailed description of our DEC-MDP model with Temporal Constraints and in section 4 we discuss how one could solve the problems encoded in our model using globally optimal and locally optimal solvers. Sections 5 and 6 discuss the two orthogonal improvements to the state-of-the-art OC-DEC-MDP algorithm that our VFP algorithm implements. Finally, in section 7 we demonstrate empirically the impact of our two orthogonal improvements, i.e., we show that: (i) The new heuristics correct the opportunity cost overestimation problem leading to higher quality policies, and (ii) By allowing for a systematic tradeoff of solution quality for time, the VFP algorithm runs much faster than the OC-DEC-MDP algorithm 2. MOTIVATING EXAMPLE We are interested in domains where multiple agents must coordinate their plans over time, despite uncertainty in plan execution duration and outcome. One example domain is large-scale disaster, like a fire in a skyscraper. Because there can be hundreds of civilians scattered across numerous floors, multiple rescue teams have to be dispatched, and radio communication channels can quickly get saturated and useless. In particular, small teams of fire-brigades must be sent on separate missions to rescue the civilians trapped in dozens of different locations. Picture a small mission plan from Figure (1), where three firebrigades have been assigned a task to rescue the civilians trapped at site B, accessed from site A (e.g. an office accessed from the floor)1 . General fire fighting procedures involve both: (i) putting out the flames, and (ii) ventilating the site to let the toxic, high temperature gases escape, with the restriction that ventilation should not be performed too fast in order to prevent the fire from spreading. The team estimates that the civilians have 20 minutes before the fire at site B becomes unbearable, and that the fire at site A has to be put out in order to open the access to site B. As has happened in the past in large scale disasters, communication often breaks down; and hence we assume in this domain that there is no communication between the fire-brigades 1,2 and 3 (denoted as FB1, FB2 and FB3). Consequently, FB2 does not know if it is already safe to ventilate site A, FB1 does not know if it is already safe to enter site A and start fighting fire at site B, etc. We assign the reward 50 for evacuating the civilians from site B, and a smaller reward 20 for the successful ventilation of site A, since the civilians themselves might succeed in breaking out from site B. One can clearly see the dilemma, that FB2 faces: It can only estimate the durations of the Fight fire at site A methods to be executed by FB1 and FB3, and at the same time FB2 knows that time is running out for civilians. If FB2 ventilates site A too early, the fire will spread out of control, whereas if FB2 waits with the ventilation method for too long, fire at site B will become unbearable for the civilians. In general, agents have to perform a sequence of such 1 We explain the EST and LET notation in section 3 Figure 1: Civilian rescue domain and a mission plan. Dotted arrows represent implicit precedence constraints within an agent. difficult decisions; in particular, decision process of FB2 involves first choosing when to start ventilating site A, and then (depending on the time it took to ventilate site A), choosing when to start evacuating the civilians from site B. Such sequence of decisions constitutes the policy of an agent, and it must be found fast because time is running out. 3. MODEL DESCRIPTION We encode our decision problems in a model which we refer to as Decentralized MDP with Temporal Constraints 2 . Each instance of our decision problems can be described as a tuple M, A, C, P, R where M = {mi} |M| i=1 is the set of methods, and A = {Ak} |A| k=1 is the set of agents. Agents cannot communicate during mission execution. Each agent Ak is assigned to a set Mk of methods, such that S|A| k=1 Mk = M and ∀i,j;i=jMi ∩ Mj = ø. Also, each method of agent Ak can be executed only once, and agent Ak can execute only one method at a time. Method execution times are uncertain and P = {pi} |M| i=1 is the set of distributions of method execution durations. In particular, pi(t) is the probability that the execution of method mi consumes time t. C is a set of temporal constraints in the system. Methods are partially ordered and each method has fixed time windows inside which it can be executed, i.e., C = C≺ ∪ C[ ] where C≺ is the set of predecessor constraints and C[ ] is the set of time window constraints. For c ∈ C≺, c = mi, mj means that method mi precedes method mj i.e., execution of mj cannot start before mi terminates. In particular, for an agent Ak, all its methods form a chain linked by predecessor constraints. We assume, that the graph G = M, C≺ is acyclic, does not have disconnected nodes (the problem cannot be decomposed into independent subproblems), and its source and sink vertices identify the source and sink methods of the system. For c ∈ C[ ], c = mi, EST, LET means that execution of mi can only start after the Earliest Starting Time EST and must finish before the Latest End Time LET; we allow methods to have multiple disjoint time window constraints. Although distributions pi can extend to infinite time horizons, given the time window constraints, the planning horizon Δ = max m,τ,τ ∈C[ ] τ is considered as the mission deadline. Finally, R = {ri} |M| i=1 is the set of non-negative rewards, i.e., ri is obtained upon successful execution of mi. Since there is no communication allowed, an agent can only estimate the probabilities that its methods have already been enabled 2 One could also use the OC-DEC-MDP framework, which models both time and resource constraints The Sixth Intl. Joint Conf. on Autonomous Agents and Multi-Agent Systems (AAMAS 07) 831 by other agents. Consequently, if mj ∈ Mk is the next method to be executed by the agent Ak and the current time is t ∈ [0, Δ], the agent has to make a decision whether to Execute the method mj (denoted as E), or to Wait (denoted as W). In case agent Ak decides to wait, it remains idle for an arbitrary small time , and resumes operation at the same place (= about to execute method mj) at time t + . In case agent Ak decides to Execute the next method, two outcomes are possible: Success: The agent Ak receives reward rj and moves on to its next method (if such method exists) so long as the following conditions hold: (i) All the methods {mi| mi, mj ∈ C≺} that directly enable method mj have already been completed, (ii) Execution of method mj started in some time window of method mj, i.e., ∃ mj ,τ,τ ∈C[ ] such that t ∈ [τ, τ ], and (iii) Execution of method mj finished inside the same time window, i.e., agent Ak completed method mj in time less than or equal to τ − t. Failure: If any of the above-mentioned conditions does not hold, agent Ak stops its execution. Other agents may continue their execution, but methods mk ∈ {m| mj, m ∈ C≺} will never become enabled. The policy πk of an agent Ak is a function πk : Mk × [0, Δ] → {W, E}, and πk( m, t ) = a means, that if Ak is at method m at time t, it will choose to perform the action a. A joint policy π = [πk] |A| k=1 is considered to be optimal (denoted as π∗ ), if it maximizes the sum of expected rewards for all the agents. 4. SOLUTION TECHNIQUES 4.1 Optimal Algorithms Optimal joint policy π∗ is usually found by using the Bellman update principle, i.e., in order to determine the optimal policy for method mj, optimal policies for methods mk ∈ {m| mj, m ∈ C≺} are used. Unfortunately, for our model, the optimal policy for method mj also depends on policies for methods mi ∈ {m| m, mj ∈ C≺}. This double dependency results from the fact, that the expected reward for starting the execution of method mj at time t also depends on the probability that method mj will be enabled by time t. Consequently, if time is discretized, one needs to consider Δ|M| candidate policies in order to find π∗ . Thus, globally optimal algorithms used for solving real-world problems are unlikely to terminate in reasonable time [11]. The complexity of our model could be reduced if we considered its more restricted version; in particular, if each method mj was allowed to be enabled at time points t ∈ Tj ⊂ [0, Δ], the Coverage Set Algorithm (CSA) [1] could be used. However, CSA complexity is double exponential in the size of Ti, and for our domains Tj can store all values ranging from 0 to Δ. 4.2 Locally Optimal Algorithms Following the limited applicability of globally optimal algorithms for DEC-MDPs with Temporal Constraints, locally optimal algorithms appear more promising. Specially, the OC-DEC-MDP algorithm [4] is particularly significant, as it has shown to easily scale up to domains with hundreds of methods. The idea of the OC-DECMDP algorithm is to start with the earliest starting time policy π0 (according to which an agent will start executing the method m as soon as m has a non-zero chance of being already enabled), and then improve it iteratively, until no further improvement is possible. At each iteration, the algorithm starts with some policy π, which uniquely determines the probabilities Pi,[τ,τ ] that method mi will be performed in the time interval [τ, τ ]. It then performs two steps: Step 1: It propagates from sink methods to source methods the values Vi,[τ,τ ], that represent the expected utility for executing method mi in the time interval [τ, τ ]. This propagation uses the probabilities Pi,[τ,τ ] from previous algorithm iteration. We call this step a value propagation phase. Step 2: Given the values Vi,[τ,τ ] from Step 1, the algorithm chooses the most profitable method execution intervals which are stored in a new policy π . It then propagates the new probabilities Pi,[τ,τ ] from source methods to sink methods. We call this step a probability propagation phase. If policy π does not improve π, the algorithm terminates. There are two shortcomings of the OC-DEC-MDP algorithm that we address in this paper. First, each of OC-DEC-MDP states is a pair mj, [τ, τ ] , where [τ, τ ] is a time interval in which method mj can be executed. While such state representation is beneficial, in that the problem can be solved with a standard value iteration algorithm, it blurs the intuitive mapping from time t to the expected total reward for starting the execution of mj at time t. Consequently, if some method mi enables method mj, and the values Vj,[τ,τ ]∀τ,τ ∈[0,Δ] are known, the operation that calculates the values Vi,[τ,τ ]∀τ, τ ∈ [0, Δ] (during the value propagation phase), runs in time O(I2 ), where I is the number of time intervals 3 . Since the runtime of the whole algorithm is proportional to the runtime of this operation, especially for big time horizons Δ, the OC- DECMDP algorithm runs slow. Second, while OC-DEC-MDP emphasizes on precise calculation of values Vj,[τ,τ ], it fails to address a critical issue that determines how the values Vj,[τ,τ ] are split given that the method mj has multiple enabling methods. As we show later, OC-DEC-MDP splits Vj,[τ,τ ] into parts that may overestimate Vj,[τ,τ ] when summed up again. As a result, methods that precede the method mj overestimate the value for enabling mj which, as we show later, can have disastrous consequences. In the next two sections, we address both of these shortcomings. 5. VALUE FUNCTION PROPAGATION (VFP) The general scheme of the VFP algorithm is identical to the OCDEC-MDP algorithm, in that it performs a series of policy improvement iterations, each one involving a Value and Probability Propagation Phase. However, instead of propagating separate values, VFP maintains and propagates the whole functions, we therefore refer to these phases as the value function propagation phase and the probability function propagation phase. To this end, for each method mi ∈ M, we define three new functions: Value Function, denoted as vi(t), that maps time t ∈ [0, Δ] to the expected total reward for starting the execution of method mi at time t. Opportunity Cost Function, denoted as Vi(t), that maps time t ∈ [0, Δ] to the expected total reward for starting the execution of method mi at time t assuming that mi is enabled. Probability Function, denoted as Pi(t), that maps time t ∈ [0, Δ] to the probability that method mi will be completed before time t. Such functional representation allows us to easily read the current policy, i.e., if an agent Ak is at method mi at time t, then it will wait as long as value function vi(t) will be greater in the future. Formally: πk( mi, t ) = j W if ∃t >t such that vi(t) < vi(t ) E otherwise. We now develop an analytical technique for performing the value function and probability function propagation phases. 3 Similarly for the probability propagation phase 832 The Sixth Intl. Joint Conf. on Autonomous Agents and Multi-Agent Systems (AAMAS 07) 5.1 Value Function Propagation Phase Suppose, that we are performing a value function propagation phase during which the value functions are propagated from the sink methods to the source methods. At any time during this phase we encounter a situation shown in Figure 2, where opportunity cost functions [Vjn ]N n=0 of methods [mjn ]N n=0 are known, and the opportunity cost Vi0 of method mi0 is to be derived. Let pi0 be the probability distribution function of method mi0 execution duration, and ri0 be the immediate reward for starting and completing the execution of method mi0 inside a time interval [τ, τ ] such that mi0 τ, τ ∈ C[ ]. The function Vi0 is then derived from ri0 and opportunity costs Vjn,i0 (t) n = 1, ..., N from future methods. Formally: Vi0 (t) = 8 >>< >>: R τ −t 0 pi0 (t )(ri0 + PN n=0 Vjn,i0 (t + t ))dt if ∃ mi0 τ,τ ∈C[ ] such that t ∈ [τ, τ ] 0 otherwise (1) Note, that for t ∈ [τ, τ ], if h(t) := ri0 + PN n=0 Vjn,i0 (τ −t) then Vi0 is a convolution of p and h: vi0 (t) = (pi0 ∗h)(τ −t). Assume for now, that Vjn,i0 represents a full opportunity cost, postponing the discussion on different techniques for splitting the opportunity cost Vj0 into [Vj0,ik ]K k=0 until section 6. We now show how to derive Vj0,i0 (derivation of Vjn,i0 for n = 0 follows the same scheme). Figure 2: Fragment of an MDP of agent Ak. Probability functions propagate forward (left to right) whereas value functions propagate backward (right to left). Let V j0,i0 (t) be the opportunity cost of starting the execution of method mj0 at time t given that method mi0 has been completed. It is derived by multiplying Vi0 by the probability functions of all methods other than mi0 that enable mj0 . Formally: V j0,i0 (t) = Vj0 (t) · KY k=1 Pik (t). Where similarly to [4] and [5] we ignored the dependency of [Plk ]K k=1. Observe that V j0,i0 does not have to be monotonically decreasing, i.e., delaying the execution of the method mi0 can sometimes be profitable. Therefore the opportunity cost Vj0,i0 (t) of enabling method mi0 at time t must be greater than or equal to V j0,i0 . Furthermore, Vj0,i0 should be non-increasing. Formally: Vj0,i0 = min f∈F f (2) Where F = {f | f ≥ V j0,i0 and f(t) ≥ f(t ) ∀t<t }. Knowing the opportunity cost Vi0 , we can then easily derive the value function vi0 . Let Ak be an agent assigned to the method mi0 . If Ak is about to start the execution of mi0 it means, that Ak must have completed its part of the mission plan up to the method mi0 . Since Ak does not know if other agents have completed methods [mlk ]k=K k=1 , in order to derive vi0 , it has to multiply Vi0 by the probability functions of all methods of other agents that enable mi0 . Formally: vi0 (t) = Vi0 (t) · KY k=1 Plk (t) Where the dependency of [Plk ]K k=1 is also ignored. We have consequently shown a general scheme how to propagate the value functions: Knowing [vjn ]N n=0 and [Vjn ]N n=0 of methods [mjn ]N n=0 we can derive vi0 and Vi0 of method mi0 . In general, the value function propagation scheme starts with sink nodes. It then visits at each time a method m, such that all the methods that m enables have already been marked as visited. The value function propagation phase terminates when all the source methods have been marked as visited. 5.2 Reading the Policy In order to determine the policy of agent Ak for the method mj0 we must identify the set Zj0 of intervals [z, z ] ⊂ [0, ..., Δ], such that: ∀t∈[z,z ] πk( mj0 , t ) = W. One can easily identify the intervals of Zj0 by looking at the time intervals in which the value function vj0 does not decrease monotonically. 5.3 Probability Function Propagation Phase Assume now, that value functions and opportunity cost values have all been propagated from sink methods to source nodes and the sets Zj for all methods mj ∈ M have been identified. Since value function propagation phase was using probabilities Pi(t) for methods mi ∈ M and times t ∈ [0, Δ] found at previous algorithm iteration, we now have to find new values Pi(t), in order to prepare the algorithm for its next iteration. We now show how in the general case (Figure 2) propagate the probability functions forward through one method, i.e., we assume that the probability functions [Pik ]K k=0 of methods [mik ]K k=0 are known, and the probability function Pj0 of method mj0 must be derived. Let pj0 be the probability distribution function of method mj0 execution duration, and Zj0 be the set of intervals of inactivity for method mj0 , found during the last value function propagation phase. If we ignore the dependency of [Pik ]K k=0 then the probability Pj0 (t) that the execution of method mj0 starts before time t is given by: Pj0 (t) = (QK k=0 Pik (τ) if ∃(τ, τ ) ∈ Zj0 s.t. t ∈ (τ, τ ) QK k=0 Pik (t) otherwise. Given Pj0 (t), the probability Pj0 (t) that method mj0 will be completed by time t is derived by: Pj0 (t) = Z t 0 Z t 0 ( ∂Pj0 ∂t )(t ) · pj0 (t − t )dt dt (3) Which can be written compactly as ∂Pj0 ∂t = pj0 ∗ ∂P j0 ∂t . We have consequently shown how to propagate the probability functions [Pik ]K k=0 of methods [mik ]K k=0 to obtain the probability function Pj0 of method mj0 . The general, the probability function propagation phase starts with source methods msi for which we know that Psi = 1 since they are enabled by default. We then visit at each time a method m such that all the methods that enable The Sixth Intl. Joint Conf. on Autonomous Agents and Multi-Agent Systems (AAMAS 07) 833 m have already been marked as visited. The probability function propagation phase terminates when all the sink methods have been marked as visited. 5.4 The Algorithm Similarly to the OC-DEC-MDP algorithm, VFP starts the policy improvement iterations with the earliest starting time policy π0 . Then at each iteration it: (i) Propagates the value functions [vi] |M| i=1 using the old probability functions [Pi] |M| i=1 from previous algorithm iteration and establishes the new sets [Zi] |M| i=1 of method inactivity intervals, and (ii) propagates the new probability functions [Pi ] |M| i=1 using the newly established sets [Zi] |M| i=1. These new functions [Pi ] |M| i=1 are then used in the next iteration of the algorithm. Similarly to OC-DEC-MDP, VFP terminates if a new policy does not improve the policy from the previous algorithm iteration. 5.5 Implementation of Function Operations So far, we have derived the functional operations for value function and probability function propagation without choosing any function representation. In general, our functional operations can handle continuous time, and one has freedom to choose a desired function approximation technique, such as piecewise linear [7] or piecewise constant [9] approximation. However, since one of our goals is to compare VFP with the existing OC-DEC- MDP algorithm, that works only for discrete time, we also discretize time, and choose to approximate value functions and probability functions with piecewise linear (PWL) functions. When the VFP algorithm propagates the value functions and probability functions, it constantly carries out operations represented by equations (1) and (3) and we have already shown that these operations are convolutions of some functions p(t) and h(t). If time is discretized, functions p(t) and h(t) are discrete; however, h(t) can be nicely approximated with a PWL function bh(t), which is exactly what VFP does. As a result, instead of performing O(Δ2 ) multiplications to compute f(t), VFP only needs to perform O(k · Δ) multiplications to compute f(t), where k is the number of linear segments of bh(t) (note, that since h(t) is monotonic, bh(t) is usually close to h(t) with k Δ). Since Pi values are in range [0, 1] and Vi values are in range [0, P mi∈M ri], we suggest to approximate Vi(t) with bVi(t) within error V , and Pi(t) with bPi(t) within error P . We now prove that the overall approximation error accumulated during the value function propagation phase can be expressed in terms of P and V : THEOREM 1. Let C≺ be a set of precedence constraints of a DEC-MDP with Temporal Constraints, and P and V be the probability function and value function approximation errors respectively. The overall error π = maxV supt∈[0,Δ]|V (t) − bV (t)| of value function propagation phase is then bounded by: |C≺| V + ((1 + P )|C≺| − 1) P mi∈M ri . PROOF. In order to establish the bound for π, we first prove by induction on the size of C≺, that the overall error of probability function propagation phase, π(P ) = maxP supt∈[0,Δ]|P(t) − bP(t)| is bounded by (1 + P )|C≺| − 1. Induction base: If n = 1 only two methods are present, and we will perform the operation identified by Equation (3) only once, introducing the error π(P ) = P = (1 + P )|C≺| − 1. Induction step: Suppose, that π(P ) for |C≺| = n is bounded by (1 + P )n − 1, and we want to prove that this statement holds for |C≺| = n. Let G = M, C≺ be a graph with at most n + 1 edges, and G = M, C≺ be a subgraph of G, such that C≺ = C≺ − { mi, mj }, where mj ∈ M is a sink node in G. From the induction assumption we have, that C≺ introduces the probability propagation phase error bounded by (1 + P )n − 1. We now add back the link { mi, mj } to C≺, which affects the error of only one probability function, namely Pj, by a factor of (1 + P ). Since probability propagation phase error in C≺ was bounded by (1 + P )n − 1, in C≺ = C≺ ∪ { mi, mj } it can be at most ((1 + P )n − 1)(1 + P ) < (1 + P )n+1 − 1. Thus, if opportunity cost functions are not overestimated, they are bounded by P mi∈M ri and the error of a single value function propagation operation will be at most Z Δ 0 p(t)( V +((1+ P ) |C≺| −1) X mi∈M ri) dt < V +((1+ P ) |C≺| −1) X mi∈M ri. Since the number of value function propagation operations is |C≺|, the total error π of the value function propagation phase is bounded by: |C≺| V + ((1 + P )|C≺| − 1) P mi∈M ri . 6. SPLITTING THE OPPORTUNITY COST FUNCTIONS In section 5 we left out the discussion about how the opportunity cost function Vj0 of method mj0 is split into opportunity cost functions [Vj0,ik ]K k=0 sent back to methods [mik ]K k=0 , that directly enable method mj0 . So far, we have taken the same approach as in [4] and [5] in that the opportunity cost function Vj0,ik that the method mik sends back to the method mj0 is a minimal, non-increasing function that dominates function V j0,ik (t) = (Vj0 · Q k ∈{0,...,K} k =k Pik )(t). We refer to this approach, as heuristic H 1,1 . Before we prove that this heuristic overestimates the opportunity cost, we discuss three problems that might occur when splitting the opportunity cost functions: (i) overestimation, (ii) underestimation and (iii) starvation. Consider the situation in Figure Figure 3: Splitting the value function of method mj0 among methods [mik ]K k=0. (3) when value function propagation for methods [mik ]K k=0 is performed. For each k = 0, ..., K, Equation (1) derives the opportunity cost function Vik from immediate reward rk and opportunity cost function Vj0,ik . If m0 is the only methods that precedes method mk, then V ik,0 = Vik is propagated to method m0, and consequently the opportunity cost for completing the method m0 at time t is equal to PK k=0 Vik,0(t). If this cost is overestimated, then an agent A0 at method m0 will have too much incentive to finish the execution of m0 at time t. Consequently, although the probability P(t) that m0 will be enabled by other agents by time t is low, agent A0 might still find the expected utility of starting the execution of m0 at time t higher than the expected utility of doing it later. As a result, it will choose at time t to start executing method m0 instead of waiting, which can have disastrous consequences. Similarly, if PK k=0 Vik,0(t) is underestimated, agent A0 might loose interest in enabling the future methods [mik ]K k=0 and just focus on 834 The Sixth Intl. Joint Conf. on Autonomous Agents and Multi-Agent Systems (AAMAS 07) maximizing the chance of obtaining its immediate reward r0. Since this chance is increased when agent A0 waits4 , it will consider at time t to be more profitable to wait, instead of starting the execution of m0, which can have similarly disastrous consequences. Finally, if Vj0 is split in a way, that for some k, Vj0,ik = 0, it is the method mik that underestimates the opportunity cost of enabling method mj0 , and the similar reasoning applies. We call such problem a starvation of method mk. That short discussion shows the importance of splitting the opportunity cost function Vj0 in such a way, that overestimation, underestimation, and starvation problem is avoided. We now prove that: THEOREM 2. Heuristic H 1,1 can overestimate the opportunity cost. PROOF. We prove the theorem by showing a case where the overestimation occurs. For the mission plan from Figure (3), let H 1,1 split Vj0 into [V j0,ik = Vj0 · Q k ∈{0,...,K} k =k Pik ]K k=0 sent to methods [mik ]K k=0 respectively. Also, assume that methods [mik ]K k=0 provide no local reward and have the same time windows, i.e., rik = 0; ESTik = 0, LETik = Δ for k = 0, ..., K. To prove the overestimation of opportunity cost, we must identify t0 ∈ [0, ..., Δ] such that the opportunity cost PK k=0 Vik (t) for methods [mik ]K k=0 at time t ∈ [0, .., Δ] is greater than the opportunity cost Vj0 (t). From Equation (1) we have: Vik (t) = Z Δ−t 0 pik (t )Vj0,ik (t + t )dt Summing over all methods [mik ]K k=0 we obtain: KX k=0 Vik (t) = KX k=0 Z Δ−t 0 pik (t )Vj0,ik (t + t )dt (4) ≥ KX k=0 Z Δ−t 0 pik (t )V j0,ik (t + t )dt = KX k=0 Z Δ−t 0 pik (t )Vj0 (t + t ) Y k ∈{0,...,K} k =k Pik (t + t )dt Let c ∈ (0, 1] be a constant and t0 ∈ [0, Δ] be such that ∀t>t0 and ∀k=0,..,K we have Q k ∈{0,...,K} k =k Pik (t) > c. Then: KX k=0 Vik (t0) > KX k=0 Z Δ−t0 0 pik (t )Vj0 (t0 + t ) · c dt Because Pjk is non-decreasing. Now, suppose there exists t1 ∈ (t0, Δ], such that PK k=0 R t1−t0 0 pik (t )dt > Vj0 (t0) c·Vj0 (t1) . Since decreasing the upper limit of the integral over positive function also decreases the integral, we have: KX k=0 Vik (t0) > c KX k=0 Z t1 t0 pik (t − t0)Vj0 (t )dt And since Vj0 (t ) is non-increasing we have: KX k=0 Vik (t0) > c · Vj0 (t1) KX k=0 Z t1 t0 pik (t − t0)dt (5) = c · Vj0 (t1) KX k=0 Z t1−t0 0 pik (t )dt > c · Vj0 (t1) Vj(t0) c · Vj(t1) = Vj(t0) 4 Assuming LET0 t Consequently, the opportunity cost PK k=0 Vik (t0) of starting the execution of methods [mik ]K k=0 at time t ∈ [0, .., Δ] is greater than the opportunity cost Vj0 (t0) which proves the theorem.Figure 4 shows that the overestimation of opportunity cost is easily observable in practice. To remedy the problem of opportunity cost overestimation, we propose three alternative heuristics that split the opportunity cost functions: • Heuristic H 1,0 : Only one method, mik gets the full expected reward for enabling method mj0 , i.e., V j0,ik (t) = 0 for k ∈ {0, ..., K}\{k} and V j0,ik (t) = (Vj0 · Q k ∈{0,...,K} k =k Pik )(t). • Heuristic H 1/2,1/2 : Each method [mik ]K k=0 gets the full opportunity cost for enabling method mj0 divided by the number K of methods enabling the method mj0 , i.e., V j0,ik (t) = 1 K (Vj0 · Q k ∈{0,...,K} k =k Pik )(t) for k ∈ {0, ..., K}. • Heuristic bH 1,1 : This is a normalized version of the H 1,1 heuristic in that each method [mik ]K k=0 initially gets the full opportunity cost for enabling the method mj0 . To avoid opportunity cost overestimation, we normalize the split functions when their sum exceeds the opportunity cost function to be split. Formally: V j0,ik (t) = 8 >< >: V H 1,1 j0,ik (t) if PK k=0 V H 1,1 j0,ik (t) < Vj0 (t) Vj0 (t) V H 1,1 j0,ik (t) PK k=0 V H 1,1 j0,ik (t) otherwise Where V H 1,1 j0,ik (t) = (Vj0 · Q k ∈{0,...,K} k =k Pjk )(t). For the new heuristics, we now prove, that: THEOREM 3. Heuristics H 1,0 , H 1/2,1/2 and bH 1,1 do not overestimate the opportunity cost. PROOF. When heuristic H 1,0 is used to split the opportunity cost function Vj0 , only one method (e.g. mik ) gets the opportunity cost for enabling method mj0 . Thus: KX k =0 Vik (t) = Z Δ−t 0 pik (t )Vj0,ik (t + t )dt (6) And since Vj0 is non-increasing ≤ Z Δ−t 0 pik (t )Vj0 (t + t ) · Y k ∈{0,...,K} k =k Pjk (t + t )dt ≤ Z Δ−t 0 pik (t )Vj0 (t + t )dt ≤ Vj0 (t) The last inequality is also a consequence of the fact that Vj0 is non-increasing. For heuristic H 1/2,1/2 we similarly have: KX k=0 Vik (t) ≤ KX k=0 Z Δ−t 0 pik (t ) 1 K Vj0 (t + t ) Y k ∈{0,...,K} k =k Pjk (t + t )dt ≤ 1 K KX k=0 Z Δ−t 0 pik (t )Vj0 (t + t )dt ≤ 1 K · K · Vj0 (t) = Vj0 (t). For heuristic bH 1,1 , the opportunity cost function Vj0 is by definition split in such manner, that PK k=0 Vik (t) ≤ Vj0 (t). Consequently, we have proved, that our new heuristics H 1,0 , H 1/2,1/2 and bH 1,1 avoid the overestimation of the opportunity cost. The Sixth Intl. Joint Conf. on Autonomous Agents and Multi-Agent Systems (AAMAS 07) 835 The reason why we have introduced all three new heuristics is the following: Since H 1,1 overestimates the opportunity cost, one has to choose which method mik will receive the reward from enabling the method mj0 , which is exactly what the heuristic H 1,0 does. However, heuristic H 1,0 leaves K − 1 methods that precede the method mj0 without any reward which leads to starvation. Starvation can be avoided if opportunity cost functions are split using heuristic H 1/2,1/2 , that provides reward to all enabling methods. However, the sum of split opportunity cost functions for the H 1/2,1/2 heuristic can be smaller than the non-zero split opportunity cost function for the H 1,0 heuristic, which is clearly undesirable. Such situation (Figure 4, heuristic H 1,0 ) occurs because the mean f+g 2 of two functions f, g is not smaller than f nor g only if f = g. This is why we have proposed the bH 1,1 heuristic, which by definition avoids the overestimation, underestimation and starvation problems. 7. EXPERIMENTAL EVALUATION Since the VFP algorithm that we introduced provides two orthogonal improvements over the OC-DEC-MDP algorithm, the experimental evaluation we performed consisted of two parts: In part 1, we tested empirically the quality of solutions that an locally optimal solver (either OC-DEC-MDP or VFP) finds, given it uses different opportunity cost function splitting heuristic, and in part 2, we compared the runtimes of the VFP and OC-DEC- MDP algorithms for a variety of mission plan configurations. Part 1: We first ran the VFP algorithm on a generic mission plan configuration from Figure 3 where only methods mj0 , mi1 , mi2 and m0 were present. Time windows of all methods were set to 400, duration pj0 of method mj0 was uniform, i.e., pj0 (t) = 1 400 and durations pi1 , pi2 of methods mi1 , mi2 were normal distributions, i.e., pi1 = N(μ = 250, σ = 20), and pi2 = N(μ = 200, σ = 100). We assumed that only method mj0 provided reward, i.e. rj0 = 10 was the reward for finishing the execution of method mj0 before time t = 400. We show our results in Figure (4) where the x-axis of each of the graphs represents time whereas the y-axis represents the opportunity cost. The first graph confirms, that when the opportunity cost function Vj0 was split into opportunity cost functions Vi1 and Vi2 using the H 1,1 heuristic, the function Vi1 +Vi2 was not always below the Vj0 function. In particular, Vi1 (280) + Vi2 (280) exceeded Vj0 (280) by 69%. When heuristics H 1,0 , H 1/2,1/2 and bH 1,1 were used (graphs 2,3 and 4), the function Vi1 + Vi2 was always below Vj0 . We then shifted our attention to the civilian rescue domain introduced in Figure 1 for which we sampled all action execution durations from the normal distribution N = (μ = 5, σ = 2)). To obtain the baseline for the heuristic performance, we implemented a globally optimal solver, that found a true expected total reward for this domain (Figure (6a)). We then compared this reward with a expected total reward found by a locally optimal solver guided by each of the discussed heuristics. Figure (6a), which plots on the y-axis the expected total reward of a policy complements our previous results: H 1,1 heuristic overestimated the expected total reward by 280% whereas the other heuristics were able to guide the locally optimal solver close to a true expected total reward. Part 2: We then chose H 1,1 to split the opportunity cost functions and conducted a series of experiments aimed at testing the scalability of VFP for various mission plan configurations, using the performance of the OC-DEC-MDP algorithm as a benchmark. We began the VFP scalability tests with a configuration from Figure (5a) associated with the civilian rescue domain, for which method execution durations were extended to normal distributions N(μ = Figure 5: Mission plan configurations: (a) civilian rescue domain, (b) chain of n methods, (c) tree of n methods with branching factor = 3 and (d) square mesh of n methods. Figure 6: VFP performance in the civilian rescue domain. 30, σ = 5), and the deadline was extended to Δ = 200. We decided to test the runtime of the VFP algorithm running with three different levels of accuracy, i.e., different approximation parameters P and V were chosen, such that the cumulative error of the solution found by VFP stayed within 1%, 5% and 10% of the solution found by the OC- DEC-MDP algorithm. We then run both algorithms for a total of 100 policy improvement iterations. Figure (6b) shows the performance of the VFP algorithm in the civilian rescue domain (y-axis shows the runtime in milliseconds). As we see, for this small domain, VFP runs 15% faster than OCDEC-MDP when computing the policy with an error of less than 1%. For comparison, the globally optimal solved did not terminate within the first three hours of its runtime which shows the strength of the opportunistic solvers, like OC-DEC-MDP. We next decided to test how VFP performs in a more difficult domain, i.e., with methods forming a long chain (Figure (5b)). We tested chains of 10, 20 and 30 methods, increasing at the same time method time windows to 350, 700 and 1050 to ensure that later methods can be reached. We show the results in Figure (7a), where we vary on the x-axis the number of methods and plot on the y-axis the algorithm runtime (notice the logarithmic scale). As we observe, scaling up the domain reveals the high performance of VFP: Within 1% error, it runs up to 6 times faster than OC-DECMDP. We then tested how VFP scales up, given that the methods are arranged into a tree (Figure (5c)). In particular, we considered trees with branching factor of 3, and depth of 2, 3 and 4, increasing at the same time the time horizon from 200 to 300, and then to 400. We show the results in Figure (7b). Although the speedups are smaller than in case of a chain, the VFP algorithm still runs up to 4 times faster than OC-DEC-MDP when computing the policy with an error of less than 1%. We finally tested how VFP handles the domains with methods arranged into a n × n mesh, i.e., C≺ = { mi,j, mk,j+1 } for i = 1, ..., n; k = 1, ..., n; j = 1, ..., n − 1. In particular, we consider 836 The Sixth Intl. Joint Conf. on Autonomous Agents and Multi-Agent Systems (AAMAS 07) Figure 4: Visualization of heuristics for opportunity costs splitting. Figure 7: Scalability experiments for OC-DEC-MDP and VFP for different network configurations. meshes of 3×3, 4×4, and 5×5 methods. For such configurations we have to greatly increase the time horizon since the probabilities of enabling the final methods by a particular time decrease exponentially. We therefore vary the time horizons from 3000 to 4000, and then to 5000. We show the results in Figure (7c) where, especially for larger meshes, the VFP algorithm runs up to one order of magnitude faster than OC-DEC-MDP while finding a policy that is within less than 1% from the policy found by OC- DECMDP. 8. CONCLUSIONS Decentralized Markov Decision Process (DEC-MDP) has been very popular for modeling of agent-coordination problems, it is very difficult to solve, especially for the real-world domains. In this paper, we improved a state-of-the-art heuristic solution method for DEC-MDPs, called OC-DEC-MDP, that has recently been shown to scale up to large DEC-MDPs. Our heuristic solution method, called Value Function Propagation (VFP), provided two orthogonal improvements of OC-DEC-MDP: (i) It speeded up OC-DECMDP by an order of magnitude by maintaining and manipulating a value function for each method rather than a separate value for each pair of method and time interval, and (ii) it achieved better solution qualities than OC-DEC-MDP because it corrected the overestimation of the opportunity cost of OC-DEC-MDP. In terms of related work, we have extensively discussed the OCDEC-MDP algorithm [4]. Furthermore, as discussed in Section 4, there are globally optimal algorithms for solving DEC-MDPs with temporal constraints [1] [11]. Unfortunately, they fail to scale up to large-scale domains at present time. Beyond OC-DEC-MDP, there are other locally optimal algorithms for DEC-MDPs and DECPOMDPs [8] [12], [13], yet, they have traditionally not dealt with uncertain execution times and temporal constraints. Finally, value function techniques have been studied in context of single agent MDPs [7] [9]. However, similarly to [6], they fail to address the lack of global state knowledge, which is a fundamental issue in decentralized planning. Acknowledgments This material is based upon work supported by the DARPA/IPTO COORDINATORS program and the Air Force Research Laboratory under Contract No. FA875005C0030. The authors also want to thank Sven Koenig and anonymous reviewers for their valuable comments. 9. REFERENCES [1] R. Becker, V. Lesser, and S. Zilberstein. Decentralized MDPs with Event-Driven Interactions. In AAMAS, pages 302-309, 2004. [2] R. Becker, S. Zilberstein, V. Lesser, and C. V. Goldman. Transition-Independent Decentralized Markov Decision Processes. In AAMAS, pages 41-48, 2003. [3] D. S. Bernstein, S. Zilberstein, and N. Immerman. The complexity of decentralized control of Markov decision processes. In UAI, pages 32-37, 2000. [4] A. Beynier and A. Mouaddib. A polynomial algorithm for decentralized Markov decision processes with temporal constraints. In AAMAS, pages 963-969, 2005. [5] A. Beynier and A. Mouaddib. An iterative algorithm for solving constrained decentralized Markov decision processes. In AAAI, pages 1089-1094, 2006. [6] C. Boutilier. Sequential optimality and coordination in multiagent systems. In IJCAI, pages 478-485, 1999. [7] J. Boyan and M. Littman. Exact solutions to time-dependent MDPs. In NIPS, pages 1026-1032, 2000. [8] C. Goldman and S. Zilberstein. Optimizing information exchange in cooperative multi-agent systems, 2003. [9] L. Li and M. Littman. Lazy approximation for solving continuous finite-horizon MDPs. In AAAI, pages 1175-1180, 2005. [10] Y. Liu and S. Koenig. Risk-sensitive planning with one-switch utility functions: Value iteration. In AAAI, pages 993-999, 2005. [11] D. Musliner, E. Durfee, J. Wu, D. Dolgov, R. Goldman, and M. Boddy. Coordinated plan management using multiagent MDPs. In AAAI Spring Symposium, 2006. [12] R. Nair, M. Tambe, M. Yokoo, D. Pynadath, and S. Marsella. Taming decentralized POMDPs: Towards efficient policy computation for multiagent settings. In IJCAI, pages 705-711, 2003. [13] R. Nair, P. Varakantham, M. Tambe, and M. Yokoo. Networked distributed POMDPs: A synergy of distributed constraint optimization and POMDPs. In IJCAI, pages 1758-1760, 2005. The Sixth Intl. Joint Conf. on Autonomous Agents and Multi-Agent Systems (AAMAS 07) 837

locally optimal solution;rescue mission;agent-coordination problem;decentralized partially observable markov decision process;decentralize markov decision process;temporal constraint;decision-theoretic model;multiplication;policy iteration;heuristic performance;probability function propagation;opportunity cost;value function propagation;multi-agent system;decentralized markov decision process