Update README.md

README.md

@@ -18,13 +18,24 @@ tags:
 
 ## Introduction
 
-* **GENERAL**
+  *Combinatorial Optimization Problems(COPs)* has long been an active field of research. Generally speaking, there exists two main approachs for solving COPs, each of them having pros and cons. On the one hand, the *exact algorithms* can find the optimal solution, but they may be prohibitive for solving large instances because of the exponential increate of the execution time.
+  On the other hand, *heuristic algorithms* can compute solutions efficiently, but are not able to prove the optimality of solutions.
+
+  In the realistic business scenarios, some COPs are large-scale(>=1000 nodes), which have very strict requirements for the execution time and quality of solutions. To better solve these problems, we
+  proposes a generic and complete solver, called **GreedRL**, based on Deep Reinforcement Learning(DRL), which achieves better speed and quality of solutions than *heuristic algorithms* .
 
+* **GENERAL**
+  
+  The GreedRL solver makes **a high level of abstraction for Combinatorial Optimization** problems, which can solve various VRP, Batching, Scheduling and Online Assignment problems. At the same time, for the VRPs, it also supports variants of VRPs with different constraints, such as Time-Window, Pickup-Delivery, Split-Delivery, Multi-Vehicles, etc.
 
 * **HIGH-PERFORMANCE**
+
+  We have improved the RL environment simulation speed by **CUDA and C++ implementations**. At the same time, some operators are costomized to replace the native operators of pytorch, to achive the ultimate computing performance.
   
 * **USER-FRIENDLY**
 
+  We have **warped commonly used modules**, such as NN components, RL training algothrim and CO constraints implementations, which makes GreedRL easy to use.
+  
 
 ## Architecture design
 The entire architecture is divided into three layers: