Dataset Card for KoopmanRL

Dataset Summary

This dataset contains the collected experimental data used for the results of Koopman-Assisted Reinforcement Learning allowing for the full reproduction, and further use of the paper's results. To reproduce the results by running the experiments yourself, please see the source code of KoopmanRL.

Dataset Structure

The dataset of the reinforcement learning experiments for KoopmanRL contains roughly 461MB of Tensorboard files, and saved policies.

Experiment	Size	Purpose
Episodic Returns	161MB	Episodic returns of all 5 considered algorithms across all 4 environments
Interpretability	55MB	Inspection of the interpretability introduced by KoopmanRL
AblationSKVIBatchSize	3.4MB	Ablation of the sensitivity to the chosen batch size
AblationSKVICompute	21MB	Ablation of the sensitivity to the amount of compute used for the construction of the Koopman tensor
AblationSAKCMonoid	86MB	Ablation of the sensitivity to the order of the monoids used for the construction of the dictionaries of the Koopman tensor
AblationSAKCCompute	134MB	Ablation of the sensitivity to the amount of compute used for the construction of the Koopman tensor

In addition the already extracted dataframes are provided. All experiments are stored as Tensorboard files, with the extracted episodic returns stores in .parquet.gz data frames for use with Pandas, and saved policies stored in .pt files.

Reproducing Plots

All plots can be reproduced with the respective Jupyter notebooks, which can be found in the order of appearance in the paper:

• Episodic Returns
• Zoomed-in Episodic Returns of the Fluid Flow and Double Well
• Zoomed-in Episodic Returns of the Linear System
• Interpretability Plots & Numbers
• Ablation Heatmaps

Usage of the Dataset

The dataset can easiest be used with the HuggingFace Datasets Library, with which one is able to either download the entire dataset

from datasets import load_dataset

ds = load_dataset("dynamicslab/KoopmanRL")

or a desired subparts of the dataset

from datasets import load_dataset

ds = load_dataset("dynamicslab/KoopmanRL", data_dir="data/EpisodicReturns")

Licensing

The entire dataset is licensed under a CC-BY-4.0 license.

Contact Info

1. Preston Rozwood (pwr36@cornell.com)
2. Edward Mehrez (ejm322@cornell.edu)
3. Ludger Paehler (paehlerludger@gmail.com)
4. Steven L. Brunton (sbrunton@uw.edu)

How to Cite

Please cite the dataset in the following format

@misc{dynamicslab_2024,
    author={ {Dynamicslab} },
    title={ KoopmanRL (Revision fcca4b3) },
    year=2024,
    url={ https://huggingface.co/datasets/dynamicslab/KoopmanRL },
    doi={ 10.57967/hf/1825 },
    publisher={ Hugging Face }
}

alongside the paper

@article{rozwood2024koopman,
  title={Koopman-Assisted Reinforcement Learning},
  author={Rozwood, Preston and Mehrez, Edward and Paehler, Ludger and Sun, Wen and Brunton, Steven L.},
  journal={arXiv preprint arXiv:tbd},
  year={2024}
}