Official codebase for "The Generalization Gap in Offline Reinforcement Learning".
By Ishita Mediratta*, Qingfei You*, Minqi Jiang, Roberta Raileanu. [* = Equal Contribution]
@inproceedings{
mediratta2024gengap,
title={The Generalization Gap in Offline Reinforcement Learning},
author={Ishita Mediratta and Qingfei You and Minqi Jiang and Roberta Raileanu},
booktitle={The Twelfth International Conference on Learning Representations},
year={2024},
url={https://openreview.net/forum?id=3w6xuXDOdY}
}
We have two sub-folders:
- procgen: Provides the datasets and experimental code for running experiments in the Procgen benchmark.
- webShop: Provides similar resources for the WebShop benchmark.
Each of these subfolders utilizes different frameworks and libraries. Therefore, please refer to the corresponding README.md in the respective subfolders for more information on how to setup the code, download the necessary datasets, and train or test different methods.
The majority of gen_dgrl code is licensed under CC-BY-NC, however portions of the project are available under separate license terms: In procgen subfolder, code for DT and online is licensed under the MIT license. The majority of webShop code is licensed under MIT license (see webshop_LICENSE.md), with train_choice_[il,bcq,cql].py files licensed under Apache 2.0 license.