Provable Benefits of Unsupervised Data Sharing in Offline RL

This is a jax implementation of PDS on Datasets for Deep Data-Driven Reinforcement Learning (D4RL), the corresponding paper is The provable benefits of unsupervised data sharing for offline reinforcement learning.

Quick Start

For experiments on D4RL, our code is implemented based on IQL:

$ python3 train_data_sharing.py --env_name=walker2d-expert-v2 --source_name=walker2d-random-v2 --config=configs/mujoco_config.py --data_share=learn  --target_split=0.05  --source_split=0.1

Citing

If you find this open source release useful, please reference in your paper (it is our honor):

@article{hu2023provable,
  title={The provable benefits of unsupervised data sharing for offline reinforcement learning},
  author={Hu, Hao and Yang, Yiqin and Zhao, Qianchuan and Zhang, Chongjie},
  journal={arXiv preprint arXiv:2302.13493},
  year={2023}
}

Note

• If you have any questions, please contact me: yangyiqi19@mails.tsinghua.edu.cn.