LAMDA-RL / PRDC

Author's PyTorch implementation of ICML'23 paper "Policy Regularization with Dataset Constraint for Offline Reinforcement Learning" for D4RL gym and AntMaze tasks.

Policy Regularization with Dataset Constraint for Offline Reinforcement Learning

Code for ICML'23 paper "Policy Regularization with Dataset Constraint for Offline Reinforcement Learning", arXiv link.

If you find this repository useful for your research, please cite:

@inproceedings{
    prdc,
    title={Policy Regularization with Dataset Constraint for Offline Reinforcement Learning},
    author={Yuhang Ran and Yi-Chen Li and Fuxiang Zhang and Zongzhang Zhang and Yang Yu},
    booktitle={International Conference on Machine Learning},
    year={2023}
}

Install dependency

pip install -r requirements.txt

Install the D4RL benchmark

git clone https://github.com/Farama-Foundation/D4RL.git
cd d4rl
pip install -e .

Run experiment

For halfcheetah:

python main.py --env_id halfcheetah-medium-v2 --seed 1024 --device cuda:0 --alpha 40.0 --beta 2.0 --k 1

For hopper & walker2d:

python main.py --env_id hopper-medium-v2 --seed 1024 --device cuda:0 --alpha 2.5 --beta 2.0 --k 1

We use reward shaping for antmaze, which is a common trick used by CQL, IQL, FisherBRC, etc.

python main.py --env_id antmaze-medium-play-v2 --seed 1024 --device cuda:0 --alpha 7.5 --beta 7.5 --k 1 --scale=10000 --shift=-1

See result

tensorboard --logdir='./result'