Model-Based Offline Reinforcement Learning with Pessimism-Modulated Dynamics Belief

Code to reproduce the experiments in Model-Based Offline Reinforcement Learning with Pessimism-Modulated Dynamics Belief.

Installation

1. Install MuJoCo 2.0.0 to ~/.mujoco/mujoco200.
2. Create a conda environment and install requirements.

cd PMDB
conda env create -f PMDB_env.yml
conda activate PMDB_env

Usage

For example, use the following command to run Hopper-medium-v2 benchmark in D4RL.

python main.py --task=hopper-medium-v2

Detailed configuration can be found in config.py.

Logging

By default, TensorBoard logs are generated in the log/ directory.

Citing PMDB

@inproceedings{guo2022pmdb,
  title={Model-Based Offline Reinforcement Learning with Pessimism-Modulated Dynamics Belief},
  author={Kaiyang Guo and Yunfeng Shao and Yanhui Geng},
  booktitle{NeurIPS},
  year={2022}
}