This repository contains the code to reproduce the experimental results of A-Crab algorithm in the paper Importance Weighted Actor-Critic for Optimal Conservative Offline Reinforcement Learning by Hanlin Zhu, Paria Rashidinejad, and Jiantao Jiao. Most of the code and instructions are adapted from and follow the logic in the lightATAC repo.
git clone https://github.com/zhuhl98/ACrab.git
or
git clone git@github.com:zhuhl98/ACrab.git
conda create -n ACrab python=3.9
conda activate ACrab
cd ACrab
pip install -e .
bash install.sh
echo "export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:~/.mujoco/mujoco210/bin:/usr/lib/nvidia" >> ~/.bashrc
source ~/.bashrc
conda activate ACrab
pip install Cython==0.29.36
python main.py --log_dir ./tmp_results --env_name hopper-medium-expert-v2 --beta 1 --C_infty 1
More instructions can be found in main.py, and please see the original paper for hyperparameters (e.g., beta, C_infty). The code was tested with python 3.9.