PyTorch implementation of the MR.Q algorithm from Towards General-Purpose Model-Free Reinforcement Learning by Scott Fujimoto, Pierluca D'Oro, Amy Zhang, Yuandong Tian, and Michael Rabbat.
Experiments were originally run with Python 3.11, but Python 3.10-3.12 is supported.
git clone [email protected]:facebookresearch/MRQ.git
cd MRQ
pip install -r requirements.txt
Benchmark is designated by a prefix (Gym-, Dmc-, Dmc-visual-, Atari-) followed by the original environment name. A complete list of environments are contained in MRQ/utils.py.
Example usage:
cd MRQ
python main.py --env Gym-HalfCheetah-v4
python main.py --env Dmc-quadruped-walk
python main.py --env Dmc-visual-walker-walk
python main.py --env Atari-Pong-v5
- Agent and hyperparameters: MRQ/MRQ.py.
- Architecture: MRQ/models.py.
- Replay buffer: MRQ/buffer.py.
- Environment preprocesing: MRQ/env_preprocessing.py.
Results are formatted in human-readable .txt files under /results. There is a code snippet in MRQ/utils.py to process the .txt files into a dictionary of arrays.
MRQ is licensed under the CC BY-NC 4.0 license, as found in the LICENSE file.