Framework to carry out both Single-Agent and Multi-Agent Reinforcement Learning experiments at a relative scale, with an initial focus on Self-Play. Developed by PhD students at the University of York. This framework has been in constant development since December 2018, and will continue to evolve to add new features and algorithms for many more years!
- PyTorch implementation of: DQN,Double DQN,Double Dueling DQN,A2C,REINFORCE,PPO...
- Every implementation is compatible with OpenAI Gym and Unity environments.
- State-of-the-art Self-Play trainings scheme for Multi-Agent environments, as introduced here.
- Emphasis on cross-compatibility and clear interfaces to add new algorithms. See Adding a new algorithm.
This project evolves fast, and is mostly maintained by a single developer, meaning that documentation will most likely be outdated. We have made a point of heavily documenting most of the codebase. So please refer to the source code for extra documentation. Some documentation can be found in the docs.
This project has not yet been uploaded to PyPi. This will change soon!
Firstly, clone this repository:
git clone https://github.com/Danielhp95/Generalized-RL-Self-Play-Framework
Secondly, install it locally using the -e flag in the pip install command:
cd Generalized-RL-Self-Play-Framework/
pip install -e .
Python dependencies are listed in the file setup.py. This package enforces Python version 3.6 or higher.
If you would like Python 2.7 or other Python versions <3.6 to work, feel free to open an issue.
Read License
List of papers that used this framework.