This repository is an adaptation of alpha-zero-general containing information-imperfect board games variants, and other code adjusted to allow for information imperfection.
The corresponding paper can be found here.
For easy environment setup, we can use nvidia-docker. Once you have nvidia-docker set up, we can then simply run:
./setup_env.sh
to set up a (default: pyTorch) Jupyter docker container. We can now open a new terminal and enter:
docker exec -ti pytorch_notebook python main.py
Some contributions that would be interesting to see:
- More imperfect-info games
- Making games work with human players
- Making MCTS work with non-placement games (Kriegspiel & such) - important if adding new games
- Games that allow one agent imperfect, one perfect (obviously one-sided but could lead to interesting results)
- Surag Nair and the other contributors to the original alpha-zero-general repository. Their code has been reused under the MIT licence.
- Kobe Knowles adjusted the neural networks, trained the models and obtained most of the results seen in the paper.
- Bhargava Gowda adapted MuZero in a similar way. His repository can be found here.