This is a class project for AA228 at Stanford. We aim to train an agent using Proximal Policy Optimization on ray.rllib's distributed computing infrastructure to play Gardner minichess.
.
├── games # Implementation of minichess variations' board and game logic
├── mcts_old # Initial experiments with mcts (hence the repo name)
├── mcts_for_train.yml # Up-to-date conda requirements for macos
├── sample_games.txt # Sample games from log output
├── train.py # Main training script
Many variations of Gardner minichess exist. We preserve most well-known chess concepts (e.g. castling, en-passant, fifty-move), but for purposes of incremental development of our agent, we may add in these rules one by one. Following the implementation of meta-minichess, we define game termination as when either side's king is captured (including by the opposing king too).
Meta-learning experiments for the game of minichess and related rule variants. - Michael Hiebert (code) Proximal Policy Optimization Algorithms - John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, Oleg Klimov (paper)