suragnair/alpha-zero-general

Implements self-play reinforcement learning via Monte Carlo Tree Search (MCTS) combined with neural network training in a modular architecture where games and frameworks are pluggable through subclassing `Game.py` and `NeuralNet.py`. The core training loop (`Coach.py`) alternates between self-play episodes guided by MCTS and neural network optimization, supporting PyTorch and Keras backends with configurable hyperparameters for simulation depth, batch size, and learning rates. Includes pretrained models and enables direct evaluation against baseline opponents through the pit interface.

4,388 stars. No commits in the last 6 months.