I built a complete, game-agnostic implementation of AlphaZero in JAX.
repo: https://github.com/wtedw/nanoAlphaZero
demo (NN + MCTS run locally in your browser): https://nanoalphazero.wtedw.com
It uses no human data, can train grandmaster-level chess models, and supports a variety of games: Chess, Go 3x3 - 9x9, Hex 4x4 - 9x9, Connect Four
You can also use this repo to train AlphaZero on any custom 2-player, perfect-information game.
How does it work?
At a high level, the entire AlphaZero algorithm gets compiled into a single jitted run_fn that repeatedly performs self-play and model updates:
state = make_alphazero()
def run_fn(state):
games = selfplay(state) # using Gumbel MuZero
# Move active games into the self-play buffer
# Move completed games into the replay buffer
state = train(state, replay_buffer.sample())
return state
while True:
state = run_fn(state)
There are no threads, queues, or distributed workers to manage. It is just one large JAX function.The repo is primarily focused on making large-scale AlphaZero experimentation fast and easy to run. Training strong models is secondary and mostly serves as a sanity check that the underlying logic is sound.
Documentation on training custom / complex games is sparse, so if you have any questions feel free to message me.