metadata
license: mit
library_name: pytorch
tags:
- chess
- alphazero
- reinforcement-learning
- mcts
- self-play
Genesis-1
An AlphaZero-style chess engine trained from self-play. It's a 15-block / 192-channel residual network (~10M parameters).
Download
To download the model: genesis_1.pt
How it works and how to use it
Genesis-1 is an AlphaZero-style network paired with PUCT Monte Carlo Tree Search.
- Input: the board encoded as 20 planes of 8x8, from the side-to-move's perspective (piece positions, castling rights, en passant, fifty-move counter, repetition).
- Body: a 15-block, 192-channel residual tower (~10M parameters).
- Two heads:
- Policy over the 4672-move AlphaZero action space (8x8x73).
- Value in [-1, 1], the expected game outcome for the side to move.
- Search: moves are chosen by MCTS guided by the network.
You can download the inference script from here
Stats
It's fundamentally a weak engine. Its Elo rating is estimated to be between 300 and 500.
Training
Genesis-1 was trained from self-play on:
- Hardware: GPU: RTX 3090 | CPU: Ryzen 9 5950X | RAM: 32GB
- OS: Windows 11
- Training time: ~12 days