santoshchandu/chess-alphazero-pytorch

Chess AlphaZero (PyTorch)

Chess policy & value network trained from scratch on 1.2M positions from Lichess.

Architecture

• Residual CNN (6 blocks, 128 filters, 10M parameters)
• Policy head: move probability over 4096 possible moves
• Value head: win probability in [-1, +1]
• MCTS: 100 simulations per move

Training

• Dataset: Lichess (20,000+ games, 1.2M positions)
• Epochs: 20
• Final validation accuracy: 42.7%
• Framework: PyTorch