♟️ Chess AlphaZero — OpenEnv
AlphaZero-style chess agent using MCTS + ResNet self-play.
Architecture
- ResNet: 5 residual blocks, 64 channels
- Input : 19 × 8 × 8 board tensor
- Output: Policy (4096 moves) + Value (-1 to +1)
Training Results
- Final Policy Loss : 1.4836
- Final Value Loss : 0.0030
- Avg Game Length : 34.0 moves
- Total Iterations : 20
Usage
import torch
net = ChessNet(in_channels=19, channels=64, num_res_blocks=5)
ckpt = torch.load("chess_alphazero_final.pt")
net.load_state_dict(ckpt["model"])
net.eval()