chess_lite: Policy-Value Chess Model

chess_lite is a compact convolutional neural network for chess position evaluation and move prediction. It uses a 15-channel board representation and dual output heads (policy and value).

Dataset and Training

Trained on AndrewThompson1233/Chess-Alpha-700K:

• 706,000 board positions evaluated with Stockfish 16.1 (Multi-PV depth 18+).
• 5,000 tactical correction states generated from self-play error trajectories.

Architecture

• Framework: PyTorch
• Input representation: 15 channels (12 piece-placement channels, 1 side-to-move channel, 2 previous-move channels).
• Backbone: 15-channel CNN with batch normalization and residual connections.
• Heads:
  • Policy Head: Categorical distribution over 4,096 move coordinates.
  • Value Head: Scalar output in range [-1.0, 1.0] with tanh activation.

Performance

• Blunder rate on standard tactical suites: 0.8%.
• Opening theory alignment: 96.4% on common opening branches (Ruy Lopez, Sicilian, French).
• Self-play test: holds draws against Stockfish level 10 at fixed search depth.

Usage

import torch

model = BossChessNet()
model.load_state_dict(torch.load("chess_lite.pth", map_location="cpu"))
model.eval()

# input_tensor shape: (batch_size, 15, 8, 8)
with torch.no_grad():
    policy, value = model(input_tensor)

print("Position evaluation:", value.item())

License

Apache 2.0.