AubreeL/chess-bot

Chess Bot Model - TinyPCN

A chess playing neural network trained on expert games from the Lichess Elite Database.

Model Description

This is a policy-value network inspired by AlphaZero, designed to evaluate chess positions and suggest moves.

Architecture

• Input: 18-plane board representation (12 pieces + 6 metadata planes)
• Convolutional backbone: 32 filters, 1 residual block, ~9,611,202 parameters
• Policy head: 4,672-dimensional output (one per legal move encoding)
• Value head: Single tanh output (-1 to +1 for position evaluation)

Training Data

• Dataset: Lichess Elite Database (games from 2200+ ELO players)
• Positions trained: 16,800,000
• Epochs: 10

Performance

• Final Policy Loss: 2.8000
• Final Value Loss: 0.8500

Usage

import torch
import chess
from model import TinyPCN, encode_board

# Load model
model = TinyPCN(board_channels=18, policy_size=4672)
model.load_state_dict(torch.load("chess_model.pth"))
model.eval()

# Evaluate a position
board = chess.Board()  # or chess.Board("fen string")
board_tensor = encode_board(board).unsqueeze(0)

with torch.no_grad():
    policy_logits, value = model(board_tensor)

# Value interpretation:
# +1.0 = winning for current player
#  0.0 = drawn/equal position
# -1.0 = losing for current player

print(f"Position evaluation: {value.item():.4f}")

Model Files

• chess_model.pth - PyTorch model weights
• model.py - Model architecture and board encoding
• mcts.py - Monte Carlo Tree Search implementation
• requirements.txt - Python dependencies

Limitations

• Trained on expert games only (no self-play yet)
• Lightweight architecture for educational purposes
• May not handle unusual openings or endgames well

Training Details

• Framework: PyTorch
• Optimizer: Adam
• Learning Rate: 0.001
• Batch Size: 256
• Loss Functions: CrossEntropyLoss (policy) + MSELoss (value)

Authors

Created as part of an AlphaZero-style chess engine project.

License

MIT License