Chess AI with Deep Reinforcement Learning

Architecture:

• SE-Residual Network with 6 blocks and 128 filters

• Dual-head design: Policy Head (move selection) + Value Head (position evaluation)

• 12-channel board encoding including castling rights and turn indicator

Training Pipeline:

• Supervised pre-training on Lichess games dataset

• Self-play reinforcement learning with PPO

• Stockfish-guided learning with depth 10 evaluations

• Policy distillation from Stockfish best moves

Adaptive Optimization Features:

• Learning rate warmup & cosine annealing

• Dynamic gradient clipping (global norm, per-parameter, adaptive)

• Entropy scheduling for exploration-exploitation balance

• Auto-freeze mechanism to prevent model collapse

Key Technical Challenges Solved:

• Fixed BatchNorm issues causing policy degradation during RL training

• Solved draw loops with asymmetric self-play strategy

• Improved checkmate execution using tactic puzzles dataset

Evaluation:

• Win rate vs random opponents: 93-99%

• Win rate vs Stockfish depth 0: 15-35%

Deployment:

• Flask-based web interface for human vs AI gameplay

• Deployed to Hugging Face Spaces as backend API

• Mobile-responsive UI with click-to-move support

Tech Stack: Python, PyTorch, python-chess, Flask, Stockfish, Matplotlib, TensorBoard