♟️ Chess Engine with CNN Evaluation & Negamax Search

A full-stack chess engine that combines a Convolutional Neural Network (CNN) for position evaluation with a classical Negamax search algorithm. The project focuses on integrating machine learning into a traditional game-search pipeline rather than competing with professional engines.

---

🚀 Features

• Play a full game of chess against an AI
• CNN-based position evaluation
• Negamax search with:
  • Alpha–beta pruning
  • Quiescence search
  • Transposition table
  • Iterative deepening
• Legal move generation using python-chess
• FastAPI backend with REST API
• React-based frontend chessboard UI
• Interactive UI built with Streamlit (v1.0)
• Clean separation between ML, engine logic, and UI

---

🧠 Architecture Overview

Backend/
├── api/ # FastAPI server
├── engine/ # Chess engine (Negamax + quiescence search + move selection + Game flow logic)
├── ml/ # ML inference + board encoding (Tensorflow)
├── models/ # Trained CNN model (.keras)
├── ui/ # Streamlit UI (v1.0)
├── utils/ # Normalize FEN
├── chess.ipynb # Training & experimentation notebook
└── requirements.txt

Frontend/ ├── React chess UI

Core Components

CNN Evaluator

• Input: 8×8×17 tensor representation of the board
• Encodes:
  • 12 piece planes (6 white + 6 black)
  • 1 side-to-move plane
  • 4 castling-rights planes
• Output: Scalar evaluation
  • Positive → White is better
  • Negative → Black is better

The CNN is only used at leaf nodes (and in quiescence), not during the full search.

Engine

• Negamax formulation (single-perspective minimax)
• Alpha–beta pruning
• Quiescence search for capture stability
• Transposition table with normalized FEN keys
• Iterative deepening with time control

This design mirrors how real chess engines are structured, though simplified.

Backend (FastAPI)

• Endpoint:

  POST /move

• Input:

  {
    "fen": "...",
    "move": "e2e4",
    "max_time": 1.0
  }

• Output:

  {
    "fen": "...",
    "ai_move": "c7c5"
  }

• CNN model is lazy-loaded
• Evaluation results are cached for performance

UI

• Streamlit-based interactive chess interface
• Dropdown-based move selection (legal moves only)
• Real-time board updates

Frontend (React)

• Interactive chessboard
• Legal-move enforcement
• Promotion handling
• AI responds automatically after user move
• Communicates with backend via REST API

---

🏋️ Model Training

Training and experimentation are done in chess.ipynb:

• PGN parsing
• Stockfish-based position labeling
• Handling class imbalance
• CNN training and evaluation
• Confusion matrix analysis

The notebook outputs a trained model:

models/chess_cnn_evaluator.keras

---

🚀 Live Demo

Streamlit:
👉 https://aichess.streamlit.app/

Frontend:
👉 https://chess-one-sable.vercel.app/

Backend:
👉 https://chess-d5p9.onrender.com/

---

▶️ How to Run

Backend

cd backend
pip install -r requirements.txt
uvicorn main:app --reload

Run the streamlit application

cd backend
streamlit run ui/app.py

Run the frontend application

cd frontend
npm install
npm run dev

---

⚙️ Configuration

• AI thinking time is configurable per move
• Longer time → deeper search → stronger play
• Shorter time → faster but weaker play

---

📌 Limitations (v2)

• Written in Python, not C++
• CNN evaluation is slower than NNUE-style models
• No opening book or endgame tablebases
• Not intended to compete with Stockfish or Chess.com engines

This project prioritizes clarity, correctness, and learning over raw playing strength.

---

🔮 Future Improvements

• Opening book integration
• Killer moves & history heuristics
• NNUE-style evaluator
• Endgame tablebases
• C++ engine core
• Self-play training
• Stronger evaluation targets (centipawn regression)

---

👨‍💻 Author

Built as a learning-focused project to explore how machine learning models can be integrated into classical game-search algorithms, with an emphasis on clean architecture and correctness.

---