Knightmare

Knightmare is a minimal chess engine written by Dennis Vink (drvink.com) combining deep learning and Monte Carlo Tree Search (MCTS) to play full games of chess through a browser interface.

This project features:

• A PyTorch-based neural network that evaluates board positions.
• A Monte Carlo Tree Search (MCTS) planner for strong move selection.
• A Flask web interface to interact with the engine.
• A complete training pipeline to generate and learn from chess data.
• Optional Docker containerization for easy deployment.

---

Demo

• https://huggingface.co/spaces/dennisvink/knightmare
• https://chessbot-uo7slb3u7a-uc.a.run.app/

---

How It Works

1. Board Representation: FEN strings are encoded into 17-channel tensors (12 piece types, 4 castling planes, 1 en-passant).
2. Neural Network: A deep residual convolutional network estimates:
   • Policy: probability distribution over legal moves.
   • Value: predicted outcome (win/draw/loss).
3. MCTS: The network guides a Monte Carlo Tree Search to simulate positions and return the best move.
4. Flask API: Communicates between frontend and engine to serve moves interactively.

---

Getting Started Locally

1. Clone the repository

git clone https://github.com/dennisvink/knightmare.git
cd knightmare

2. Set up a virtual environment

python -m venv .venv/
source .venv/bin/activate  # On Windows use `.venv\Scripts\activate`

3. Install dependencies

pip install -r requirements.txt

4. Ensure model and index exist

Ensure the following files exist in the project root:

• model.pt – trained model weights
• move_to_idx.json – dictionary mapping UCI moves to index positions

5. Run the server

python app.py

Navigate to http://localhost:7860 to play!

---

Docker

Build the image

docker build -t knightmare .

Run the container

docker run -p 7860:7860 knightmare

Visit http://localhost:7860 in your browser.

---

🎓 Training Pipeline

You can train Knightmare from scratch using your own labeled games.

Step 1: Prepare Your Data

Under the pipeline/ directory:

• fens/: Each file contains one game with a list of FENs, one per line. First line must be # Result 1-0, 0-1, or 1/2-1/2.
• moves/: Each file has the corresponding best move per position (UCI format), one per line. Use none for terminal positions.

Example:

fens/game000000000.txt
moves/game000000000.moves.txt

Step 2: Preprocess into shards

cd pipeline/
python data_preparation.py

This creates shards/ with training and validation .pt files.

Step 3: Train the model

python train.py

• Will resume from checkpoint_last.pt if available.
• Saves best model as checkpoint_best.pt.

---

📁 Project Structure

├── app.py
├── model.pt
├── move_to_idx.json
├── requirements.txt
├── Dockerfile
├── templates/
│   └── index.html
├── static/
│   └── favicon.ico
└── pipeline/
    ├── fens/
    ├── moves/
    ├── shards/
    ├── data_preparation.py
    └── train.py

---

Requirements

• Python 3.8+
• PyTorch
• Flask
• python-chess
• NumPy
• tqdm

Installable via pip install -r requirements.txt.

---

Contributing

Knightmare is an educational chess engine meant to grow. Feel free to fork, modify, and experiment.