zero to checkmate: RL Based Chess Engine

Overview

zero to checkmate is a chess-playing repository that leverages reinforcement learning and Monte Carlo methods to make strategic decisions. The repository provides various features that simulate self-play, evaluate positions, and make intelligent moves based on a neural network model trained through various techniques, including AlphaZero with Stockfish reinforcement, basic AlphaZero, Leela architecture, Monte Carlo Tree Search (MCTS), Q-learning on MCTS, and a CNN architecture for self-play.

Resources

• Chess Games from https://www.pgnmentor.com/files.html#players
• Stockfish 17 Used (https://stockfishchess.org/download/)

Features

• Reinforcement Learning - Implements Reinforcement Algorithms like QLearning and Self Learning using Stockfish
• Minimax Algorithm: Implements a minimax algorithm with alpha-beta pruning for optimal move selection.
• Value Estimation: Utilizes a neural network for board state evaluation, alongside classic heuristic evaluations.
• Self-Play: Capable of playing against itself for training and testing purposes.
• PGN Export: Saves the game in PGN format for analysis and record-keeping.
• Transfer Learning: Implemented Transfer Learning on different models.

Installation

1. Python 3.x: Ensure you have Python 3 installed on your machine.

2. PyTorch: Install PyTorch following the instructions from the official website.

3. python-chess: Install the python-chess library.

   pip install python-chess

Architectures

1. alpha_stockfish: Implements RL based Alpha Zero architecture with self learning using Stockfish (https://nikcheerla.github.io/deeplearningschool/2018/01/01/AlphaZero-Explained/)

2. MCTS: Monte Carlo Tree Search (https://en.wikipedia.org/wiki/Monte_Carlo_tree_search)

3. Leela: Implements the LEELA architecture (https://lczero.org/dev/backend/nn/)

4. Alpha Zero: Implements the base AlphaZero architecture (https://nikcheerla.github.io/deeplearningschool/2018/01/01/AlphaZero-Explained/)

5. Q-Learning based MCTS: Added Q Learning to improve MCTS (https://en.wikipedia.org/wiki/Q-learning)

6. CNN.py: Implements a CNN architecture for self-play. (https://geohot.com/)