tuhinaww / Chess-Engine

Python-based chess engine developed using Monte Carlo Tree Search (MCTS) algorithm coupled with Transformer models for reinforcement learning.

Chess Engine: MCTS and Transformer Implementation

This repository contains a Python-based chess engine developed using Monte Carlo Tree Search (MCTS) algorithm coupled with Transformer models for reinforcement learning.

Introduction

The goal of this project is to create a robust chess-playing program employing MCTS and leveraging Transformer models for state representation and reinforcement learning.

Tech Stack

Language	Deep Learning Framework	Algorithm	Models
Python	PyTorch	Monte Carlo Tree Search	Decoder-only Transformers

Approach

1. Piece Implementation:

   • Individual pieces coded:

   Piece	Quantity
   Pawn	8
   Knight	2
   Bishop	2
   Rook	2
   King	1
   Queen	1

   • Piece representations:

   Black as negative, White as positive, empty squares as 0
   

2. Class Structure:

   • Classes designed for pieces, display, and the chess board
   • Game rule-specific classes (e.g., Pawn Promotion, Fifty-move rule)

3. User Interaction:

   • Input system for moves or potential GUI implementation

4. MCTS Implementation:

   • End-game condition checked
   • Neural Networks to evaluate value + action probabilities
   • Upper Confidence Bound used for move selection

5. Transformer Models:

   • Decoder-only architecture
   • Attention mechanism for feature learning
   • Tokenized piece representation with positional encoding

6. Output and Learning:

   • Model outputs: Outcome probabilities and actions
   • One-hot encoding for classification
   • Reinforcement learning using Binary Cross Entropy loss minimization

To-Do List:

• Implement pieces: Pawns, Knights, Bishops, Rooks, King, Queen
• Define Class Structures: Piece classes, Display, Chess Board, Rule implementations
• Integrate MCTS: End-game checks, Neural Network integration
• Develop Transformer Model: Decoder architecture, Attention Mechanism, Tokenization
• Output and Learning Setup: Model output configuration, Reinforcement Learning implementation
• UCI Protocol Integration

References

1. AlphaZero: Shedding New Light on Chess, Shogi, and Go
2. Monte Carlo Tree Search - Wikipedia
3. Transformer Architecture & Positional Encoding
4. The Illustrated Transformer