Neural Network Chess Engine

Final Project @ Spiced Academy | Sept. 2022

This Chess Engine was created in one week as the final project of the 2022 Spiced Academy - Data Science Bootcamp in Berlin by Valentin Lorenzen.

It is fairly basic and makes a lot of mistakes, especially at the beginning of the game. It is currently running at a depth of two moves, meaning it calculates two moves into the future. Higher depth would require longer calculation time or a more efficient Alpha-Beta pruning search.

See if you can beat it and read up on how it was created below.

---

Tech-Stack:

• Tensorflow/Keras
• Flask
• CSS/Javascript/Python/HTML

---

Resources:

• DeepChess: End-to-End Deep Neural Network for Automatic Learning in Chess by Omid E. David, Nathan S. Netanyahu, and Lior Wolf

• A TensorFlow implementation of "DeepChess: End-to-End Deep Neural Network for Automatic Learning in Chess" by oripress

• python-chess: a chess library for Python by Niklas Fiekas

• chessboard.js by Chris Oakman

• FlaskChess by brokenloop

---

Tensorflow Model:

The engine is trained on the CCRL computer chess database consisting of 1'450'489 chess games. From those games 2'000'000 positions were extracted at random, 50% from games that white won at the end and 50% from games that black won at the end of the game.

The positions were first converted into the commonly used FEN format and then transformed into a Bitboard - 769 bit representation.

The Tensorflow model takes two of these positions as input (x1 and x2) and gives an output (y) between 0 and 1. (1 meaning the position in input 1 is more winning for the white side than the position in input 2.)

The model consists of 7 layers. Two input layers, two hidden layers extracting high level features from the input positions, two hidden layers computing the actual evaluation and one output layer with a sigmoid activation.

The Tensorflow structure looks as follows:

Four dropout layers were additionally used for adding noise to the training-process to prevent overfitting as well as L2 Kernal-Regulization at each layer. All layers, except the output layer, were activated through a Rectified Linear Unit (ReLu) - activation.

"Adam" was chosen as the optimzier as well as Binary Crossentropy as the loss-function.

The model reached a validation accuracy of 87% through 300 epochs of training with a Learning Rate that started at 0.001 and was multiplied by 0.99 at each epoch.

---

Things to improve:

• Implement an autoencoder structure

• Position hashing and draw-avoidence in winning positions

• Improved Alpha-Beta Search

• Write custom training loop for better hyperparameter optimization

---

Files:

• train.py: Train the A.I
• chess_engine.py: Evaluating and chosing a move
• app.py: Flask app