Deep learning and computer chess (Part 1): using neural networks for chess evaluation functions

Abstract

This report presents the implementation of two different chess evaluation functions based on the Giraffe and DeepChess papers. In the first implementation, the evaluator network architecture from Giraffe’s evaluation function was adapted into a multiclass classifier designed to predict 7 classifications of Stockfish evaluations through supervised learning. Experiments were conducted to gauge the effectiveness of input feature representations and dropout regularisation. The second implementation, based on DeepChess, uses a different approach to evaluation, through comparison of two chess positions in a Siamese network and outputs which of the two has a more advantageous position, evaluating board positions through binary classification. The network was trained in a two-stage process with a combination of unsupervised and supervised learning. Experiments were conducted to observe the effect of freezing pretrained layer weights as well as changing layer activation functions to LeakyReLU.