A comparison of generalization techniques on supervised chess evaluation functions

Abstract

This dissertation was created in pursuit of comparing different deep neural network general-ization techniques on neural networks that were trained by supervised optimisation to be used as chess evaluation functions. Investigations were performed on training material, network anatomy and on specialized generalization techniques. These experiments were evaluated firstly as normal classifiers, and then as part of a fully fledged chess agent. It was found that deep convolutional neural networks perform exceptionally well in both of the criteria men-tioned, but that their hyperparameters may be harder to optimize. A strong correlation be-tween classifier performance and chess prowess was also noted, suggesting that classification metrics can in a broad sense be used as a gauge of chess performance.