Deep learning and computer chess (part 1)

Abstract

This paper focuses on two novel implementations of modern chess engines that use neural networks: Giraffe and DeepChess. Both models aim to evaluate chess positions without relying on handcrafted heuristics that are common in traditional chess engines. Giraffe consists of a three-layer feed-forward network that uses a low-level feature representation of game positions and makes predictions about the advantage of players (White and Black) in a given position. On the other hand, DeepChess employs a Siamese network architecture, integrating two models that are based on autoencoders, to compare game positions and predict the winner between two game states. In this paper, we implement, train and evaluate these models while introducing modifications to improve their performance and test the effectiveness of certain model parameters. Different architectures which include creating deeper networks as well as alternative activation functions were tested. Our results show that in the problem space of multiclass classification, a modified deeper Giraffe architecture with more hidden layers significantly improved evaluation accuracy. Our DeepChess implementation with the Leaky Rectified Linear Unit activation function achieved the best performance, albeit with significant overfitting. These experiments help provide further insight into the strengths and limitations of these models and neural network based chess engines in general, as well as the potential for future improvements.