Chess position evaluation using neural networks

Abstract

The game of chess is the most widely examined game in the field of artificial intelligence and machine learning. There are many approaches where implementations attempt to substitute parts, or the whole functionality of a chess engine. In this Thesis we propose a method for obtaining the evaluation of a chess position without using tree search and examining each candidate move separately, like a chess engine does. Instead of exploring the search tree in order to look several moves ahead, we propose to use the much faster and less computationally demanding predictions of a properly trained neural network. Such an approach offers the benefit of having a prediction for the position evaluation in a matter of milliseconds, while a chess engine may need even minutes to achieve the same result. The proposed approach introduces a novel set of input features, in conjunction with models which are based on the radial basis function (RBF) neural network architecture and trained with the fuzzy means algorithm; two different methods of network training are also examined and compared, involving the multilayer perceptron (MLP) network architecture. All methods were based upon the same dataset which was derived by a collection of over 1500 top-level chess games. A Java application was developed for processing the games and extracting certain features from the arising positions in order to construct the training dataset, which contained data from 81967 positions. Various networks were trained and tested as we considered different variations of each method regarding input variable configurations and dataset filtering. Ultimately, the results indicated that the proposed approach using the RBF method was the best in performance. The models produced with the proposed approach are suitable for integration in model-based decision making frameworks, e.g. model predictive control (MPC) schemes, which could form the basis for a fully fledged chess playing software.