Chess Results Analysis Using Elo Measure with Machine Learning

Abstract

An Elo score is a known measure of past performance in chess and other games. This paper investigates the impact of Elo ratings on chess game results and whether this measure can be used to predict future performance in matches. To achieve the aim, various machine learning classification techniques have been evaluated using chess data sourced from an online chess server. We examine how much influence the Elo score has on the prediction power of these techniques based on classifiers they derive. The prime objective of this experiment is to accurately predict the winner of a chess game from attributes that are available before the game starts. We are particularly interested in how large an impact the Elo score has on the prediction when compared with other features. Empirical results reported that classifiers derived by artificial neural network (Multilayer Perceptron), Decision Tree (J48/C4.5), Rule Induction (JRip/RIPPER) and Probabilistic (Naïve Bayes) showed how useful the Elo is at predicting chess results, at least on the dataset considered, improving classifiers’ performance with respect to accuracy, precision, recall and area under curve, among others.