Online Prediction of Chess Match Result

Abstract

In this work we propose a framework for predicting chess match outcome while the game is in progress. We make this prediction by examining the moves made by the players. For this purpose, we propose a novel ensemble based learning technique where a profile-based segmentation is done on the training dataset, and one classifier is trained from each such segment. Then the ensemble of classifiers is used to predict the outcome of new chess matches. When a new game is being played this ensemble model is used to dynamically predict the probabilities of white winning, black winning, and drawing after every move. We have evaluated our system with different base learning techniques as well as with different types of features and applied our technique on a large corpus of real chess matches, achieving higher prediction accuracies than traditional classification techniques. We have achieved prediction accuracies close to 66% and most of the correct predictions were made with nine or more moves before the game ended. We believe that this work will motivate the development of online prediction systems for other games, such as other board games and even some field games.