Predicting Problem Difficulty in Chess

We investigate the question of automatic prediction of task difficulty for humans, of problems that are typically solved through informed search. Our experimental domain is the game of chess. We analyse experimental data from human chess players solving tactical chess problems. The players also estimated the difficulty of these problems. We carried out an experiment with an approach to automatically estimate the difficulty of problems in this domain. The idea of this approach is to use the properties of a “meaningful search tree” to learn to estimate the difficulty of example problems. The construct of a meaningful search tree is an attempt at approximating problem solving by human experts. The learned difficulty classifier was applied to our experimental problems, and the resulting difficulty estimates matched well with the measured difficulties on the Chess Tempo website, and also with the average difficulty perceived by the players.