A Moderately Successful Attempt to Train Chess Evaluation Functions of Different Strengths

In this paper, we report the results of experiments in which we trained four chess evaluation functions on games of weak chess players in four different rating groups, with the goal of reproducing computer players of that strength. Although the differences in playing strength between the players loosely correlates to the playing strength of the training data, this goal could not be achieved because the differences are much too small, and sometimes spurious. Nevertheless, the results are interesting because the learned functions exhibit a clear and systematic difference in some of the learned positional parameters (e.g., the importance of open ranks and lines), and show some unexpected but consistent differences to previous results (e.g., a much lower value for the queen). 0 A somewhat shorter version of this paper appeared in the Proceedings of the ICML-10 Workshop on Machine Learning and Games.