Automatic Learning of Evaluation, with Applications to Computer Chess

A new and fast learning method is described in the context of teaching a program to play chess. A theory of the meaning of a position evaluation is developed, and is then confronted with a large collection of games played by masters or other programs. The program learns by fitting its evaluation to better predict the results of the games. The method has been employed by a top-rated program for the past 10 years, and has earned several world championships and successful matches against the world's best grandmasters for the program. The effectiveness of the method is demonstrated by showing its successful prediction of known playing strength of the programs.