@MISC{McConnell_tuningevaluation, author = {Chris McConnell}, title = {Tuning Evaluation Functions for Search}, year = {} }
Share
OpenURL
Abstract
This paper examines the problem of applying machine learning techniques to improving the performance of actual game playing programs in complex domains like chess. This is a challenging problem because chess is a domain where a great deal of human effort has been spent and the performance level of programs is already very high. Current programs play chess by focusing on accumulating a material advantage that eventually becomes so crushing that a mate can be found. This advantage is accumulated through the application of tactics--a specific sequence of moves leading to a material gain. One of the current weaknesses in computer programs is that they have a weak sense of what to do when there are no tactical plans that can be found through search. What humans do in these positions is to play moves that provide strategic opportunities that eventually accumulate to the point that they can be converted to a tactical plan. By the time the computer sees the tactical plan it is too late.