Learning middle-game patterns in chess: A case study

Abstract

Despite the undisputed strength of today’s chess-playing programs, the fact that they have to evaluate millions, or even billions, of different positions per move is unsatisfactory. The amount of “computation” carried out by human players is smaller by orders of magnitudes because they employ specific patterns that help them narrow the search tree. Similar approachs hould in principle be feasible also in computer programs. To draw attenion to this issue, we report our experiments with a program that learns to classify chessboard positions that permit the well-known bishop sacrifice at h7. We discuss some problems pertaining to the collection of training examples, their representation, and pre-classification. Classification accuracies achieved with a decision-tree based classifier are encouraging.