Generalising Closed World Specialisation: A Chess End Game Application

The chess endgame domain provides a continuing challenge as a benchmark for the study of concept learning. The King and Rook against King (KRK) endgame is simple in chess terms but provides concept learning tasks which can be demanding, as evidenced in previous studies by a number of authors. For learning systems these tasks have highlighted problems of representation, such as the ability to express the structural relationships to be found in training examples , and other issues like correctness, compression and comprehensi-bility. Our focus is on Inductive Logic Programming (ILP) methods which are based on the generalisation and specialisation of normal logic programs. We are principally concerned with improving these methods to be able to handle more examples during learning. The main contribution consists of a new method of incremental learning, together with new results on a set of concept learning problems taken from the KRK endgame domain.