Representation of Knowledge for Chess Endgames Towards a Self-Improving System

This thesis describes an investigation of the problems involved in representing knowledge within the task area of elementary Chess endgames. Two major criteria are taken for the choice of a model of & the chessplayer's knowledge : firstly, that algorithms constructed using the model should be natural from the viewpoint of a chessplayer and commensurate with his, view of the complexity of the task, and secondly that the algorithms should be capable of refinement in the light of experience in a manner which preserves the previous property. Elementary chess endgames are studied as a field in which programs based on tree-searching and traditional evaluation functions have achieved poor results and where tree-searching seems to play little or no part for people. It is therefore possible to examine problems of knowledge representation and program refinement largely independently of the tree-searching paradigm. A long term aim of the research is to develop a representation suitable as the basis for a fully automatic system of algorithm refinement, whilst maintaining the criteria given above. A model is proposed and algorithms are given for two endgames, King and Rook against King (KRK) and King and Pawn against King (KPK) using this model. It is argued that both algorithms are reasonably natural and compact representations and experiments in refining these algorithms are described in detail. In both cases, the process of refinement is shown to be a reasonably straightforward one (for people) and one which maintains the properties of naturalness and compactness. The possibility of automating this process is considered