Knowledge discovery in endgame databases

Abstract

In many application domains we have seen an explosive growth in the capabilities to both generate and collect data. Representative examples are business, medical, and scientific databases. In the game of chess (and similar games), we have a similar situation. Moreover, in chess we have gigantic databases with perfect information available. An endgame database is a very rare case of an information source with complete knowledge. However, this information, although complete, is not in a form which is particularly useful to human beings. Therefore, the raw data inside an endgame database have to be transformed into knowledge in understandable form. This task concerning chess endgame databases is, in principle, the same as in KDD in general.

This paper summarizes a few results from literature where knowledge (especially patterns) has been discovered for simple chess endgames — in most cases this has been done empirically, by intuition, and interacting with the computer, but not automatically. Up to now, there are almost no satisfying results for practical endgames to extract knowledge in an automatic way.

The main contribution of this paper is to present a new chess endgame database from which we have extracted knowledge automatically using decision trees — a method well-known from Machine Learning. Additionally, a mechanism for automatic feature discovery is introduced.