Darkthought Goes Deep

Article type: Research Article

Authors: Heinz, E.A.^{; 1}

Affiliations: Karlsruhe, F.R. Germany

Note: [1] Institute for Program Structures and Data Organization (IPD), School of Computer Science, University of Karlsruhe, P.O. Box 6980, D-76128 Karlsruhe, F.R. Germany. Email: [email protected]. http://wwwipd.ira.uka.de/~heinze/.

Abstract: This contribution follows up on the work by Hyatt and Newborn (1997) who let the program CRAFTY “go deep”. We repeated their experiment with our own chess program DARKTHOUGHT and obtained similar experimental results. Both experiments provide strong empirical evidence for the surprising observation that even at high search depths of 11 to 14 plies modern chess programs steadily discover new best moves in still 16 percent of all searches on average. Moreover, the experiments do not reveal any conclusive trend towards fewer new best moves at search depths beyond 14 plies. Hence, the available experimental results do not really fuel the intuitive notion that such “changes of mind” taper off continuously with increasing search depths. If at all, only the behaviour of DARKTHOUGHT with a drop to 13.7 percent new best moves on average in iteration #14 hinted at decreasing “changes of mind” for search depths of 15 plies or more. Additionally gathered data about the 14-ply searches of CRAFTY and DARKTHOUGHT allowed us to study the behaviour of both programs in greater detail. This led to the astonishing finding that regardless of the actual search depth a sizeable 30 to 50 percent of all new best moves on average represented “fresh ideas” which the programs never deemed best before. The finding adds support to Newborn’s (1985) hypothesis about the playing strength of chess programs. Furthermore, the additional numbers educated us about continuing odd/even instabilities of modern chess programs. These instabilities decreased notably only at high search depths of 9 to 14 plies in positions with reduced material as they mostly occur in endgames and late middle games.

DOI: 10.3233/ICG-1998-21403

Journal: ICGA Journal, vol. 21, no. 4, pp. 228-244, 1998