Affiliations: School of Computer Science, Carnegie-Mellon University
Note: [1] This article is an edited version of Chapter 7 of the author’s Ph.D. thesis, September 1992, Carnegie-Mellon University (report CMU-CS-92-192). The editing was mainly done in order to make the contribution self-contained.
Abstract: In computer game-playing research, the influence of characteristics pertaining to the opponent, rahter than to the position or the game in general, has been largely neglected. Instead, most research has focused on techniques for searching, evaluating, and solving positions or certain classes of positions, independently of any specific game-playing configuration. In this article we examine the question whether there are configurations in which optimal play, as defined by characteristics of the game only, may be inferior to speculative play, i.e., play that uses strategies which take the opponent into account. The simple 4-piece chess endgame of King and Queen vs. King and Rook (KQKR) is used as a testbed. Two different types of strategies, one optimal and one speculative, are compared in terms of their results against human players. The optimal strategy chooses only optimal moves, as taken from a KQKR database. The speculative strategy makes use of plausible human heuristics for KQKR to assess the difficulty of successor positions for the human opponent, and uses this difficulty in selecting its move. The test results reported in this article show that speculative play induces more and larger inaccuracies in the play of human opponents, and may lead to a better overall performance. This confirms the belief that, even in games as simple as KQKR, knowledge about some characteristics of the opponent may be used advantageously.
DOI: 10.3233/ICG-1993-16102
Journal: ICGA Journal, vol. 16, no. 1, pp. 3-17, 1993
