Workshop Report: Theory and Practice in Computer Chess

Abstract

The moderator opened the workshop with a summary of the good and bad points in a simple computer chess model. At present most chess programs use a progressively deepening full-width search to some fixed depth, say N-ply, following it with a selective search for the next K-ply to reach a horizon. From the horizon nodes, a narrow quiescence search to arbitrary depth assesses tactically threatening captures. The quality of the quiescence search ultimately controls the calibre of play, so in that phase some programs also consider checking and pass pawn moves. Unfortunately quiescence searches are expensive, and yet studies of their failure always show that even more moves should have been included, for example, those that threaten check mate[5]. As an alternative, some programs use an exchange analyzer at the horizon. These two methods can also work effectively in conjunction, since an exchange analysis can identify those cases when a quiescence search is vital. Some programs, conventionally using brute force, have no selective layers, and a few have no initial full-width layer. The latter, the truly selective search programs, are now uncommon, perhaps because in the past they have been less successful. Since selective search is one of the few ways to limit the exponential growth of the tree, from time to time it enjoys a revival in popularity. But since there is always some non-zero probability that the correct move in a key position is not examined deeply enough, these programs tend to make at least one critical error in the course of every game. One purpose of the workshop is to determine from the participants their experience with this model, keeping in mind that perhaps a major aim in the field of computer chess is to develop variable-depth search algorithms. Another aim is to assess the effectiveness of various memory tables. All competitive programs incorporate a number of memory functions or tables to help them speed their search. Most popular is the transposition table which allows search reduction by not examining again positions that have occurred before. Another function is the refutation table[1], which is used to guide an iteratively deepening search. Other tables account for the values of pawn structures, king safety, and the history heuristic identifies frequently successful moves. From transposition tables, the best available move entry is used frequently, since taking it often saves an expensive move generation[14]. Transposition tables are fundamentally sound and are invaluable in endgames, where they allow the effective depth of search to be extended, perhaps to as much as twice the horizon distance. They are