Learning in Bebe
Most chess programs are deterministic. They will always play the same move for a given position, and so may experience problems in tournaments when an opponent follows a line of play that produced a loss for the chess program in an earlier round. In 1982 we noticed this phenomenon when Bebe lost the same game repeatedly during extended (8 hours) speed chess sessions against humans. Bebe has depth controlled access to a transposition table (T-table) (Slate and Atkin 1977), but during a speed game it would only use about 25% of the table, because the depth limit was 5 plies. Thereafter the T-table was left intact between games, hoping to startle opponents when Bebe played moves different from the previous game. The psychological effect on the opponent of a computer that appeared to be learning was often devastating enough to give Bebe a victory. The effect was short-lived because after 5 games little remained in the T-table from the first game. We did not explore this idea further until 1986, when the danger of losing the same game twice during the 5th World Computer Championship was large enough to warrant action.
KeywordsRoot Node Game Tree Principal Variation Original Game Hash Code
Unable to display preview. Download preview PDF.