Deep Thought for winning chess.

Deep Thought for winning chess

The computer Deep Thought has earned the highest chess rating yet achieved by a machine, putting it in the top ranks of all chess players. Built and programmed by a team of graduate students at Carnegie Mellon University in Pittsburgh, Deep Thought last month also tied for first place in a major chess tournament featuring some of the top human chess players in the United States.

Deep Thought is a direct descendant of Chiptest, an experimental machine that last year won the North American computer chess championship (SN: 11/21/87, p. 335). Developed by Feng-Hsiung Hsu and Thomas Anantharaman and their collaborators, Deep Thought consists of two Chiptest processors. It incorporates a new searching algorithm known as "singular extension," which allows the machine to probe deeper along promising tracks rather than stay with a general search. By replaying completed games backwards the machine uses hindsight to learn from its mistakes. Nevertheless, it has trouble generalizing its new knowledge to situations that are similar but not identical to chess positions already encountered.

In the world of chess, Deep Thought has now earned a rating of roughly 2545, putting it at the grandmaster level. The world champion has a rating of more than 2700. By maintaining a rating of more than 2500 for 25 consecutive games, Deep Thought qualifies for the $10,000 Fredkin intermediate prize awarded to the first machine achieving a grandmaster rating.

Deep Thought's principal computer rival is another Carnegie Mellon chess machine named Hitech, developed by computer scientist and chess expert Hans Berliner and graduate student Carl Ebeling. Hitech, which has a rating just over 2400, doesn't search as deeply or as intelligently as Deep Thought, but tries to make up for this drawback by incorporating more chess knowledge to guide its play. Hitech is particularly good at recognizing special chess patterns.

In the only meeting between Deep Thought and Hitech, during this year's North American computer chess championship, Deep Thought won and went on to capture the title. Hitech lost two games in that tournament. The computer happened to encounter chess positions it didn't really understand, says Murray Campbell, who has worked on both Hitech and Deep Thought. "It knew it was winning, but it didn't know how to win and ended up losing."

Hitech, however, played superbly last September in an exhibition match against grandmaster and former U.S. champion Arnold Denker. Winning three games and drawing one, Hitech became the first chess computer to beat a human player ranked as high as a grandmaster. In last month's tournament, Deep Thought alos defeated one player at the grandmaster level but lost to another -- its only loss in the tournament.

Hsu and his group now plan to put together an eight-processor version of Deep Thought, further increasing the chess computer's speed and capabilities. However, even in its present form, Deep Thought gets the respect of its human opponents, who sometimes term its play surprisingly creative.

"The seemingly creative behavior of this computer leads one to speculate whether or not there are other human endeavors in which creativity could be simulated by a clever, fast search," comments Carnegie Mellon computer scientist Daniel D. Sleator.

In their efforts to build and program chess-playing computers, researchers are also learning new things about the game itself. Systematic studies of how to end chess games when both players have only two or three pieces on the board have already revealed many flaws in conventional wisdom. Chess computers running through published games are also finding errors in many books describing various game openings.