The soul of a chess machine; lessons learned from a contest pitting man against computer - Deep Blue loses to chess champion Garry Kasparov

Science News, March 30, 1996 by Ivars Peterson

Deep Blue showed obvious weaknesses in its ability to evaluate certain types of chess positions, such as not recognizing when pieces needed to be sacrificed. Such deficiencies can be easily corrected by adding more knowledge to the program, Marsland says.

But there is a tradeoff. Complicated evaluations slow down the searches, so a balance must be struck between depth of search and complexity of evaluation. So far, depth of search has proved more significant than sophistication of positional analysis in th e success of high-level chess computers.

In recent years, however, programmers have made great strides in creating surprisingly competent chess programs that run on personal computers. They have done it by carefully refining and tuning the chess knowledge component to make up for the smaller com puters' lack of computing power compared to machines like Deep Blue.

Programs such as Chess Genius and Fritz 4 have shown the way. "I've played some of the micros," Berliner says. "It's amazing how well versed they are in almost all phases of the game.

"The best way to improve the evaluation [by the computer] is to keep playing-make some changes and then play the new program against the old one to see what happens," he advises. "That's what the people with the micros have been doing."

Some researchers are investigating alternative ways of supplying chess knowledge to a computer. One possibility is to see if they can program computers to learn, just as human players improve their play with experience and study. A few years ago, Robert A . Levinson and his coworkers at the University of California, Santa Cruz developed a computer program, called Morph, that learned to play chess starting only with a list of legal moves. They pitted their novice system against a conventional chess program known as Gnu Chess, which plays about as well as the average tournament player.

After thousands of such games, Morph identified enough patterns to play a reasonable game against a beginning tournament player, even though it looked ahead only to the next move. "It's not really impressive compared to existing chess programs," Levinson says. "But it is impressive given that it was all learned from experience."

Levinson is now working on a new, improved version of Morph. The program is capable of looking ahead several moves and has access to a database of essentially all the games ever played by top chess players.

"It finds the chess position it considers most similar to its own position and tries to reason by analogy," Levinson says. "If that position was good, then this position is good.

"I think we have a promising model," he adds. "But there's something about a grand master staring at a chessboard that's hard to capture in a computer."

Kasparov's key advantage over Deep Blue was that he could learn, both as a game progressed and between games.

Because Deep Blue had no track record as a chess player, Kasparov could not prepare for this match as he has for other matches by studying his opponent's previously played games. Instead, he built up in his mind a portrait of his computer opponent as they played.