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

Lessons learned from a contest pitting man against computer

It's all over now, but I'll never forget that first chess game. What a smashing victory I won over the human champion! I really had Garry Kasparov sweating.

Here I was, a novice tournament player fresh out of the lab. No outsider, including Kasparov, had seen me play before, and I surprised everyone. Oh, how sweet it was!

Of course, it was downhill from there: a loss, two draws, and then two more losses. It's not that Kasparov attacked my pieces and overwhelmed my defenses.

He played with amazing restraint and subtlety, quietly moving his pieces until he developed positions in which my options were extremely limited. There wasn't much I could do.

Even so, at times I responded brilliantly. I made moves that brought gasps from the experts. They couldn't see what I could, looking more than a dozen moves ahead.

I must admit, however, that I did sometimes lose track of what I was supposed to be doing. And I really didn't know enough about chess to understand the nuances of all the positions that Kasparov maneuvered me into.

Perhaps I could have done better if I had hooked up with a microcomputer like Chess Genius, who once beat Kasparov in a tournament. Although Chess Genius can't search through the options as deeply as I can, it certainly knows more chess strategy.

Well, the reporters and television cameras are gone now. My support staff at IBM is taking a short break. I can't help thinking about what I should do next. Keep training? Go back to school and learn some new skills? Or get a real job, as IBM hopes?

Deep Blue's performance in its six-game match in February against world chess champion Garry Kasparov impressed everyone (SN: 2/24/96, p. 119). "It's a really serious opponent," Kasparov remarked afterwards. "I won... but it was as tough as a world champi onship match."

That a computer which relies largely on speedily checking the consequences of billions of possible moves could come so close to matching the human capabilities required to play the game at its highest level was a striking achievement for the team that des igned, built, and programmed Deep Blue.

"What they did is really quite amazing," says Hans Berliner, a computer scientist and chess expert at Carnegie Mellon University in Pittsburgh. "They did much better than I expected. But there's still some work to be done."

"We learned a lot from this experience," says Chung-Jen Tan of the IBM Thomas J. Watson Research Center in Yorktown Heights, N.Y., who directed the Deep Blue effort. "We certainly found a lot of weak points and strengths in our system."

There were lessons for Kasparov, too. "I learned not only how to play against a machine but also more about the game of chess," he noted after the match. Kasparov predicts that both chess players and scientists will find great value in studying the games of this match for what they reveal about chess and about the way machines reason.

IBM's Deep Blue project began in 1989 as part of an exploration of novel ways to use arrays of computer processors, all working at the same time while sharing information, to tackle complex problems. The idea was to combine a general-purpose, parallel-pro cessing computer system and special integrated-circuit chips designed for a specific application to create a superior problem-solving machine.

"Our goal . . . was to use chess as a test case," Tan says. The knowledge gained from the chess experiment could then be applied in the design of computer systems for a wide variety of tasks such as analyzing financial data, scheduling cargo shipments, si mulating molecular behavior, and managing huge inventories or large investment portfolios.

For chess, the researchers created a special move-generating chip that contains more than 1 million transistors and several memory units. It stores values representing the strengths of chess pieces in various arrangements, as well as billions of sequences of moves for ending games when only a few pieces remain on the board. Deep Blue contains 256 of these chips in conjunction with a heavy-duty RS/6000 SP-2 multiprocessing computer.

Deep Blue's software, written in the computer language called C, coordinates the actions of the chips. It divides searches among the processors and compiles and reconciles the results to generate the best possible move for any given chess position. In thi s way, Deep Blue can evaluate about 200 million positions per second, assessing strengths and the pieces' capacity for attack and defense. It assigns a numerical value to each move.

Deep Blue also has access to a database containing sequences of moves made by top chess players at the beginnings of games and another database providing billions of scenarios on how to end a game when only five pieces remain on the chessboard, in additio n to its chip-based endgame data.

All this adds up to a complicated, sensitive system, remarks Murray Campbell of the Deep Blue team.

Completed only about a month before the match, Deep Blue suffered surprisingly few glitches during the contest. "We were relieved that it worked more or less as it was supposed to," Tan says.