Review / Books:
A Match for All Seasons
By Christopher F. Chabris
27 December 2002
The Wall Street Journal
(Copyright (c) 2002, Dow Jones & Company, Inc.)
BEHIND DEEP BLUE
By Feng-Hsiung Hsu
(Princeton, 298 pages, $27.95)
RECENTLY the world chess champion contested a close match against a formidable opponent. After taking an early lead, he let his guard down and lost two games, one by making a horrendous blunder, the other by resigning in a position he could have drawn. The match ended with his opponent scoring better than expected, although not well enough to win.
All this happened two months ago in Bahrain, in the Persian Gulf, where grandmaster Vladimir Kramnik counted himself lucky to escape with a 4-4 tie. His opponent was a software package called Deep Fritz running on a Compaq computer. After the match, expert opinion held that, despite the result, the human player was still the stronger one.
The verdict was not so clear five years ago. An eerily similar scenario played out in May 1997, in New York, in a match between Garry Kasparov, the world chess champion, and IBM's Deep Blue supercomputer. It was easily the most widely publicized chess event since the world-championship match, 25 years before, between Bobby Fischer and Boris Spassky. In 1997 the championship was not on the line, but in the popular mind the battle was even more significant -- a case of "man vs. machine." Most of those following the match rooted for the man to win.
A detailed account of this match is the most dramatic episode of "Behind Deep Blue," Feng-Hsiung Hsu's engaging memoir of the 12 years during which he and his colleagues developed a series of chess-playing computers, culminating in Deep Blue. Mr. Hsu manages to make seemingly dry, technical material vivid and gripping, even for readers without a background in chess or computers. And his story is a fascinating study, of men as well as machines.
Virtually from the moment that he started thinking about computer chess, Mr. Hsu wanted to build the first computer to defeat the world champion, a goal set by the earliest researchers in artificial intelligence. Those pioneers believed that chess was an ideal way of measuring the progress made by computer science and cognitive psychology in penetrating "to the core of human intellectual endeavor," as three of them famously put it in 1958.
As a graduate student (and chess amateur) at Carnegie Mellon University in the mid-1980s, Mr. Hsu was skeptical of this high-minded approach. He wondered whether, by contrast, simply speeding up existing chess computers -- without trying to reach the core of anything -- would be enough. He tested this "brute force" thesis by designing a series of chess-specific microprocessor chips and gathering other students to help write the software that controlled them.
The computers created by Mr. Hsu and his fellow researchers broke several performance records in short order. After about three years of development, Deep Thought became the first computer to play as well as a grandmaster, defeating several in the process. Although it lost its first match against Mr. Kasparov in 1989, it earned a sort of moral victory by nearly drawing a game against former world champion Anatoly Karpov in 1990.
Until this point, Mr. Hsu and his colleagues had been free-spirited researchers dabbling in a chess project as time allowed, imposing their own goals and deadlines. Their success, however, caught the attention of IBM Research, which hired several of them to continue the effort, with the express goal of defeating Mr. Kasparov (or his successor).
As the plan to speed up the rechristened Deep Blue moved forward, Mr. Hsu realized that speed alone might not be enough to win. The chess processors needed to be "trained" to pay attention to the right features in a chess position and to weigh them appropriately. The final critical step was to get human grandmasters involved. Joel Benjamin, at one time the youngest chess master in U.S. history, became the computer's personal tutor, playing game after game against it, probing for weaknesses and suggesting ways that its knowledge could be tuned for better performance.
By February 1996 the first incarnation of Deep Blue was ready, though just barely, to take a shot at Mr. Kasparov. To the surprise of nearly everyone, it won the first game of a six-game match. But Mr. Kasparov kept his cool and came back to win convincingly in the end, 4-2.
What about 1997? Here the situation was reversed. Mr. Kasparov won the first game, in part by seeking blockaded positions in which strategic maneuvering -- the domain of human knowledge -- is more important than tactical skirmishing, the computer's forte. In the second game he tried the same approach but was outplayed from the start in a masterpiece that any grandmaster would have been proud to play -- until the last move. Later analysis revealed the penultimate position to be more complicated than Mr. Kasparov, or anyone else, realized. Too complicated even for Deep Blue, which blundered and would have allowed Mr. Kasparov to force a draw if he had chosen to do so. But Mr. Kasparov, perhaps demoralized, resigned the game.
With the score now tied, the next three games were all drawn, after long struggles, in an atmosphere made tense by Mr. Kasparov's allegations that IBM had cheated during the second game. He believed that Deep Blue's moves were simply too good to have been found by a mere machine. (I was present in the Deep Blue control room during the entire match and can report that there was no human intervention.)
After getting an agreement that logs of Deep Blue's calculations would be printed and sealed, the fight seemed to go out of Mr. Kasparov. He snapped like a twig in game six, losing in just 19 moves -- one of the worst performances of his career -- and losing the match. IBM reaped a publicity windfall with such a momentous victory and even featured Deep Blue in TV ads.
What is the moral of Mr. Hsu's story? Deep Blue won fairly, proving that a computer could in fact defeat the world chess champion in a match. But many regarded the 1997 result as a fluke -- a one-game victory margin is not definitive -- and saw the failure of both sides to agree on a rematch as a lost chance to measure the computer's ability and to discover whether humans could adapt to its unique style of play.
One thing is certain: Although the specific events of 1997 were unusual, their pattern was not unique, as October's Kramnik-Fritz match showed. When the best chess players face the strongest computer opponents, the foibles of human psychology are as important as the speed of machine calculation, and the likely result is still up in the air.
Mr. Chabris, a lecturer in the Department of Psychology at Harvard, holds the National Master title in chess.