May 13, 1997

Kasparov, Unbowed, Looks to Third Bout With Deep Blue


NEW YORK -- What would you call it? A no-brainer?

Today in CyberTimes

Articles and Columns

Activism: Nonprofit Has Inner City 'Plugged In'
By Matt Richtel

Kasparov, Unbowed, Looks to Third Bout With Deep Blue
By Bruce Weber

Flawed Strategy vs. Fierce Resistance
By Robert Byrne

Apple Sets Out to Sell Developers on New Operating System
By Lawrence Fisher

Searching the Net, Continental Style
Eurobytes /
By Bruno Giussani

Brilliant Scanning, Dumb Design
Personal Computing / By Stephen Manes

Today's Section Front

Seven-Day Index

CyberTimes Forums

CyberTimes Navigator

On move 35 of Game 2 in its six-game match with the world chess champion, Garry Kasparov, last week, the IBM computer Deep Blue took a knight with its bishop. It was expecting, along with just about every grandmaster on the planet, that after Kasparov took the computer's bishop in return, it would slide its queen into a threatening position deep into Kasparov's side of the board. The move Qb6, queen to square b6, was all but assured.

Not quite. Looking way out into the hypothetical future, Deep Blue noticed that Qb6 was not the path to victory, but the road to oblivion. Indeed, as it looked ahead, ply by ply (a ply is computer chess jargon for a move by one player), Qb6 looked worse and worse. So it decided to search for an alternative.

Using almost two minutes of what its programmers call "panic time," the machine came up with a pawn exchange it liked better.

Thus began the sequence of moves that flummoxed Kasparov enough that he blamed it for upsetting his concentration in the four games to follow, leading to his historic defeat. Ever since Game 2, which was won by Deep Blue, he had been asking for a printout of the computer log of the vital sequence, so he could see for himself how the computer made its extraordinary choice, and perhaps to palliate his suspicion that a human had intervened.

On Monday, the scientists who programmed Deep Blue finally gave him what he was asking for. But the champion's side remained unsatisfied.

"The log tells you, 'O.K, I don't like this line, I like a better line,"' said Frederick Friedel, Kasparov's computer adviser. "But why does it like that line better? That it doesn't tell you, and of course that's the crucial question."

Kasparov, who resigned the final game of the match in a losing position on Sunday, saying he was ashamed of his play, declined to be interviewed on Monday. But Friedel said Kasparov intended, later this week, to challenge Deep Blue to a third match -- Kasparov won the first contest against a less sophisticated and less powerful version of the computer in February 1996 in Philadelphia -- to be played in late summer or fall.

He talked of that possibility on Sunday, suggesting a 10-game match spread over 20 days. Friedel also said that in such a rematch, Kasparov would not be considered a partner with IBM, furthering the cause of computer science. Instead, he said, "Garry would come as a hostile opponent."

Friedel said that the match would have to be sponsored by an organization other than IBM. Whatever the prize fund, Friedel said, Kasparov would suggest that the winner take all. In the match that ended Sunday, Kasparov got $400,000, with $700,000 to the IBM side, to be put back into research.

"He wants his revenge," Friedel said. "Yesterday evening, it was, 'I can't be bothered with this.' But this morning, it was, 'O.K., I want to get them."'

Kasparov's defeat has catalyzed the ornery and competitive chess world, and offers from grandmasters (and others) to play against Deep Blue and salvage the pride of humanity have been coming out of the woodwork. Susan Polgar, the women's world champion, issued a challenge immediately after the match, saying she would stake her "woman's intuition" against Deep Blue. On Monday, speaking from Moscow, Kasparov's longtime adversary, Anatoly Karpov, said: "I'm ready. I would like to have the chance to uphold the honor of mankind's chess players."

Karpov, who is the current champion of the International Chess Federation and is generally regarded as Kasparov's closest rival, said he would stand a better chance against Deep Blue because his strength is the positional style of play that is most effective against a computer, as opposed to the aggressive and attacking style that Kasparov favors. Karpov said Kasparov altered his style of play, but he would not have to.

"In this match you saw a different Kasparov than you see in tournaments," Karpov said. "Here I could see on his face during TV reports that he was not confident. When you are not confident, even when you play against human beings, this is dangerous. Against a computer, it is extremely dangerous."

Karpov agreed that in Game 2, the computer's choice on move 36 was a surprise. But he said Kasparov's fatal mistake had preceded it. That occurred on move 34, when Kasparov moved a pawn to square f6.

"Kasparov made a terrible move, f6," Karpov said. "After f6 he shouldn't have complained. After f6, Black is dead."

For his part, Kasparov has said that Game 2 unsettled him, because for most of it, the computer played brilliantly, better than at any other time in the match, but that at the end it blundered, allowing for the possibility of a draw. And though he should have seen the draw, he did not; he explained that he presumed the computer would have obviated it.

That is sour grapes, perhaps. But Kasparov's point was the computer's strikingly inconsistent play. It is human to be unpredictable. But how could a machine be so incisive and prescient one minute, so dunderheaded and blind the next?

"It was a very good move," said C.J. Tan, the manager of IBM's Deep Blue project, referring to the unexpected pawn exchange on move 36. "He just doesn't want to admit it could play so well. That's my interpretation."

Murray Campbell, an IBM research scientist, added: "The machine didn't search deeply enough to see the draw, which was at least 20 ply ahead. He's assuming it's playing perfect chess. But if it was perfect, why would we be playing at all?"

It actually works the other way around. "There's no such thing as the perfect move unless you have infinite time," Campbell said. "The computer presumes its opponent sees everything it sees. It tries to play the best move, and assumes its opponent will too."

The two men were sitting in a midtown Manhattan hotel, at a table piled high with the controversial printouts -- 441 pages of tightly bunched notation, gibberish to the untrained eye. In an hourlong interpretation on Monday, they explained what had happened in the time between move 35, when Deep Blue took the knight with its bishop, and move 36, when it exchanged pawns instead of positioning its queen for an attack.

First, as is its wont, the computer guessed at Kasparov's move. It guesses correctly about half the time, and it did so in this case. By the time Kasparov made the expected move -- it took him about 20 seconds -- the computer had looked 9 ply ahead at the move it, and most human players, favored: Qb6. It had established that by playing those 9 moves after playing Qb6, it would be ahead by the numerical factor 79. (On this scale, a pawn is worth 100 points, so Deep Blue felt it would be ahead, but by less than the value of the weakest man on the board.)

After Kasparov moved, Deep Blue continued its search. At 10 ply, the move had been devalued to 74. In the midst of its search of possibilities at 11 ply, the alarm went off, and it clicked into the mode known as "panic time."

"This happens when the move has gone bad, but you don't know why," Campbell said. The computer allotted itself 651 seconds to evaluate the 11th ply, and 100 seconds into that allotment of time, the value of the move had dropped to 48. By this time it had already been searching for more than five and a half minutes, and it kicked into another mode, "extra time."

"That happens when it's unhappy, if I can use that word, with the move it has," Campbell said, "and it has decided it wants to find something else."

After 71 seconds it found the pawn exchange at square b5, which it scored, after 11 ply, at 63. That recovered enough of its advantage for the computer to make the move, given the amount of time it had expended.

"The further you search, the closer you get to 'truth,"' Campbell said, though he said it was possible that a move that looked good through a dozen ply might well turn sour several moves later. The converse is also true, he said.

"If it can't search deeply enough to verify what it thinks early on in the search, it might make the wrong move," Campbell said.

It also might make the right one, which is what Tan meant when, referring to the printout, he said: "Kasparov asked, 'How can a computer not play Qb6 in that position?' This is how."

Home | Sections | Contents | Search | Forums | Help

Copyright 1997 The New York Times Company