IBM's Deep Blue
It was a classic match of man versus machine: In February 1996, world champion chess player Gary Kasparov pitted his wits against Deep Blue, a computer designed by a team of computer scientists from IBM. Deep Blue took the first game, but Kasparov recovered and ended up winning the match 4 games to 2. Three weeks later, John Horgan of Scientific American interviewed the Deep Blue team at the IBM Thomas J. Watson Research Center in Yorktown Heights, New York.
Present during the interview were Chung-Jen Tan, manager of the Deep Blue team; Feng-hsiung Hsu, who designed Deep Blue's predecessor, Deep Thought, when he was a graduate student at Carnegie Mellon University in the mid-1980s; Murray Campbell, who also worked on Deep Thought at Carnegie Mellon and is the IBM team's best chess player; Joseph Hoane, a programmer; and Marcy Holle, an IBM public relations manager. Absent were Jerry Brody, another programmer who had been delayed by an ice storm; and Deep Blue itself, a brace of refrigerator-sized IBM SP 2 computers, each containing 16 parallel processors, which remained housed in its room elsewhere in the building.
A slightly edited transcript of the interview follows.
SCIENTIFIC AMERICAN: Some news reports suggested that Deep Blue might have lost the match because of human error, such as your tinkering with Deep Blue's program between match one, which Deep Blue won, and match two, which the computer lost.
Campbell: "That had no influence... there were no errors of any consequence in the second game."
Hoane: "Yeah, the only human error was the draw in game five." [Early in game five, Kasparov offered Deep Blue a draw. The computer's managers rejected the offer, and Kasparov went on to win the match.]
Campbell: "But that's not a human error. The computer doesn't ever accept draws. If we want to, we can accept a draw, but it will never accept a draw. It's in the rules."
Tan: "Offering of draws generally happens when things are very close, unless it's a repetition of some sort. Otherwise the computer will just keep going. [Kasparov] offered a draw because he saw he had no chance of winning and he wanted to do better next game... We just made a decision."
Campbell: "If we'd won everybody would have said we were brilliant."
SCIENTIFIC AMERICAN: So you are not second-guessing yourself about your strategy during the match with Kasparov?
Campbell: "At the time we did what we thought was the right thing to do... "
Tan: "It was really early in the game, move 23, and there was no indication one way or the other about which side had the advantage."
Hoane: "You've got to understand how tough it is for us to get a good game out of a grandmaster. They don't give you their best unless there's money on the line."
Campbell: "When you're playing against the world champion, and you're not losing, you might as well play it out."
SCIENTIFIC AMERICAN: Was Kasparov finding weaknesses in Deep Blue over the course of the match?
Tan: "Not only weaknesses. He found how strong the machine was, therefore he changed the style of play and the strategy of play. He said the first game he lost was the best thing that ever happened to him, because otherwise [his initial strategy] would have come back to haunt him in later games."
Campbell: "He might have lost at a critical time later on instead of in the beginning when he could recover."
SCIENTIFIC AMERICAN: So if the match had continued, would Kasparov have kept getting winning more and more easily?
Tan: "It's hard to tell, because he could get more tired and nervous... and [Deep Blue] will stay very strong. And he may get a couple of games where, if he loses to Deep Blue, he would feel tremendous pressure on him."
Hoane: "Plus he may have learned everything he could learn [in the first six games]. It's not clear."
Campbell: "It's also possible that we could change the program too [to alter its strategy between games]. We were very conservative about changing the program, because in such a short time when you make a change you can't really test it thoroughly. But in theory, if we see the same thing happening over and over, we can change it. We can patch holes. We didn't do that, but we are allowed to."
SCIENTIFIC AMERICAN: Did you think you would win the match?
Campbell: "Some people were predicting Kasparov would win six-zero, and some people were predicting we would win by a large margin. I had strong feelings going into the match that we would be as good as Kasparov. The problem going into the match is, the computer and the world champion just have such different strengths, different weaknesses. It's hard to see how this is going to match up. The computer has the power of calculating..., being 'like God,' in certain positions, to use Kasparov's words. And Kasparov has all this intuition and understanding of chess and all the things that he's good at. It wasn't clear how it was going to come together, what strengths would play against what weaknesses. And the answer is, it happened many different ways, and it looks like a respectable, perhaps equal challenge we gave him."
SCIENTIFIC AMERICAN: In an article published in Scientific American in October 1990, Hsu, Campbell and two colleagues that are no longer involved in the chess project predicted that the computer would have a rating of 3,200, much higher than Kasparov's rating of 2,800, by 1992. Why hasn't that happened yet?
Tan: "The truth of the matter is, our project did take longer than expected."
Campbell: "Also, we never predicted a 3,200 rating for ourselves. There's a graph in the Scientific Americanarticle which a lot of people misinterpreted... There were some very careful words, we thought, qualifying that [prediction]. 'If present trends continue, and assuming this and assuming that.'"
Tan: "We went into this match with the goal of playing somewhere like 2,800... And I think we did pretty well, as far as getting to that level... We didn't win the match, but it was a very close encounter, especially the first game. If we patched up some potholes, we could have easily turned the match around."
SCIENTIFIC AMERICAN: Does the computer always make the same move when faced with a given position?
Hsu: "There is some randomization by the fact we are doing a parallel search... Sometime the same position, the same program, the same hardware, everything the same, start it again, it might behave differently."
Campbell: "Minor changes in initial conditions produce large changes in outcome."
[MARCY HOLLE, the IBM public relations spokesperson, suggests that the Deep Blue team explain why Deep Blue made certain moves in its first game against Kasparov.]
Campbell: "It's hard to explain what it did, because it is a 20 billion move search, or something like that."
Hsu: "And we can only see... one line of moves."
Tan: "... It's hard to even just record and look at 20 billion board positions. So the computer just prints out the principal line of its search. The machine is very accurate [in its analysis of certain situations], because it looks at all possible moves up to 12 ply and others down further. So when Gary was trying to checkmate Deep Blue at the end of game one, the computer just saw every move and was one step ahead of him [and could ignore the threat]... A human player... would have been scared."
Hoane: "The machine is always blithe."
Campbell: "You can't bluff it."
Hsu: "You can say that Kasparov lost the first game, maybe, because of lack of experience in playing with a machine. And we lost the last two games because of lack of experience..."
SCIENTIFIC AMERICAN: Will Deep Blue and Kasparov play again anytime soon?
Holle: "We're negotiating that with him at the moment. He wants it, he asked for it."
Tan: "He got more exposure out of the match than any other match."
SCIENTIFIC AMERICAN: Why has it taken so long for a computer to surpass all humans in chess, which seems like such a perfect game for a computer to play?
CAMPBELL: "The reason chess is interesting is, it's not perfect for a human and it's not perfect for a computer. It's somewhere in the middle. Games like checkers... at this point have been mastered by a machine, [whereas in the game of Go] people are so far ahead it's just not interesting. But chess is an interesting game because there are elements of calculation, there are elements of intuition and pattern recognition, all the traditional AI [artificial intelligence] stuff, and all sort of clashing. It hasn't been resolved one way or the other yet."
SCIENTIFIC AMERICAN: Does the difficulty that you have had in beating human players in chess suggest something about the limits of artificial intelligence (AI)? Aren't other problems in AI, such as speech recognition, much harder than chess?
Tan: "This chess project is not AI. In the early days people used chess to demonstrate AI techniques, but in the late 1970s and early '80s people started to realize that [if you reduce chess to a problem in pure computation rather than one of judgment and intuition], then you can really get much better results from a computer. Before you understand the problem, you call it AI, but once you really understand it you can reduce that to a computation algorithm."
SCIENTIFIC AMERICAN: If chess can be reduced to computation, why has it been so hard to beat a human like Kasparov?
Campbell: "Because there are some things that people can do that we haven't been able to reduce to computation."
Tan: "And it's also taken this long for the computer technology to give us the computational power to do that. Without the architecture, without the parallel processor we developed, you cannot do it."
Campbell: "The amazing thing is that it requires that much computation to get close to beating the best people... Nobody would have guessed in the mid-1950s that it would take 100 million positions per second to play close to the level of the world champion."
Hoane: "The techniques that tried to mimic human judgment failed miserably. We still don't know how to do that at all."
SCIENTIFIC AMERICAN: How long will the Deep Blue project continue?
Tan: "This is really part of the overall research to understand how to use parallel processing's computational capability to solve complex problems. We have many activities going on, and chess is one of them. When we get to a point where we think we understand enough from chess to derive benefit from it for improving our understanding of parallel processing, we will stop. Gradually, when chess is no longer interesting, we will completely stop and move onto other areas. But it will be a smooth transition. It will not be something that stops tomorrow."
Campbell: "When we say, 'Chess is no longer interesting,' we mean, 'Chess is no longer interesting from a computer-science research point of view.' Obviously we have no intention of making chess uninteresting for chess players!"
SCIENTIFIC AMERICAN: How do you feel about the press portraying Deep Blue as the modern equivalent of the machine that took over the job of John Henry, the legendary railroad worker?
Hoane: "A lot of miners don't have black lung disease because machines do the work for them."
Hsu: "... This is not just a simple tool like the steam engine... This is a big challenge from a computer science point of view. When Gary won the match, everybody applauded, and we applauded with a different perspective, because we knew what Gary went through..."
Hoane: "I got one little gee whiz number. How long has [Kasparov] been playing chess? Twenty-eight years or so. And how long have we been playing chess [in total]? About 25 years... The human effort that went into this event was the same, in a sense..."
Tan: "It is really just one person using a brain versus many people using a tool."
SCIENTIFIC AMERICAN: When will Deep Blue beat Kasparov?
Tan: "We already did once!"
Hsu: "There's always an element of chance, In fact, even in this match we had some good chance of winning....It could happen next time even with the same machine..."
SCIENTIFIC AMERICAN: Is it just a matter of time before Deep Blue defeats Kasparov easily? [Everyone nods]
Tan: "In this match we used a 32-node parallel processor... We have in this building a 128-node parallel processor. We did not use that one, because it is very expensive... [We would have] to stop all other research in this building. That's one thing. Secondly, our goal is just to get to the level where we can play Gary... We did that. If we really wanted 100-percent certainty that we will beat him in a match, we would have used 128 or 512[-node] parallel system to get extra insurance. But we didn't do that. So the technology is here today, in terms of computational strength."
Campbell: "If we hired ten grandmasters for two years to help us and we got the biggest machines and spent hundreds of millions of dollars, I think we could have done it. But what we're learning just as much from what we're doing."
Holle: "Chess is the perfect model for taking parallel processing and understanding it, for taking it to various applications. That's what this is all about. Because, this is IBM."
Tan: "That's why we didn't go for the big machine. That wasn't our goal."
SCIENTIFIC AMERICAN: Have any of you beaten Deep Blue?
Hoane: "I saw Campbell beat Deep Thought [Deep Blue's predecessor] once, in our entire career. And I don't need to be beaten by Deep Blue. Any cheap commercial program can thrash me."
SCIENTIFIC AMERICAN: Do any of you have any strong thoughts on whether computers can think?
Campbell: "I don't think chess playing has any bearing on whether computers can think. Just because a computer can play chess doesn't mean it can think. Thinking is a very difficult word."
SCIENTIFIC AMERICAN: Do you think there are any upper bounds to what computers can do?
Hoane: "Computers keep doubling [in speed] every 18 months."
Campbell: "But you can't go below the size of an electron [when constructing data-storage devices, and] an electron can't go faster than the speed of light."
Hoane: "Yeah, but 20 years ago I could have made up some limit that you can't pack too many devices in a certain space..."
Campbell: "It's obvious to me that exponential growth won't continue for the next 10,000 years." [Everybody laughs.]
Tan: "You can talk about the so-called upper limits, the electron and so forth... But advances will be made in other areas, such as software... It's hard to say where the limit will be."
SCIENTIFIC AMERICAN: Will computers someday be able to emulate every human attribute?
Campbell: "Of course, in principle. You just emulate [the brain] neuron for neuron, device for device, but that's like centuries in the future."
Tan: "It's more than that. Yeah, you can substitute electrons for neurons, but the brain is more than hardware. It's all the software and everything else. I'm not a psychologist or a neurologist, but I'm sure they don't understand those problems either."