One-on-one with Murry Campbell

Q: There's a historical route in the IBM in chess tournament, dating back to '78 or '79 that began with Deep Thought.

Murray Campbell: Yeah, Deep Thought played a short two-game match against Gary Kasparov in 1989, when we first came to IBM.

Q: OK. Let's talk about the historical perspectives -- not necessarily from Deep Thought, but from last year, and how this year compares to last year, relative to the technology. Technology has been enhanced for purposes of a rematch but also for purposes of commercial applications.

M.C.: All right. There are two main differences between Deep Blue this year and Deep Blue last year. First of all, we're going to be running on a faster system, we're using the latest version of the SP, which has the P2SC or Power Two Super Chip processors. That will give us, in the end, a factor of two speed-up over the system that played last year. And, in chess programs, speed is very important, the faster you are, the stronger you play.

SOUND BITE: WAV(564k) | Bamba(30k)

The second aspect of Deep Blue that is stronger this year is the chess knowledge. We spent a lot of time, several months, working with a grandmaster, Joel Benjamin, basically teaching Deep Blue more about chess. There's sometimes things that a grandmaster knows that it's sometimes difficult to put into a computer program. We are working hard to get to know as much about chess as possible.

Q: How many moves a second can Deep Blue calculate?

M.C.: Last year, Deep Blue was averaging about 100,000,000 chess positions per second. That is, it would examine and evaluate 100,000,000 chess positions every second. This year we're hoping for the factor of two, or 200,000,000 chess positions per second.

Q: Relative to Garry Kasparov?

M.C.: Well, most human chess players, including Gary Kasparov, are very limited in the number of positions they can examine. Maybe two positions per second, but they have this talent, this intuition, which they aren't very good at describing how it works. That is, they can look at a position and almost always know the right move to play, without having to do any searching through possibilities.

If we understood how a player like Kasparov can do this, then we could obviously make Deep Blue much stronger, but it's a mystery at this point how the human brain is able to play chess as well as it does.

SOUND BITE: WAV(614k) | Bamba(32k)

Q: Is that the only human edge? The intuition factor? Or are there other factors to take into account?

M.C.: Each player in this match will have certain advantages. In addition to the intuition and the strong evaluation capabilities that Kasparov has, he is able to learn and adapt very quickly. In the last match, Kasparov adjusted his play. He learned particularly from the first game in the match, which he lost.

He adjusted his style of play when he won the last two games of the match. He was playing a more conservative, more controlled style of play and that was giving Deep Blue a lot of trouble. So he had adapted his play during the match and we have to be able to deal with that. As it stands right now, Deep Blue is not a learning system.

And so, any changes that take place in the system have to be done by us, the programmers. So, we have to make the system as flexible as possible so that during the match if we see a problem come up, we are able to fix it.

Now, it's easy enough to fix, if a problem comes up, it's easy enough to fix that given problem. But you have to make sure that by fixing the one problem, you don't create a thousand other problems, in other positions. And that is why it is so difficult for us to make changes to Deep Blue during a match. We hope we'll do that a little better this time around.

Q: How long since the end of last year's match has the team been working on Deep Blue in terms of upgrading its speed, upgrading its knowledge?

M.C.: We've been working since the last match, apart from taking Deep Blue Jr., which is a smaller version of Deep Blue, to various places around the world and playing matches and exhibitions. But, other than that, we've been working continuously. The three scientists on the team, plus the grandmaster who joined us in August of '96. We've been working since then on improving Big Blue.

Q: What's your role?

M.C.: Well, there are two main jobs. The first one is working with the evaluation function. That's the part of Deep Blue that given a chess position, tells you how good the position is.

That's very important, because even if you can search many moves deep into a position, search forward all the possible moves, you still have to evaluate those positions at the end of the sequence of moves that you've looked at; and if you evaluate them incorrectly, you're going to play poor chess. And so, I'm working on helping Deep Blue evaluate the positions more correctly and understand more about chess.

And the second thing is, I'm working with the grandmaster a lot taking some of his suggestions and trying to incorporate them. And also, working with him to prepare the opening, also called the opening book. It's a repertoire of openings that Deep Blue will be playing in the upcoming match.

Q: You mean opening moves?

M.C.: Opening moves.

Q: Are you a chess player?

M.C.: When I was in high school I used to play. I used to play chess competitively, but I haven't played much since then.

Q: All this aside, when the last pawn is moved, what are the real world applications for this beyond the chess board?

M.C.: Well, we're taking some of the lessons we learned from building this system and applying it to other complex and difficult problems that require a tremendous amount of computational power. The computer that we've brought, Deep Blue, is capable of doing extraordinary amounts of computation in order to chose a good chess move. And, applying and creating a system for other problems would be the ultimate goal of a system like this.

We work with both a parallel computer like the SP and special-purpose hardware to accelerate particular aspects of a problem that can be put on a computer chip. And so, for example, we're looking at developing a system that can accelerate the molecular dynamics problem, that is the interaction between atoms and molecules, in order to predict the behavior of those molecules.

For example, the pharmaceutical companies would use this in synthesizing drugs, predicting the behavior of drugs, before they even have to go to the trouble of actually synthesizing them and testing them. They can predict some of their properties beforehand. And accelerating that process would be a very valuable addition to the technology of today.

SOUND BITE: WAV(1119k) | Bamba(58k)

Q: Following Deep Blue, and recognizing its commercial applications, are there people knocking at our doors saying, "Hey, can we get a piece of this?"

M.C.: Well, we've already begun. A group here at Yorktown has already begun developing the molecular dynamics system that I was talking about, and we're all looking at other applications as well: data mining, financial modeling. These may be applicable to this approach as well.

Q: You mentioned earlier that Kasparov was able to eventually win last year's match primarily because he was able to shift strategies mid-game. Do you anticipate him being able to do that again this year?

M.C.: He will certainly be able to do that this year. That's one of his great strengths, being able to evaluate the opponent, sense their weaknesses, and try to take advantage of those weaknesses. We believe this is going to happen, so we have to be able to adjust as best we can to his adaptability and try and counter it.

Q: Last year's event was historical - not only because Deep Blue was the first computer to beat a world champion under regular chess controls, but because the worldwide media attention was tremendous. Can you describe how it was for you as a developer at the event?

M.C.: It was somewhat overwhelming. We didn't expect as much attention as it got. And of course, after the first game, where Deep Blue managed to win, we felt like we had accomplished what we set out to do -- to prove that our approach to creating this type of powerful computer was on the right track and that we had created a system capable of beating a world champion. Now, it turned out, in the match, that his adaptability gave him the edge, but we think we proved our point. And we were very happy about that.

SOUND BITE: WAV(427k) | Bamba(23k)

Q: Do you and the team have a good relationship with Kasparov? Do you interact?

M.C.: Yeah, we're in contact with him from time to time, and he has actually visited here at the labs. He also gave a lecture, giving his thoughts about the match and what he sees for the future in computers and chess and the interaction with people.

Q: This question is optional: Do you have a prediction for this year's match?

M.C.: Well, let me say that I think it's going to be a higher quality match. We know that Deep Blue is going to be playing at a higher level than last year. Deep Blue -- the current version -- has already beaten the last version in several test games that we've played. So, we know that it's better.

We also know that Kasparov has spent a lot of time, and is spending a lot of time, preparing for this match and will have come to this match with some new ideas on how to play against computers, so that will be interesting. I think the match is going to be at a higher level, and I think it will be very close, but I'm not willing at this point to call a winner.

Q: When you were at Carnegie Mellon developing Chiptest and Deep Thought, did you ever envision being able to compete against a world champion? Did you envision that this would be the outcome of your work?

M.C.: Well, you believe that the approach that you're taking is eventually going to lead to the ultimate goal, otherwise, there's not much point. You have to believe there's some value in what you're doing. Yeah, I did think that some day it would happen, that we'd be competing with a world champion, and I, in fact, thought it would be about now, about 10 years after we began, so I was hoping it would happen.

Q: What got you initially interested in computers?

M.C.: Well, interestingly enough, when I was in high school, in Edmonton, Canada, I went to the University of Alberta for an open house -- the computer and science department there had an open house – and at that point I saw a computer playing chess. I was a chess player at the time and that fascinated me and I ended up going into the computing science dept. there and getting my degree, and going on to get my Ph.D. In Computer Science at Carnegie Mellon. So you could say that got me started.

Q: Deep Blue itself, what sort of language is it written in?

M.C.: The program itself is written in C. But obviously there's a special-purpose chip that is the heart of Deep Blue.

Q: And is Deep Blue's general approach a breadth-first search as opposed to a depth-first search?

M.C.: Deep Blue uses … the heart of Deep Blue you could say is a depth-first, what we call an alpha-beta search, which allows it to search very deeply into a position. But that alone wouldn't be enough. We use what we call selective extensions, which allow it to identify the critical lines of play, the important lines of play, and search those lines much more deeply. Without that, we believe Deep Blue wouldn't have much of a chance against Kasparov.

Q: How does Deep Blue select the opening for play if for example, it is playing white?

M.C.: Well, normally before the game the team, including the grandmaster, will choose a set of openings for it to play that day. It'll be tailored for a match situation. If we're looking to win that day as opposed to drawing, we'll choose a different set of openings. So Deep Blue doesn't yet choose its own openings.

Thanks.