One-on-one with F.H. Hsu

Q: Well, let's start with the easy question: Do you play chess?

Feng-hsiung Hsu: Not really, I only played Garry on TV.

Q: You played Garry on TV? How did you do that?

FH: During last year's match, I was usually the one who moved the pieces. I was able to shake Garry's hand when he resigned last year.

Q: Oh, that was you? So I saw you on the "Tonight Show" when Jay Leno poured the water on the computer?

FH: Yes.

Q: Let's look at last year. Last year was a historic landmark, not just for IBM, but for the world of computing and even for the applications that go beyond that. But, ultimately, Deep Blue lost the match and now Kasparov has issued a re-match – how has this year's version of Deep Blue been re-tooled for the upcoming match?

FH: A year has passed since the last match, and given time, the computing always makes progress. So, there will be a new version of the SP. The RS/6000SP will be twice as powerful as last year's machine. And, given the host is twice as fast, we can put twice as many special purpose chess processors in it. In turn, this means we can run the machine about twice as fast as the previous year.

Instead of 100,000,000 moves per second, we can do about 200,000,000 moves per second. But, that's not the big point. The big one is that the chess knowledge is going to be much more refined than last year. Last year, what we brought in was a new baby. The machine was just barely completed. We didn't have enough time to prepare the machine as far as giving it what is important in the chess world.

Joel Benjamin, grandmaster Joel Benjamin, was with us last year for about two months before the match. He did some preparation and told us something about every function, but not enough to actually make a big impact on the match. And this match, Joel has been with us for over half a year now, close to seven months now. And, during this period, he was essentially taking the computer, the baby, to chess school, teaching the machine what's important, showing it the ropes and so on. And that's very important.

Right now, it appears that the new machine, the new program, is playing much better. It's playing positions far closer to human style. Sometimes this machine plays strange moves, but sometimes those strange moves are beautiful moves when you start taking a closer look at it.

SOUND BITE: WAV(999k) | Bamba(52k)

Q: How do you decide on an opening move?

FH: Opening moves are usually pre-programmed by the grandmasters and us and are pre-selected. We try to get position where the machine will be better. Well, not necessarily better, but the machine knows how to progress.

Q: What's your role on the team? You started, maybe in earnest, back in August, and now that we're leading up to the match, what's your day-to-day look like?

FH: Actually, I started the project when I was in graduate school, back in 1985, the same year that Garry won the World Championship. But that was just a coincidence. During the past year, my real day-to-day things are going through the hardware to determine which things are wrong. Also, I spend time tuning the variation function. How you adjust the variation terms makes Deep Blue play better.

Q: Let's elaborate on the origins of Deep Blue. You started the entire project Carnegie Mellon with Chiptest and later, Deep Thought. You beat the world champion last year in the first game, and this year there's a good chance you'll win the whole match. How does that feel for you?

FH: It feels great. I think that last year, going into the match, we knew that the machine could search so many positions per second, and so on. And, based on previous projections, you think you have a good chance to win. But, to tell the truth, I did not really think we could win last year. This time I believe we can win. And that's a good feeling to have.

Q: On May 12 when the match is over and we all wake up, what's next? What does this mean?

FH: There are a few things that some other people have been looking at. Just two doors down from me someone is working on something called molecular dynamics. And there's a Japanese corporation that looks at the many diseases of the body. You look at what you can do if you put different molecules together, and what happens. That could be useful for producing new drugs.

We may be able to determine whether or not a molecule is going to have the right interaction with a particular protein within a drug. That could be quite important. One thing the people are talking about, in relation to AIDS research, is trying to figure out what the proteins in your body are doing and how you can stop the AIDS virus from proliferating.

With Deep Blue technology computing things much faster, you could possibly develop drugs in five years that would normally would take 15 years. If you could do it in five years, that would be tremendous. Because within that ten years, the number of lives you save is going to be up in the hundreds, maybe hundreds of thousands, maybe even millions.

SOUND BITE: WAV(478k) | Bamba(13k)

Q: Can you describe the difference between the way Kasparov thinks and solves problems and the way Deep Blue "thinks"?

FH: Grandmasters have a lot of experience, so they have better judgment about position. Also, they don't search as many positions as the computer for obvious reasons -- they are doing it more selectively. They have some idea about which moves might be the right ones.

So the very good player can probably make two moves per second. Garry may be slightly faster than that, but who knows. Now, what happens is that usually computers can go very deep in their searches when they see it's important. The computer can do part of that. We have mechanisms and one of the mechanisms that we developed, back in 1987 was the Cossinger extension. It allows you to see a position, and when you only have one good move, it searches deeper.

That produces behavior similar to what a human does. Except that the mechanism we used back in those days were kind of crude. Now we use something a little bit more refined, but still, nowhere near what Garry's capable of doing. On the other hand, [the computer] never gets tired and it can search as wide as it can. When you get in a position and there's one good move, the human can go very deep. But when you get in the position where there are several good moves, it becomes very hard for the human to see very far, and the machine is not intimidated by that.

Q: When did you first get involved in computers?

FH: I first got involved with computers back in my college years. We had a class in programming that was just a regular course. And, then, in the second year -- in my sophomore year -- I got involved with a microprocessor project with one of the professors. I actually started in Engineering as an undergraduate. But the experimental project triggered my interest in computer science in general. When I came to the States, I was still leaning a little bit about Electrical Engineering, but I ended up getting a Ph.D. in Computer Science.

Q: When did you first play chess?

FH: It must have been when I was maybe 6 years old – no, older than that, maybe 10? My brother taught me how to play, except he didn't know all the rules. The castle move he taught me was wrong, for example.

Q: You've won a number of awards for the chess machines that you've designed in the past. Would winning this match be the ultimate reward for all of your hard work?

FH: Winning the match is important, not just personally for me, but also to solve a long-standing problem in computer science. But if you look at the larger picture, it's not just a computer science problem either. It's a matter of where we are going as human beings. Because when you view the match, you can view it this way: on one side you have Kasparov who is a great performer, and the other side, you have a bunch of guys who are essentially tool-makers.

SOUND BITE: WAV(243k) | Bamba(13k)

Now, these tool-makers don't have the chess understanding that Garry has, but they're competing with Garry now. And the reason we are competing with Garry is because we can use the computer as a tool to compete with him. Now, what that symbolizes is that if you don't have a talent in something, that doesn't necessarily mean it automatically limits you from doing something, because the computer may be able to help.

Q: One of the ideas that C.J. discussed was that the computer is sort of an extension of human capabilities. For example, a car can make you "run" faster, or a computer extends human capabilities…

FH: …Makes you smarter, yes.

Q: Makes you smarter, right (laughs).

Let's shift gears for a second. Last year's event in Philadelphia, can you describe your feelings during the match? The atmosphere, the excitement?

FH: Going into the match I had some apprehension -- lots of apprehension -- because, as I said, I thought we could win but I didn't really believe we could win. So, one big concern was how we were going to do. And after game one, the feeling of apprehension went away because, yes, we made history there, and we knew we could compete.

SOUND BITE: WAV(292k) | Bamba(16k)

Game two we lost, but that was an OK loss because we fought hard during the game. Games three and four were quite important psychologically for us. They demonstrated to us that the machine, even though it wasn't given a good position -- in both games the machine didn't get good position after the opening -- managed to survive. It managed to draw both games. In fact, in the fourth game, Deep Blue almost won the game.

Now game five was a different story. Garry offered a draw, but we refused it. Now, from the match point of view, that may not have been the best idea. But from a scientist's point of view, that was the right thing to do. Of course, the troubling thing is that up to that point, he knew certain weaknesses of the computer that he naturally explored the next game.

Q: Do you think he'll be able to use a similar strategy that he used last year?

FH: It will be harder. First, we have been working with a grandmaster longer, and there are fewer holes. Last year's machine was strong in some aspects, but it had many weaknesses that he found over the span of six games. This time, there are a lot fewer mistakes that are going to be made, and Kasparov will have a lot harder time trying to locate them. Even if he finds a mistake, he may not be able to exploit it. It's going to be much harder for him I think.

Q: How much do you expect to interact with the computer during the game itself?

FH: During the game we're not allowed to interact with the computer… really interact. You can just enter in the moves into the machine. But between games you can do whatever you want.

Q: And how much do you think you'll have to do between games?

FH: It depends on how the match goes. If we find a problem with our opening, then that may trigger another opening preparation that will have to be done. Which means some of us will have to work long hours, and the machine will work even longer hours during the period and hopefully, we'll find something good for us that we can use.

Q: What's the next really hard problem you're going to try and solve after this year's match?

FH: One of the things we are looking at is data mining. That seems to be an interesting subject and there's some similarity with chess programming in the sense that data mining is something that has bi-level manipulation. That makes it a little easier to make special purpose hardware that gives you better bang for the buck.

Q: What kinds of real-life applications might data mining be appropriate for?

FH: Data mining is being used by companies like Wal-Mart and other organizations. It's one thing they use to predict how customers buy things, and so on. This might affect how they arrange their shelves for example. And insurance companies might decide who is a better insurance risk through data mining.

There's another example, a guy here at IBM is working on a program that allows an NBA team to be able to predict what the best line-up is against the Chicago Bulls. There was some interesting thing I was discussing with him, I don't remember precisely, but the program was showing that there was some specific combination with some particular team that tended to work quite well against the Bulls.

Q: Oh yeah. It was the Seattle Supersonics in last year's playoffs. Perhaps the next project for the development team could be the construction of a computer that can play against the Chicago Bulls.

FH: That will be a little hard, since the machine cannot move. (laughs)

Q: Is it nerve-wracking to know that the match is coming up in a month?

FH: No, I don't think it's nerve-wracking now. You know what you need to do. You know what you need to do and how you have to do it. So you make a plan and make a schedule and just prepare for it.

Q: Do you feel that you're ever ready? Or, say the game wasn't in May, but in June. Do you think you could even further prepare for it?

FH: You can always use more time. Chess is a fairly big field. Somebody said that chess is like a sea and you would never be able to drink it. When we started, we wished that we could have done some things differently and so on, but Deep Blue was still pretty good. It could have been better, yes, no doubt about that one. A few more months, we could have done something slightly different, and made things quite a bit better. That's possible. But now the match is set, and we just have to be ready. We'll see what happens this year.