Stork: What is it like playing Deep Blue?
Hoane: [Laughs] I could play any chess program and get crushed so I don't play Deep Blue very often. For me there's this sense of inevitability. I have no way out.
Stork: You mentioned that the Deep Blue team tweaked Deep Blue Between games. If you hadn't, what would have happened to the full tournament?
Hoane: One of Mr. Kasparov's strategies was playing this soft opening where there was a lot of maneuvering. We hadn't anticipated that he would play so far out of normal lines. So we had to deal with that. That's what people do between chess games: you figure out what the other guy's playing and you make some lines around what he's playing. That's a very human grandmaster sort of action. It would have been a close thing. Yeah, it could have turned out differently, sure.
Stork: Can you tell us about each game.
Hoane: The 1st game we lost but we were very optimistic. He threw us his best curve. At one point we had the game drawn, despite his efforts with the surprise opening. We were very satisfied. We knew we were in the ball game.
In the 2nd game Deep Blue played a beautiful game. It played an opening completely naturally that we hadn't planned for. At one point there was a move where Mr. Kasparov was expecting a typical materialistic move and Deep Blue didn't make that move. He didn't expect the computer to understand the attack he was making. This was the game we won. I don't remember much about 3 or 4ÉI think there was a really nice rook ending in 4, we were very happy with the play. That's one thing we knew he was going to be surprised about. In our grandmaster play it was playing endgames beautifully. It was a weakness in the earlier version of Deep Blue and a distinct strength in the final version of Deep Blue. We really really did a good job in the endgames. It really seemed to understand the use of the pawn in the endgame. The 5th game it played really well. There was just one really suboptimum move that we didn't like and it made a draw essentially. But we didn't have chess covered. We haven't got all the chess concepts in the world entered into the machine yet. The 5th game he tried something he thought the computer would self destruct on, and it didn't and so it won. It was a sort of all or nothing try on Mr. Kasparov's part to win the final game as black which has the disadvantage.
Stork: When you refer to deep blue you use the word ÔUnderstand', can you explain more what you mean by that?
Hoane: Deep Blue's understanding takes two forms. One understanding is of specific chess concepts that we have articulated to it in the form of here's a chess idea, how do we encode it? And this is the process of many many years' work. Two, operational understanding given by the search. The search is a very powerful tool which takes a little bit of knowledge that you put in and it amplifies it just like an amplifier. That's what the search does for chess. It makes the computer understand consequences. When I say understand I mean computationally execute. It would be a great advance if Deep Blue could make a good move and tell you what the themes were and tell you why it made that move. That would mean understanding in a human sense. When I say understanding for Deep Blue I mean operational sense. I mean it executes as if it understood [laughs] which it doesn't Ð I mean in a human sense.
Stork: Would it be a worthwhile research goal to try to make Deep Blue able to explain its moves?
Hoane: Yeah, explaining is a nice next step, well before understanding [laughs] right you can't get to Hal yet. You can't get to a Hal 9000 with what we know. But that would be an interesting step past Deep Blue.
Stork: Hal found his game with Frank enjoyable. Did Deep Blue find his game enjoyable?
Hoane: Deep Blue doesn't even think. Emotionally there's nothing there, not at all. It doesn't even think much less emote.
We came off the first match with the idea that the machine, the engine was big enough to dominate the world champion tactically. It really was. That's what happened in the first game in the first match. And he reacted to that and brought in other chess skills and won based on strategic things. So we understood after that match that what was lacking was strategy, adaptability, and maybe chess knowledge. I remember at one point there was this list of grand challenge things that we could do to try to fix Deep Blue in a year so that we could win a match. One of the items on this was really studying the grandmaster games. But we'd had some failures with that before. It's really hard to say here's a chess position. This one is similar. Humans can do that. That's something that humans are great at is saying ok here's the themes that go on in this game. I think I'll use this theme. Oh that one won't work because the pawn's over here but you know people can do this sort of reasoning. And we're sort of stymied at coming up with some sort of similarity measure like that this one is similar to that one therefore you should play this theme. Very difficult. So what we went for were the rules of chess that you would find in a chess book that are obviously true, and put those into the machine and that's what worked. We worked with Joel Benjamin, former and current U.S. champion. He'd worked with us in the 96 match as a consultant about openings and we realized that we needed him a lot more and we brought him in almost right away and had him a start working with the machine meaning playing games and saying hey, that looks funny lets pounce on that. Real sort of school of hard knocks training for a year. We redid the hardware and as many important concepts. In the end what we did was we stuffed a whole bunch of chess knowledge into the machine a tested it so that there were no major holes as far as we knew. It was sort of an engineering approach unfortunately. Some of the big AI things are hard. Finding the themes in a chess game and playing that theme is hard. Better just to find the nuggets of truth. You know, past pawn is good sort of things, and the fine gradations around that concept and just get it right.
MOC: Take yourself back to game six, you're sitting opposite Gary Kasparov. Tell us what the mood was, and how you felt as the game went on.
Hoane: For me, as the operator of the machine, game six, the final game of the match, I'm always nervous because I'm not a very good chess player. It's a very nerve wracking thing to sit as the operator in this situation. There came a point in the game where Mr. Kasparov tried this gambit and he knew that he'd lost and I knew that he'd lost looking at the score and he was very upset by this. I think, and he said this after the match, if he would just prepare for Deep Blue like it was a regular, strong opponent he would have done much better. Instead his preparation, and this is the flaw in his preparation, he prepared for it as if it were a typical computer chess player. That is the flaw in his months of preparation. That he didn't understand how strong it was going to be. That it would really understand various aspects of chess that computers typically don't Ð maybe not the same aspects that people do, I mean it's not a human, it has different strengths. But I think, as he was sitting there and her realized he was losing the game and the match, it was very emotional. You could see him getting very upset. I was getting very upset in reaction to this. So the world champion lost a match to a chess computer and it was quite a trauma for him.
MOC: How did you feel?
Hoane: I felt traumatized because of his obvious discomfort, because of the symbolic weight and meaning attached to this match.
MOC: Was there some sense of achievement?
Hoane: I think I felt a sense of achievement starting to sink in maybe three weeks after the match. I was trying to come to grips with the fact that the way we had created this world champion level machine was with a lot of engineering and putting in the knowledge. I didn't understand what that meant to computer science. I think it means there's a benchmark in the progress of computer science and computing but I don't know in and of itself whether there's useful knowledge in what we did. As I said I don't necessarily think Deep Blue is a stronger player than the world champion. It's just that the world champion is in the same league as Deep Blue or vice-versa.
What doesn't come out when you talk about the second match, the '97 match, is what happened in the first match. We really, in the first game of the first match, hit him (Kasparov) with a sledgehammer. He just wasn't expecting a machine that could totally dominate him tactically. We won over the board and just wiped him out. Reportedly he was up all night thinking about this. He came back in the next match and won. This great human achievement of reorienting himself to using his other chess skills to win the first match. And seeing how delighted he was with that, he was so happy to be a part of it. He said, "Who would've thought at 33 that I'd learn something new about chess?" It was a wonderful experience to go through that with him, a personal triumph Ð as opposed to the difficult experience of the 2nd match. |
