Kasparov knows more about Deep Junior than we do

2/15/2003 – Two days after the X3D/FIDE Man-Machine World Championship match between Garry Kasparov and Deep Junior had finished, Mig talked for over two hours with Junior co-programmer Shay Bushinsky. From his home near Tel Aviv, Israel, Bushinsky spoke about the joys of a big match, Junior's preparation, the program's style of play, and the touchy subject of man-machine draw offers in a fanscinating interview.

Congratulations on Junior's 2847 performance in New York. This high-profile media circus was a big change from computer events at universities. Was it fun?

Garry said it would be lots of fun. I would call it anything but fun. Even more than that, it was the climax of non-fun. It's not fun to play against Garry Kasparov, it's not fun to try and refute the Anti-Meran. I mean, downstairs [at the New York Athletic Club] the Lakers were there playing 3 on 3! That's better than life! And I was up there with Boris and Amir, changing things in Junior. It's like a high-class prison. My brother Aviv was going to the pool. I literally never saw the pool, not even a glimpse.

What changes were made to Junior during the match, beyond the opening book?

After game one, we felt that everything needed to be changed. As Amir said, if Garry can do that to us on a regular basis, we don't deserve to be there, we're tourists, to use an expression of Garry's. We had an arsenal of different weapons that we came with, and we used them quite efficiently. The secret, not exactly deliberate, was that we gained small wins. Not game wins, but little things here and there. Like in game three, getting out of a position like the one we lost in game one.

In game three, Boris hadn't seen 9.Bd2 as a possibility, and we're in the g4 thing again and it's looking even worse. And it's refuted by Junior playing brilliant defense. It started somehow in game two and we were gaining throughout the match.

Isn't it a little dangerous making changes in the program with so little time to test the changes?

You're right, changing the program can be dangerous, but it was absolutely necessary. You don't want to be in a position in which you can't stop it from playing something, either an opening like the Anti-Meran, or 22.b4 from game two. We have a set of methods, a sort of meta-programming. We introduced certain changes that for example as a by-product eliminated b4 and resulting in it playing Nf3 instead.

How do you tune and test Junior? In games against other versions, against humans, with test suites?

We have practices for tuning and we do various things. Some is what you say, and Boris's comments, it's a mixture. You never get a full consensus, something telling you yes, great. But then you get a set of new problems, it's a matter of experience. When you do all these things. Sometimes you find out only during the game that it wasn't such a great job.

What's it like sitting across the table from Garry Kasparov for five hours?

There's a good angle from all of Garry's facial expressions. He's an honest person. He doesn't want to bluff you, he plays with honest facial expressions. After 10...Bxh2+. He was looking to right and then to the left and then he took the bishop. I caught him looking at me on occasion, like "how's that for your program, Shay?" or "what do you think about that?" I felt the looks in game four when there was a chance for him to win a piece with the ..b5 move. But as he said in the press conference after the match his facial expressions have not much to do with the high percentage of wins he scores.

You see the advantage of the programmers not having good chess skills. You can look at me, no problem, what are you going to learn from a patzer like me?

But you're a competent player, don't you spend some time wondering what he's going to play, or what you would play?

Sometimes I've found myself trying to get away from the position. I didn't want to give away anything with my face. I've been accused of doing that by Boris and Amir. Maybe, so I try not to look at the evaluation sometimes, not to look at the lines. It gets into a statement you made about our being cautious about the evaluations. We feel they are very accurate and if you get the people smiling about a position you can give away points. They can give away the story of the game. [Arbiter] Geurt Gijssen was fascinated by comparing Junior's evaluation and Garry's face. He was amazed by the correlations he noticed.

About those evaluations. Much has been made, at least by Garry and the organizers, of the scientific aspect of this match. Are Junior's logs or other data going to be released?

I don't know, we're thinking about different options. It depends on what's next. We want to publish, but it depends on what's ahead of us in terms of competition, commercial possibilities, and publishing more information. I differentiate between having it available in real time, during the match, and having it out in sustained release.

Not giving out the information has caused us some grief. At one point someone from Garry's camp said we were trying to torture him in a clear-cut draw. [Game four.] Well, the computer was showing a very positive eval and while me and Amir might have the knowledge, it would be a disservice to tell the computer to take a draw.

We couldn't tell Garry about the evals in game one. When did Junior's king-safety features kick in, for example, would be very important for him to know. All the expert committee received the exact program version and the game log so they could recreate the moves. And they saw the eval and time usage info on the monitor during all the games in real time. All of that was provided.

I'd like to mention that we never had a reboot, never a malfunction, and we're proud of that. There were some long thinks, but no problems.


Kasparov, Ban, Bushinsky, commentator Maurice Ashley

But now that the match is over, can't some information be released? Not to give away the secret formula but something needs to come out.

About releasing future logs we need to coordinate amongst ourselves and decide what we want to do next. We were very happy with many of the moves, and an effect we were happy about was that lot of the commentary was way off because they didn't understand what Junior was thinking. Not happy to be condescending , but it shows the difference between human thought in chess and computers.

Game four puzzled a lot of people, and 10..Bxh2+ in game five as well. Was it sane, are these moves believable or not? We heard the biases that we wouldn't have heard if people had access to the evals. If they can see the eval they tend to believe it after a while. It's a common phenomenon in this arena. But many of Junior's moves, conspicuous moves like a3 and h3 got a lot of ridicule, while Garry's moves like Re8 "oh how subtle!" Maybe if people could see the evals they wouldn't think that.

You say happy about some moves, what about the games in general?

Game four was outstanding in my opinion. If you look in the database for a game in which Garry's pieces didn't cross the fourth rank you probably won't find one! It was pretty cool from out point of view. Garry admitted that we were a little unlucky not to win that one (38.Bxe5). As far as his waiting moves, probably this is the way one plays the hedgehog. Maybe he had experience in letting it make errors by making waiting moves, I'm not sure what he expected. Probably our playing f3 and h3 weren't good from a time perspective, but time didn't matter so much in that position.

You have to analyze that game deeply to see if he actually did have a good ..d5 break there. I tend to trust Garry in these situations. I don't think he ignored it on purpose. Would it have worked at one point, I don't know. It might be combination of elements. Psychology, state of mind, I don't know. You have to be careful with the generality versus the specific analysis. You can't say you would play ...d5 just like that. You have to know whether Black's passivity was a by-product or a policy. It might just be a result of game three, when he played for a win and lost. Maybe he decided to wait for our blunder this time.

As we have seen in its recent computer events, Junior plays a very open, aggressive style of chess. Is this something you and Amir encourage?

Junior's style is evolving, it's not premeditated or fixed. So to hear Junior plays like Kasparov and Fritz like Karpov is flattering of course but no intention meant. It 's interesting that every program has these characters, but the style comes from how it plays, not visa-versa. We don't know how to make the program become a Leko or a Judit Polgar. We don't have a secret mode to turn it into a wild attacker or draw-maker or anything. It's the same Deep Junior.

We discover the results from our changes, we don't specifically target these results. The fact that we were able to successfully change the program was very good for us, if only for our confidence level. From a human standpoint we need to have some confidence building. It's very difficult to obtain a precise Elo rating for a program on short notice.

How about the opening book? Were additions made during the match or was it more a question of telling it what to play?

The changes in the opening book are massive as far as Boris Alterman's work. (Photo) He's done a tremendous job. Deep Boris was sitting there working all the time. I would doze off and he would still be there finding new variations. He did an excellent job there and we couldn't dare walk into the match without him. He's just invaluable. A lot of our changes went into that. And each time he made changes we would want to know how the book meshed with the changes we made. We made improvements regularly, not all were accepted, but there were adjustments on an ongoing basis throughout the match.

Do you envision any big leaps in computer chess programming or is it all optimization and tweaking from here on out, with the big advances over time coming from Intel and AMD? Do you still spend any time working on very experimental versions in which such leaps might occur?

If you look at versions in retrospect, you would see the various milestones and a lot of work. Amir and I did a lot of work between the WC [the computer world championship won by Junior in Maastrich, 2002 -ed.] and today to make it very different. It's difficult to talk about leaps and breakthroughs. I think we have the arsenal of tools that make big advances possible, but I don't think we can talk about breaking new grounds in terms of computer science, introducing new concepts, etc. Or that every version of Junior is a pearl. It's a constant effort and I don't see any reason why it should change dramatically. You always have interesting ideas and you are always tempted to try them, but you find they really don't work too often. But I pretty much assume that this is the way of life in other projects. But at the end of the day it's the routine that governs the long-term development.

In human chess, world champions often have a lot of influence among other players as far as style and openings. Do you see a similar legacy for Junior? What message does it carry for other chess programs and programmers?

I don't think anybody is really standing still right now. Shredder, Brutus, Fritz, everyone is improving all the time. Funnily enough, Junior 7 might be a little old in terms of championships and this match, but I think it has been very successful in carrying a new message. One can quote Garry when he says it is a technological advance as compared to other programs. I assume he means Deep Blue and other dinosaurs of that era. But also with other programs in how it evaluates abstract positional advantages.

When we won the WCCC in 2001 we were behind in material in every game! Yet we won undefeated with 8/9, which was tremendous. It was a new direction for the Deep Junior project. It's not just speculative evaluation, like playing ..Bxh2+ just to see what happens. We do not have innately speculative play. I think this is very important when you talk about Junior's contribution, or legacy. But it's a certain trademark and Kasparov says this is progress and I take his word.

And it's become a trademark of Junior, even playing this sort of thing against other computers. And now against the best calculating human player.

Look at Garry Kasparov, there is room here to allow him to enter into a computer championship! I don't think that he has too many problems playing through a computer test suite! Do you remember him doing that before his London 2000 match? I'm not going to ask him for a blood test, but there might be some silicon in there. He could have a good chance to win the WCCC, he's unbelievable. It does give all the more pleasure to play a move like ..Bxh2+ against him in particular. And for him to go for a draw was a worthy acknowledgement for Junior. Maybe against others he continues, but he made a decision.

Are you happy with the level of recognition your achievements have received in Israel? How was the coverage back home? Are you guys signing autographs in restaurants yet?

Disproportion in Israel as far as news coverage. It's a troubled time here and with the Columbia disaster it was so terrible. This country had one thing going for it and when this happened it put everybody into grief and sorrow and in comparison a chess match is a joke. I think we're just too used to bad news here and don't know what to do with good news. I think with the results of Israeli athletes and the losses by Maccabee [top Israeli sports team] that it should have gotten more attention. I wish we had had the match in Jerusalem, it would have made a big difference. It would have focused attention a great deal. There were so many postponements I think people weren't even sure it was going on.

Then there is the public perception that computers should be winning these matches. "Is it fair to have a match against a human?" People don't understand that the story isn't over, that it's far from being over. Computer have a hard time with chess and humans have learned a lot since 1997 and 1997 wasn't scientific. You find yourself apologizing for drawing with Kasparov! But that's the general perception, that the computers should always be winning. That's bad in general, and bad for chess.

What can be done about changing this perception?

It should be corrected, and in the right way. We need to have more humans playing against more computers. On one hand we had the bad result for the sport in 1997 and then a long history of stopping humans from playing against computers. (Dutch championship, boycott). So it was a double loss. And the third thing was that many humans didn't want to play against them. So these three things have mislead the public to a great deal. I've discussed this with Kirsan, pointing out the very visible chess event Kasparov-DJ was proposing to get the programs into the tournaments and onto the rating list." That would really pave the way to clarifying the relationship between computers and humans in chess. These sporadic events over the past few years aren't enough and it's been a dark period, the black plague years as I see it.

Losing to a computer used to be somewhat embarrassing to a human, but at now I think the players should be immune to that. Israeli Grandmaster Ilya Smirin was almost a hero for winning his match. So Gulko lost to computers, now so did Garry Kasparov! It was a PC, no big deal. So they should be over that stigma by now, and Garry has helped in this a great deal by playing again after six years. These programs have personalities, and some people can play very well against them and others cannot. But it was a fair chance and it was very interesting. He's saying that there is an interesting aspect to study here, that this can bring a lot of interest. Human-human play is of course interesting, but man-machine is a good opportunity to attract the general public. And when you play exciting games like we did, it should be embraced and not, "we don't want to play them at all."

Yes, it was impressive how much media interest the match generated. The man-machine thing never fails.

This phenomenon is in every sport, you need to broaden things as much as possible to attract the widest audience. Look at the NBA and Yao Ming. He's not Michael Jordan but he's very popular and very good. You need to try introducing new elements to add interest. Do we have to wait another decade for the next match? Why wait so long? Meanwhile, many human players do very well against computers and that's interesting to look at. The rules are the same, it's still chess. You can't stop progress at the end of the day and maybe in 100 years people will be used to it.

But I must state that FIDE has taken a huge step forward and they need to continue. One must congratulate Kirsan Ilyumzhinov both for taking the step and setting up the means to back it. President Ilyumzhinov is making efforts in the right directions to turn chess into being more popular and he is constantly changing much of the methods to find the right manner to attract the masses. IMHO if Piket beats DJ then he should get Elo points for doing a good job.

Should there be an open championship or a closed championship? Should man and machine be separate until the final match like this or should there be mixed championships?

I would say that the framework to have the computer as a regular opponent would 1) help evaluate its strength, 2) allow players learn how to play them, 3) improve chess as a whole and attract a lot of interest. In a way it's as tricky a subject as how to resolve the FIDE cycle.

And computer involvement could create many new high class events. There aren't so many out there these days and sponsors might be interested in tournaments with computers. I think man-machine events can attract more sponsorship. I also proposed to Kirsan to allow a computer team in the Olympiad.

How well did the final result match your expectations going in?

We didn't have any concrete expectations before the match. The impact of giving Garry the six-month-old version was underestimated by us, particularly its effects on the opening prep. He probably knows more about Junior than we do by this point, its strengths and weaknesses. We commend Boris for being able to cope with that in real time. We didn't have any predictions, we just didn't know. We had a good record, we came in thinking we had a strong program. But not knowing his strategy, his form, it was hard. Knowing his great record, his strength, made it hard to know what to expect. After game one we obviously expected the worst.

What sort of preparation did you do? Did you play training matches?

We had very good results in training games against other Grandmasters before match. (But I cannot disclose who they were or the results.) But it was misleading. In retrospect we understand that it was a very different setup in the match itself. The training games were also designed to check certain aspects, not really predict what would happen in the match. We played approximately 10 training games.

Obviously we wanted to win, not draw. It seemed as if we were improving throughout the match and were getting a good feeling about our chances to win. It would have been nice to win that last one. However, the big satisfaction for my point of view is how the games were conducted. They were exciting and hard-fought. I didn't want to go into this with boring lines, torture for both sides. Or catching Garry in bad shape and playing badly. These were some concerns I had before the match.

People said we missed our chance when we didn't play Garry last year after Moscow [the Russia vs. the World match, in which Kasparov had one of his worst-ever results], when Kasparov was in bad shape, but I disagree. You want to play a great opponent in his best shape. And Junior wasn't disappointing either and that made for a good match. That was the best outcome. Of course the result was also important and losing would have been very hard. But in essence, the quality and interest and adrenalin and tension before every game was what made it great, in retrospect. But being there at the time, it wasn't that great. A lot of work and then the tension at the board.

How does that feel, watching it play when it's all out of your hands?

After it starts there's nothing much you can do. You know it's not going to blunder, but you still worry. You don't feel helpless but you have a lot of interest and concern. It's also exciting. And to watch Garry find the ways to refute Junior's play was amazing to me, personally. It was great to watch him in action. I've seen him play and seen other people play against Junior and we wouldn't have had the same quality of play against anyone else.

Did he change at the board during the match?

In the second half we saw a different Kasparov. He was more timid. In the first half he was going for it.

So let's talk about the confusing finish of game six. Kasparov makes a very strong move with 23...Rxc3 and then offers a draw. You refuse and a few moves later you offer a draw and he accepts in what looked like a good position for him.

At the point he offered the draw, Junior's eval was positive. And note that Grandmaster Alex Shabalov also liked White. Garry did admit missing DJ's 25.Bc1. I continued play, refusing his draw offer, and made a few moves. But that he found the formation of ..f6 and ..Ne8 , which was very impressive in Boris's opinion. As he did all match, Garry found so many great moves. At least from Deep Junior's perspective Garry found the best lines to continue again and again. That's really impressive, that really shows his genius at close range.


The secret? Note the cross-handed method!

The consultation on the Junior team regarding draw offers has stirred up a lot of interest. Most people didn't know that you guys could consult not only Junior's eval, but Boris Alterman using Junior on another machine. What's your take on this?

On one hand there is the undisputed best player in the world and he's offering a draw to whom, Shay and Amir? We can be robotic and follow the eval blindly, but there are so many exceptions in each position. We can't tell it what to do, we aren't chess players. Garry suggested that it must accept if it's in a certain radius around the eval, but that doesn't make sense either. There are sporting considerations, like who is leading the match, and if the position is complex. Thank god we had the possibility to consult with Boris and with Junior when considering.

I don't think you want to see the computer making its own draw decisions. You'd have another four or five hours of game, the human falls asleep... The computer wins, the humans get angry. I don't think that's what you want to see. Chess is a very pragmatic game played over the board, you cannot really isolate a position or two for an example. And you have parameters and ethics around these matches. Garry tires and the computer doesn't. And you have human objectivity that needs to be inserted. We don't want to win and be accused of torturing the human in an equal position, but that wasn't the case and it was a great match.

If you left it up to the machine there would be almost no reason to ever accept a draw. You would just have the computer refuse every draw offer, unless it the eval was, say, -1.50 or worse. I don't think people want to see such matches and I don't think it's a good idea.

Are there some comments you'd like to make about the match, or some misconceptions you'd like to clear up?

There were conflicting reports about the hardware. We were using the four-processor machine with 1.9 GHz Pentium 4 processors. It had an effective three gigabytes of memory, and all tablebases were installed although never accessed. We never had a single crash, never rebooted once.

So this was the fasted/strongest Junior ever?

The fastest Junior ever, sure, definitely better than the one we used in the WCCC. Not too much of a difference in practical performance. I'm not sure our dual Athlon at home would have done much worse. We were somewhat tempted to use the dual because of some initial difficulties we had with the new hardware, but it was a little better on the quad. The eight-processor machine wouldn't have made much of a difference, although there were a few indications that it would have been a bit better in a few situations. We went with the quad for stability and because we were happy with its performance.

I definitely want to emphasize the hard work of Boris Alterman. Garry had his team there with him, Yuri Dokhoian and Mikhail Kobalia. I'd like to talk with them all about Junior, they are probably experts.

There are many thanks owed to many people. First of course Amir Ban. we've been working together for a decade and it's tremendously rewarding working with him. I really have a high appreciation for his work. It's great that cooperation can work this well. I remember we went to a tournament with Junior "Zero" and we beat a Grandmaster. That was our first or second tournament. It was the Kfar-Saba Open and we played Leonid Gofshtein. For a PC program to beat a human at a normal time control was considered a great achievement in 1994. It was a landmark that told us, hey, we have something good here.

And there were a few landmarks in this match: first machine to beat Kasparov with black, first PC program to ever play him (and beat him) at a standard time control. [It turns out Kasparov lost two blitz games to Fritz in 92 and 94 and a rapid game to Genius in 1994 all on the white side. -ed.]

We also need to thank our anonymous GM trainers, our loyal beta tester Jonas Bylund of Denmark, who has always been there for us. He's a heck of a guy. The X3D technical team, Cheryl and Pascal. Scott from Canvas Systems for his help and management. They made our life much easier with the setup. The rest of the X3D staff, especially Bill Drury and Armand of course. And the organizers Serge Grimaux and Katerina Tornerova. It was great to have someone like Milos Foreman make the first move! Special thanks to Matthias Wüllenweber/ChessBase for the interface support that made operating a cinch.

And of course people back home, our families who sacrificed a lot. Testing from friends like Alon Rozen, many others. There was a lot of human effort behind Deep Junior, that's really what I would like to magnify. It's humans driving technology and that's a point that gets lost with all the "man versus machine" stuff. It's human effort and it can't be otherwise.

You won the computer world championship, you drew with Kasparov. What goals do you have for Junior now? Can you foresee a point at which the project is no longer so interesting to you? And on a related note: Once computers inevitably dominate humans, will the interest of computer-computer chess be enough for you, do we all take up Go?

Chess programming is very personal. I love chess and love to see the thrill of the fight. Seeing something like ..Bxh2+ is great. I feel lucky to feel associated with this and I'll continue while I enjoy the game. From a scientific point of view it's hard to say. It's a big question. Much has to do with what we touched on before, with introducing Junior and computer chess into human chess life on a regular basis. Some predicted that if we won this match that it would be the last game for Junior and we don't want to see that. We want to see more matches.

With computers on the PC level it made the fight more interesting, and I think that will continue. I don't think I'll ever change to Go, I'm just not a Go player. It's not like "we're done with chess, we'll move on."

Photo: Shay and Amir with their bag of tricks?


Feedback and mail to our news service Please use this account if you want to contribute to or comment on our news page service



Discuss

Rules for reader comments

 
 

Not registered yet? Register