In cooperation with:
|INTERVIEW WITH DAVID LEVY|
International Master David Levy has been a central figure in computer chess practically since the term was invented. After many books and events, he is now the president of the International Computer Games Association (ICGA), which is auspicing the upcoming Kasparov-Junior man-machine match in New York City. On January 8 we sat down for a virtual interview.
Mig: What do you see as the main differences between the Kasparov vs. Deep Junior match and the Kramnik vs Deep Fritz event in Bahrain?
Levy: From the sporting point of view the match in New York is the very best against the very best. Kasparov is stronger than Kramnik, by 40 Elo points at the present time, a significant difference. And the Junior engine is the reigning World Computer Chess Champion program and has proven itself stronger than the Fritz engine in the past few World Championship tournaments. The public wants to see the champions play each other, particularly in a country like the USA where everyone loves a winner.
Mig: And what about the rules. There has been a lot of discussion on the rules for New York but we haven't seen them yet. Are the rules very different from those in Bahrain?
Levy: The discussion on the rules has been going on for some time but, because of the postponement of the match, this debate has not been the most urgent topic. In fact we had the rules about 95% complete as long ago as mid-September, but then there was a hiatus and the subject only rose to near the top of the pile as the definitive dates for New York became known.
I should say at the outset that the aim has been to create rules that will stand the test of time and can be adopted by FIDE, virtually unchanged, for future official FIDE man vs machine events. For this reason we sought the opinions of a number of experts: prominent computer scientists who are also chess programmers; Grandmasters; the past Chairman of the FIDE Commission on Computer Chess; as well as the participants of course. I believe we have struck a good balance and created rules that are very fair to both sides, though experience at the New York match may dictate one or two refinements. And, as with the FIDE Laws of Chess, which are subject to regular consideration and change by FIDE's Rules Commission, the man vs machine rules will also develop with time and with experience.
As to the main differences in the rules between New York and Bahrain, these were necessitated by Kasparov's desire to avoid either side having an unfair advantage. For Bahrain the rules were, to be frank, rather human-friendly. The program code had to be frozen several months before the start, with no changes being made thereafter. And Kramnik was given a copy so he had the luxury of being able to practice against the exact same version of the program against which he played. He even had an opportunity to play against Deep Fritz on the same hardware used for the Bahrain match. All this allowed Kramnik to test the strengths and weaknesses of the match version of the program, a far cry from what Kasparov went through before he played Deep Blue in 1997.
There were other points in the rules which could have worked to Kramnik's advantage but did not arise. If a game reached move 56, Kramnik would have the right to adjourn until the next day and to analyse the adjourned position using Fritz itself, in effect enabling him to "ask" what it intended to do against whatever variations he was considering playing after the resumption. In our rules there are no adjournments - we use the "classical" time control that ensures a game is finished after no more than 7 hours play. This is just as in top class human chess nowadays, but it does make life more difficult for Kasparov because a computer program does not tire after 4, 5 or 6 hours.
Mig: You say that Kramnik had all of these advantages in Bahrain, so why didn't he beat the program?
Levy: I must confess that before the match I felt that Kramnik would win easily. For one thing, Kramnik's style is well suited to playing against a program. But most of all I believed that the advantages I've just mentioned would make life very easy for him. And indeed they did at the start of the match. Forgetting about game one, where I got the impression that Kramnik was just feeling his way a little, to give himself confidence, he destroyed the program in the next two games and most chess enthusiasts, myself included, thought it was all over.
But in the second half of the match Kramnik was unrecognisable. He said afterwards that he was already exhausted by the end of game three and that explains a lot. Because when playing a program even the very strongest human players must stay on their toes, lest they suffer from Grand Canyon poisoning (one slip will kill you). And when Kramnik made that horrible one-move blunder in game 5 the whole match really was over, but not in the way we expected. The psychological impact of that blunder clearly left him below par for game six, otherwise he would never have sacrificed on f7 when he had a very strong and simple continuation (Bd5) which would almost certainly have won the game.
To answer your question more succinctly, I would say that Kramnik failed himself. But please don't think by this that I mean to disparage what Deep Fritz achieved - it is an excellent result for a computer program to achieve a 2800 performance over 8 games, especially against a player of Kramnik's style. And Frans Morsch is an outstanding chess programmer.
Mig: When will we be able to see the rules for the New York match?
Levy: As soon as they have been signed off by both sides.
Mig: Is there anything else about this match that will be significantly different from Bahrain?
Levy: You mean, apart from the near-freezing weather in New York?
Mig: Yes! Not that I'm missing the 100% humidity of Bahrain and the air-conditioning colds we all got. (Not to complain, the hospitality and organization were excellent.)
Levy: The New York match is much more like a human world championship in the sense that we have a very experienced arbiter - Geurt Gijssen - I think he must be the most experienced arbiter in the world for top class chess events. And we have a Rules and Arbitration Committee, rather like the appeals committees that oversee a human world title match. Because of these factors the match is being sanctioned by FIDE, by the ICGA and by the USCF. So it has all the status of a world championship, hence the official FIDE title: "Man vs Machine World Chess Championship".
Mig: What do the "computer people" hope to learn or gain from the match?
Levy: Three very important answers. Firstly, who (or which) is stronger - The world's strongest human player or the world's strongest computer program? This is important for benchmarking progress in this field.
The next question is, and many people are asking it, do we know how Deep Junior compares in strength with Deep Blue? The really interesting thing, from the AI point of view in general and for computer chess researchers in particular, is that Deep Junior examines something like 1 percent of the number of positions per second of Deep Blue. But despite this Deep Junior may well play better chess because its "understanding" of the game is better. It appears to have more chess knowledge and understanding in its evaluation function than Deep Blue did, and this compensates for the difference in positions-per-second. Unless there is some new or dramatic aspect to the program's play, most of us lesser mortals will not be able to answer the question even after the match is over, partly because we are not strong enough to detect all the differences. But Kasparov will know. He is going to be the only human to have played matches against both programs and he'll be able to speak from experience.
The other interesting question is going to be, what are the program's discernable weaknesses? No matter how well it may perform in New York, Deep Junior is bound to have its own area of vulnerability. Kasparov may have found this already by training against a 6-month-old version of the program, but the true test will come in New York, when the latest version of the program is revealed. Amir and Shay will almost certainly learn some useful things about their program from studying the game logs. [Amir Ban and Shay Bushinsky of Israel, Junior's creators and programmers, pictured above. Junior, like Fritz, is published by ChessBase. -ed.]
Mig: How does computer chess spill over into other areas?
Levy: A great deal has been learned from computer chess techniques and applied to other games. The science of how to search game trees has benefited enormously from the research into computer chess. And the endgame databases used in chess to play certain configurations perfectly, for example queen and pawn against queen, and all the other endings with 6 men or fewer on the board. The techniques used in this field have enabled programmers to "solve" some games, i.e. play them perfectly. Awari, or Mancala, is one example; Nine Men's Morris is another; and Connect-4. But there are also some more general applications of game playing programming that can be employed outside the field of strategy games. For example, I am convinced that some of the techniques from our field can be used in stock market prediction programs much more advanced than the current ones. And in other long-range planning problems, including those in manufacturing, airline scheduling, the freight (shipping) business, … there must be countless big-money businesses that could benefit from programs which use adversary planning techniques derived from computer chess.
Mig: What will the impact be on the human game when, in 5-10 years from now, computers are far too strong for humans to compete with? Will this super-analysis slowly destroy the game? My concern is that we will just end up going out and trying to regurgitate what our computers told us to play. We sort of do that now, with databases and openings, but with 3000-level computers coming up in a few years, they will push deeply into the middle-game.
Levy: Many people have asked the same question over the years. But I am far more optimistic. I believe very strongly in the future of top class chess as being based on Kasparov's "Advanced Chess" concept. Garry's original idea is tested nowadays in the annual matches in Spain, and there have been some other experiments. I feel there is huge scope for expanding this concept. Let me give you just two examples, two variations on Garry's original theme as it were.
My first idea is that each human player is allowed to access his program only a limited number of times during the game. Say 3 times or 5 times. This raises the question: "When should I use my program?" If your opponent uses up most or all of his accesses relatively early in the game then you will try to create complex, tactical positions so that you can use your own program in the most critical moments of the game, when tactics are paramount. Then the struggle takes on a more psychological aspect.
Another idea is to award points partly on the basis of which player uses their program the most (or least) in a game, and by how much. For example, You beat me but you use your program 20 times and I use mine only 5 times. The difference is 15 times. So instead of you scoring 1 and me scoring 0 you would score 0.85 (i.e. you lose 1% for every extra use of the program) and I score 0.15. Or it could be that you score 0.7 and I score 0.3 (i.e. a 2% adjustment for every extra use). By changing this factor (1%, 2% or whatever) one could create events with very different characters. I hope you can see where I am coming from on this. The range of possibilities is enormous.
Mig: How accurate have expert predictions been about computer chess strength over the years?
Levy: You're inviting me to boast. First of course there was my bet, which started off this man vs machine thing. (Visit www.icga.org and then click on "ICGA Media Briefing Pack" followed by "Chess Champion vs Computer"). In the early days I was right that I would win the 10-year bet; I was right that I would win the subsequent 5-year bet; and I was right to stop betting after that!
On the other hand, a few AI luminaries from the late 1950s and 1960s were of the opinion that a program would be playing world championship level chess within a decade or so. And even Botvinnik, who should have known better, told me in 1970 that: "I feel very sorry for your money".
By the time I stopped making bets on myself, in 1985, this whole business was becoming serious. I wrote an article in the ICCA Journal in 1986 called: "When Will Brute Force Programs Beat Kasparov?" in which, based on a graph of the best rating performances from recent years, I concluded that the earliest it could happen was 1990. But Garry's rating at that time was "only" 2720 and he got stronger, much stronger! As I later showed in another ICCA Journal article (June 1997), using that same rating performance data, together with Garry's true rating for each year, and allowing for a difference of 150 Elo points between USCF ratings and FIDE ratings, a prediction made in 1986 for the year in which a program would first achieve a performance rating consistent with beating Kasparov, was ….. 1997!!
A different method of prediction came a few years later, at the World Computer Chess Championship in Edmonton, in 1989. I carried out a survey amongst 43 of the computer chess cognoscenti present - most of the top programmers plus some other AI experts. I asked each of them when, they believed, a program would defeat the human World Chess Champion. Only one, John Stanback, the author of the program Zarkov (and later a co-author of Gnu-chess), picked the year 1997; Kevin O'Connell was close with 1998 and then came a host of experts guessing either 1995 (including Murray Campbell who was part of the Deep Blue team) or 1999 (including Claude Shannon, whose seminal 1950 paper on chess programming showed the way).
Mig: And what do you predict for New York? Who will win and by how much?
Levy: First I must say that because of my position on the Rules and Arbitration Committee for the match I must remain neutral throughout. It would be wrong of me to say what I expect to happen and, since I predicted a 6-0 win for Garry in 1996 in the first match against Deep Blue, my prediction today is almost certainly not worth anyone betting on. But I will tell you what I consider to be the most important factor in this match. It is psychology.
Let me explain. Kasparov has far more knowledge and understanding of the game than any program can ever hope to achieve. (Well, "ever" is a very long time, so let us say: within the next 50-100 years.) His intuitive feel for the game is second to none. He is, quite simply and without a doubt, the strongest chess player in history, human or silicon. But against this the program has certain advantages: it never tires; it never overlooks a tactic within its search horizon (and this horizon can be very deep indeed in tactical variations); and in compensation for its lesser intuitive abilities the program has an extremely good grasp of the relative values of the different features that combine to make up an evaluation of a chess position - not as good as Garry's but good nonetheless. And by being able to evaluate millions of positions each second the program can largely compensate with quantity of evaluation for what it lacks in quality of evaluation.
As is well known there are certain types of chess position that are easier for a human Grandmaster to play than for a program. Conversely, there are certain types of position in which a program can excel, relative to a human player. To a large extent, therefore, the struggle may revolve around the attempts by each player to create the type of position with which he (or it) is most comfortable. It is the winner of this psychological struggle who (that) will have the best chances.
Mig Greengard runs ChessNinja.com
and writes for ChessBase. He directed the website and content of the Kramnik-Deep
Fritz match in October, 2002 and will be providing live web commentary
for Kasparov-Deep Junior.
For comments and questions regarding the website please contact the webmaster.
All content at ChessNinja.com and distributed by ChessNinja.com is copyright Mig Greengard.