Rajlich, Research, Rybka

ArnieChipmunk
CM ArnieChipmunk
|
0 | Chess Event Coverage
An interview with the Rybka team on the future of computer chess

Computer chess is everywhere. Chess engines reign supreme. Their expertise ranges from endgame tablebases to amazing tactical resources in the middlegame to surprising opening novelties in well-known variations. There is only one aspect of chess that is still mainly dominated by humans: the understanding of the fundamental opening moves. This is often too easily dismissed by defenders of computer chess. Computers are simply given enormous opening libraries to compensate this lack of understanding. In human vs. computer matches, we then see things like Kasparov playing (against Deep Blue) 1.d3 to avoid the computer's book knowledge, or humans in general avoiding sharp opening theory. Intuitively, this doesn't feel entirely 'right'. One of the most basic aspects of human chess knowledge is made unimportant by giving chess computers these huge libraries. This was one of the reasons why I recently wrote a deliberately provocative article on Chessvibes called 'Turn off those opening libraries!'

I received a lot of feeback on this article, also from the chess computer world itself. For instance, I was told that this month, Grandmaster Jaan Ehlvest would play a match against the strongest chess engine in the world, Rybka, without an opening library. The match was privately organised by IM Larry Kaufman, who is a member of the Rybka team himself.

Now this was great news! Not only would humans have a much fairer chance, but it would also be extremely interesting in principe to see what the engine would 'think' of some basic chess opening positions. Finally, we would get a glimpse of the Truth on these fundamental human concepts. It would be a bit like Leonardo giving his opinion on the latest Boeing 777, like Shakespeare analysing Nabokov's Lolita, or Bach interpreting Kind of Blue by Miles Davis. Or so I mused until the match started last week.

By the way, it wasn't the first match Ehlvest and Rybka had played. In March, they had already played a match at pawn odds which Rybka had won very convincingly. But this would be different, of course. Or would it?

Again, it turned out Rybka was simply too strong for Ehlvest. True, Rybka did have a small (3 moves) opening book, but still the result was quite convincing. Everything about the match, including the games and comments by the Rybka-programmers, can be found here. I would also like to draw your attention to Ehlvest's own blog, where he comments on the games. In general, he put up very tough resistance, but not surprisingly, he was tactically too often outwitted.

Of course, the match raised more questions than it answered. So instead of trying to figure it all out myself, I decided to interview the inventor of Rybka, Vasik Rajlich, and the match organiser, Larry Kaufman about the match, its implications, and the future of computer chess.

Mr. Rajlich, congratulations on Rybka's victory! First, I would like to ask you some questions about the match in general. It is often suggested that humans vs. computer matches only serve commercial purposes and that they are not really ?¢‚ǨÀúmeasuring' anything. This match, however, seemed perhaps a little different, because of its unusual and provocative setup. What, if anything, was ?¢‚ǨÀúmeasured' in this match? Was there a special purpose, or was it just ?¢‚ǨÀúfun'?

VR: Actually, Rybka team member IM Larry Kaufman organized this match, so he would be better qualified to answer this question. From my point of view, these matches are fun and also a chance to get some new stimuli. Humans do play differently, and this can give us some new ideas and insights. For instance, Ehlvest seemed to be relatively good at getting small advantages. Even engines which are on paper much stronger than Ehlvest have much more trouble doing this against Rybka.

Vasik Rajlich

LK: For me, there were several purposes. First, I wanted to prove to doubters that any future matches against humans (at least humans we can afford to play) need to be at explicit handicaps if there is to be any doubt about the result of the match. I think this point was proven here. Second, I wanted to see if I could create an opening book only three moves deep that could allow Rybka to survive with Black without getting into serious trouble against a renowned theoretician and top class player. At least in this match, I was successful. Third, I wanted to find out whether Rybka would be able to create winning chances from the inferior positions she would normally get out of these openings as Black. In general, the answer to this question was no. Rybka's wins came in two games where she equalized completely in the opening and in another where White blundered; in the games where White got an opening advantage he was able to draw.

There was a special possibility in this match for Ehlvest to take ?¢‚ǨÀúrest breaks'. While this definitely solves some inevitable concentration-problems for the human player, still there were quite a few gross errors from the Grandmaster that had a decisive influence on the result. Were you in fact ?¢‚ǨÀúhappy' with these blunders (because Rybka scored a bit better than it perhaps deserved), or do you think they ?¢‚ǨÀúspoiled' the match in a way?

VR: We don't have a good way to quantify it. In general I think Ehlvest did quite well to limit his blunders. 33. Rxc7 in game 1 is a mistake which no decent chess engine would make, but for a human it is surely difficult to calculate the consequences of 33. Re2. In game 4, Ehlvest made a number of blunders in a deteriorating position where his attack was gradually petering out ?¢‚Ǩ‚Äú again, no engine would make such moves, but for a human squeezing the maximum drawing chances from that position is going to be extremely difficult. Anyway, in general, mistakes in human chess can be quite a harsh reality. It's painful to destroy your position with oversights, and it's not always the greatest feeling to watch your opponent do the same.

It is always difficult to measure the machine's strength when it's playing humans. Suppose Rybka normally plays at 3150 strength. Do you think it was a big handicap for Rybka to play without an extensive book? In other words, could you say it played, for example, at 3000 or less now?

VR: The first comment here is that performance ratings in a six-game match are going to have a huge variance. One more half-point and the performance rating goes up by 100 Elo; one fewer half-point and it goes down by 100 Elo. Larry did various experiments and calculations before the match and in computer vs computer play, the limited opening book seemed to hurt Rybka by around 75 Elo. Arguably, the damage would be worse against a human, because unlike Rybka's computer opponents, Ehlvest is aware of the handicap and can try to exploit it. In the games of this match, Rybka did not seem to suffer hugely in the openings. There were little issues here and there - in game 2, she did fail to equalize, and Ehlvest could stay in control the entire game. It's hard to quantify the effect of all of this, especially with just six games. Keeping in mind that even a half-point more would mean a 100-Elo higher performance rating, I'd probably say that the effect of the missing opening books was in the range of what we expected.

What do you try to learn from these kind of matches, and are there going to be concrete changes to the software because of what you've learned?

VR: What we learn from this is more of a general, big-picture nature. We see Rybka react in a new environment, and this helps build our picture of what she can do, where she needs work, etc. We're not planning any single specific change as a result of this match.

What are your plans of organising other matches of this kind? Do you think it is more interesting than matches ?¢‚ǨÀúwith book'?

VR: Larry is the organizer of all of these matches. He has already arranged for a match vs Joel Benjamin in August. This match will be played with pawn odds and a longer time control. The rest of the conditions are still being discussed. As for ?¢‚ǨÀúwith book' and ?¢‚ǨÀúwithout book', it's pretty clear to me that any match between a human and Rybka where Rybka doesn't give a material handicap must be played with a tiny Rybka book just to keep things reasonably competitive. Even this doesn't seem to be enough, so probably, we will look in the direction of material handicaps now. We'd like to have matches where the grandmasters have reasonably good chances to win.

Now I would like to ask some questions about Rybka and her opening books. In this match, Rybka only had an opening book of 3 moves. Why still this small opening book? Suppose Rybka had an opening book of 0 moves, would we then have seen any ?¢‚ǨÀúembarrassing' first moves?

VR: This was done for three reasons. First, to introduce variety. Second, to prevent Ehlvest from preparing special variations at home. In theory, Rybka is deterministic and will always play the same move in a position, although the multi-processor implementation and the time control can lead to different moves when the different moves are close enough in score. The third reason was to keep Rybka from entering some sharp theoretical variation where Ehlvest could use his theoretical knowledge to beat her.

As it turned out, Rybka's ?¢‚ǨÀú3 move book' seemed to contain only ?¢‚ǨÀúoff-beat variations' like the English Defence (1. c4 b6), 1?¢‚Ǩ¬¶Nc6 against 1.d4, Scandinavian Defence against 1.e4, and some obscure lines of the Queen's Gambit and Slav Defence. I must say I had hoped to see Rybka go for a Sicilian or perhaps re-invent the Marshall Gambit! Why did your team choose these obscure openings? Was it just a coincidence, or was there a ?¢‚ǨÀústrategy' behind it (for example that Ehlvest couldn't prepare for it very deeply, or the Grandmaster could put less trust his ?¢‚ǨÀúgeneral knowledge' of the mainstream openings)?

VR: This was definitely not a coincidence. Expecting Rybka to make it through the minefield of some modern theoretical variations like the Najdorf or Marshall, rejecting along the way various logical moves which theory over time has proven to fall short, would be suicide. In some of these theoretical variations, opening theory is incredibly high level ?¢‚Ǩ‚Äú let's say (as a wild educated guess) 3700 Elo. You have many grandmasters, working with computers, testing their ideas in over-the-board games, returning to their pet variations again and again over months and years, evolving with the changing fashion and taking note of the discoveries of others, etc ... A 1-minute Rybka search simply cannot compete with this.

Still, I must admit I find you surprisingly negative about the powers of the computer in such an early stage of the game! Remember the last match game of Deep Fritz - Kramnik: the machine made a very surprising opening move (10.Re3 I think it was) there on its own power, which would probably seem bad if it were played by a human, but which turned out pretty strong when played by Deep Fritz! Don't you think such a scenario would also be possible a few moves earlier?

LK: The openings have been researched with computers searching for hours and aided by grandmasters. Why should we expect even a very powerful computer searching only for a minute or two to find moves as good as those that have been found that way?

Did Ehlvest know beforehand that he would meet only 'offbeat' lines or was he ignorant about your book strategy? In other words, do you think he also prepared for regular Sicilians and mainstream-stuff like that as well?

LK: He said that he was warned that we would probably avoid mainlines. In any case he would have no need to prepare for mainlines, because if we had played them we would surely have made some inferior move before leaving known theory, and as a top GM he would surely know how to react to that. He said he did expect Scandinavian, but did not expect 3...Qd6 (which was suggested by Jeroen Noomen [the Rybka opening library expert - ed.]). He did not anticipate any of my replies to the other first moves. In general, he played the opening 'straight', pretty much as if he were playing Kramnik or Anand.

Well, we've seen that Grandmasters can still be beaten with these obscure lines! Do you think this shows that strong grandmasters should concentrate more on these lines, rather than study the main topical variations until move 30? And what about the book programmers of Rybka? Do they also analyse ?¢‚ǨÀúminor' openings for the Rybka book?

VR: I think the way to phrase this conclusion is that such lines are not so bad that Ehlvest can beat Rybka in a six-game match with them. This match doesn't really tell us that main lines are not better, although this is itself an interesting discussion. There are top GMs who are a bit more original in their opening choices.

Finally, I have some questions about computer chess and opening theory in general that I'm curious about. I remember some twenty years ago I bought my first chess computer. It didn't contain any opening book theory. After 1.e4 it thought for a while and then it always came up with 1?¢‚Ǩ¬¶e5. After 2.Nf3 it played 2?¢‚Ǩ¬¶Nf6 or sometimes 2?¢‚Ǩ¬¶Nc6. Good moves, but it never came up with the Sicilian! Now, twenty years later, my new Rybka engine still doesn't consider the Sicilian Defence after 1.e4 in its first five alternatives. How do you explain this, and do you think this will change some day?

VR: This is a matter of taste. In the Sicilian, black concedes various ?¢‚ǨÀúclassical' trumps such as space and development, and Rybka simply values these more highly than they deserve to be valued in that position. As Rybka's understanding of chess improves, I hope that she will give a slightly more appropriate evaluation of the Sicilian. However, this is an incredibly hard thing to do when you are looking at this starting position ?¢‚ǨÀúfor the first time'. It's also not that important, because black does have alternatives defenses to 1. e4. In other words, I think that there will always be a legitimate scope for style in chess play.

Do you, as a programmer, spend time at all trying to adjust your algoritms to this 'appropriate evaluation', so that the engine will see the value of, for example, 1?¢‚Ǩ¬¶c5 over 1?¢‚Ǩ¬¶e5? In other words, what do you think is necessary for Rybka to ?¢‚ǨÀúappreciate' the strength of the Sicilian Defence?

VR: I myself would never tune Rybka to specific positions. This is just not the way I believe in working. If Rybka ever does change her mind about her opening preferences, it will be on the grounds of various general principles. For instance, in the Sicilian, black typically gets an extra central pawn, so the evaluation of this type of pawn structure might change a bit and then Rybka will prefer the Sicilian.

A final philosophical question. Do you think computers will ever be able to tell us which move is better: 1.e4 or 1.d4, or perhaps, as Jonathan Rowson once half-jokingly suggested, 1.a4?

VR: Computers definitely can be (and in fact already are) an essential tool in answering this question, but I don't see anybody in computer chess working in the direction of directly answering this question.

Larry and Vasik, thank you very much for this interview and good luck with future matches!
More from CM ArnieChipmunk
Why chess will never be popular

Why chess will never be popular

In praise of draws

In praise of draws