An Interview With Anand, Part 2, And Topalov Interview Excerpts
This part is a little less interesting, but there is a brief discussion of Topalov's strongest "second": Rybka 4 on a computer cluster with 114 cores - and not only that but access to an IBM super-computer capable of running 50 trillion floating point operations per second. (If my quick online research is correct, that's about a thousand times faster than even a really souped-up home system.)
Have a look here, too. After a recap of the first part of the Anand interview, there's a brief interview with Topalov. Kind of amazingly to my mind, he boasts about what he takes to have been his superior preparation against Anand, as if having access to a super-computer reflects favorably on his abilities. He also reiterates his triumphalist story about game 1, as if it wasn't just decided by a one-move blunder which Anand claims was the product of mixing up his moves.
Anyway, Topalov aside, the idea of chess preparation moving to the super-computer stage is slightly nauseating to me. I'm no Luddite and I find the progress of opening theory interesting, but is it really the game we play and are trying to understand when Blue Gene blinks on to tell us at depth 55* that our favorite opening variation loses unless we find 27 only-moves in a row? I guess it's not that bad yet, since Topalov, for all his (alleged) dominance in the openings, (allegedly) better nerves, relative youth and better physical condition still couldn't beat Anand, but how long do we have before machine prep renders the gap between those with access and those without unbridgeable?
* Depth 55 is a made-up figure, but can anyone out there tell us what sorts of depths such a machine would reach in a given time period, using Rybka or Fritz or some other contemporary engine on a desktop computer as a benchmark?
Reader Comments (17)
I did a quick plot of Rybka benchmarks from superchessengines.com:
http://picasaweb.google.com/lh/photo/Inp5OxmyZ8YoZ37AVmvjrw?feat=directlink
It looks like it fits the expected exponential behavior quite well. Assuming the supercomputer you mention can explore 1000 times as many nodes as these computers, we get to about 30 ply, compared to the 18 or 19 ply you can get on consumer hardware. Probably that's over-optimistic: things don't usually scale so well. But I don't know much about chess engines.
So it's only finding 5 or 6 only-moves that your home computer couldn't see, not the 27 you're worried about :)
Anyway, things might not be as dire as you think. If Dr. Evil came along and gave us a computer 1 meeellion times faster than a home system, our calculation here gets us to 42 ply--two and a half times the depth of today's home systems. Nothing to sneeze at, but maybe a bit less than Dr. Evil was hoping for.
It will be slightly better, but not nearly as good as you might think. First of all we have the exponential "wall" that makes each new ply much more expensive. Second - the "alpha-beta" algorithm that is the basis of all chess engines does not scale well to parallell processors. It is much happier with faster, not more CPUs. That being said, the extra cpus can of course be used. Deep Blue was massively parallell.
When chess engines compete, a small depth advantage tends to be disproportionally amplified, because that is the most important thing that differentiates engines. A large advantage in engine vs engine rating will not translate into an equally large advantage as analysis partner. I guess Topalov gets mostly a psychological advantage (Anand wad afraid of the engine and avoided lines because of it)
I find his commentary on the openings bizarre as it seemed every commentator consistently remarked on how Topalov was behind in the openings. In the second half, sure, he got the kinds of games he wanted, but that is something very different, right? Of course, nothing is as creepy as his inability to talk about himself and NOT include Danailov in a sentence!
The computer point is well taken. I liked Anand's comment in the second CB interview, saying that a danger of the computer analysis is that it can steer you away from your favorite (and more playable) openings. Regardless, it may help, but Topalov made plenty of mistakes in the mnatch and computers can't really keep you from that last time I checked.
The computer has to search the move tree, so to get 1 ply deeper you need a 10-20 times faster computer.
So if the computer is 10.000 times faster (10x10x10x10) it gets you 4 plys. Ok, it's some advantage, but is
it really that big a deal for preparation ? Also, Rybka doesn't run on the supercomputer, so some heavy
reprogramming has to be done. To me this seems like some strange publicity gag.
I can't speak to a lot of the points made above, but andy, the work on Rybka for at least a year has been to make it run on a cluster. (They're planning to sell online subscriptions for users to use Rybka-cluster, or whatever they're going to call it.) It's not a publicity stunt, or at least there's no special reason to think it is.
I think the speed is 1,000,000x rather than 1000x
Uff Da: I had seen a site claim of an 8-core system something like 47 gigaflops. Maybe it was a made-up number, but if not it's in the neighborhood of 1000x.
I find it interesting that despite Topalov's use of Rybka 4 and a supercomputer, Anand still seemed to get slightly more comfortable positions (for the most part.)
I think there's a danger in the reliance on computer programs for opening analysis. For instance, when I downloaded Firebird I let the engine run all night for a couple of nights to see what it thought the best opening was for each side. The result was the French Exchange Variation. Nothing against the French Exchange Variation, but, well...not my cup of tea, really. In my amateur opinion, Topalov's preparation smacked too much of computer stuff instead of good playable positions.
Anand was also a victim of the computer virus, I believe. In game one, for example, he played a line that I'm sure Rybka or Firebird could defend without breaking a sweat or frying a nerve ending, but for a human, as happened in the game, just get one thing wrong or out of order and the whole defense collapses.
That's just my opinion, I could be wrong...
J.A. Topfke
The answer to the problem of computers in chess is quite simple: switch to go.
Go is hundreds of ORDERS of magnitude more complex than chess, at a minimum.
Now, it is certainly conceivable that computers will beat humans at chess; however, the extraordinary complexity of go will make it impossible for pure memorization of openings to work. That's not even mentioning the problem of local vs. global play in go.
I do like chess, but I've been rather shocked at the growth of chess due to computers. For someone learning the game, the value of a world class sparring partner is invaluable, but that doesn't change the fact that chess at the top is moribund.
Now, one can try to modify chess to make computers less relevant. There are two problems with this approach. First, any small modification will not increase the complexity enough to take computers out of the game. Second and more seriously, it's difficult to find a modification which results in a good game. Modern chess is very delicately balanced. Put the same pieces on a larger board, or change the pieces on the 8x8, and you may get a very inferior game.
Go has the advantage of thousands of years of history which demonstrates its richness, interest, and strategic complexity.
Chess is a beautiful game, and has some elements which are unmatched by go. However, go has it's own intricate beauty; more importantly, it is demonstrably a more successful competive game, where draws are almost unheard of (only a couple rare situations with repetition result in draws)
For the record, I'm a master, who used to play a lot on ICC---in the mid 90's I was one of the top blitz players there. (I haven't played for years---I just enjoy following top tournaments)
I was told the same things about go and resisted for many years, but when I finally took it up, the truth of go's superiority became evident.
Hi Mark,
I'm sure go is a wonderful game, but I don't play chess solely because of its complexity, and I'm sure that's true of many others as well. Enjoy it, but the point of this blog isn't to solicit arguments for its dissolution. This isn't the place for your go-spel.
As for the claim that chess at the top is moribund, I think the match suggests something else. First, both players forgot their prep in different games, so even with insane quantities of computer power at their disposal, human limits still come into play. Second, even with all the computer and awesome mental power applied to the match, five of the twelve games were decisive, and several of the draws were nearly wins/losses as well. Finally, worst comes to worst, there's always Chess960!
Dennis, you're right. I'm showing my age, still thinking of MHz instead of GHz. Sigh.
All modern chess engines use variations of minimax search with alpha-beta pruning. Best case for this sort of algorithm is O(n^d/2) where d is the depth and n is the branching factor. For chess we can assume n is around 30 for a typical position. (Note: most modern programs use other kinds of pruning, but that doesn't really change what I'm about to say -- the ratio is still the same.)
So, to search 2 ply further, we have to spend 30 times as much time or have a machine 30 times as fast. For example, if we assume that each core scales perfectly (which it doesn't), if a single core CPU can search 25 ply in 3 minutes, then an 8 CPU, 32 core system can probably do 27 ply in 3 minutes. Kind of underwhelming, eh?
Always bet on better algorithms over faster machines. Today's programs with null move, futility pruning, razoring, etc., just look much deeper than yesterday's programs.
Dennis,
How's the checkers world championship scene?
Checkers used to be a live game too.
I don't understand the intellectual point of going through the motions of showing one can memorize computer lines 30 moves deep.
It's really disheartening to see someone like Topalov at the top of chess---someone proud that he can get to the top by using the best programs, coupled with some talent and a huge work ethic. Ugh.
Switching to Chess960 would be good. Actually, that alone would be enough to kill the memorization issue, because not even a freak of nature could memorize thousands (10's of thousands?) of lines for each of 960 positions.
Personally, I think the variant where the first 8 moves consist of placing pieces on the back rank is appealing.
Another possibility is assigned openings. This was actually done in checkers before the advent of computers, because there were many lines which led to forced draws; in order to create competition, players were forced to play other openings by lot.
I'd bet on Ivanchuk as WC in this format!
The problem is that players who are currently at the top have zero motivation to approve of any change, and every reason to stick with the game that put them on top.
Fans, on the other hand, are anything but excited by watching a game which they know was decided away from the board, and not even by a human.
Also, people who have the interests of the younger generation ought to be pushing for changes which will make chess a viable competitive game in 30 years or more, instead of a game which is looking at serious problems when computers start finishing off opening lines entirely.. Example: do you think anyone will play the Marshall attack or allow it in 10 years? I bet not, because we'll have a final answer about that opening.
.
MarkL, you're way too pessimistic! Topalov first made it to the upper echelons in the mid-90s, when computers (certainly home computers) stunk in comparison to top players. Second, there are relatively few lines that can be mastered by memorizing 30 moves and calling it a day, and fewer still the players who can keep it all in the memories. Many lines require a lot of understanding, too, and still others that are purely tactical demand a high level of skill, so that if somewhere between, say, moves 12 and 30 the opponent deviates, the memorizer must be strong enough to exploit it.
So sure, there is a ton of material out there, and it is annoying, but it's not enough to kill top chess and not relevant to players at the IM level and down. (Maybe not even to "ordinary GMs" and down - look at the US Championship, for example.)
And to answer your question about the Marshall, I think players probably will continue to play and allow it, but I'm wrong there are still several thousand other variations for players to go to. By the way, if the Marshall gets "solved", it's quite possibly because Black knows how to draw certain pawn-down endings - and that's a trend that goes back to Spassky in the 1960s, before computers were competent to play beginners. Of course, that's a solution that's relevant to very strong GMs; at lower levels, certainly at club levels, those endings will still be losable.
Complaints about the draw death of chess are not new, as I'm sure you know. Capablanca complained of it, and there have been complaints in other eras as well. So far, so wrong, but if it really comes to that expedients like faster time controls, Chess960, and your suggestion of forced openings will prolong its life indefinitely. I think, though, that merely human work isn't enough to solve chess, while even if computers solve it we're incapable of understanding or memorizing more than a tiny fraction of their solutions. There are some 7-piece endings where the strong side can win in several hundred moves, the longest one running 517 moves, I believe. Run through some of those solutions sometime and let me know if you think humans have a prayer of using that information in any practical way. (Next, suppose that one of the "solutions" to some opening involves that 517-move ending, which arises, let's say, after 120 accurate moves. Good luck with that.)
IF your nightmare scenarios all come to pass, though, I might join you in the land of go. But you know the clock is ticking there, too...
Dennis, I am strong enough that I can understand that top players are well into the 20 movess and even 30's with their pet openings.
It certainly is a fact that players can memorize tons of 30 move openings. I know I'm not alone as a fan in finding it distasteful.
To your taste, this kind of chess may be exciting, but I find it the antithesis of what chess should be.
As an intellectual exercise, sure it's fascinating to know whether or not this opening or that is sound.
I will be tickled to death to find out that the Sveshnikov is a forced win for White.
The point is that humans playing chess is completely irrelevant to this inquiry.
As far as go goes, sure computers will probably beat humans, although with the complexity of the game, I do not see how that is a sure thing.
My argument is that computer superiority at go will be irrelevant to tournament play: first, because openings are too complex; second, because any strong player can sidestep complicated opening variations at will. I don't know if you are aware of this, but almost ever go game is unique within 10 whole moves (20 moves as they count in go). I can' t think of an example of two games which are the same for 25 whole moves, and I've played over thousands of professional games.
Another factor is that openings are local (in one corner), but their evaluation depends on looking at the whole board; hence, there is no absolute measure of who has "won" an opening battle just by looking at the moves.
To categorize by computer whether an opening line is favorable would be like having 4 chess games going on in 4 corners, with the Marshall in one, the QGD in another, etc, then deciding how to play alll these openings to get a favorable result.
Sure, a computer will be able to answer this question eventually, but not in a way that will let the memorizers rattle off complete games to win---there is no chance of that happening, ever.
By the way, I just looked up complexity estimates for go and chess. The total number of legal chess positions is on the order of 10^47; the total number of legal go positions is estimated at 10^170. If you count the number of legal games, the disparity is even greater.
Personally, I don't understand why people don't play checkers on larger boards. It scales immediately, like go and unlike chess, and would be incredibly exciting.
It's not in the interest of the younger generation to stick with plain chess. Older players who teach ought to encourage alternatives.
Of course it's not impossible for strong players to memorize long lines. I've done it myself without being anywhere near the world-class level. But I also read, in practically every single event, of this or that player forgetting his preparation. It happened at least twice in the world championship, for instance. I also expressed my lack of enthusiasm for this sort of thing - though not because I'm worried about the death of chess. Anyway, I'm not interested in making this a forum for chess variants and other games, so let's close this discussion, please.