Slashdot Mirror
Chess - 2070 CPUs vs 1 GM
jvarsoke writes "ChessBrain.net broke the world's record for 'largest number of distributed computers used to play a single game' by holding a chess match between Danish GM Peter Heine Nielsen and the equivalent of SETI@home (which similarly, has some people looking for a Mate). 2070 CPU's from 56 countries aided Black by running the chess program Beowulf, including a couple of University clusters. Their supernode ran Linux, and MySQL. The game was relayed by FICS. Results can be viewed here(1) and here(2)."
It's too bad that chess has become a matter of memorizing a series of opening moves rather than a game of strategy. GMs don't even play to mate anymore, they just play out an opening move and whoever has the upper hand at the end takes the game.
This kind of 'training' will eventually catch up to the chessmasters. Computers can do this MUCH better than a human ever could.
I have been pwned because my
The problem with this is that it seems to assume that chess is a difficult problem. It isn't. Modern chess algorithms are really simple search-and- prune systems, relying on the computer's immense number-crunching ability to overcome the more heuristic human mind. Unfortunately, this isn't very interesting. What's the point? We know that computers can search faster than a human. See: Google. All these projects (DeepBlue, Fritz, this) accomplish is trivializing the game of chess, which is rather sad. Now, I'll be really annoyed when Go programs start improving to a 'decent amateur' level...
To give credit to Danish GM Peter Heine Nielsen, I would have to say if there were only 2069 CPUs then he might of just won... :P (J/KING)
More interestingly, would the ChessBrain.net team would of won with more CPUs?
With two people, there are some elements that can not be programmed into a chess game. I remember in high school playing chess, there was a differance between playing a math academy team and a school best known for its basketball program. Expectations were different, the pressure was different. I remember the pressure of the state finals. There is the look the other person has, almost like poker. Can I bluff this person? Can I trick this person? What about the clock, can I manipulate that to cause an emotion in the other person.
Maybe Spock can play a PC and have no differance in quality of play. But I prefer humans.
Rosco: "If brains were gunpowder, Enos couldn't blow his nose."
I want to see this cluster take on IBM's system!
Has anyone ever written a system by which a large number of average chess players could collaborate to play a single game? The individuals could vote for the best move, and the majority would rule. Would a group like this be able to beat a high ranking player?
While this is true and I definitely agree with your sentiments, it should be noted that players at the GM level spend a considerable amount of time in preparation for their specific opponents. They spend countless hours analyzing the games of the person that they will be playing tomorrow. In this sense, a computer will and already is better facilitated to analyzing styles/methods/openings/etc. to play against a human than any human being is capable of. A computer could easily go through every game someone has ever played and at least know which opening(s) to present and which variations based on statistics. While a human might have some intuition, the computer should have a more comprehensive view of this.
Why do you believe that over a 5 match series the GrandMaster will win handily?
If ChessBrain refused some normal traps that computers normally fall for, then could it be the case that the computer is better than you realise. What if the drawn match was a bad one for the computer?
If I point out that you are incorrect, making me a foe does not make you any more correct.
There are approximately 35 moves per position in Chess (average value). Thus, the branching factor of the search tree is ~35 with a simple min-max search. Assuming that the program is always picking the best move to search first -- which is obviously not systematically the case -- alpha-beta pruning allows us to get a branching factor equal to approximately the square root of 35, that is: close to 6.
Assuming that 2070 CPU are able to do the calculations 2070 times faster than 1 CPU -- which, again, is not the case -- it appears that the resulting supernode is able to 'see' up to 4 or 5 half-moves deeper than a single CPU in the same amount of time:
6^4 < 2070 < 6^5
It doesn't seem to be *that* useful. For most strategical positions, thinking 5 half-moves deeper just doesn't make any difference. Game 3 of 'Kasparov vs X3D Fritz' is a good example: I'd be willing to bet that 2070 X3D Fritz playing together would have lost the game the same way, since the serious troubles caused by the pawns diagonal are still far beyond the resulting analysis depth. (Well... At least, I think so. I'm not a Chess expert!)
Anyway, this is quite an interesting project. I hope to see it grow up in the future.
-- Arnauld
The problem with Slashdot memes is that YOU INSENSITIVE CLOD!
Well, you got to understand. Closed source for a distributed project is a good thing, because people can tamper with the client and make it give bad results, or the results you do not want.
You raise an interesting point. Most commentators believe that chess in the limit is not a forced win for either side. Current belief is that if chess is played (by optimal players) to the end, it's a draw.
That makes chess fundamentally just a zero-sum game somewhere between tic-tac-toe and poker (two other zero-sum games when played by perfect players). Chess Grandmasters typically can play for the draw if they want, but that makes the game like tic-tac-toe. Human players are fallible, humans instead play for the win and play the player not the game itself. (Consider this: Why do people switch openings?)
Hence, chess is also like poker, insofar as in both games all upsets among great players are psychological.