Knowing your opponents

[Uri Blass]

When a human plays, he/she often modifies the playing style based on his opponent, taking into accounts an opponent's weaknesses. I know in developing my program, I have come up with versions that do well against some computer opponents at the expense of other opponents. The UCI spec does not feed the name of the opponent to the chess engine. I think in most tournaments programmers are able to fine tune values between matches, change books, etc, so they could tune to the opponent. Would it make sense to expand UCI to feed the opponent program name to the engine to help automate this? Or do you feel that such tuning is "gaming" the system?

Mark

I think most humans do not modify the playing style based on the opponent.

I also doubt if it is possible to earn much by modifying the playing style based on the oponnent.

see the FRC rating list
http://www.computerchess.org.uk/ccrl/40 ... t_all.html

usually the performance of programs after a match of 100 games is very close to the rating of the programs.


When I click on rybka3 and look at the perf column I see numbers between +32 and -50

http://www.computerchess.org.uk/ccrl/40 ... a_3_64-bit


clicking on shredder12 gives numbers between +28 and -37

clicking on Naum4 gives numbers between +86 and -69

looking at many programs I found no number that is bigger than +86
and usually the numbers are even smaller than 50.

Uri

Are you saying that there is not very much intransitive between most players? In other words, if A beat B and B beats C, then we can also expect A to beat C.

If that is what you are saying, I agree. I am not saying there are NO intransitive relationships, but I believe they are fairly weak and almost certainly much weaker than people imagine. Most people claim some kind of intransitive relationship after playing only a few hundred games and trying to draw conclusions where the error margins are too high. I think a lot of people would jump to the conclusion that a 60 out of 100 result is a proof of superiority when it isn't.

But I do stress this - I think there is some intransitivity between certain programs. If your program is particularly good at something and your opponent is particularly bad at that same thing, it MIGHT make your program score lower than your ELO ratings would predict.

ELO is a great system, but it is just a model and it's not a perfect model. It assumes a 1 dimensional relationship space that defines chess strength.

60-40 is practically a significant result in FRC CCRL 40/4 list

You are not going to find a single case in the CCRL list when the weaker program(based on rating) scored 60 out of 100 games or more than it.

I can add that in all cases when the result is 60-40 or higher the winner is at least 23 elo stronger based on the rating list.

Uri