Author: Uri Blass
Date: 22:24:38 12/28/01
Go up one level in this thread
On December 28, 2001 at 18:27:34, Tom Kerrigan wrote: >On December 28, 2001 at 17:36:53, Uri Blass wrote: > >>I agree that it is not a good idea to optimize when you have not enough >>positions so if you have only one position or a very small number of positions >>with white knight at b8 then it is a bad idea to try to investigate changes in >>the piece square table for knight at that sqaure. > >I doubt you can find a test set large enough that every term of an eval function >is represented thoroughly. I mean, how many positions are there with white >knights on b8? > >You can do what the DT guys did--make a test suite from GM games, search the >positions deeply, and discard the ones where the deep score didn't agree with >the static eval by a certain margin (indicating that something tactical >happened). Optimizing against these deep scores has the same local maxima >problem that you were all worried about a few posts ago, though. Ignoring tactics also does not seem to be a good idea because if there is a good chance that tactical tricks are going to work then it should be considered in the evaluation. > >I have also thought of finding positions where you want your program to play a >particular move. The program searches the position, and then it searches that >move, and it takes the derivative of the evaluation function at each endpoint >node and changes the weights of the differing terms until it plays the desired >move. This scheme might work well, and AFAIK nobody has tried it, but it would >probably screw up for some reason that I haven't thought of. > >>It is possible that they are useful for chess and the problem is that you need >>first to waste hundreds of years of computer time >>that is impossible to do by one programmer. > >If you have to spend hundreds of years doing it, I'd argue that it's not useful. It can be useful because a team of some hundreds of people may waste only one year. > >>I agree that by one change you cannot do it but it is possible that by many >>small changes it is possible to do it when the problem is to find the right >>changes. > >Yeah, but you have no proof of this statement. Why do you believe it, exactly? > >>The case that you describe is a case when the program has not the right >>knowledge to explain the position. >> >>It is not the case for the top programs. > >I think you're giving WAY too much credit to top programs. > >>I found that Junior5 lost against Fritz3 in the ssdf games because it did not >>know that a white knight at h8 is bad and it's knight was trapped at h8. > >There are positions where having a knight on h8 is perfectly reasonable. If all >you do is jack up the penalty for a knight on h8, you will end up with a program >that will lose the game trying to force its opponent's knights onto h8. What you >really want is a knight trapping term, which no amount of evaluation weight >optimization will give you. > >-Tom I agree that it may be better to have a trapping term but even when the program does not have it then I tend to believe that changing the piece square tables may be productive. Of course if you change it too much the program may lose games in the way that you describe but my observation is that practically there are more games when programs are losing because they put the knight at the corner than the opposite case. Uri
This page took 0 seconds to execute
Last modified: Thu, 15 Apr 21 08:11:13 -0700
Current Computer Chess Club Forums at Talkchess. This site by Sean Mintz.