Author: Marc van Hal
Date: 17:13:00 04/29/03
Go up one level in this thread
On April 29, 2003 at 15:17:32, Amir Ban wrote: >On April 28, 2003 at 11:30:08, Charles Worthington wrote: > >>That's an interesting question but I have one of my own. Is not the engine's >>choice of continuations based soley on its eval of the line? Many times I have >>seen Shredder follow its over optimistic evals to a dead draw as white. I would >>rather my program know where it really stands and whether or not it has >>realistic winning chances. Also...I personally would like to know where it >>stands as well. There is nothing quite so frustrating as to have an over >>optimistic eval turn on you and bite you. I have seen numerous situations where >>The King has an eval of +2 against Deep Fritz's -1 and it RARELY works out in >>The Kings favor. It's hard for me to muster faith in an engine that is clueless >>about it's own position and it is the one thing that keeps me from giving >>Shredder a fair shot at becoming my main engine. Enough times of seeing Fritz >>pull the rug out from under my Shredder eval has spooked me I guess. I just do >>not see how it is possible to handle a position properly when you cannot even >>eval it properly and in fast blitz time controls (3+2), I might also add that >>the "optimistic" engines almost always perform more poorly than the well >>balanced engines. The truth seems to perform better than the lie in this case. >>Of course this discussion is about self-inflicted optimism by tampering with the >>parameters so it seems even more extreme. You are taking an already-optimistic >>engine and furthering it's optimism. It would likely take many thousands of >>games to come to a conclusion as to which works best. Honestly though, I think >>that in this case having your program lie to you (and itself) will not prove >>beneficial. The program will play better if it has a full understanding of its >>actual chances and not imagined ones...Optimism causes the program to >>overestimate it's chances and play far too aggressively (ie. unwarranted >>sacrifices) in certain positions where caution is warranted. Against a human >>this may prove beneficial but against a program firmly grounded in reality it >>may prove fatal. So, like in life, I think the truth has to be better than the >>lie. And I think that extensive testing of these settings would show that, more >>often than not, the lie would come back to bite you. >> > >I agree. > >I believe overvaluation is the most common reason for engine losses. It's much >more common than undervaluation, a less fatal problem. This is especially true >in tactical situations, where the program with overvaluation seems not to see >tactics, because its search is meaningless. > >Amir I have created a setting in Chessmaster wich also is to optimistic. In qucik games it sacrefice for almost nothing. But with longer time controls it gives the right evaluation. Much comparable with the other comercial programs Means it has to be played with at least 60 min/sec or preferable tournament games to get the best score I only write this to bring you to the idea that it's better to test with longer time controls. You might have tost the strongest settings away because it did not preform well in blitz games. But is it not so that in tournaments and the SSDF the program only plays with tournament control? I even tried to triple the valeu of all pieces to solve some null move problems. With the understanding that the computers became a lot faster and the valeu has stayed the same so it might miss the right move order because the diference between candidate one an two was too small. Hoping it would prune more agresive. But I can't say that it realy worked so far. Still I think it my be worth checking. But the real thing to focus on ofcourse remains knowledge. Especialy the problems Chessprograms find against the Botvinik setting of Chessmaster. ( Or in anti computer games Though I don't agree that every game posted as one is one.) Intresting about this is that the winning games against it manytimes are played in the style of the personalety it beats With humans it is most of the time the oposite But I also think it is ok to not exchange pieces to easely. Only when the endgame is clearly better but then it again needs lot of knowledge. But I believe you already tried this Because junior7 is even pesimistic about winning end games. Especialy pawnstructures are crucial and not only in endgames Also in openings and middle games Junior7 is not bad only the search methode is to risky. Just at the last moment it chances it's mind playing a move which is worse then the move it pondered all the time. I also think the way Mobilety is measured by most programs is not intirely corect And I think space is confused with mobilety Space can be fine but also can be a weaknes when Black still can develop easely. and after that setting up a counter attack. Like for instance many programs give -0.50 and above in the Kings Indian Which is not a realtic evaluation. Junior is the exeption here though. But you also have to take in acount the mobilety of the pawns chain. and this kind of knowledge is not inserted deeply enough. But I can understand the problems to program that. Which looks natuaral for a human in that perspectiv is hard to evaluate for programs It would be nice to see a program play very well in clasical closed games though. Marc Marc
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.