Author: Robert Hyatt
Date: 20:33:08 07/29/99
Go up one level in this thread
On July 29, 1999 at 19:44:55, Amir Ban wrote: >On July 29, 1999 at 09:29:32, Robert Hyatt wrote: > >>On July 29, 1999 at 08:25:58, Chris Carson wrote: >> >>>On July 29, 1999 at 07:16:32, Amir Ban wrote: >>> >>>>On July 28, 1999 at 18:16:24, Dann Corbit wrote: >>>> >>>>>On July 28, 1999 at 17:50:51, Kristo Miettinen wrote: >>>>> >>>>>>The position is the opening array, all pieces in their initial positions. The >>>>>>explanation about the eight pawns makes sense, intending to steer Crafty into >>>>>>open waters (on the assumption that the opponent is human?) >>>>>> >>>>>>I was looking into this on a whim, as I use the advantage of White in the >>>>>>opening position as my quantum of positional value (on which scale the value of >>>>>>a pawn is 6 quanta for me). >>>>>Here is the C.A.P. record for that position. >>>>> >>>>>rnbqkbnr/pppppppp/8/8/8/8/PPPPPPPP/RNBQKBNR w KQkq - acd 15; ce -7; pv e4 e6 Nf3 >>>>>Bb4 Nc3 Ne7 Bc4 Nbc6 O-O O-O d4 Bxc3 bxc3 Na5 Bb5; pm e4; id "C.A.P. 4028"; >>>>> >>>>>I bet you never knew crafty was French. >>>>> >>>>>Crafty thinks it is behind by 7 one hundredths of a pawn. This is obviously >>>>>conservative because white has a tempo at least. But I don't think that it is >>>>>grossly inaccurate. >>>> >>>>A correct evaluation is one that matches the winning percentages of the >>>>position. I think white has about 54% in serious play, and if so the evaluation >>>>should be about +0.20. >>>> >>>>Amir >>> >>>Amir, >>> >>>Interesting point. If I read you correctly, the "Evaluation" should match >>>the winning changes. This is not the way most programs "Evaluate" a position. >>>Granted that a higher "Eval" by a program should mean a higher "Chance" to >>>win, it is normally not a "Percentage" based on results. >>> >>>I have thought that this might be a better method of "Evaluation", some >>>programs do use a "Percentage" (Crafty) for opening book moves, but not >>>for middle game or end game positions. >>> >>>Any thoughts on how to incorporate "Percentage" into the "Evaluate" function >>>of a program (knowledge)? Perhaps a "Percentage" "Evaluation" for positions >>>and endgames as a part of the learning (Crafty might be able to do this) >>>would be useful. Any comments? >>> >>>Best Regards, >>>Chris Carson >> >> >>I disagree. Evaluations are not 'absolute' any more than FIDE Elo ratings are >>absolute. The correct evaluation is the one that lets you _win_ 54% (or better) >>of the games from the opening position. Whether the starting score is +1.00 or >>-1.00 is immaterial so long as you choose the best move(s) by using those >>scores... > >This is to answer several posts in reply to my original comment: > >Evaluations need to represent winning chances in some way, or else there's not >much use for them. It's true that the object of all this is to play good moves, >but to say that is to beg the question of how to evaluate positions so as to >play good moves. > >There are many ways to do this mapping. Obviously you can multiply the eval by a >factor to choose your scale, and you can also add a constant without changing >much, but an additive constant is suspect if you define a 0 evaluation to be >equivalent to a draw or 50% outcome. As long as your mapping is monotone in >winning chances, and your draw score is calibrated correctly, it's good. > >Practically, almost everyone agrees on scale by calling a pawn advantage about >1.00 (on average). Assuming some smooth mapping (there are exponentials that are >natural to use), to say that 54% maps to +0.20 is not so arbitrary as some >commented, though if someone insists it's +0.15 or +0.30, I won't argue. A minus >score, though, obviously doesn't fit because it has the wrong sign. > >The problem with having incorrect evaluations (not monotonic, or wrong sign) is >obvious with some thought: the program may prefer a bad position to a good >position (which always involves playing a bad move ...), or may accept a draw >when ahead. > >Our evaluations may be bad regardless, because our knowledge of the game is >incomplete, but there's no reason to accept a logical inconsistency in the >evaluation. > >When I talk about winning chances I'm not referring to any specific database >information that is available, but about an objective (and usually unknown) >outcome of the position. > >Amir I wouldn't begin to argue... but two points... (1) a minus score for white's first move can be considered bad... in that we all agree that white has better winning chances than black. So for those that want to take a program's 'evaluation' of a position as gospel, then - is confusing at best, and misleading at worst... (2) a minus score for white's first move can be reasonable, if a program has some asymmetry in the eval. IE I use Belle's idea of '8 pawns is bad' and until one is exchanged, crafty gets a small penalty whether it is playing black or white. And if white, it is possible that it can find a PV where no pawn exchange occurs and yet it can't find a developmental/positional advantage to offset that 8-pawn penalty. I do, on occasion, look at positional scores and ramp down or ramp up if I believe the number is 'out of line'. But I am more interested in the move the eval leads it to play, because that is the critical issue in winning games... I think most programs are pretty naive positionally, mine included, and often see scores that make little sense. For any program...
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.