Author: Uri Blass
Date: 19:18:48 02/12/04
Go up one level in this thread
On February 12, 2004 at 09:48:20, José Carlos wrote: >On February 12, 2004 at 06:42:47, Uri Blass wrote: > >>On February 12, 2004 at 06:16:53, Daniel Clausen wrote: >> >>>On February 12, 2004 at 06:03:37, Uri Blass wrote: >>> >>>[snip] >>> >>>>The main point is that programs should not evaluate unclear position as winning. >>>> >>>>Black has a big advantage without capturing the pawn so there is no reason to >>>>capture h3 and get unclear pawn endgame. >>>> >>>>evaluating Rxh3 Rh2 as a win for white without search is a msitake but >>>>evaluating it as a win for black is also a mistake that cause Crafty to miss a >>>>simple win by not capturing the pawn. >>> >>>And again I'm reminded of the posts some months ago, in which someone suggested >>>to have some kind of an uncertainty-value together with a score. In this case, >>>the machine could decide that it's better to play a move with score +1 (5% >>>uncertainty) than a move with score +2 (70% uncertainty). >> >>No >> >>score should consider the uncertianty inside of it. >>The score is in pawns but can be also translated to an expected result. >> >>Uri > > I do an uncertainity level in Anubis. I use it mostly to allow/disallow >pruning or stand-pat, but it also works fine for comparing values. It's tricky >and should be used carefully, but it can work. > > José C. My point is that you need to decide which position is better so you get some final score to compare. If you claim that +1 pawn is better than +2 pawns because you are less sure about the +2 pawns then +1 and +2 are not the final scores of the position but only numbers that are used to evaluate the position. I agree that it is better to evaluate uncertainty but this should be used both for the final evaluation and extensions or reductions. The final evaluation can be in probabilities or expected result but it is possible to translate pawns to probability so there is practically no difference if you say final evaluation of +2 for white or final evaluation of 84% expected result for white. 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.