Author: Robert Hyatt
Date: 15:14:06 12/04/02
Go up one level in this thread
On December 04, 2002 at 03:43:09, Bas Hamstra wrote: >On December 03, 2002 at 22:45:16, Robert Hyatt wrote: > >>On December 03, 2002 at 18:59:20, Jon Dart wrote: >> >>>On December 03, 2002 at 17:16:58, Robert Hyatt wrote: >>> >>>> >>>>the idea is to do anything you can to exclude moves. Because below each move >>>>you search, you grow a tree, and that is expensive. If alpha is 0.00, and >>>>the current material score is -15.0, then capturing a pawn is not going to >>>>help, and avoiding that capture avoids the tree below it. >>> >>>I've noticed that Crafty uses the material value as a basis for calculating >>>futilty. In other words, if material + SEE gain + safety margin < alpha, you cut >>>the move (I do something similar at present). >>> >>>What Mike is suggesting is to use the total position eval, which is also >>>something I've experimented with. So: if eval + capture + SEE gain + safety >>>margin < alpha, you cut. If you have large positional scores, this is possibly >>>safer (the "safety margin" in Crafty is only 1 pawn). >> >>If you follow the logic carefully, this is also what crafty is doing. At >>the top of Quiesce() I grab the positional (stand-pat) score, and if this is >>> alpha, then I set alpha to that value. All the pruning is then based off >>of this positional (material+positional returned by Evaluate()) score plus the >>capture gain, plus a fudge as removing a piece could increase the score but >>probably not by a lot... > >It is not the same IMO. I think Jon means this: suppose Eval is below alpha. >Then in Crafty you calculate Delta=Alpha-Mat-Margin. However more accurate might >be Delta=Eval-Margin. IE you prune based on Eval rather than material. >I tried this, and the result is it will throw out less captures while I see no >noticable improvement in accurracy. > >Bas. I am not quite following. I prune on the eval value _before_ the capture, adjusted by the value of the captured piece plus an "error estimate". I think my error estimate is very safe except for one circumstance, that where the last piece is removed and it can allow a pawn to run...
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.