Author: Bas Hamstra
Date: 01:08:03 10/06/99
Go up one level in this thread
On October 05, 1999 at 12:05:41, Robert Hyatt wrote: >On October 05, 1999 at 05:14:23, Bas Hamstra wrote: > >>I have been thinking about "Delta pruning". The way I saw it (a while ago) in >>Crafty is like this: >> >> Delta = Alpha - Material - Margin; >> >>Then for each capture: >> >> if(CaptureValue < Delta) skip this capture. >> >>Now I would like a raction to the following statements: >> >>(Material - Margin) is a cheap estimation of the current eval. If positional >>bonusses get frequently more than a pawn (or even more) "Margin" should be set >>to near the maximum the positional part of the score can be, or else this >>estimation can become very inaccurate and introduces much errors. Reaction? >> >>If positional bonusses get frequently much more than a pawn, maybe it is more >>accurate to use this: >> >> Delta = Alpha - Eval; >> >>Because with one move the Eval won't change *that* much. Reaction? >> >> >> >>Regards, >>Bas Hamstra. > > >You are overlooking something _important_ in crafty. the "alpha" value >_is_ the real evaluation for the current position, and includes both the >material score _and_ the positional evaluation. So my test is safe in that >I am betting that this single capture can't add more than 100 positional >points. It can be wrong, such as when the last piece is removed, because >maybe a pawn can run then. But overall it is very safe and I haven't found >cases where it screws the search up at all.... Please help me understand, for now I get lost. The Alpha value is the real evaluation? I just looked in Crafty's quisce and Alpha is just Alpha. It seems to me the current position can have a totally different Evaluation than Alpha. Only if the current position is *better* than Alpha, then Alpha = Eval, says Crafty. Like normal. What this type of pruning does is just see if the CaptureValue is not enough to fill the gap between the current Eval and Alpha, right? Then it will be beta-cut on the next ply anyway. And it seems to me it makes a big difference if you use a) Gap = Alpha - Material - Margin (which estimates Gap, accuracy strongly depending on Margin) or b) Gap = Alpha - CurrentEval (-Margin) (which assumes just that the Eval *difference* is no greater than Margin. Something quite different than a and in my opinion more accurate) Regards, Bas Hamstra.
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.