Author: Robert Hyatt
Date: 18:34:00 01/17/99
Go up one level in this thread
On January 17, 1999 at 18:56:20, Gerrit Reubold wrote:
>Hi all,
>
>I remember several post about the following problem, but can't remember the
>details or the solution:
>Consider you search at the root with a window [a,b] and the search returns the
>score x, x < a, a fail low. Now you do a re-search with a less optimistic
>window, maybe [x-100,x] or even [-MATE,x]. It can happen now that the search
>fails high immediately. I call this the "Fail-Low-Fail-High" problem. I have
>ignored the posts on this problem so far, because I didn't understand the
>problem until today, while debugging. It seems to happen when the
>NullMove-Search calculates a score and maybe returns that score (or beta)
>*depending on beta*. When re-searching after a root fail-low, alpha and beta
>have changed, and the NullMove-Search of the same node in the tree may or may
>not behave as in the first search. My NullMove-search code (R = 2) is standard
>code, like this:
>
>if(bTryNullMove /* && some other conditions*/)
>{
> int nNullMoveScore = GetNullMoveScore(nBeta,nBeta+1,...);
> if(nNullMoveScore >= nBeta)
> return nNullMoveScore;
>}
>
>Is this a known problem, or just a bug in my code?
>How do others solve it?
>Do I have to *avoid* using the alpha and beta values to make my search
>consistent (I use them frequently, for triggering extensions, deciding about
>lazy eval, etc.)?
>If I remember right, Bob Hyatt's advice on this problem was to ignore it and
>simply to search fast and deep. Is it that simple?
>
>Any comments are welcome,
>thanks in advance,
>
>Gerrit Reubold
So long as you 'hash' there is little you can do to this. It has been seen
many times.... I see it here and there in Crafty. And I just ignore it...
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.