Computer Chess Club Archives




Subject: Re: Fail high reductions

Author: Tony Werten

Date: 04:30:32 07/02/03

Go up one level in this thread

On July 02, 2003 at 05:47:27, martin fierz wrote:

>On July 01, 2003 at 18:24:43, Omid David Tabibi wrote:
>>On July 01, 2003 at 18:19:44, Ralph Stoesser wrote:
>>>On July 01, 2003 at 18:10:04, Uri Blass wrote:
>>>>On July 01, 2003 at 17:34:47, Russell Reagan wrote:
>>>>>From "Fail High Reductions by Rainer Feldmann"
>>>>>"...a fail high node is a node 'v' with a search window of [alpha,beta] at which
>>>>>a static evaluation function 'c' produces a cutoff. The FHR-algorithm reduces
>>>>>the search depths at these fail high nodes thus searching their subtrees with
>>>>>less effort."
>>>>>Their subtrees? I thought fail high nodes didn't have subtrees, and that you
>>>>>return beta at a fail high node. I must be misunderstanding something. Could
>>>>>someone give a simple explaination of how fail high reductions work?
>>>>If I understand correctly the idea is that you evaluate a position that is not a
>>>>leaf and the static evaluation is not in the window of [alpha,beta] so you
>>>>reduce the depth.
>>>>Example:when you analyze e4 Nf6 Qh5 Nxh5 your evaluation is a queen advantage
>>>>for black and you can be almost sure of fail high so instead of searching to
>>>>remaining depth of 7 you may decide to search to a smaller depth.
>>>I think that's right. This is similar to the Nullmove technique, but not that
>>>radical, because you only reduce the search depth instead of cutoff the whole
>>... which is called verified null-mvoe pruning :)
>to me FHR sounds very similar to ProbCut / MultiProbCut, except that it uses the
>static eval at a node instead of a shallow search.

Not quite. FHR reduces depth when static eval>>beta and then continues search as

Probcut uses a shallow search (several actually ) to generate a cutoff.



This page took 0.01 seconds to execute

Last modified: Thu, 07 Jul 11 08:48:38 -0700

Current Computer Chess Club Forums at Talkchess. This site by Sean Mintz.