Computer Chess Club Archives




Subject: Re: Fail high reductions

Author: Gerd Isenberg

Date: 11:28:27 07/02/03

Go up one level in this thread

On July 02, 2003 at 11:09:21, Omid David Tabibi wrote:

>On July 02, 2003 at 10:49:41, Gerd Isenberg wrote:
>>On July 02, 2003 at 07:22:00, Georg v. Zimmermann 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?
>>>IMHO Rainer Feldmann uses bad terminology. A fail high node is - at least by my
>>>definition - indeed a node where one subtree returns a value above beta, you
>>>therefor "fail high" and return (value or beta, depending on if you use fail
>>>What he intends to say is probably : " a fail high REDUCTION node is a node 'v'
>>>with a search window of [alpha,beta] at which
>>>>a static evaluation function 'c' produces a cutoff. "
>>>The technic he describes sounds a lot more error prone than null move to me, at
>>>least in tactical situations.
>>Hi George,
>>If i remember well, Rainer Feldmann's FHR is based on the NullMove observation.
>>Instead of foreward pruning, FHR reduce depth if a NullMove fail high occurs.
>FHR reduces the depth if static eval >= beta. I think you are confusing FHR with
>verified null-move pruning; the latter reduces the depth when null_move_score >=
>beta (and cuts off immediately if null_move_score >= beta in the subtree of a
>fail-high reported node).

Hi Omid,

I had no access before to "Advances In Computer Chess 8" (1997) page 111-127:

Fail-High Reductions

on Page 118:

2. The FHR is based on the NMO. Therefore one could argue that special attention
should be given to nodes where the reduced search fails to compute a cut-off. We
tested a version of the FHR that re-searched such nodes without depth reduction.
This version turned out to be inferior to the original FHR version."

3. The question, whether a better approximation for threats should be used than
the static function t, e.g., by doing a null-move quiescence search, is
addressed. It turned out that the overhead for the extra searches does not pay
off, even if these extra searches are restricted to nodes at leat 3 ply from the

So the point in question seems NMO, whether one use a sophisticated static eval
with some threat detections against the side to move or a reduced nullmove
search (OK, Rainer wrote null-move quiescence search).


This page took 0.02 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.