Computer Chess Club Archives


Search

Terms

Messages

Subject: Re: Verified Null-moving

Author: Tony Werten

Date: 08:02:21 08/12/04

Go up one level in this thread


On August 12, 2004 at 10:52:36, martin fierz wrote:

>On August 12, 2004 at 09:28:24, Tord Romstad wrote:
>
>hi tord,
>
>thanks for the clarification! so you really RETURN when your
>almost_certainly_fail_high function returns true? isn't this a bit dangerous? i
>mean, probably it helps, but i can imagine that there are positions where
>gothmog will never find the right move because your almost_certainly_fail_high
>function returns 1 in a position where it shouldn't.
>
>my philosophy (in my checkers program, my chess program is not in a stage of
>having a philosophy yet...) was always to make big reductions in depth possible,
>but never to allow a complete return - because then you can have such effects
>that you can never solve a position.

Not if almost_certainly_fail_high factors in the remaining searchdepth. ie
multiply the margin with the remaining ply left or something.

I don't use it myself but I can understand something like that can work.

Tony

>
>
>>That recursive null move pruning is not the only kind of pruning I do.
>>At fail low nodes, most of the moves are not searched with full depth.
>>When the first 3 moves have failed low, I search all remaining moves with
>>reduced depth except when they look especially interesting or forcing.
>>If a move searched with reduced depth surprisingly turns out to fail high,
>>it is re-searched with full depth.
>>
>>For some reason, most of the published research on selective search
>>mechanisms has concentrated on ways to reduce the work at fail-high
>>nodes.  Nullmove pruning, Multicut pruning and ProbCut are a few examples.
>>Pruning techniques for fail-low nodes are much less commonly seen.  The only
>>ones I can think of is futility pruning and razoring, but these are too
>>dangerous to use except when the remaining depth is very small.
>>
>>Reducing the search depth towards the end of the move list is my own
>>attempt to do selective search at expected fail-low nodes.  It is
>>probably possible to invent better techniques (and it wouldn't surprise
>>me if it has already been done in some of the professional engines), but
>>at least in my engine, what I do works far better than nothing.
>
>and thanks for sharing this - another interesting idea of yours :-)
>
>cheers
>  martin



This page took 0.04 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.