Computer Chess Club Archives


Search

Terms

Messages

Subject: Re: singular extension

Author: Robert Hyatt

Date: 16:46:39 09/15/04

Go up one level in this thread


On September 15, 2004 at 18:15:33, Gerd Isenberg wrote:

><snip>
>>>>Hsu's paper defined singular extensiosn for PV nodes and CUT nodes.  His paper
>>>>said "we have found no useful definition for a singular move at an ALL node
>>>>however.."
>>>
>>>Hmm isn't the definition simply that there is one move which is by a certain
>>>amount better than all others, independent on pv-, cut- and all-node?
>>
>>No.  There is a precise definition of a singular move, but the test is only
>>defined for PV and CUT nodes.  At an ALL node there is absolutely no way to
>>determine if one move is better than all others.
>
>Maybe with fail soft and altered zero window bounds to check for one
> score > alpha - SMALL_MARGIN
>while all others are far below
> score <= alpha - BIG_MARGIN
>



Won't work.  At fail-high nodes (the next one down in the tree) we don't look
for the _best_ move so that you get good scores backed up for each refutation at
the previous ply, we just find a "good move".  Trying to pick out the best move
at an "ALL" node is impossible for that reason, unless you do a true minmax
search rather than alpha/beta



>
>>
>>
>>
>>>
>>>Probably the forced situation, that there is only _one_ move  (e.g. becomes
>>>alpha while all others are less alpha-margin), makes it worth to look deeper to
>>>look whether the all-node becomes a singular cut node or improves alpha, with
>>>possible influence at the root.
>>
>>That test won't work.  Tests where the value is <= alpha or <= alpha-margin are
>>really meaningless in the context of alpha/beta/minmax search.  That is why they
>>could not define a workable test for singularity at ALL nodes.
>
>I see.
>
><snip>



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.