Computer Chess Club Archives


Search

Terms

Messages

Subject: Re: Q: Singular extensions

Author: Robert Hyatt

Date: 19:50:20 01/06/04

Go up one level in this thread


On January 05, 2004 at 19:36:50, José Carlos wrote:

>On January 05, 2004 at 19:24:31, Bas Hamstra wrote:
>
>>On January 05, 2004 at 08:48:23, José Carlos wrote:
>>
>>>  Hi, I'd like to try singular extensions in my program. I've been trying to
>>>think about it and all my tries so far result in worse performance. After some
>>>web search I haven't been able to find anything but the general idea described.
>>>  Is there any good description somewhere? Some pseudo-code? I guess there must
>>>be something interesting in the archives but I can't download all of them.
>>>  Thanks in advance,
>>>
>>>  José C.
>>
>>Here is what I tried (idea from Bruce I believe):
>>
>>Before you search all moves (after nullmove) you search all moves (unsorted)
>>with D-R
>
>  Good point I had missed :)
>  I was doing an expensive D search with (Alpha-Margin,Alpha-Margin+1) and if
>only one move failed high, search it again with (Alpha,Beta) extended 1 ply.
>
>>and check if there is one and only one move with a score above alpha
>>*and* above Alpha-Margin.
>
>  So if I get it right, you search all moves with D-R and a null window around
>Alpha-Margin. If only one move fails high, then search it again with D-R and a
>null window around Alpha. If it fails high again, then that move is extended.
>

See my previous post.  If you search _all_ moves, you are asking for mucho
overhead.  Suppose this is an ALL node where you have to search every move
and each one fails low.  There is no reasonable definition of singularity in
such a position, and searching everything is a waste, even to depth D-R.  We
were playing with searching only the first N moves, and if _none_ failed high,
we quit.  If one did fail high, then we searched all the remaining moves to
make sure they failed low...  And then we passed this move back to search and
said "when you search this move, add a ply (or whatever you like) to the
depth first."

>>If so, in the normal search to depth D extend that
>>move. I did not do this in nullwindow searches, or else performance crashed.
>
>  I've had problems when making this difference. For example, I used to not try
>null move in PV nodes. At some point, I realized that a non PV move failed high,
>then was researched with an open window and failed low due to not applying a
>null move threat extension!
>
>>No great results. In particular I didn't see much spectacular long pv's show up.
>>
>>I am interested in refinements or other approaches. And are there good
>>testpositions to check the effect of SE?
>
>  Thanks for the tips. I'll report my results if I manage to make it work.
>
>  José C.
>
>
>>Best regards,
>>Bas.



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.