Computer Chess Club Archives


Search

Terms

Messages

Subject: Re: Conspiracy Numbers Search and followups

Author: Stuart Cracraft

Date: 15:36:06 01/04/06

Go up one level in this thread


On January 04, 2006 at 15:49:39, Dann Corbit wrote:

>On January 04, 2006 at 14:52:06, Stuart Cracraft wrote:
>
>>
>>http://www.robinupton.com/research/phd/cp_intro.html
>>
>>So, has anyone actually used these CNS and CNS-derived ideas in
>>a chess program? I know about B* and Berliner. I want to know
>>about things more recent that at least *someone* can speak to
>>in the commercial or semi-commercial world.
>>
>>If you don't have actual data, how about theories on how these
>>would be used?
>>
>>It seems that determining the least number of nodes that would
>>have to change to affect the root score or growing the tree
>>in such a way that the conspiracy is as large or as small as
>>possible, could yield some interesting deeper searches.
>>
>>What is actual in regards to the above for computer chess?
>
>Conspiracy search:
>Alice by Sven Reichard
>
>Proof number search:
>Older versions of Sjeng by Gian-Carlo Pascutto
>Defeo (chess variant based on Sjeng)
>
>There is a checkers program called Dragon-draughts that uses proof number search
>
>There is a Go program called PubGo that uses proof number search
>
>It appears that Popeye and Natch (mate solvers) may be using proof number search
>
>There is a Connect 4 program called Velena by Giuliano Bertoletti that uses
>proof number search

The result has been poor - not many big names in there with good results.

Rybka is intriguing but little is known now if it is CNS/CNP-like thought
I believe mention was made of probabilistic evaluation by the author which
is not the same as CNS/CNP.

For example, as Brian Richardson pointed out, I can map a function on top
of my evaluation so that instead of returning centipawns, it takes that
and maps it to a sigmoid or tanh and returns a number between 0 and 1 or
-1 to +1.

It is an interesting experiment but in this case would take your evaluation
and make it probabilistic at the expense of creating yet another computation
for every terminal node.

I do not see any great advantage in this except that if you cutoff the
probability to some fixed set of decimal points, you will begin to see
different evaluations mapping to the same probability score and so
various terminal nodes in the tree will be identical. I do not know what
effect that has on PVS, MTD(f), Alphabeta, and so on to have more nodes
with identical scores.

I see no advantage to probabilistic evaluation without something else
helping it.

Stuart



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.