Computer Chess Club Archives


Search

Terms

Messages

Subject: Re: pv score oscillation

Author: Robert Hyatt

Date: 17:14:46 10/20/97

Go up one level in this thread


On October 20, 1997 at 09:45:52, Reinhold Gellner wrote:

>On October 20, 1997 at 08:52:02, Chris Whittington wrote:
>
>>
>>On October 20, 1997 at 02:29:25, Reinhold Gellner wrote:
>
>>>To eliminate that problem, try increasing the search depth by 2 ply
>>>every iteration.
>>
>>This is a surprisingly well used solution.
>>But increasing by 2 each iteration loses much of the advantage of
>>iterating (guided search and 'small' incrementally deeper researches).
>>You need to jump by two every two iterations, and find a method for what
>>to do in the 'gap' iteration.
>>
>>Its also an advantage for an exchange-evaluator to stop on odd plies,
>>rather than even. Its decisions that an as yet untaken piece is actually
>>lost will be better taken if the piece is a computer piece (odd ply)
>>than a human piece (even ply). Asymetric safety factor skandal for
>>Komputers !.
>>
>>Best then is to jump by 2 ply every two iterations and use a pruning
>>function over the final plies that is strong pruning the first iteration
>>and weaker pruning the second - then jump two plies again and so on.
>>
>
>2-ply-jumping is not neccessarily more stable than 1-ply-steps, because
>the jumping "distance" is simply one ply more. If you are doing
>sophisticated prunings, e.g. part-ply prunings, together with
>part-ply-extensions, then perhaps half-ply-iterations might be worth
>considering.
>
>At the moment I can live with smooth oscillations jumping allways one
>ply.
>
>Reinhold


The only known "production" program doing this was "BeBe" by the late
Tony
Scherzer.  he did odd ply searches only, as he liked the "aggressive"
behavior
of odd plies.  But it costs quite a bit, particularly if you have enough
time
to do 8, but not 9.  So you settle for 7...



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.