Author: Rémi Coulom
Date: 05:42:50 10/28/99
Go up one level in this thread
On October 28, 1999 at 06:30:37, Ulrich Tuerke wrote: >On October 27, 1999 at 15:38:54, Frank Schneider wrote: > >>Hi all, >> >>this is a question about the searchalgorithms used by other programs. >>Many programs and also Gromit seem to use PVS or MDT(f) which both use >>nullwindow-searches. >>Is there anyone using a search that doesn't do nullwindow searches? >>Why? Why not? > >I am sure that there are some programs not using minmal window searches, also >among the top programs (Hiarcs ?). >In principle (AFAIK), the AB standard algorithm works correctly, PROVIDED the >result of the search does not depend on the choice of alpha and beta. >However in practice, this condition may be violated in most cases. Think for >instance of evaluation window techniques, null move cuts, futility cuts, >extensions, ... . >The violation of this precondition can make the search instable, e.g. you may >get a fail-low in a fail-high verification search, because alpha & beta have >changed when doing the re-search and the conditions for your tactical extensions >will not be reproduced in some cases. > >When it now comes to a minimum window search, you have generally many more >verification searches with shifted aspiration window. Thus, the above described >instabilities may become far more dramatic. Of course, this happens not only on >the root of the search, where it becomes obvious but also internally (in PVS) >all the time. In my prog, I observed that I do not get stable evaluations at all >when using PVS, though it fastened up the search a little bit. I think that this >was a consequence of the very strong alpha-beta dependencies in my prog. >Therefore, i decided not to use PVS. > >Uli >> >> >>Frank I have the same problem in The Crazy Bishop. One idea that might be worth trying would consist in using two windows in a search : the real alpha-beta window and the aspiration window. Beta cutoffs on the aspiration window would be allowed only in case there is a real move >= AspirationBeta, the null move being tested against RealBeta only. This is on my list of things to try. Remi
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.