Author: Robert Hyatt
Date: 16:43:51 10/13/03
Go up one level in this thread
On October 13, 2003 at 19:19:27, Dieter Buerssner wrote: >On October 13, 2003 at 17:09:00, Omid David Tabibi wrote: > >>Actually I have found checks in the first ply of quiescence to be quite helpful >>when used with null-move pruning. > >You mean helpful in test positions? > >>Assume that you are at depth = 3 and use R = >>2, the null-move search will be conducted by depth 0, i.e., direct call to >>quiescence. Also assume that no matter what you do, you are mate in 1. If you >>have checks in quiescence the opponent will checkmate you immediately, returning >>a mate score which in turn will trigger a mate trheat extension in the main >>search. But if you don't do checks in quiescence you will miss the checkmate and >>have a good chance of failing high on the null-move search resulting in a >>cutoff. I.e., you think that the position is good enough to justify a cutoff >>when in fact you are mate in 1! > >I totally agree with your analysis. > >>I think mainly for this reason adding checks in the first ply of quiescence >>results in such an improved performance in solving test suites. But it doesn't >>seem to be as helpful in actual games. > >And I made the same experience ... in reality it does not seem to help. I also >tested to only call a qsearch with checks (that are not captures, the capturing >checks I can handle in normal qsearch at first ply) after a null move. It was >great in some test positions, but not in games. > >Perhaps my tests were not conducted well enough (too few games, too little >tuning, whatever). It seems too logical, to at least try a qsearch with checks >after a null move. I may try harder, soon. > >Regards, >Dieter I think this is one of the things that simply has to "hang around" and be tried from time to time. I can't count the number of times something failed for me, then after lots of other changes, suddenly it seemed to work on the N+1th test. And, on occasion, something that has worked for years suddenly hurts and gets removed, after some significant change alters the tree shape. Just look at my main.c comments on how many times I tried R=2 with null-move before I finally stuck with it. Doing 5 plies, R=2 is horrible and exposes so many horizon blunders it isn't funny. By the time you reach 8-10 plies it suddenly doesn't look as bad, and by the time you reach 11-12-13-14 it looks downright outstanding...
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.