Author: Robert Hyatt
Date: 07:16:49 09/03/02
Go up one level in this thread
On September 03, 2002 at 01:22:38, Uri Blass wrote: >On September 02, 2002 at 19:29:17, Robert Hyatt wrote: > >>On September 02, 2002 at 17:15:54, Dave Gomboc wrote: >> >>>On September 01, 2002 at 13:26:44, Robert Hyatt wrote: >>> >>>>On September 01, 2002 at 09:40:34, Vincent Diepeveen wrote: >>>> >>>>>pawn=32 in fritz seemingly. that's all you need to know to consider >>>>>it works for it. >>>> >>>>What does that do? I have seen large positional scores out of fritz, >>>>which suggests (to me) that mtd(f) could cause some problems... >>> >>>It means there's a lower granularity of values, so less (re-)searches are >>>required to arrive at the score. I have believed for some time that this is a >>>correct decision, because I don't have confidence in the precision of even >>>centipawn evaluations (much less the millipawn evaluations that Diep years ago >>>when Vincent based mtd techniques heavily). I didn't know that Fritz actually >>>had done it, though. >>> >>>The range of positional scores is relevant for lazy evaluation schemes, but I >>>think that issue is pretty well handled by Don Dailey's recommendation. >>> >>>Dave >> >> >>I understand that. But that wasn't the issue I was addressing. It was "fritz >>uses a paw value of 32 or whatever..." And I don't see how that influences >>mtd(f) as much as an evaluation that produces positional scores that vary from >>+3.00 to -3.00, no matter what 1.00 means. If an evaluation can produce wild >>positional scores, then mtd(f) causes problems... It makes it hard to hone in >>on the right final score, particularly with lazy eval.. > >I do not know if Fritz is using lazy evaluation and there are alternatives to >lazy evaluation like incremental evaluation. > >You also assume that lazy evaluation is based only on material and can be >significantly different than the full evaluation. No i don't. In Crafty, the lazy evaluation is the _big_ terms, ie the passed pawn evaluation, and so forth... And that has nothing to do with the mtd(f) discussion. The issue is that if one additional ply significantly changes the evaluation in a lot of positions, that will cause mtd(f) to "hunt" excessively and drive efficiency down. That is why it didn't work for me when I spent a few months working on it when it first was reported. > >It is not clear for me and lazy evaluation can be an estimate for the real >evaluation with difference of at most a pawn. It depends on where you do it. If you have _lots_ of large positional terms, then lazy evaluation is going to be far more than a pawn off, no matter where you put it. > >Big positional factors may be evaluated also in lazy evaluation. > Certainly. I already do this. But there are other large positional terms that can not be lazy evaluated. My king safety stuff, for example. I need to know some basic info about all the pieces, and this is computed in the individual piece evaluation blocks... If I lazy evaluate after computing those, I save nothing as all the work is done. If I lazy evaluate before those, I don't include king safety which can swing the score +/- 3 pawns or more. the point... >Uri
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.