Author: Robert Hyatt
Date: 10:54:28 04/26/04
Go up one level in this thread
On April 26, 2004 at 11:32:26, Tord Romstad wrote: >On April 26, 2004 at 10:39:42, José Carlos wrote: > >> An interesting experiment, of course. But I think your conditions are rather >>different from 'most' programs. I mean: >> - You allow any number of null moves in a row (most programs don't do even >>two) > >This has no importance, I think. My experience is that I almost always get the >same score and PV when I enable/disable several null moves in a row, and that >the difference in number of moves searched is *very* tiny. > >> - You do pruning and reductions that are not public domain >> This is important because your results mean: 'in a non-standard null move >>implementation (where you try more null move searches than most others) and with >>a lot of heavy pruning and reductions (I assume they're heavy according to your >>posts here and the high depths Gothmog gets) and in a MTD(f) root search, >>limiting null move application seems to benefit". This conclusion is, apart from >>interesting of course, very intuitive. > >This is a very good point. During our discussion last week, several people >claimed that checking the static eval before doing a null move search wasn't >really necessary, because the cost of a null move search would be tiny >compared to the cost of searching the moves anyway. This isn't really >true for Gothmog, because most of the moves are usually not searched to >full depth. > >I am not sure it is fair to describe my pruning and reductions as "not >public domain", though. It is true that I haven't published any paper >and that my engine is not open source (at least not yet), but I am happy >to describe the tricks I use to anybody who wants to listen. :-) We all listen. :) > >Whether the use of MTD(f) is important is hard to say, but my guess is >that it isn't. Mine is that it is a _big_ issue. I played with mtd(f) for quite a while, and I found that it changed most everything. Needed 2 hash bounds. Needed fail-soft. Needed to accelerate the convergence. It really was a different animal from normal PVS-type approaches. > >> Thanks for the data. You sure inspired many others to do similar tests (I'll >>do when I have time). > >Yes, it would be interesting to see the results of similar experiments for >other engines. It seems like a sure bet that this is a YMMV issue. :-) > >Tord I agree. I did it one way, Bruce (see my other post in this thread) had conflicting data, turned out his data was better and I changed. :)
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.