Author: Ross Boyd
Date: 03:40:48 08/12/04
Go up one level in this thread
On August 12, 2004 at 00:21:42, Tor Lattimore wrote: >I have been fiddling arround with my nullmove pruning lately. I have tried both >R=2, R=3 and R=3 with verifications. > >Average ply on WAC, 1 second a move P IV, 512 MB ram: >R=2 9.5 >R=3 10.5 >R=3 with verification 10.2 > >However R=3 had a terrible success rate (135 or so), R=2 got 147 and R=3 with >verification got 143. What do other people use here? and what sort of >improvments do you get? The extra ply may be worth the puzzle sacrifice, but i'm >not convinced i'm really *getting* the extra ply. >Thanks >Tor Hi Tor, Be careful with R=3. It has the potential to make your engine go blind. I lost ~50 (!!) elo when using pure R=3 in TRACE. I ran the experiment again two days ago and it confirmed my previous findings. Currently, pure R=2 works best for me... fewer OTB blunders. Mind you, the test machine was a 450Mhz P3 so that would skew the results in favour of R=2... and I don't test at slow time controls because life is too short. :-) You could try an adaptive approach which is R=3 by default and R=2 when remaining depth is < 6 ply.... or something like that. As always, YMMV... It obviously depends on a whole lot of factors eg. do you use futility pruning, delta pruning, SEE pruning, hash pruning? Do you try quiet checks in qsearch if captures don't return a score above alpha? All these things will interact with the null reduction factor... I believe the engines which perform okay with R=3 are compensating by, for example, not qsearch/futility pruning moves which dramatically reduce king safety or affect other positional factors. That's just a theory. I guess all you can do is test, test and test... and keep what works... ditch what doesn't. :-) Best regards, Ross
This page took 0.01 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.