Author: Vincent Diepeveen
Date: 21:19:54 12/28/02
Go up one level in this thread
On December 28, 2002 at 19:40:57, Alessandro Damiani wrote: >On December 28, 2002 at 19:03:33, Tony Werten wrote: > >>On December 28, 2002 at 16:03:46, Uri Blass wrote: >> >>>On December 28, 2002 at 15:39:08, Tony Werten wrote: >>> >>>>On December 28, 2002 at 14:17:17, Alessandro Damiani wrote: >>>> >>>>>[snip] >>>>> >>>>>> >>>>>>the problem of most 'reductions' is the hard fact that you lose a full ply >>>>>>near the root. >>>>> >>>>>That's why reductions are not done in every node, but under certain conditions. >>>>>The quality then depends on those conditions, of course. Therefore, reductions >>>>>are not bad per se. >>>> >>>>This was about recursive reductions as FHR. What happens is at a ply you decide >>>>to reduce depth, but 2 ply later, the conditions are still met and you reduce >>>>another ply etc. >>>> >>>>I dumped them because they cost to much tactical strenght. Ed's nonrecursive way >>>>seem to give me a 5% node reduction. Not bad for 2 minutes work. >>>> >>>>Tony >>> >>>The question is still if it does not cost too much tactical strength. >>> >>>It is not clear if being 5% faster in 95% of the cases and seeing tactics one >>>ply later in 5% of the cases is a good idea. >> >>I had no case where I saw tactics one ply later. But then again, I didn't have >>the 15% speedup Ed mentioned either. >> > >We should not forget that we are talking about *marginal benefit*. For instance, >we have two pruning systems. If the first pruning system already prunes a lot, >adding a second pruning system may give only few additional percents of gain. >While measuring the second system alone may give a big speedup. > >The same applies to move ordering, for example. If one has near perfect move >ordering there is not much to gain with additional heuristics, but the same >heuristics may be a big win compared to randomly ordered moves. > >Alessandro With regards to pruning heuristics you are correct that if you already prune with nullmove away all the nonsense that reductions won't save you much more and in contradiction cause other problems possibly. However with regards to move ordering the opposite is true. If a program already has a very good branching factor then a small change from a 3.0 b.f. which diep currently has (but that's with SE without it's more like 2.8 at the bigger depths), then that is of course giving me more than an improvement of 0.1 will give to a branching factor that's already huge. Even if you think of it in %: it gives a 1% better cutoff rate for example, then that will not influence b.f. in a % way but simply get you a ply deeper easily. Of course the problem usually is that something doesn't help much then :) Best regards, Vincent
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.