Author: Robert Hyatt
Date: 20:03:49 10/16/02
Go up one level in this thread
On October 16, 2002 at 18:10:18, Vincent Diepeveen wrote: >On October 16, 2002 at 08:12:00, Joachim Rang wrote: > >>On October 16, 2002 at 07:06:30, Vincent Diepeveen wrote: >> >>>On October 16, 2002 at 06:52:17, Gian-Carlo Pascutto wrote: >>> >>>>On October 16, 2002 at 06:39:08, Vincent Diepeveen wrote: >>>> >>>>>On October 16, 2002 at 06:37:40, Thorsten Czub wrote: >>>>> >>>>>>On October 16, 2002 at 06:33:39, Gian-Carlo Pascutto wrote: >>>>>> >>>>>>>It uses (I wouldn't say wastes) quite a bit of resources, >>>>>>>but there is also a gain. I find that it is almost perfectly >>>>>>>breaking even for me. I search less deep, but I am seeing >>>>>>>some deep tactics and avoid getting into some dangerous >>>>>>>situations. >>>>>> >>>>>>really ?! >>>>>> >>>>>>interesting. >>>>>> >>>>>>i gave up using SEX in rebel XP.... >>>>> >>>>>the problem of SE shows when search depths get big. >>>>>GCP is basically testing at bullet levels and then you don't >>>>>get search depths of 12+ ply soon. >>>> >>>>I don't notice them worsening things at longer time controls >>>>either, but then again, I've never been a very deep searcher, >>>>though things are changing. >>>> >>>>I also understand that you do the full Hsu thing, so your >>>>overhead must be a lot bigger. >>> >>>I do way more. I can do way more. The nonsense lines get >>>pruned by nullmove. In Hsu's creations they were not. >>> >>>>If I find that with increasing hardware speed/slower timecontrols >>>>it does start getting worse, I will uncheck the 'singular extensions' >>>>checkbox and problem solved :) >>>> >>>>>In DIEP the price of SE is basically getting too big at those depths. >>>> >>>>I gather from this you won't be using them this weekend? Or is >>>>90m SD also 'bullet'? >>> >>>I am talking about the crucial 12 ply search depth for DIEP. >>>I will be getting about 10-12, so the weakest spot is those 10 ply >>>searches. There i will be using SE. >>> >>>My overhead at 10 ply is about half a ply for SE. But at 12 ply >>>it can be easily 2 ply. Doesn't need to be so. Nullmove sometimes powerful >>>prunes away the stuff in quiet positions. >>> >>>It's the tactical positions where sometimes even a 11 ply search depth >>>has plies overhead. SE is a very stupid extension in fact. I am sure >>>in the far future it's too stupid to let your b.f. worsen with it. >>> >>>It makes my b.f. simply a lot worse. With 3 ply reduction however >>>the worsening starts not directly. Because in the root i don't do 'em >>>and the last 3 ply i don't do them. So i save 4 ply. >>> >>>It's trivial that at 12 ply you feel it for 8 ply and at 10 ply you >>>'just' feel it for 6 ply. >>> >>>In other words it's % from 10 ply is 60% of search depth. At 12 ply it's 75% >>>of total search depth. >>> >>>That triggers more. >>> >>>Suppose in a few years we all get 15 ply. For me within 1 year perhaps >>>(if i get some processors from NWO to use). I can't afford SE then simply. >>> >>>DB never was tested very well, let's be realistic here. Otherwise he >>>might have kicked it out too and would have gotten more nodes a second >>>then. >>> >> >>So your testings showed, that SE create more overhead at deeper depth? What does >>that cause? Why is the overhead at depth 10 about half a ply and at depth 12 >>two? >> >>regards >> >>Jochim > >This is very easy to proof. Because you now and then extend a move, >your overhead becomes bigger. So your branching factor becomes bigger, >say from 'bx' to 'by'. > >So if you lose 0.5 ply now with that worse branching factor 'by', >then 2 ply deeper from here you lose extra in plies: > = (by^2 / bx^2) / bx = by^2 / bx^3 plies > >A major additional problem is that where bx is pretty constant, >'by' because of singular extensions is not constant. > >it's trivial that if at 10 plies of search DIEP has 6 plies where it >could do a singular extension. This trigers at 60% of the search depth >extensions. That's of course getting more when the search depth gets bigger. > >So there is more chance for those extensions to happen. So also observed >with all such programs is that you do more and more of those extensions >when depths get bigger. Also extensions overlap. In short sometimes >'by' gets radically worse. > >Each extra ply the program without SE needs just 'bx' to get deeper >and a worsening 'by' which is already bigger than 'bx' of course >loses to program using 'by' another ply. > >So when search depths in future get real big, then i can't use SE at all. Then you must have the _same_ problem with the check extension, the recapture extension, and so forth... The deeper you go, the more times a particular extension can fire, regardless... It is controllable however... You just have to think about how...
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.