Author: Uri Blass
Date: 09:29:17 07/08/02
Go up one level in this thread
On July 08, 2002 at 11:39:40, Robert Hyatt wrote: >On July 08, 2002 at 00:21:23, Christophe Theron wrote: > >>On July 07, 2002 at 23:53:16, Robert Hyatt wrote: >> >>>On July 07, 2002 at 23:42:03, Omid David wrote: >>> >>>>On July 07, 2002 at 21:43:47, Robert Hyatt wrote: >>>> >>>>>On July 07, 2002 at 16:47:33, Omid David wrote: >>>>> >>>>>>On July 07, 2002 at 16:36:57, Robert Hyatt wrote: >>>>>> >>>>>>>On July 07, 2002 at 11:48:27, Omid David wrote: >>>>>>> >>>>>>>>On July 06, 2002 at 23:23:28, Robert Hyatt wrote: >>>>>>>> >>>>>>>>>On July 06, 2002 at 22:29:44, Omid David wrote: >>>>>>>>> >>>>>>>>>>On July 06, 2002 at 10:20:17, Robert Hyatt wrote: >>>>>>>>>> >>>>>>>>>>>On July 06, 2002 at 01:07:36, Ricardo Gibert wrote: >>>>>>>>>>> >>>>>>>>>>>> >>>>>>>>>>>>Okay, but so what? >>>>>>>>>>>> >>>>>>>>>>>>> >>>>>>>>>>>>>So perhaps the idea of "forward pruning" is foreign to us as well... >>>>>>>>>>>> >>>>>>>>>>>>I see no logical difference between deciding which moves are interesting and >>>>>>>>>>>>worth looking at and deciding which moves are not interesting and not worth >>>>>>>>>>>>looking at. It looks to me like 2 sides of the same coin, so your speculation >>>>>>>>>>>>that "perhaps the idea of "forward pruning" is foreign to us as well..." does >>>>>>>>>>>>not seem to be of any consequence. >>>>>>>>>>>> >>>>>>>>>>> >>>>>>>>>>>However, that has been _the point_ of this entire thread: Is DB's search >>>>>>>>>>>inferior because it does lots of extensions, but no forward pruning. I >>>>>>>>>>>simply said "no, the two can be 100% equivalent". >>>>>>>>>> >>>>>>>>>>Just a quick point: The last winner of WCCC which *didn't* use forward pruning >>>>>>>>>>was Deep Thought in 1989. Since then, forward pruning programs won all WCCC >>>>>>>>>>championships... >>>>>>>>> >>>>>>>>> >>>>>>>>>In 1992 no "supercomputer" played. In 1995 deep thought had bad luck and lost >>>>>>>>>a game it probably wouldn't have lost had it been replayed 20 times. No >>>>>>>>>"supercomputer" (those are the programs that likely relied more on extensions >>>>>>>>>than on forward pruning due to the hardware horsepower they had) has played >>>>>>>>>since 1995... >>>>>>>>> >>>>>>>>>I'm not sure that means a lot, however. IE I don't think that in 1995 fritz >>>>>>>>>was a wild forward pruner either unless you include null move. Then you >>>>>>>>>would have to include a bunch of supercomputer programs including Cray Blitz >>>>>>>>>as almost all of us used null-move... >>>>>>>> >>>>>>>>I personally consider null-move pruning a form of forward pruning, at least with >>>>>>>>R > 1. I believe Cray Blitz used R = 1 at that time, right? >>>>>>> >>>>>>> >>>>>>>I believe that at that point (1989) everybody was using null-move with R=1. >>>>>>>It is certainly a form of forward pruning, by effect. >>>>>> >>>>>>Yes, and today most programs use at least R=2... The fact is that new ideas in >>>>>>null-move pruning didn't cause this change of attitude, just programmers >>>>>>accepted taking more risks! >>>>> >>>>> >>>>>I think it is more hardware related. Murray Campbell mentioned R=2 in the >>>>>first null-move paper I ever read. He tested with R=1, but mentioned that >>>>>R=2 "needs to be tested". I think R=2 at 1980's speeds would absolutely >>>>>kill micros. It might even kill some supercomputers. Once the raw depth >>>>>with R=2 hits 11-12 plies minimum, the errors begin to disappear and it starts >>>>>to play reasonably. But at 5-6-7 plies, forget about it. >>>> >>>>So using a fixed R=3 seems to be possible in near future with faster hardware, >>>>doesn't it? >>> >>> >>>Very possibly. Or perhaps going from 2~3 as I do now to 3~4 or even 4~5 for >>>all I know... I should say that going from 2 to 3 is not a huge change. Bruce >>>and I ran a match a few years ago with him using Ferret vs Crafty with Ferret >>>using pure R=2, and then pure R=3. We didn't notice any particular difference >>>at that time. It played about the same, searched about the same depth, etc... >> >> >>Increasing R is pointless after 3. >> >>Because instead of having a null move search using 5% of your time (just an >>example, I do not know the exact value), it will use only 2% or 3%. >> >>The speed increase is ridiculous, and the risks are getting huge. >> >>The only thing you can get by increasing R after that is having a percentage of >>search spent in null move close to 0. So a potential of 2% or 3% increase in >>speed. >> >>And an big potential to overlook easy combinations everywhere in the tree. >> >>That's why I believe that working on R>3 is a waste of time. >> >> >> Christophe > > >You are overlooking _the_ point here. At present, doing 12-14 ply searches, >R>3 doesn't make a lot of difference. But in the future, when doing (say) >18 ply searches, R=4 will offer a lot more in terms of performance. Same as >R=3 did when we got to 12-14 plies... _then_ it might make sense to up R >once again. I do not know. I did not investigated different R's but I suspect that constant R may be a bad idea and R should be function of the position. I do not see a reason to use R=4 in the future and not to use it today at the same conditions. 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.