Author: Uri Blass
Date: 12:11:56 10/08/02
Go up one level in this thread
On October 08, 2002 at 14:33:05, Robert Hyatt wrote: >On October 08, 2002 at 13:26:52, Uri Blass wrote: > >>On October 08, 2002 at 13:16:22, Robert Hyatt wrote: >> >>>On October 08, 2002 at 12:22:30, Uri Blass wrote: >>> >>>>On October 08, 2002 at 12:10:41, Robert Hyatt wrote: >>>> >>>>>On October 08, 2002 at 10:55:29, Uri Blass wrote: >>>>> >>>>>>On October 08, 2002 at 10:50:20, Robert Hyatt wrote: >>>>>> >>>>>>>On October 08, 2002 at 07:08:51, Uri Blass wrote: >>>>>>> >>>>>>>>On October 08, 2002 at 00:52:38, Eugene Nalimov wrote: >>>>>>>> >>>>>>>>>Wrong. >>>>>>>>> >>>>>>>>>Today I visited the talk by Feng-Hsiung Hsu he gave at Microsoft. Here are some >>>>>>>>>points from memory: >>>>>>>>> >>>>>>>>>They used forward pruning in the hardware, and according to Hsu it gives them >>>>>>>>>5x-10x speedup. He wrote about that in the book, too, but without any details. >>>>>>>> >>>>>>>> >>>>>>>>Can you ask him if 12(6) really means 12 plies in the software and 6 plies in >>>>>>>>the hardware? >>>>>>>> >>>>>>>>A second question is if the plies in the hardware were selective search from the >>>>>>>>first ply. >>>>>>>> >>>>>>>>>In the talk he named that pruning as "analogy cutoff" and mentioned that "if the >>>>>>>>>move is useless in some position, it is also useless in the similar position". >>>>>>>>>In the book he writes "it can be done in the hardware as long as it does not >>>>>>>>>have to be 100% correct". >>>>>>>>> >>>>>>>>>They used null-move thread detection, as well as not only singular extension, >>>>>>>>>but also extension on only 2 or 3 good replies. They used fractional extensions. >>>>>>>>>He also says that their Q-search is much more powerful than the one that is >>>>>>>>>usually used in the software-only programs. >>>>>>>>> >>>>>>>>>Hsu gave some details why they don't use null-move: >>>>>>>>>(1) He thinks that singular extensions and null-move gave more-or-less the same >>>>>>>>>rating difference (100-200 points IIRC). >>>>>>>> >>>>>>>>I think that he underestimates null-move pruning. >>>>>>>> >>>>>>>>I believe that for long time control null move pruning gives more than 100-200 >>>>>>>>points. >>>>>>>> >>>>>>>>People may try Fritz with selectivity=0 to find it's rating without null move. >>>>>>>> >>>>>>>>Uri >>>>>>> >>>>>>> >>>>>>>I can assure you it doesn't. Several of us ran this experiment in the past. It >>>>>>>produced a 50-100 >>>>>>>point improvement at most. Bruce ran it first. I then repeated it to see if >>>>>>>his result held for me >>>>>>>as well. 50-100 is nothing to sneeze at of course... But that is all it will >>>>>>>give... >>>>>> >>>>>>What was the time control and the hardware. >>>>>> >>>>>>I believe that the improvement is bigger >>>>>>at slower time control. >>>>>> >>>>>>If the experiment was some years ago and >>>>>>in time control that is faster than 120/40 >>>>>>then the results may be different today. >>>>>> >>>>>>Uri >>>>> >>>>> >>>>>I don't know about Bruce. I used 40 moves in one hour followed by 20 moves in >>>>>30 minutes, >>>>>with no sudden-death at all. >>>>> >>>>>I ran it on several computers here for several weeks... >>>> >>>>We have a factor of 2 in the time control. >>>> >>>>What was the hardware that was used? >>>>If the games were played 5 years ago then today we have clearly >>>>faster hardware. >>>> >>>>Uri >>> >>> >>>Pentium pro 200's. I played 2 games at a time on my quad, plus more games in a >>>linux lab we >>>had set up. >>> >>>yes hardware is faster. No I don't believe that going deeper and deeper >>>eliminates all the problems >>>with null-move. I only saw bad problems at depths of 5-6-7-8. I _never_ saw >>>them pop up at >>>depths of 12 and beyond, which Crafty could reach in the pentium pro at 1 minute >>>a move... >>> >>>R=2 used to gain at most 2 plies. Yet the overall performance improvement from >>>my testing >>>was in the 50-60 point range. Far less than what you would normally expect from >>>gaining 2 >>>plies. The conclusion? You aren't _really_ gaining two plies of search depth, >>>just two plies >>>reported in the output... >> >>I did not say that you get 2 plies of search depth from 2 plies of output but >>only that I believe that the difference is bigger than 100-200. >> >>It may be interesting to repeat the experiment today >>(you use R=2/3 and not R=2 and it is also important). >> >>Uri > > > >OK... that means that you think that a null-mover will win 3 of every 4 games >vs a non-null-mover, >_everything_ else being the same? > >I can run that test. > >I'll let you pick the time control. Pick something that xboard/winboard will >like and I'll fire it up on >a quad and let it play two games at a time using one processor for each side) >for a while... > >How about 40 moves every 60 minutes or something similar, repeated until the >game is drawn or >ends normally... I guess that even Crafty(R=3) is good enough to get 75% at 60 minutes/40 moves against Crafty(R=0) (I think that programs that do checks in the qsearch like Tao can get even better results). My guess is based on my experience with movei when my results suggest that the latest movei earn more from time relative to the public movei when one of the differences is that I use R=3 in the latest version when the public version is using R=2. I admit that I did not do comparison of only R=2 against R=3 except test suites when R=3 scored better. Uri
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.