Author: Frank Quisinsky
Date: 07:50:47 07/29/01
Go up one level in this thread
On July 28, 2001 at 10:59:36, Robert Hyatt wrote: >On July 28, 2001 at 04:17:18, Uri Blass wrote: > >>On July 28, 2001 at 01:12:41, Terry McCracken wrote: >> >>>On July 27, 2001 at 18:58:30, Dann Corbit wrote: >>> >>>>On July 27, 2001 at 18:44:45, Roy Eassa wrote: >>>> >>>>>On July 27, 2001 at 16:55:55, Dann Corbit wrote: >>>>> >>>>>>On July 27, 2001 at 16:51:08, Roy Eassa wrote: >>>>>> >>>>>>>Be careful, Dann. Dr. Hyatt has argued strongly that ponder should always be >>>>>>>on, even with a single CPU. (It seemed counter-intuitive to me too, but you >>>>>>>should check out his recent postings -- over the past couple days, I think.) >>>>>> >>>>>>Not when both engines play on the same machine. >>> >>> >>>>>>No way. >>>>> >>>>> >>>>>Yes way! I reacted the same way (in my head), but I respect Dr. Hyatt and his >>>>>arguments are worth reading at least! >>>> >>>>If you run two engines which are both pondering on a single CPU machine, then >>>>you are simply out of your mind. >>>> >>>>I think you must have misread the arguments. >>> >>>No Dan he's read the arguements correctly! >>> >>>Posted by Robert Hyatt (Profile) on July 27, 2001 at 13:28:16: >>> >>>In Reply to: Re: Permanent Brain ON vs Permanent Brain OFF posted by Uri Blass >>>on July 27, 2001 at 11:58:57: >>> >>> >>>On July 27, 2001 at 11:58:57, Uri Blass wrote: >>> >>>>On July 27, 2001 at 11:18:08, Miguel A. Ballicora wrote: >>>> >>>>>On July 26, 2001 at 14:41:54, Robert Hyatt wrote: >>>>> >>>>>>On July 26, 2001 at 12:55:06, Miguel A. Ballicora wrote: >>>>>> >>>>>>>On July 26, 2001 at 10:43:45, Robert Hyatt wrote: >>>>>>> >>>>>>>>On July 26, 2001 at 09:56:24, Matthias Gemuh wrote: >>>>>>>> >>>>>>>>> >>>>>>>>>Hi Robert, >>>>>>>>> >>>>>>>>>I think you just wanted to make a joke. We all know that PONDER OFF hurts nobody >>>>>>>>>(Fritz used its full time). PONDER ON on one CPU is very appropriate to arrive >>>>>>>>>at wrong engine comparasons. >>>>>>>>> >>>>>>>>>Regards, >>>>>>>>>Matthias. >>>>>>>> >>>>>>>> >>>>>>>>>>>>>both will get 1/2 of the machine and the time controls won't be screwed up. >>>>>>>> >>>>>>>>ponder=off exposes the opportunity for a program to get into time trouble >>>>>>>>because it assumes it will save time with ponder=on when it really can't since >>>>>>>>it is disabled... >>>>>>> >>>>>>>Why the program should assume that it will save time in with ponder=on when >>>>>>>it knows that it is off already? >>>>>>>Shouldn't a program take this into account? >>>>>>>If ponder=off is an option for the program, it should notice the difference >>>>>>>and act accordingly IMHO. >>>>>>> >>>>>>>Regards, >>>>>> >>>>>> >>>>>>because in my case, 99.9% of all games played have ponder=on. I only disable >>>>>>pondering to debug so that I can reproduce the same searches over and over >>>>>>when necessary. Since almost all real games are played with ponder=on, I don't >>>>>>have a special time-allocation formula for ponder=on and another one for >>>>>>ponder=off. I just have one that _assumes_ ponder=on. >>>>>> >>>>>>I see no reason to waste what little time I have working on something that is >>>>>>hardly going to be used... >>>>> >>>>>Well, it is used a lot actually by lots of people already. Most of the people >>>>>are running matches with ponder=off for some reasons. >>>>>If both engines were tuned for ponder=off, it will be the best condition to >>>>>optimze the resources since time used pondering is never as good a time used >>>>>thinking. For instance, you have a competitive mindset in your answer but if my >>>>>purpose is to run a match between engines to learn a particulat opening, I want >>>>>my resources to be used as efficient as possible. That is ponder=off for both. >>>>>Maybe you could consider making Crafty to be able to process "ponder=off" >>>>>accordingly because there will be users that would benefit from it. >>>>> >>>>>Regards, >>>>>Miguel >>>> >>>>Crafty is not a commercial program so I guess that the way that users use it is >>>>not important for Bob. >>>> >>>>I also do not think that the difference between ponder on and ponder off is more >>>>than 20 elo in most of the practical cases. >>>> >>>>Uri >>> >>> >>>I disagree. When this first cropped up a couple of years ago, I ran a test, >>>gnuchess on a single-cpu, vs crafty on a quad pentium-pro 200, but using only >>>one cpu. The quad was far slower than the 500mhz single cpu by a significant >>>amount. The match was pretty even (ponder=on) even though crafty had a much >>>slower processor. With ponder=off, the match was way more lop-sided in favor >>>of GNU because crafty would get into time trouble and near the end of the >>>time period it would have to move too quickly to avoid losing on time. >>> >>>It was _clearly_ worse. And by more like 100 points, too. I adjusted it a bit >>>to help, but I have _never_ invested as much time in the ponder=off timing as >>>I have in the ponder=on timing. And I never will, any more than I am going to >>>try to tune my son's mustang to run on a road coarse when we only take it to >>>the 1/8th and 1/4th mile drag strips... >>> >>>I think you should spend the time making it optimal in the way it will normally >>>be used, not in oddball configurations... >>> >>> >>>I ran a bunch of games, everything equal but the hardwar >>> >>> >>>Quote/Hyatt; Nope... no joke at all. Two programs, one machine, my preference >>>is ponder=on. >> >>Note that it is hyatt's opinion. >>I said that in most of the practical cases the difference between ponder on and >>ponder off is not more than 20 elo. >> >>The practical cases are usually with the commercial programs. >>I am not talking about amatuers like GNUchess. >> >>I agree that there may be a big difference with amatuers. >>One reason is simply the fact that part of the amatuers do not know to ponder so >>ponder off is clearly better for them than ponder on. >> >>I do not know if this was the reason for GNU chess but when I read 100 elo >>difference I suspect that this is the reason. >> >>I do not believe that program are handicapped by more than 100 elo if they do >>not know to ponder. >> >>I believe that the difference is even less than 70 elo(70 elo is the difference >>from doubling the speed of the program and it is better to be twice faster and >>play without pondering). >> >>Uri > > >No... the problem was definitely in crafty. I was playing games of 15 minutes >for 40 moves. At move 35-40, crafty was in time trouble and made bad moves in >many (but not all) games because it had far less time than GNU did. You can't >afford to do a 6-8 ply search and move when your opponent is doing 10-12 plies >on that same move because he has used his time more reasonably. > >I fixed the problem in a simplistic way by simply spreading out the time over >the time-control-moves in a more uniform way. Very unlike the way a human plays >chess of course, but it was one quick-and-dirty way to avoid getting blasted >near the end of the time control. Hi Bob, the same discuss then for 2 years here :-) I, the ex non ponderer, will say the following: 01. You have right if you say that with ponder = off engines have time manangment problems, not all and not in all games but it's an important point. I know that this is for statistics not sooo important (ELO statistic). 02. Match without ponder on single system with Athlon 1Ghz: Engine A = 1Ghz Engine B = 1Ghz Match with ponder on a single system with Athlon 1Ghz: Engine A = ~ 497-500Mhz Engine B = ~ 497-500Mhz Now we have 25-40% ponder hits if the engines play on the same level (after my experiments). Result: 500Mhz + 25-40% ponder hits = ~700MHz. With ponder = on, the matches are running with 700Mhz on 1Ghz Athlon! With ponder = off, the matches are running with 1Ghz on 1Ghz Athlon! I believe better is to play with ponder = off on single system. OK, the time managment is a good point so say please with ponder but 300 MHz is a better point to say please without ponder. 03. 4-piece and 5-piece tablebases and engine-engine matches on a Dual system ... This is not very clear. We have engines witch used tablebases very aggressive or not very aggressive. Example: Gromit - Patzer with 5-pieces ... Gromit played move 50 in 2:45 with ponder = on and 5-pieces. Only 20-30% processor time in this game if Gromit used 5-pieces. Patzer played move 50 in 1.25 with ponder = on and 5-pieces. Now the same positions with 4-pieces ... Gromit play move 50 in 1.58! with ponder = on and *4-pieces*. Patzer played move 50 in 1.13 with ponder = on and *4-pieces*. Without 4-piece tablesbases are the differents not very important. In my opinion it is better to play eng-eng matches with ponder on a dual system with 4-piece tablebases. Best Frank
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.