Author: Bas Hamstra
Date: 15:01:54 01/09/02
Go up one level in this thread
On January 09, 2002 at 04:31:54, Uri Blass wrote: >On January 08, 2002 at 18:54:20, Bas Hamstra wrote: > >>On January 07, 2002 at 14:52:49, Uri Blass wrote: >> >>>On January 07, 2002 at 14:42:35, Bas Hamstra wrote: >>> >>>>On January 07, 2002 at 12:06:24, Uri Blass wrote: >>>> >>>>>On January 07, 2002 at 11:44:44, Bas Hamstra wrote: >>>>> >>>>>>On January 07, 2002 at 10:29:12, Uri Blass wrote: >>>>>> >>>>>>>On January 07, 2002 at 07:35:14, Bas Hamstra wrote: >>>>>>> >>>>>>>>On January 07, 2002 at 07:08:08, Vincent Diepeveen wrote: >>>>>>>> >>>>>>>>>On January 06, 2002 at 21:17:43, Bas Hamstra wrote: >>>>>>>>> >>>>>>>>>>...because the fastest hardware simply wins. You can invent all kind of >>>>>>>>>>ingenious tricks, but it's nothing compared to faster hardware. On 2x faster >>>>>>>>>>hardware Tao just crushed GT 3x in a row and won the latest 10 15/0 games at >>>>>>>>>>FICS against strong opponents on slower hardware. Come on, the only fair way to >>>>>>>>>>compete is on equal hardware. I don't want to buy a computer twice a year just >>>>>>>>>>for CC tournaments, that's ridiculous. IMO the competition would be much more >>>>>>>>>>satisfying on equal hardware. Factor 2 hardware difference means hard to win for >>>>>>>>>>any program against a not too bad opponent. Anything above that makes the >>>>>>>>>>chances *way* too small to be fair. Yet that is quite normal in tournaments and >>>>>>>>>>you won't hear anyone about it. Program X played this AMAZING knight sac >>>>>>>>>>againtst program Y!! Hardware differences seem to be simply ignored. And that's >>>>>>>>>>crazy, in fact. >>>>>>>>>> >>>>>>>>>>Bas. >>>>>>>>> >>>>>>>>>Then why didn't Zugzwang win IPCC99. It is not a bad program at all. >>>>>>>>>In tests it completely annihilated the commercial programs they tested >>>>>>>>>against. >>>>>>>>> >>>>>>>>>Zugzwang was at like 512 alpha processors and getting millions of nodes >>>>>>>>>a second. I can't even remember how much, but zugzwang is already a slow >>>>>>>>>program on a PC... >>>>>>>>> >>>>>>>>>...it was using (making use of message passing and thereby losing many >>>>>>>>>factors of speed but it is worth it) global hashtable and was having more >>>>>>>>>Mhz and bigger hashtables than anyone else. >>>>>>>>> >>>>>>>>>This though the 17 ply searching Cilkchess at like 256 (or 500?) >>>>>>>>>sun processors wasn't searching undeep either. Yet it lost chanceless from >>>>>>>>>8 ply searching Lambchop. >>>>>>>>> >>>>>>>>>How do you explain that? >>>>>>>> >>>>>>>>Yes, I expected this. >>>>>>>> >>>>>>>>a) The probability that the "best" program wins a tournament is far smaller then >>>>>>>>you would expect. We once did some math and simulation, the shocking conclusion >>>>>>>>was that the probability that the best program wins the tournament was only 50% >>>>>>>>or something near that. >>>>>>>> >>>>>>>>b) CilkChess is a bad program. If you practically only do piece square then even >>>>>>>>17 ply won't help you. >>>>>>>> >>>>>>>>The server or the SSDF gives better data (more) to draw conclusions from. Both >>>>>>>>indicate that hardware is an enourmous factor. >>>>>>>> >>>>>>>>Best regards, >>>>>>>>Bas. >>>>>>> >>>>>>>The ssdf does not say that hardware is an enourmous factor >>>>>> >>>>>>Are you serious? When has there EVER been a program in the top-3 that was not on >>>>>>the fastest chip? >>>>>> >>>>>>>A1200 against K6-450 is more than2 times faster and >>>>>>>Crafty on A1200 is not better than the best programs >>>>>> >>>>>>What does this prove other than that it is hard to overcome a factor 2? It >>>>>>hardly occurs! Clearly a BIG factor... >>>>>> >>>>>>>on K6-450. >>>>>>>If your program is really better than GambitTiger >>>>>>>when the hardware difference is only 2:1 then >>>>>>>you may do it a commercial program. >>>>>> >>>>>>On 2 to 1 hardware advantage I fear not a single program. But that's easy to say >>>>>>because that holds for everyone with a not too crappy program. Better at 2:1? I >>>>>>don't know, it would certainly be an interesting experiment to test it. >>>>>> >>>>>>Bas. >>>>> >>>>>By your definition most of the programs >>>>>are too crappy programs >>>>> >>>>>see http://home.hccnet.nl/leo.dijksman/index.html >>>>> >>>>>The programs in the second devision and lower devisions are >>>>>most of the programs and I expect all of them to lose a match >>>>>of 10 games against Tiger when the hardware difference is 2:1 >>>>>against Tiger. >>>>> >>>>>I expect even most of the programs in the first devision >>>>>that includes Crafty to lose a match against Tiger >>>>>in the same conditions. >>>>> >>>>>Uri >>>> >>>>I have seen the programs. Tell me why you expect that, maybe we can solve the >>>>problem. Are you willing to take a bet? My money is on YACE bigtime. Let's say >>>>8-2 for YACE if the books are about equal. >>>> >>>>Bas. >>>> >>>>Bas. >>> >>> >>>8-2 for yace is not ralistic. >>> >>>Even if you take the same program against itself on >>>2 times slower hardware you are not >>>going to get an average result of 8-2 >>> >>>7.5-2.5 is 200 elo difference and being 2 times faster >>>gives only 70 elo advantage. >>> >>>Uri >>> >>>Note that yace is a strong program >> >>Well, sorry I picked a program from you list :-) >> >>And yes, a result 8-2 with one program playing itsself at 2:1 is quite possible, >>why do you say that? The ELO formula is nice, but A will see everything that B >>does, and MORE. >> >>Let's say we take the latest Fritz and play 100 5/0 games. Now, with the ELO >>formula in the back of our heads, what do you predict? >> >> >>Bas. > >I predict 59-41 for the twice faster hardware >that means 72 elo difference. > >Uri Note you are using a standard ratingdifference to predict a Blitz result. Wouldn't you think 2x speed is more important at Blitz? Besides: I do not believe the rating difference based on play against a "pool" of opponents can reliably predict the outcome of a 2:1 selfplay match at *any* time control. I haven't looked at SSDF for quite a while, but does the same version of Fritz play itsself at SSDF at different hardware? If so it would be interesting to see the scores... Bas.
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.