Author: Uri Blass
Date: 22:23:14 11/08/00
Go up one level in this thread
On November 08, 2000 at 19:45:25, Christophe Theron wrote: >On November 08, 2000 at 03:08:27, Uri Blass wrote: > >>On November 08, 2000 at 00:15:28, Christophe Theron wrote: >> >><snipped> >>>All this was about the best program of 1990 at 500GHz against the best program >>>of today at 500MHz. >> >>There is a misunderstanding >> >>This discussion began because of the question if the best program of 1990 at >>500GHz are enough to equal kramnik(You claimed that 500mhz is enough with >>software improvement). >> >>I claimed that 500Ghz of 1990 is not enough to equal kramnik because it is even >>not enough to equal 1Ghz of today(the point is that 1Ghz is weaker than >>kramnik). >> >>> 77 Chess Genius 1.0 486/50-66 MHz 2182 23 -23 931 54% 2154 >>> 83 Fritz 3.0 486/50-66 MHz 2157 20 -20 1226 45% 2190 >> >>So it seems that the difference in rating between Fritz6a and Genius1 is >>274-(2182-2157)=274-25=249 elo. >> >>>>I believe that the difference is smaller for older programs and my guess is that >>>>programs like genius3 are going to earn less elo from doubling speed from >>>>500 mgh to 1gh relative to the new programs so 70 elo per speed doubling is not >>>>the actual value even without diminishing returns. >>> >>> >>> >>>But 30 elo per doubling speed is probably a lower bound, that's why I did the >>>two evaluations (with 30 and with 70). >> >>I am not sure if the estimate of 30 is a lower bound. >>My impression is that there are a lot of things that programs do not understand >>when time is not going to help. >> >>Thorsten posted a position that black is losing from a correspondence game when >>most programs do not evaluate correctly the king attack and tiger give more than >>+2 evaluation. >> >>At blitz it may be less important because tiger can do mistakes and miss the win >>against better hardware but I believe that if gambittiger gets 3 minutes per >>move on 1Gh machine nothing can stop it from winning. >> >>Here is the relevant position again. >> >>[D]rn1q1rk1/6bp/p2p4/1p1Pp2n/6b1/2NBB3/PP1QN2P/2KR3R w - - 0 1 >> >>Some programs do not see here even after hours that white can get a wining >>advantage after the opening. >> >>My impression is that programs with wrong evaluation have no way to stop gambit >>from getting this position and wining the game and 24 hours per move are not >>going to help because seeing in this position that white is winning is too late >>and they had to see it some moves before. >> >>Fritz6a has not this knowledge but I believe that it probably has also decisive >>knowledge against old programs of 1990. >> >>This is the reason for my impression that the 300 elo per 10 doubling may be too >>optimistics. > > > >There is something that could well compensate for the effect you describe. > >I think that some programs for 1990 (or around) had a lot of knowledge to >compensate for the lack of depth they could reach at that time. > >If we take the example of black trapped bishop on a2 or h2, you know very well >yourself that some modern program amongst the top ones do not have this >knowledge. Some older programs including Rebel for example have this knowledge. > >The same for pins. I even believe that king attacks were treated more seriously >by programmers in the 80s than by programs of today. For example ChessGenius for >Palm, which is the Mephisto Roma code (1987) seems more aware of king attacks >(and evaluates them rather high) than Genius5 itself. > >Programs of 1990 knew that they could be outsearched, and the programmers used >to add knowledge in the program to compensate for this. Between 1990 and 2000, I >think the most successful programs have been the one that had less knowledge but >were able to outsearch their opponents. For them it was outsearch the opponent >or die (because of lack of positional understanding). > >That's why I fear that modern programs on slower hardware could suffer more from >the fact that they will be outsearched. > > > > >>It is also possible that the 70 elo per doubling is not wrong but 70*10 is not >>700. >>I mean that if you play a match between equal programs when one hardware is >>twice faster you get 70 elo difference but when you play a match between equal >>programs when one hardware is 1024 faster you get less than 700 elo difference >>when part of it is not because of diminishing returns. > > >It will all depend on the time controls. Give a modern program 1 second per move >and a very old one 1024 seconds and I'm sure the modern one will not win one >game. > >But give the modern one 3 minutes and the old one 3072 minutes (51 hours and 12 >minutes) and probably the winning rate will be lower. > >In any case, the difference might well be above 400 elo points, and >unfortunately the evaluation systems I know are not able to measure an elo >difference bigger than 400 points between direct opponents... > > > > Christophe There are tables to give the difference in rating between programs based on the score(assuming that there are enough games not to get 100% score). I remember that the assumption behind the rating is that the ability of players have a normal distrivution with standard deviation of 200 and the abilities are independent variables. It means that the difference between them has standard deviation of 200*(2^0.5)=282.xxx(I will assume for the discussion that it is 282) It means that you can divide the difference between the players by 282 and get a number and when you look in the tables you can see the expected score in one game based on this number. 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.