Computer Chess Club Archives


Search

Terms

Messages

Subject: Re: chess computer ratings

Author: Robert Hyatt

Date: 08:28:13 09/24/01

Go up one level in this thread


On September 24, 2001 at 09:56:50, Sonja Tiede wrote:

>>>>Against humans, there is no real data.  If you are talking about taking a
>>>>program on an XXX megahertz machine, and playing it against a program on
>>>>a 2*XXX megahertz machine, then the faster machine will be rated about 60
>>>>points higher, using the typical Elo formula.  I don't think that +60 will
>>>>be true for games vs humans, however.  It might be 1/2 of that or even less.
>>>
>>>
>>>Take 2 programs, one with 2000 SSDF ELo and one with 2600 SSDF Elo,
>>>assume the 2000 value is true against humans.
>>>Do you really think the other programm has a human-strength of less than 2300 ,
>>>since all SSDF elo-values are based on comp-comp matches ?
>>
>>
>>It is certainly possible.  But more likely, take a 2500 SSDF program and a
>>2650 SSDF program and play them against humans, it is likely they will produce
>>similar results...
>
>Ok. Your example is true, but in my opinion this (your example 2500&2650) is
>a result of a 'tuned' book  and book learning, and not a result of a generall
>'lack' in measuring the playing strength only with comp-comp matches.
>
>What i want to say is,
>(1) that book learning does improve the comp-comp
>'winning rate'  but the overall playingstrength is not going to be improved,
>even not in the opening!!!


However, this has nothing to do with a +60 Elo improvement due to hardware
speed.  I am assuming that the _program_ is the same in both cases, which is
the only reasonable way to evaluate hardware speed improvements.



>The convergence-rate of todays book-learning algorithm is _extremly_ slow, and
>if you observe a overwhelming 150-50 result with similar comp-comp-opponents
>(prg A vs. prg B, prg A book-learning enabled) you are witness of some other
>'effects' but you cannot be sure that program A is better.

Again, If you take the same two programs, and vary _only_ the hardware, then
any difference _must_ be the hardware.  SSDF is not passing learning info
around between the testers.  Most programs that I know of (Crafty being the
one exception) don't have an external portable learning file...



>
>In generall , when both programs are using book-learner the variance is rising
>up , and you need thousands of games to get an result that reflects
>the true playing strength.
>
>
>(2) the book that is shipped originally with a programm is not tuned
>to be good in general (in chess theory), it is tuned to play 'succesful' against
>the most important competitors

Again, true.  But it won't affect the results of different hardware matches.



>
>(3) comp-comp matches without the 'lack' of (1)+(2) are to be determined
>better been suitable the playing strength than comp-human games, since a
>single human would apply point (1).



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.