Computer Chess Club Archives


Search

Terms

Messages

Subject: Re: chess computer ratings

Author: Robert Hyatt

Date: 06:58:38 09/24/01

Go up one level in this thread


On September 23, 2001 at 11:51:02, Mike S. wrote:

>On September 23, 2001 at 11:13:14, José Carlos wrote:
>
>>(...) Besides, I belive the "scale" (if we can
>>use this term) is smaler in human world than in comp-comp world, meaning that a
>>small difference between two versions of a program (or two different programs,
>>of course) can appear over and over in games, translating into more and more
>>rating points.
>
>This would be the case only if it's not just a small, but also a decisive
>difference, furthermore many programs have position learning.
>
>But there's another aspect of this type of speed/strength estimation: The faster
>program can't be +60 elo (or whatever the value is) stronger at every move in a
>game. Actually, many moves will be the same on 400 MHz and on 800 MHz, if this
>is the only difference of the conditions. In some positions, moves will be
>different but not necessarily stronger (i.e., if there is a number of moves
>within a narrow margin of evaluations). Btw. I've heard experiments have been
>made, showing that approx. 16% of moves are different if the search is one ply
>deeper.




You are thinking about "Elo" wrongly.  Elo is a _game outcome_ predictor.  There
is no "60 elo on every move".  There _is_ a 60 elo when applied to a game.  It
is possible that on the right move, a 2x faster machine is 500 elo stronger.
But overall, it is 60 elo better.  that 60 is what the formula measures...




>
>So I think there is a very small number of moves, maybe only one or two per
>game, where the higher speed is decisive for the result. Humans are certainly
>more flexible in time consumption, and probably can adapt better to that than
>programs. Furthermore, the higher speed affects only parts of a programs
>performance. Opening book, egtbs, and parts of what can be described as
>knowledge (in human terms) of the program, is the same at all cpu speeds and
>time settings.
>
>Regards,
>M.Scheidl

In the "Crafty goes Deep" experiment, we found that Crafty would change its
mind about 15% of the time, if it is given more and more time.  This held true
for machines that could do in 3 minutes what we did in 24 hours 2 years ago.
It will be a _long_ time before the hardware to make that happens, and causes
us to have to run another set of 24 hour positions again..



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.