Computer Chess Club Archives


Search

Terms

Messages

Subject: Re: A theory of ratings drift for the SSDF

Author: Uri Blass

Date: 13:26:48 04/09/02

Go up one level in this thread


On April 09, 2002 at 16:22:36, Uri Blass wrote:

>On April 09, 2002 at 16:04:01, Dann Corbit wrote:
>
>>I think it likely there is ratings drift for the SSDF.  Consider that for the
>>most part, these programs do not learn.  So "what you see is what you get" no
>>matter when you play them.  It is interesting to look over the programs on a
>>given platform in this list.  It appears to be a general trend that the ratings
>>have been pushed down over time.  Let's consider one example:
>>Ch.Machine 30 MHz King 2.0 aggr/R30 off(1996) 2296   21   -21  1153   65%  2185
>>Ch.Machine 30 MHz King 2.0 aggr/R30 off(1997) 2294   21   -20  1202   64%  2194
>>Ch.Machine 30 MHz King 2.0 aggr/R30off (1998) 2293   21   -20  1202   64%  2193
>>Ch.Machine 30 MHz King 2.0 aggr/R30off (1999) 2293   21   -20  1202   64%  2193
>>Ch.Machine 30 MHz King 2.0 aggr/R30off (2000) 2195   21   -20  1219   63%  2099
>>Ch.Machine 30 MHz King 2.0 aggr/R30off (2002) 2195   21   -20  1219   63%  2099
>>
>>It appears to have lost 100 ELO over the span from 1996 to 2195.  And yet it is
>>the exact same hardware that was used over each test.
>
>There is an obvious reason for it.
>The ssdf decided to reduce all the numbers by 100 elo because the rating of the
>top programs was too high.
>
>Uri

I forgot to add that they  decided that the rating is too high because they
wanted to have a better estimate for the rating of the top programs against
humans and they did not care about the old programs.

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.