Computer Chess Club Archives


Search

Terms

Messages

Subject: Linear Increase in Elo Rating With Depth (Revisited)

Author: Michael Neish

Date: 00:40:58 02/01/00



I read some posts on this subject in the recent thread, and saw the table
which was reproduced from the Dark Thought Web site.  I can't say I fully
understand it.  It seems that the table listed the performance of various
computers at various depths, not the performance of one computer as it searched
deeper.

Just wondering, what rigorous tests have been done to measure Elo rating as a
function of ply depth for one particular program?  And is it possible to see the
results?  By rigorous I mean consisting of a large enough number of games (at
least several hundred for each search depth tested).  How deep did these
analyses go?

Nothing seems easier than to set a program to play itself over several hundred
games and then see how the performance varies with depth, so it must have been
done many times before.  At the moment I'm doing such a run for CM6000, which is
the only program I have that allows me to define different search depths for
White and Black  If anyone is interested I will post the results after I've
accumulated enough games.  At the moment, though, it appears that a 5-ply
searcher will score about 80% against a 4-ply searcher, and likewise for a 7-ply
searcher against a 6, although I haven't gathered enough games to say anything
conclusive.

Is there anything wrong with gathering data in this way?  Since I'm testing a
program against itself, I wonder whether the fact that both sides think in the
same way (albeit to a different depth) can skew the results somehow.  Would it
be better if I chose two different personalities for the same program?

Also, I suppose the use of search extensions will probably affect the results.
Should they be turned off?

Additionally, how does Elo increase with depth reflect on a program's
performance against humans?  A whole new kettle of fish altogether, I suspect.

Any comments would be welcome, and apologies for attempting to resuscitate a
recent thread.

Cheers,

Mike.




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.