Author: Christophe Theron
Date: 10:30:45 01/30/04
Go up one level in this thread
On January 30, 2004 at 09:28:38, GuyHaworth wrote:
>
>Several self-play experiments have been run with the two sides set to different
>plies of 'basic search depth', a concept somewhat open to debate given that one
>would not wish to prevent an engine chasing a line to stability if it wanted to.
>
>The most comprehensive experiment was by Ernst Heinz: his results have been
>published in his book and (a later and larger experiment) in papers.
>
>
>I re-analysed the 'significance' of the results because I believed that the
>claim of 'diminishing returns' could be made with more statistical_confidence
>than Ernst was attributing to the data.
>
>Ernst' exp went to 12 plies, not 16. There is a possibility that it is more
>meaningful to compare a match of 12-11 games with a match of 10-9 games (i.e. a
>whole move different) rather than comparing 12-11 with 11-10 and 11-10 with 10-9
>- because of the stm/sntm bias.
>
>From memory, I think the benefit was about 50 ELO (not plies) were ply at the 12
>level ... and decreasing.
>
>
>A related question that I'm pondering at the moment: how many ELO are lost by
>halving the time available to the computer?
>
>g
As Uri said, it is most probably in the 50-70 range. Some time ago I used the
SSDF data and found that doubling the speed of the computer resulted in a 70 elo
points increase.
Naturally, it must depend on the chess program. But there was a kind of
uniformity in the result for various chess programs. I think that the 70 elo
figure applies to state of the art branching factor chess programs.
I suspect that with faster computers the doubling of speed could give only 50
elo points (I believe in dimishing returns) but I don't think this has been
proved yet.
If you want to measure this yourself, I would strongly advise you to take the
SSDF data and work on it. It is the most reliable data and comprehensive data
that we have at this time, and I'm sure it is possible to extract this info from
it. It would allow you to study slow and fast computers, and probably to find a
significant difference between the two categories (for slow computers, doubling
the speed probably gives a better elo increase than for fast computers, if the
law of dimishing returns is true).
Christophe
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.