Computer Chess Club Archives


Search

Terms

Messages

Subject: Re: A New Self-Play Experiment -- Diminishing Returns Shown with 95% Conf.

Author: Dann Corbit

Date: 15:16:05 05/24/00

Go up one level in this thread


On May 24, 2000 at 17:46:55, blass uri wrote:
>On May 24, 2000 at 17:27:19, Dann Corbit wrote:
>>I have a bit of a puzzle.
>>
>>In these experiments, you varied the depth from 5 to 12 plies, but in the "goes
>>deep" experiments, it was deeper, if I recall correctly.
>
>It was deeper but it was not based on comp-comp games and I think that the
>positions do not demonstrate what happens in comp-comp games(I read that the
>positions were taken from kasparov-deep blue games and if this is the case
>kasparov chose positions that computers do not understand so the fact that they
>changed their mind even at big depthes is not surprising(you change your mind
>when you have no idea what to do).
>
>>
>>In the earlier experiments, it seemed to level out.
>>
>>I would like to see control matches where the plies are equal.
>>IOW, 5:5, 6:6 up to 12:12.
>>In that way, you will know what the gain of the extra ply is.  I don't think you
>>can tell clearly strictly from the data given.
>>
>>For example, in your 12 <==> 11 experiment, you show 35.71% wins, 52.38% draws,
>>and 11.9% losses.  What would the result be for 12:12?
>
>I guess that the result will be 50% because Fritz6 played against itself.

I will be shocked if it turns out that way.  For instance, no computer chess
program has a perfectly symmetrical eval.

>You can learn from 12<==>12 experiment about the number of draws but I do not
>see what you can learn about diminishing returns from more plies.

I probably do not understand exactly how the experiment was run.  I will have to
read in detail and study it later.  But when you change both variables at once,
you don't have a control.



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.