Computer Chess Club Archives


Search

Terms

Messages

Subject: Re: Hello from Edmonton (and on Temporal Differences)

Author: James Swafford

Date: 10:16:34 08/02/02

Go up one level in this thread


On August 02, 2002 at 12:45:39, Sune Fischer wrote:

>On August 02, 2002 at 12:11:48, Rudolf Huber wrote:
>>
>>Any respectable result (>=25%) of TDLeaf would be a big surprise
>>to me. In my experience, automatic learning methods quickly learn the
>>simple things but have a hard time with the finer aspects of chess.
>>
>>So for me it is not really necessary to use a common search engine
>>as a base to prove the point that the TDLeaf method has potential.
>
>The problem is, that if you take a new engine that isn't very optimized to begin
>with, you are not proving it can be done _better_ than hand tuning.
>I think we can agree TDLeaf is better than nothing as the KnightCap project
>showed, but would it work well enough to also improve on Crafty?

Right!  That is one of my biggest questions.

>
>>And one should also note that hand tuning will profit from public
>>source code which gives it a big advantage.
>
>Hand tuning has it's limits. You can't adjust every weight to 1/100th pawn
>accuracy, there are thusands of weights with non-linear relations.

Yes... see the Sutton quote in a previous post.

>
>Given the computerpower, I think it would be possible to do it better than a
>human tuning.
>
>For instance, do you have any idea what the right material values are?
>Some use 1, 3, 3, 5.5, 9 and others 0.8, 3, 3.3, 6, 11 or what ever.
>I think the *real* values are impossible to guess, it's probably something like:
>1.08, 3.12, 3.35, 5,87, 10.64 or in other words - impossible to find without
>machine tuning.

I'm not sure there are any correct absolute values.  I think the true
values change with the position.



>
>-S.
>
>> Rudolf



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.