Computer Chess Club Archives


Search

Terms

Messages

Subject: Re: Puzzled about testsuites

Author: Michael Yee

Date: 12:47:39 03/10/04

Go up one level in this thread


On March 09, 2004 at 17:48:19, Gian-Carlo Pascutto wrote:

>On March 09, 2004 at 16:37:46, Michael Yee wrote:
>
>>Wow. I was partly just being facetious with my initial comments. But I actually
>>was mostly serious. I certainly don't think the analogy is "completely" flawed
>>(since I think that would invalidate a lot of reasonable ideas in machine
>>learning).
>
>Machine learning is not flawed, your analogy of what he's doing to what
>you're talking about is!
>
>>If you parameterized your whole program, I don't see why a global search
>>technique couldn't find the same weights that you hand-coded or even better ones
>>(given a nice large training set). For example, let f(x) = DS's performance in a
>>tournament given param vector x. Then a search technique (e.g., tabu search)
>>could be used to optimize f(x) over x. I admit that it could take a long time,
>>but I don't think it's impossible. (Also, I think it would still work if f(x)
>>was based on the ability to match GM moves form a large set of training
>>positions.)
>
>Yes, this would work. No, this isn't what he's doing. He's making changes in
>(the parameters of) one part of the program, and then measuring something
>entirely different.
>
>Strength of evaluation and the results of a tactical testset are not correlated
>enough to be useful for optimalisation. As he already noticed in his first post,
>they can be negatively correlated, even.
>
>--
>GCP

Hi Gian-Carlo,

You're right--I wasn't really referring to what the first poster had tried to
do. Instead, I was responding to your general statement: "Tuning your evaluation
on a testsuite is fundamentally wrong."

I just wanted to give some examples of approaches (i.e., parameter estimation
based on training examples) that seem reasonable to me. In other words, I wanted
to qualify your statement a bit. If I ever get positive results (similar to some
already existing in papers that Dave hinted at), I'll be sure to share them.

Michael



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.