Author: Dann Corbit
Date: 14:16:39 04/16/04
Go up one level in this thread
On April 16, 2004 at 16:53:40, Dave Gomboc wrote:
>On April 14, 2004 at 19:48:53, Dann Corbit wrote:
>
>>On April 14, 2004 at 19:46:18, Dave Gomboc wrote:
>>
>>>On April 14, 2004 at 18:17:07, Dann Corbit wrote:
>>>
>>>>I have downloaded the document and skimmed it.
>>>>
>>>>I will read it carefully this weekend.
>>>
>>>In that case you might as well re-download it when you sit down to read it.
>>>I've already made some tweaks that aren't yet posted.
>>
>>I am curious about the "wandering" evaluation terms.
>>
>>I suspect two things about these terms:
>>1. That [usually] the correlational coefficient will be much smaller for these
>>2. That they occur in pairs {or vectors} (e.g. if wildly different sets can
>>both have nearly optimal results, then likely there are interactions)
>
>I'm not sure what you mean by #1.
The correlational coefficient is the goodness of the fit, for whatever gradient
you are following. If (for instance) the gradient was tracing out a straight
line and all the points were exactly on the line, then the correlational
coefficient will be 1.0. That means that there is a powerful relationship and a
direct determination. On the other hand, if most of the points were well off of
the line, then the correlational coefficient will be small. A correlational
coefficient of about 0.3 is considered as insignificant such that we cannot tell
if there is any relationship between the variables.
E.g. strongly correlated:
*
/
*
/
*
/
/
*
E.g. not correlated:
*
*
*
*
*
*
*
>
>Regarding #2, my experience is that there are indeed many interactions.
>
>Some features contribute only a small amount to the evaluation function quality
>-- the weights of such features can have a wider range without penalizing
>effectiveness too highly. It'd probably be better to find better features,
>though. :-)
>
>Dave
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.