Re: Why RBF classfier Vs sigmoidal MLP for Statistical classification?

[EMAIL PROTECTED] (Peter J. Acklam) wrote in message 
news:<[EMAIL PROTECTED]>...
> [EMAIL PROTECTED] (Greg Heath) wrote:
> 
> > Thick-tailed Cauchy has no variance.
> 
> Sure it does, but the variance is infinite.  It's the mean (and
> all other odd-numbered moments) that don't exist.

I'm not a mathematician (and I don't play one on TV), but
my engineering sense tells me that: 

Improper integrals with odd bounded integrands have
a Cauchy Principal Value of zero.

> > Theoretically, Student-t may be more appropriate.
> 
> Although a Student-t with one degree of freedom is a Cauchy
> distribution.

A random N-case sample from a Gaussian distribution has
a student-t distribution with N-1 degrees of freedom.

I don't think Paul has tried to train any RBFs with
a two-case design set.

Hope this helps.

Greg