Newsgroups: comp.ai.neural-nets
Path: cantaloupe.srv.cs.cmu.edu!das-news2.harvard.edu!oitnews.harvard.edu!newsfeed.rice.edu!bcm.tmc.edu!cs.utexas.edu!swrinde!tank.news.pipex.net!pipex!warwick!bsmail!usenet
From: Max Ratcliffe <m.j.ratcliffe@bris.ac.uk>
Subject: Re: RBF Question.
Message-ID: <DG0tB6.38v@uns.bris.ac.uk>
Sender: usenet@uns.bris.ac.uk (Usenet news owner)
Nntp-Posting-Host: pc19.aer.bris.ac.uk
Organization: University of Bristol, England
References:  <44tser$g54@scotsman.ed.ac.uk>
Date: Fri, 6 Oct 1995 09:35:30 GMT
Lines: 15

> Is there any relevant literature available which discusses the choice of
> the widths of the Gaussians or addresses the no. of centres vs. width question ??

Hello there, I know that most RBF networks tend to use the Gaussian 
function.  However, other functions can be used as the non-linearity,
and for some reason the thin-plate-spline function does not need
a spread constant. 
             2    
     phi(z)=z log(z)

I don't know the reason for this fact (most people simply state that
a spread constant is not necessary). If someone can sort this out, 
I'd be grateful.  

Thanks, Max.
