Newsgroups: comp.ai.neural-nets
Path: cantaloupe.srv.cs.cmu.edu!rochester!cornellcs!travelers.mail.cornell.edu!news.kei.com!news.mathworks.com!tank.news.pipex.net!pipex!usenet.eel.ufl.edu!usenet.cis.ufl.edu!usenet.ufl.edu!zeno.fit.edu!zach.fit.edu!ram82001
From: ram82001@zach.fit.edu (Gary Russel /ADVISOR L. FAUSETT)
Subject: radial-basis function network
Message-ID: <DF7LBw.731@zeno.fit.edu>
Sender: news@zeno.fit.edu (USENET NEWS SYSTEM)
Nntp-Posting-Host: zach.fit.edu
Organization: Florida Institute of Technology
Date: Wed, 20 Sep 1995 14:52:44 GMT
Lines: 33

I've been working with a radial-basis function network in which
each input is an n-vector, the output is a scalar, and there are
m centers.  So

     m                            n
    ---                          ---
    \                            \            2
y = /  z w ,  where z  = exp(-s  /  (x  - c  ) )
    --- j j          j         j ---  i    ij 
    j=1                          i=1                    

Each s  is the "spread" associated with center c .
      j                                         j

                                      1
QUESTION:  Instead of s , Kosko uses ----  
                       j                2
                                     2s
                                       j

                                    ("one over two s sub j squared")

What are the pros & cons?  In the first form, one never worries
about dividing by zero.  In Kosko's form, the * of exp(*) is never
positive and therefore exp(*) is never greater than 1.
Does anyone have any thoughts on this?

Gary Russell
Florida Institute of Technology
ram82001@zach.fit.edu

"The boss said to change my sig quote so I did."

