Newsgroups: comp.ai.neural-nets
Path: cantaloupe.srv.cs.cmu.edu!bb3.andrew.cmu.edu!newsfeed.pitt.edu!godot.cc.duq.edu!newsgate.duke.edu!news.mathworks.com!newsfeed.internetmci.com!usenet.eel.ufl.edu!bofh.dot!warwick!bsmail!usenet
From: M.J.Ratcliffe@bris.ac.uk
Subject: Can't classify a previously successful regression...
Content-Type: text/plain; charset=us-ascii
Message-ID: <Ds876M.LH9@uns.bris.ac.uk>
Sender: usenet@uns.bris.ac.uk (Usenet news owner)
Nntp-Posting-Host: pc53.aer.bris.ac.uk
Content-Transfer-Encoding: 7bit
Organization: University of Bristol, UK
Mime-Version: 1.0
Date: Thu, 30 May 1996 16:01:34 GMT
X-Mailer: Mozilla 1.1N (Windows; I; 16bit)
Lines: 20

Catchy title, huh?

This isn't a huge problem, but it's something I don't understand. I have 
trained a radial basis function net (using Gaussians) to estimate two 
parameters. The net generalizes reasonably well. However for various 
reasons, I also want to use a classifier on the same data, to say "Well 
the data looks more like output #1 than output #2". 

The networks really do not generalize well on this classification task, 
despite the fact that the problem seems intuitively to be much simpler 
than the regression problem. I have had to do a lot of tuning of the 
spread constant and error goal to get the net to work right.

Incidentally, a thin plate spline function RBF net works really well, 
presumably because of the more global nature of the TPSF creating a 
better decision boundary. I'm still baffled by the Gaussian net's 
behaviour, though...

Max

