Newsgroups: comp.ai.neural-nets
Path: cantaloupe.srv.cs.cmu.edu!bb3.andrew.cmu.edu!newsfeed.pitt.edu!gatech!newsfeed.internetmci.com!in2.uu.net!redstone.interpath.net!sas!mozart.unx.sas.com!saswss
From: saswss@hotellng.unx.sas.com (Warren Sarle)
Subject: Re: RBFNN and distance measures
Originator: saswss@hotellng.unx.sas.com
Sender: news@unx.sas.com (Noter of Newsworthy Events)
Message-ID: <Dn8yDp.JAt@unx.sas.com>
Date: Fri, 23 Feb 1996 21:01:01 GMT
X-Nntp-Posting-Host: hotellng.unx.sas.com
References:  <3129DFEC.2A64@rio.esm.vt.edu>
Organization: SAS Institute Inc.
Lines: 38


In article <3129DFEC.2A64@rio.esm.vt.edu>, Mauro Jorge Atalla <mauro@rio.esm.vt.edu> writes:
|> ...
|>      For one of the cases I am working on, the input data are the values of 
|> the frequency response function of my system at certain frequencies.  The output
|> are target parameters of the model that need to be identified.  The problem I 
|> have is as follows:  I can't get an uniform improvement in the generalization 
|> accuracy of the network when varying the smoothing parameter sigma.  Some of the
|> target parameters are more accurate with a small sigma and the others with a 
|> large sigma. 

You could:
 1) Use a separate network for each target.
 2) If you have sufficiently flexible software, there's no reason you
    can't use different sigmas for different targets in the same
    network. I've just spent the last hour confirming that this
    can be done with my software, but it's a bit of a pain.
 3) Use separate, trainable sigmas for each rbf unit. I have found
    this to be advisable whenever using ordinary RBF networks, as
    opposed to normalized RBF networks.
 4) Use a normalized RBF network, and you are likely to get better
    results in general.

|> I tried using a metric other then the Euclidean (a weighting 
|> matrix based on the variance of the input data) but I got bad results.  I think 
|> my problem is related to how I measure the distance between 2 vectors.

It is quite plausible that you would need different sigmas for
different targets. That would be expected if the target functions
differ in smoothness or amount of noise. But I don't see how this
ties in with needing a different distance measure; maybe you do,
but that's a separate issue.

-- 

Warren S. Sarle       SAS Institute Inc.   The opinions expressed here
saswss@unx.sas.com    SAS Campus Drive     are mine and not necessarily
(919) 677-8000        Cary, NC 27513, USA  those of SAS Institute.
