Newsgroups: comp.ai.neural-nets
Path: cantaloupe.srv.cs.cmu.edu!rochester!udel!news.mathworks.com!news.duke.edu!agate!howland.reston.ans.net!pipex!sunic!umdac!cs.umu.se!jan
From: jan@cs.umu.se (Jan T}ngring)
Subject: Re: [Q] LVQ Classification Difficulties
Message-ID: <D4rJ2J.FM2@cs.umu.se>
Sender: news@cs.umu.se (News Administrator)
Cc: jan, r1t@icf.hrb.com
Organization: Dept. of Computing Science, Umea Univ., 901 87 Umea, Sweden
References: <1995Feb25.094157.22915@hrbicf>
Date: Wed, 1 Mar 1995 13:06:02 GMT
Lines: 32

In <1995Feb25.094157.22915@hrbicf> r1t@icf.hrb.com (Richard Timblin) writes:

>I have been working on developing a pattern classification NN with 3-6
>inputs and 80 to 100 possible output classes depending on the variant
>of the problem with which I am working.  I am using the LVQ method
>provided with Matlab and have made some observations relating to the
>limits of success I have encountered:

>(1) The LVQ algorithm does not seem to converge in the sense that the
>other NN methods do.  The 'circles' denoting the cluster centers converge
>to approximately the center of the clusters and then continue to move
>about - in some cases even leaving the cluster entirely.  This then begs
>the question - when does one stop.

I have studied the interface of Matlab LVQ, but I didn't run it.

If there is no randomization of the sequence in which the input vectors
are presented, the centers should converge, if you look at them after a
full presentation. You say they don't, which means Matlab uses some
randomization. On the other hand, there is one parameter called
"Maximum number of presentations" (and not just "number of
presentations"), wich seems to imply that the constructors imagined
that the centres can converge.

TRAINLVQ seems to control the scale of the weight adjustment by the
constant "learning rate" parameter TP(3), so it doesn't seem to allow
for the Learning rate to gradually decline to zero, which also would
make cluster centers converge. You should be able to tackle this by
calling TRAINLVQ with successively smaller values on TP(3).

/Jan Taangring

