Stability problems with LVQ algorithms.

[Martin]

We have observed  stability problems with LVQ algorithms in our
application on EEG data. Analysis of these problems resulted in
general conditions under which the codebook vectors drift away from
the data with both, LVQ3 and LVQ1 training. 

A paper which demonstrates the problems on simple 2-dim data is in
preparation. A first draft can be downloaded form:

FTP-host: 		fdpmial03.tu-graz.ac.at
FTP-filename: 	/pub/outgoing/lvq_stab.ps.Z (135kB)
		     	/pub/outgoing/lvq_stab.ps	(733kB)

The problems are also outlined  in my dissertation:
"Distinction Sensitive Learning Vector Quantization", Graz, Austria,
University of Technology, 1997. 

Comments are welcome to pregenz at dpmi.tu-graz.ac.at.

Martin Pregenzer