Newsgroups: comp.ai.neural-nets
Path: cantaloupe.srv.cs.cmu.edu!fs7.ece.cmu.edu!hudson.lm.com!godot.cc.duq.edu!news.duke.edu!news.mathworks.com!europa.eng.gtefsd.com!howland.reston.ans.net!swrinde!gatech!newsxfer.itd.umich.edu!news1.oakland.edu!vtc.tacom.army.mil!ulowell.uml.edu!umassd.edu!umassd!s_vogt
From: s_vogt@umassd.cis.umassd.edu (Stefan Vogt)
Subject: Q: ANN & softeware engineering?
Message-ID: <D2z423.IKq@umassd.edu>
Sender: usenet@umassd.edu (USENET News System)
Organization: University of Massachusetts Dartmouth
X-Newsreader: TIN [version 1.2 PL2]
Date: Wed, 25 Jan 1995 18:15:38 GMT
Lines: 52


Hi,

in the discussion of this newsgroup Ben Hellstrom mentioned that there
is nearly no research in the field of engineering methods for ANN.

My impression of project managment in ANN is that erverybody tries
different architectures, learning algorithms, and data analysing
methods, but nobody does it with a specific concept. The research
spends much money and time until the net provides the desired task. It
is really no wonder that the *real world* doesn't use ANNs in their
applications.

I think there are general things in any ANN project, an engineer must
take into account. There are certain steps that occur in any ANN
project:

	1.	Requirements analysis

	2.	Input/output data analysis
		a) feature extraction
		b) relation between input and output data

	3.	Design
		a) data representation decision
		b) net architecture
		c) learning/storage algorithm

	4.	Implementation

	5.	Learn/storage phase
		a) change several parameters
		(extrem repetative phase)

	6.	Recall phase, evaluation of performance

What do you think? Didn't you start with 1. after you figured out in
6. that your ANN doesn't perform as desired? Isn't this the ghost we
learned about in the software crises?

In my master thesis I'd like to deal with this problem. Up to now I
couldn't find any literature about this topic. If your research is
close to that, or you are involved in ANN projects facing this
problem, or you know where I can find engineering methods, I would
appreciate your comments or help.


			     Bye, Stefan



		       (s_vogt@cis.umassd.edu)
