Newsgroups: comp.ai.neural-nets
Path: cantaloupe.srv.cs.cmu.edu!nntp.club.cc.cmu.edu!godot.cc.duq.edu!ddsw1!panix!news.mathworks.com!uhog.mit.edu!europa.eng.gtefsd.com!howland.reston.ans.net!Germany.EU.net!news.dfn.de!news.rwth-aachen.de!newsserver.rrzn.uni-hannover.de!gm.pc!lzh
From: lzh@ifwmv2.ifw.uni-hannover.de (Christoph Graumann)
Subject: Re: "Neural projects" - bound to fail ...?
Message-ID: <lzh.22.0012F99B@ifwmv2.ifw.uni-hannover.de>
Keywords: backpropagation, simulation of laser welding
Lines: 47
Sender: news@newsserver.rrzn.uni-hannover.de (News Service)
Organization: IFW
X-Newsreader: Trumpet for Windows [Version 1.0 Rev A]
References:  <e21_9501030445@wn.muc.de>
Date: Tue, 3 Jan 1995 17:58:22 GMT

In article <e21_9501030445@wn.muc.de> rst@tbus.muc.de (rst@tbus.muc.de) writes:
>From: rst@tbus.muc.de (rst@tbus.muc.de)
>Date: 02 Jan 95 12:41:08 -0200
>Subject: "Neural projects" - bound to fail ...?



>       Anyone around interested in this topic?

>       Admittedly, such a discussion would not affect "neural projects" only,
>but using them as a pivot may be a justification to discuss this topic here.
>And we all have to avoid this NOT trap, if we want to get most out of NNT to
>solve everyday problems. So, maybe, we could discuss this topic here, beside of
>those "What-might-be-the- right-transfer-function"-stuff...?

>       Any input or comment?

Dear Rudolph,

this is an extremly good idea since especially newcomers will get an 
impression if an NN can satisfy their demands for a particular problem. Right 
now we are facing a couple of problems while training a NN with experimental 
data for the laser penetration welding process to predict the welding depth 
for a given set of laser parameter. 

In this simulation of the real process with a backpropagation approach we have 
difficulties with 

1)    the number of data required to train a net 
2)    how to classify the data sets
3)    structure of the NN (neuron on hidden layer for a given number on the 
       input and output layer
4)    which algorithmen would give the fastest convergence and with what type 
       of prediction probability for the welding depth of an unknown laser 
       parameter set ?
5)    is there an statistical method to evaluate classifiers (compare posting 
       of Arthur Flexer
6)    what type of commercially available software is suitable for what type 
       of hardware

I think there are a lot of points we can talk about. I am new to NN research 
and would therefore like to focuss on the mentioned points. 

In case anybody has an answer to the given questions - your help will be 
greatly appreciated.

sincerely chris 
