Newsgroups: comp.ai.neural-nets
Path: cantaloupe.srv.cs.cmu.edu!das-news.harvard.edu!news2.near.net!MathWorks.Com!news.kei.com!world!wshill
From: wshill@world.std.com (Wayne S Hill)
Subject: Re: NeuroWindows/Chromatography Questions
Message-ID: <Cw86nr.1EG@world.std.com>
Organization: The World Public Access UNIX, Brookline, MA
References: <3580p3$5ur@master.cs.rose-hulman.edu>
Date: Fri, 16 Sep 1994 13:55:03 GMT
Lines: 71

erwinaw@image-5.Rose-Hulman.Edu (Anthony W Erwin) writes:

>	As part of a 2-quarter computer science project at school, my team  
>is thinking about trying to implement a neural network to solve a  
>chromatography analysis problem.  We basically want to use some pattern  
>recognition to determine what PCB's are in different samples.  We are  
>thinking about using NeuroWindows by Ward Systems to help in the solution.
>	I have several questions.  I do not have a copy of the FAQ and am  
>under a little time pressure to get some answers so please excuse me if I  
>ask a FAQ.
>	 
>1.  Has anyone had any experience with NeuroWindows?  Is it good?  Is it  
>bad?  

NeuroWindows is excellent.  It permits the generation of essentially any
NN architecture from within the Visual Basic programming environment. 
However, I don't think NeuroWindows is what you're looking for...  

I would suggest you use NeruoShell 2, also from Ward Systems.  It is a
flexible, capable, and easy to use "canned" ANN program that runs in
Windows.  It permits you to try 3-, 4-, and 5-layer backprop nets, 3-layer
nets with different recursion layouts, 3-, 4-, and 5-layer nets with jump
connections, Generalized Regression nets, Probabilistic nets, and Kohonen
nets.  In addition, it permits a variety of activation functions,
including different functions within a given network (which they call a
"Ward net").  It gives you good control over training and handy means of 
keeping track of training progress.  Best of all, once you've got a 
network trained and debugged, you can create a run-time version of the 
net as a DLL (dynamic link library) that is callable from Visual Basic or 
Windows C code.  I understand (but can't attest) that Ward Systems code 
is tightly written and quite fast, often achieving the speed of hardware 
coprocessor nets without a coprocessor card.

>2.  How hard is it for someone who is far from a neural net expert to use?   
>What kind of learning curve can be expected?
>	NOTE:  I have read a book on neural networks and have a decent  
>grasp of what some of the issues are.  I know a little about the standard  
>three-layer design and about backpropagation.  Am I in way over my head? 

Go for it.  If you understand the physical system you're trying to model 
and have some understanding of the ideas of neural nets, you're ready to 
try it out.  The learning curve is exceedingly steep:  you'll get up to 
speed in a day or so (maybe less).

>3.  Are there better PC-based neural network routines than NeuroWindows?

I'd be interested in hearing feedback on this point.  Few people have 
experience with more than one package.

>4.  Has anyone out there used neural networks to solve a chromatography  
>problem?  Other chemistry problems?

Well, yes, but I believe that our results are currently proprietary...

>5.  Any other thoughts concerning this problem?  
>  
>Anthony W Erwin, Rose-Hulman Institute of Technology 
>erwinaw@nextwork.rose-hulman.edu
>(812) 877-8870
>GO HOOSIERS!
>"In basketball, the mental is to physical as four is to one."
>			-Bob Knight

Regards,
Wayne

-- 
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
Dr. Wayne S. Hill                                         wshill@world.std.com
Foster-Miller, Inc.            ***Disclaimed***              617-890-3200
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
