Newsgroups: comp.ai.neural-nets
Path: cantaloupe.srv.cs.cmu.edu!bb3.andrew.cmu.edu!newsfeed.pitt.edu!gatech!swrinde!newsfeed.internetmci.com!in2.uu.net!bcstec!nntp
From: Mike Fouche <mikef@hsvaic.hv.boeing.com>
Subject: Re: some questions about FF-nets
X-Nntp-Posting-Host: e677744.hv.boeing.com
Content-Type: text/plain; charset=us-ascii
Message-ID: <Do98Fu.4E8@bcstec.ca.boeing.com>
Sender: nntp@bcstec.ca.boeing.com (NNTP News Access)
Content-Transfer-Encoding: 7bit
Organization: Missiles & Space Division
References: <Pine.A32.3.91.960312161055.12696A-100000@tharros.dipchim.uniss. <x1KocEp.predictor@delphi.com> <4i7aj6$6vl@airgun.wg.waii.com>
Mime-Version: 1.0
Date: Thu, 14 Mar 1996 11:11:53 GMT
X-Mailer: Mozilla 1.2 (Windows; U; 16bit)
Lines: 25

Hello,

I'll give you my 2 cents worth - based on 6 years experience. :)

If the application is new to me I generally start with a small amount of data and a 
small network.  For example, with the classic cart/inverted-pendulum case, I trained
a neural network from the behavior of the classical controller.  One training set
would be the data generated from uprighting the pendulum from say -10 degrees (and
moving the cart back to the origin).  Instead of ASSUMING the network needed several
passes with other examples, such as initial deflections from +10, -40, +40 degrees, 
I went ahead and tested the trained network as a controller.  Voila!  It not only 
worked but far outperformed the classical in every case (could even upright the
pendulum from a hanging position which the classical could not).  Also, the network
had 6 inputs, 1 hidden layer with 4 neurons, and 1 output.

So ... the point is "don't assume anything".  Start small and work your way up.  In
terms of data one can compress or represent a lot of data with "curve fitting".  If
your doing image processing than hundreds of thousands of pixels can be represented
with Wavelet transforms, Fast Fourier transforms, etc.

Enough said - I've seen neural nets do some scary things.  Good luck!

Mike Fouche
Boeing Missiles & Space Division  

