Newsgroups: ieee.general,comp.ai.neural-nets
Path: cantaloupe.srv.cs.cmu.edu!bb3.andrew.cmu.edu!newsfeed.pitt.edu!scramble.lm.com!news.math.psu.edu!chi-news.cic.net!newsfeed.internetmci.com!in1.uu.net!bcstec!nntp
From: Mike Fouche <mikef@hsvaic.hv.boeing.com>
Subject: Self-Taught Neural Controller 
X-Nntp-Posting-Host: e677744.hv.boeing.com
Content-Type: text/plain; charset=us-ascii
Message-ID: <DsFDtw.K21@bcstec.ca.boeing.com>
Sender: nntp@bcstec.ca.boeing.com (NNTP News Access)
Content-Transfer-Encoding: 7bit
Organization: Missiles & Space Division
Mime-Version: 1.0
Date: Mon, 3 Jun 1996 13:08:19 GMT
X-Mailer: Mozilla 1.2 (Windows; U; 16bit)
Lines: 29
Xref: glinda.oz.cs.cmu.edu ieee.general:2433 comp.ai.neural-nets:31823

Hello,

I've designed a neural controller for a cart/inverted-pendulum system
that was trained solely from the open-loop response of the system.  A 
sinusoidally varying force was applied to the system for a period of 
time.  Two small trajectories were selected and the neural network was
trained from these trajectories then used as a controller to upright the
pendulum and drive the cart to a target location.

Question:  Has anybody heard of this particular method before (training
the system solely from the open-loop response)?  

Some neural control systems are designed by implementing the neural 
network on-line and allowing it to learn on-line (I believe Bernard
Widrow's broom balancer used this method).  But this is very different
than teaching it by using the open-loop response.  In my particular case
it took 5 seconds to train the neural controller and only 27 training
sets (two segments of the overall trajectory data) were required for
training.

I'm going to publish a paper but have no idea if this has been done 
successfully before.  Any insight/leads/etc. would be very much 
appreciated!  

Thanks!

Mike Fouche
Boeing Missiles & Space Division

