Newsgroups: comp.ai.neural-nets
Path: cantaloupe.srv.cs.cmu.edu!das-news2.harvard.edu!news2.near.net!news.mathworks.com!udel!gatech!howland.reston.ans.net!newsserver.jvnc.net!nntpserver.pppl.gov!princeton!isler.Princeton.EDU!adaml
From: adaml@isler.Princeton.EDU (Adam E. Lichtenstein)
Subject: Sin Curve
Message-ID: <1995Apr9.194843.18446@Princeton.EDU>
Originator: news@hedgehog.Princeton.EDU
Sender: news@Princeton.EDU (USENET News System)
Nntp-Posting-Host: isler.princeton.edu
Organization: Princeton University
Date: Sun, 9 Apr 1995 19:48:43 GMT
Lines: 27

I am writing a program to develop a nueral network for robotic control.  As a
test case, I am using 10 points along a sin curve.  I modeled the back
propagation algorithm after Quickprop (epsilon = 0.1 - 1.0, maxfactor = 1.75, 
sigma prime offset = 0.1, decay = -0.001).  The network has one input and one
output and two hidden layers with sizes varying from two nodes each to ten nodes.
 The initial weights are randomly set between -1 and 1.  
	The problem is that the neural net rarely converges to anything like a
sin curve.  The graph of RMS has a lot of noise in it, and it usually settles
down around RMS = 0.4 (generally corresponding to a horizontal line between 0 and
PI).  I don't think the code is entirely wrong given that sometimes it converges
quite nicely, but I don't want to attempt a much more difficult problem until I'm
sure I'm doing everything correctly.
	I would appreciate any suggestions.  It might be helpful if someone has a
working data set that I could look at (along with the results).  Thanks for your
help.

	- Adam
--------------------------------------------------------

Adam Lichtenstein

P.O. Box 1193                            301S Dod Hall
Princeton, NJ 08542-1193          Princeton University
U.S.A.                             Princeton, NJ 08544

(609) 258-7847
Fax: (609) 258 - 2649
