Newsgroups: comp.ai.neural-nets
Path: cantaloupe.srv.cs.cmu.edu!das-news2.harvard.edu!news2.near.net!howland.reston.ans.net!agate!news.ucdavis.edu!sunnyboy.water.ca.gov!sunnyboy!nsandhu
From: nsandhu@venice.water.ca.gov (Nicky Sandhu)
Subject: Re: SNNS - overlearning?
In-Reply-To: h9151382@idefix.wu-wien.ac.at's message of 30 Dec 1994 09:23:55 GMT
Message-ID: <NSANDHU.94Dec30092117@grizzly.water.ca.gov>
Sender: news@sunnyboy.water.ca.gov
Organization: Calif. Dept. of Water Resources
References: <3e0jjb$pub@osiris.wu-wien.ac.at>
Date: Fri, 30 Dec 1994 17:21:17 GMT
Lines: 20

>>>>> Regarding SNNS - overlearning?; h9151382@idefix.wu-wien.ac.at (Joerg Schaeffer) adds:
	Joerg> How can I tell SNNS to stop learning after it reaches a certain MSE/SSE? 
	You can set the dmax = target - output (Remote panel)  to
certain  value which is not the same as MSE. I usually just watch for
SSE in the graph and stop it manually.
	Joerg> And how can I tell whether overlearning occurs? So far I think this 
	You can do that by setting aside a small subset of the
training set. When the error on that subset bottoms out or starts
increasing it could indicate overtraining. I would personally start
with the simplest net possible and add hidden units if required.
	Joerg> happening when the graph window shows a sudden rise after the curve went 
	Joerg> down steadily.
	Yeah. I never understood this one. It has happened to me on
several occasions where the error curve has either started to increase
or starts to vibrate during training. Does anyone know the reason for this?
	Joerg> TIA
	Joerg> Joerg (still baffled about SNNS and its user
friendliness :->)
	So am I...
-Nicky
