Newsgroups: comp.ai.neural-nets
Path: cantaloupe.srv.cs.cmu.edu!rochester!udel!news.mathworks.com!zombie.ncsc.mil!paladin.american.edu!news.ecn.uoknor.edu!news.uoknor.edu!ns1.nodak.edu!badlands!francl
From: francl@badlands.NoDak.edu (Leonard J Francl)
Subject: Re: Struggling with SNNS BPTT
Sender: usenet@ns1.nodak.edu (Usenet login)
Message-ID: <D78GIy.Lyx@ns1.nodak.edu>
Date: Tue, 18 Apr 1995 13:37:46 GMT
References: <D770E7.nyz@ns1.nodak.edu>
Nntp-Posting-Host: badlands.nodak.edu
Organization: North Dakota Higher Education Computing Network
X-Newsreader: TIN [version 1.2 PL2]
Lines: 29

Leonard J Francl (francl@badlands.NoDak.edu) wrote:

(Repost)

We are having problems with SNNS BPTT.  The model runs OK but the 
predicted output is invariate and the EMS does not decrement as the 
cycles iterate.  Hopefully, someone has encountered and solved this 
problem because the manual doesn't seem to address it.

Here is our setup:

10 inputs	(various weather variables
 3 hidden nodes (larger numbers, no difference in result
 1 hidden layer (2 or 3, no diff
 2 outputs	(leaf wetness at two heights
		(set at binary values or scaled from 0 to 1
0.005 learning momentum
0.1   learning parameter
1     number of backsteps (higher numbers, no diff
shuffle off or on, no diff

The standard BP _does_ work normally on this data set so we suspect we're 
missing something in the network itself rather than the data structure.

Len Francl
North Dakota State University



