Newsgroups: comp.ai.neural-nets
From: NKL@langf.demon.co.uk (Nick Langford)
Path: cantaloupe.srv.cs.cmu.edu!das-news2.harvard.edu!news2.near.net!news.mathworks.com!udel!gatech!swrinde!cs.utexas.edu!news.sprintlink.net!peernews.demon.co.uk!langf.demon.co.uk!NKL
Subject: Re: Help: Real-time recurrent learning
References: <jnicol.795480371@sfu.ca>
Organization: Ant War Tales
Reply-To: NKL@langf.demon.co.uk
X-Newsreader: Demon Internet Simple News v1.29
Lines: 32
X-Posting-Host: langf.demon.co.uk
Date: Sat, 25 Mar 1995 07:45:29 +0000
Message-ID: <796117529snz@langf.demon.co.uk>
Sender: usenet@demon.co.uk

In article <jnicol.795480371@sfu.ca>
           jnicol@selkirk.sfu.ca "Jordan James Nicol" writes:

"..."Hi,
"..."
"..."I'm tryimg to implement a recurrent net with the following structure:
"..."         1 input node
"..."         3 units (one of which serves as the output node.
"..."
"..."I want to train this network using Williams and Zipser's RTRL
"..."algorithm, but infortunately I haven't been able to get a hold of
"..."their original paper. All I have is a paper which briefly describes
"..."the algorithm without too much explanation. My problem is that the
"..."algorithm seems to require the calculation of the error at the output
"..."of the each unit, i.e.
"..."
"..."     ek[t+1] = dk[t+1] - yk[t+1]  for i=1,..number of units.
"..."
"..."where dk is the desired output value of the k'th unit and yk is the
"..."actual value. So, what do I assume for dk for units other than the
"..."output unit? How can these "hidden" units have desired values? 
"..."Any help would be greatly appreciated.
"..."
"..."Jordan Nicol
"..."School of Engineering Science
"..."Simon Fraser University
"..."jnicol@sfu.ca
"..."
"..."

-- 
Nick Langford
