Newsgroups: comp.ai.neural-nets
Path: cantaloupe.srv.cs.cmu.edu!nntp.club.cc.cmu.edu!hudson.lm.com!newsfeed.pitt.edu!dsinc!spool.mu.edu!howland.reston.ans.net!news.sprintlink.net!uunet!in1.uu.net!ftpbox!mothost!pts-nntp!tseserv!RedKnight
From: RedKnight@tseserv (RedKnight)
Subject: Q: Recurrent net with online training?
Message-ID: <D9t4uK.G0F@pts.mot.com>
Sender: mccarley@tseserv (RedKnight)
Nntp-Posting-Host: tseserv
Organization: Motorola Inc, Paging Products Group, Boynton Beach, FL
Date: Wed, 7 Jun 1995 14:42:20 GMT
Lines: 18


I'm looking for a reasonable way to train a recurrent net on line, ie. on
an unending input data stream.  The problem is how to take into account the
effects of current outputs on future activation.  A fairly obvious approach
would be to do a lookahead for a clock tick or two and make do with the
error results from these.  

If anyone has experience with this or an alternate approach I would be grateful
to know how it affected convergence and what adjustments were useful.

Thanks advancedly

-- 

  \ /      RedKnight               | Chris McCarley
 --O===<|  mccarley@pts.mot.com    | Motorola, Paging Products Group
  / \      "I have seen the future |
           and it is neural."      |
