Path: utzoo!utgpu!news-server.csri.toronto.edu!clyde.concordia.ca!uunet!mailrus!ames!decwrl!shelby!csli!mpf
From: mpf@csli.Stanford.EDU (Michael Frank)
Newsgroups: comp.ai.neural-nets
Subject: a NN for learning UNIX programs
Message-ID: <12724@csli.Stanford.EDU>
Date: 18 Mar 90 13:50:42 GMT
Sender: mpf@csli.Stanford.EDU (Michael Frank)
Reply-To: mpf@csli.stanford.edu (Michael Frank)
Organization: Center for the Study of Language and Information, Stanford U.
Lines: 25

The following is the abstract of my final project for David
Rumelhart's introductory class here at Stanford.  Anyone interested,
send me email, and I'll send you the rest of the paper, and source for
the program (in UNIX C with Curses) if you so desire.

        A user-friendly program was created to connect a recurrent PDP
network to arbitrary UNIX programs, in such a way that the network
could learn through back-propagation to predict the program's textual
response to characters in an input stream.  In numerous experiments,
the network was able to learn, for an impressive variety of simple
programs and input regimens, to predict exactly what the next text
character produced by the UNIX program would be.  Results of these
experiments are described, and a proposal is made for future work that
would see if a PDP network could be made to explore complex UNIX
programs on its own.  The UNIX environment is promoted as a good
testbed "world" that proposed mind-emulating programs could learn to
explore.

Thanks,
  , ,                       __
 /|/| .  _ l_   _   _  l   l_  _  _  ,_  l,
/ | | l (_ | | (_l (-' |   |  |  (_l | | |\

mpf@csli.stanford.edu   bugboy@portia.stanford.edu
bugboy%portia@stanford.bitnet    - or use the friendly Return-Path!