NetNews Usenet Archive 1992 #16/NN_1992_16.iso/spool/comp/ai/neuraln/2895

Text File | 1992-07-23 | 1.6 KB | 34 lines

Newsgroups: comp.ai.neural-nets
Path: sparky!uunet!cs.utexas.edu!usc!snorkelwacker.mit.edu!bloom-picayune.mit.edu!news
From: [e-mail]
Subject: how to get neural net to answer within a set of legal answers?
Message-ID: <[e-mail]>
Sender: [e-mail] (News system)
Nntp-Posting-Host: pit-manager.mit.edu
Organization: Tufts University
Distribution: usa
Date: Thu, 23 Jul 1992 18:23:08 GMT
Lines: 21


     Suppose I want to use a neural net to inductively learn the
relationship between some inputs and outputs. Suppose also that the
outputs I want from the net have to be members of a certain set,
specified with a syntactic rule. How do I get the net to output only
"legal" guesses?  More specifically, suppose I want the net to learn a
relationship between sets of real numbers and strings; however, the
strings that it offers up as its guesses need to be "legal" strings -
they need to conform to a certain syntax. How in general do people
working with neural nets get them to only output legal (relative to
some domain) guesses? In case anyone can help me, I want the net to
produce outputs of strings S, where each S is recursively defined as a
real number, the letter "z", or the string "(f S' S'')" where f is one
of a small set of function names, and S' and S'' are other strings
(which may themselves be complex strings etc.). Thus, "7" is a legal
string. So is "(+ (* 8 z) (- z 5.5))".  Basically, a legal string is
any math expression in variable z in infix (LISP-like) notation.  So,
what can I do to have it output only legal formulas? Please reply to
[e-mail].

Mike Levin