Path: utzoo!utgpu!jarvis.csri.toronto.edu!mailrus!tut.cis.ohio-state.edu!ucbvax!ucsd!sdcsvax!ucsdhub!calmasd!wlp
From: wlp@calmasd.Prime.COM (Walter L. Peterson, Jr.)
Newsgroups: comp.ai.neural-nets
Subject: Re: Training
Summary: targets of opportunity
Message-ID: <235@calmasd.Prime.COM>
Date: 22 Mar 89 17:23:01 GMT
References: <2698@sun.soe.clarkson.edu>
Organization: Prime-Calma, San Diego R&D, Object and Data Management Group
Lines: 60

In article <2698@sun.soe.clarkson.edu>, spam@sun.soe!clutx.clarkson.edu (Roger Gonzalez,,,) writes:
> Has anyone come up with a good way to train a net
> without knowing a "target" in advance?  For example,
> if my net is in an undesireable state S1, and it is
> totally clueless about how to get to state S2, I would
> like it to improve the connections that do get it to
> state S2, and impede any others.  Now, since it has
> no data at all to go on, and the only info it gets after
> 1 pass is "nope, not in state 2 yet", what can I use as
> a target?  All the learning methods I've studied so far
> are inadequate for this.
> 



State S2 is, in and of itself, the target.  That state, whatever it
might be, has to be represented to the machine in some manner.  The
fact that S2 might be a very long text string or some very large
binary number or some arbitraty bit pattern that you make up makes no
difference, if you know where you are, and you know where you want to
be, you can get there ( it might not be an easy trip, though :-) ).

The point is that you must not confuse your interpretation of the
"meaning" of the bit pattern with the bit pattern itself.

As far as you statement that "...since it has no data at all to go
on...". I submit that this is meaningless, not only for a neural net,
but for any computer program. Also, your connections and their
weights, by themselves, constitute 'data to go on'.

Given that the system starts in some arbitrary state S1, which
must be known or knowable to the system, and given that its
goal is to get to some other known or knowable state S2 there
are three ways to proceed.  First, if there is some known transform
that can be applied to state S1 to turn it into some
other valid state Sn, then that transform must be continually applied
to the current state, call it Sc, and then Sc must be tested against
S2 to see if we have found S2.  The existance of some transform
implies that there is some set of parameters upon which the transform
operates, and lacking any data on which to go, the system must perform
a systematic exhaustive search of the parameter space. The second
method is a variation on the first. Lacking a known transform, the
best that you can do is apply random values to the things that
constitute the state and wait for S2 to turn up. This is the situation
that you would have in your example, where you say that "...it is
totally clueless about how to get to S2...". Totally clueless implies
that you don't even have an algorithm for transforming S1 into ANY
other valid state, Sn.

This, of course, assumes that you have no means of determining how
"far away" the current state Sc is from S2.  If you have some way of
determining that distance, then what you've got is, essentially, the
gradient descent of back-propagation which is the third way of going
from state S1 to state S2. This "distance" information ( the error
function in back-prop ) and the connections with their weights are,
essentialy, all the data that you need.-- 
Walt Peterson.  Prime - Calma San Diego R&D (Object and Data Management Group)
"The opinions expressed here are my own and do not necessarily reflect those
Prime, Calma nor anyone else.
...{ucbvax|decvax}!sdcsvax!calmasd!wlp