Path: utzoo!utgpu!jarvis.csri.toronto.edu!mailrus!cornell!batcomputer!sun.soe.clarkson.edu!spam
From: spam@sun.soe!clutx.clarkson.edu (Roger Gonzalez,,,)
Newsgroups: comp.ai.neural-nets
Subject: Re: Training
Message-ID: <2771@sun.soe.clarkson.edu>
Date: 3 Apr 89 21:29:15 GMT
Sender: news@sun.soe.clarkson.edu
Reply-To: spam@clutx.clarkson.edu
Lines: 42

Let me restate my question on training in a slightly different
way.  Since I'm an undergraduate doing research in NN and I can
get away with it, my "problem" is not one of much importance
or practicality.  Here it is:

There is a small organism in my pseudo-universe that has a life that
consists mainly of swimming up or down, eating things, and dodging
annoying predators that nibble at it.  The shallow waters are safe, but
all the food (and predators) are down deeper.  There are other stimuli
as well, but these will suffice to give you the gist of what I'm playing
with. 

Lets assume that our friend is in deep water (literally) and a predator
is munching on him. (It's never fatal, but the "pain" and "fear" inputs
are bothering it.)

The correct solution to get out of this state is to swim up.  However, I
don't want to have to tell it this.  The solution that I'm playing with
right now (suggested to me via mail) is something I'm told is called
"weakly supervised learning".  I haven't implemented much of this solution
yet, so I don't know how well it will work.  The basic (and obvious)
method seems to be to negatively affect any attempt that didn't get it
out of the undesireable state.  I can't believe I didn't think of this
on my own.

Any other suggestions?

   - Roger










++ Roger Gonzalez ++ spam@clutx.clarkson.edu ++ Clarkson University ++

   "Just like I've always said; there's nothing an agnostic can't do
    if he's not sure he believes in anything or not!" - Monty Python