Path: utzoo!utgpu!jarvis.csri.toronto.edu!mailrus!ames!pasteur!ucbvax!decwrl!shelby!glacier!jbn
From: jbn@glacier.STANFORD.EDU (John B. Nagle)
Newsgroups: comp.ai.neural-nets
Subject: Re: Training
Message-ID: <18248@glacier.STANFORD.EDU>
Date: 4 Apr 89 04:29:04 GMT
References: <2771@sun.soe.clarkson.edu>
Sender: John B. Nagle <jbn@glacier.stanford.edu>
Reply-To: jbn@glacier.UUCP (John B. Nagle)
Organization: Stanford University
Lines: 17


      This is a vivarium approach to AI, and a good one.  See Ann Marion's
classic "aquarium" program.  Mike Travers recent MS thesis at MIT offers
some insight on current thinking in this area.

      Worth considering for a system like yours, where you really want to
cold-start with no built-in knowledge at all, is to add some analogue of
"evolution" to the system.  The simulator should cause the fish to die
if they don't get enough food or get bitten too much.  Fish that thrive
should reproduce, with some tweaking of the parameters of the offspring
to simulate mutation.  Provided that the initial environment is not so
hostile that all the fish die, which can be handled by arranging the
simulator so that the predators are very few until the population
increases above some level, the fish should become more effective over
time.  Once it's working, let it run over a weekend and see what happens.

					John Nagle