Newsgroups: comp.ai.neural-nets
Path: cantaloupe.srv.cs.cmu.edu!rochester!udel!delmarva.com!news.internetMCI.com!newsfeed.internetmci.com!in1.uu.net!halon!sybase!usenet
From: George Van Treeck <treeck@sybase.com>
Subject: Re: unsupervised learning
Content-Type: text/plain; charset=us-ascii
Message-ID: <30BB89B9.368C@sybase.com>
Sender: usenet@sybase.com
Content-Transfer-Encoding: 7bit
Organization: Sybase, Inc.
References: <peschko.817580034@cs.umn.edu>
Mime-Version: 1.0
Date: Tue, 28 Nov 1995 22:14:17 GMT
X-Mailer: Mozilla 2.0b2 (X11; I; SunOS 5.3 sun4m)
Lines: 24

Unsupervised learning falls into two categories: self-organizing
and reinforcement-based.  In self-organizing NNs, if a neuron
responds, it stimulates it's nearest neighbors and inhibits the
more distant neighbors. By adjusting weights based on which neuron
or neurons responded, neurons in a region tend to respond to a
"class" of patterns.  This makes them useful for identifying
and classifying patterns.

Reinforcement-based learning is simply learning by trial and error.
If it does the right thing, that response to the input gets
reinforced -- made more likely to give the same response to the
input pattern in the future.

Nature gives the test first and the answer later (reinforcement-based
learning). Supervised nets you give the test along with correct
answer.

For example, suppose you wanted to send a walking robot to a
distant planet. On that planet it encounters a terrain for which
it has not been trained. If the robot had reinforcement-based
neural nets, it could learn by trial and error to traverse the
new terrain (without the need for a supervisor to teach it).

-George
