DAI-List Digest       Friday, 20 September 1991         Issue Number 57

Topics:
  Putting the Taxonomy before the Horse
  Follow-Up on Horseless Taxonomies

Please send submissions to DAI-List@mcc.com.  Send other requests,
such as changes in your e-mail address, to DAI-List-Request@mcc.com.
------------------------------------------------------------------------

Date: Mon, 16 Sep 91 14:45:37 EDT
From: chopin!orpheus!gray@uunet.UU.NET ( Michael Gray)
Subject: Putting the Taxonomy before the Horse

I have been following the discussions on understanding the various DAI
systems, but have become dissatisfied with the latest exchanges.  Hassan
Laasri's message in Digest #55 triggered what it was that has been
bothering me.  I think we have slipped into creating classification
systems in the absence of a reasonable theory of DAI systems, and thus
our efforts are having a limited and scattered effect.  Without
understanding what a reasonable set of dimensions ought to be (that is,
without understanding the independent variables of DAI), the
construction of classification systems is somewhat arbitrary.

Actually, I think our original tack on talking about theories of DAI by
trying to understand better how various systems differ from each other
is a more fruitful approach.
------------------------------------------------------------------------

Date: Thu, 19 Sep 91 15:34:44 EDT
From: chopin!orpheus!gray@uunet.UU.NET ( Michael Gray)
Subject: Follow-Up on Horseless Taxonomies

As a follow-up to my comments on the need for theories of DAI as a basis
for classification of systems, my way of thinking about systems such as
the DVMT is to try and understand what theory of DAI they are based on.
I expect a theory of DAI to provide at least two elements.

The first is some notion of what is an individual in the DAI theory.
That is, how can I identify what kinds of things in a system are
individuals, and what properties, capabilities, etc. can I expect these
individuals to have.

The second is some notion of what is a collection (of individuals), how
to identify one, and what properties, capabilities, etc. I can expect
from a collection that are not present in a simple, noninteracting union
of the individuals.

Once I have a theory that gives me these two things, then I can analyze
a system as to whether it is based on the theory and, if so, how it
implements the theory.

In Jeff Rosenschein's communication-free agents, there seems to be a
clear notion of an individual, specifically a thing that performs
actions and is interested in maximizing its own utility.  It has access
to knowledge about how the other agent present in the system will act in
certain situations and it knows that the other agent is also rational.
It does not communicate with the other agent. This seems to me to be
characteristic of multiagent approaches.  There is usually a strong
notion of an individual and its properties.

It is less clear to me what the notion of the individual is in the DVMT
system, since nodes themselves are built out of knowledge sources that
act independently.  The agenda mechanism run by the node makes the
situation even more cloudy.  Is the DVMT an individual of many separate,
interacting pieces?  Or is it a collection of individuals?

For both systems I have a very incomplete understanding of what the
collection's identity or properties are.  Is there a way to distinguish
the collection from the individuals?

Because of the questions above, it seems to me to be somewhat premature
to attempt classification of these systems.  I just can't see why one
"dimension" should be preferred over another.