Abstract for AISB 2000: How to Design a Functioning Mind

Abstract for the
Symposium on How to Design a Functioning Mind
17-18th April 2000
At the AISB'00 Convention

AUTHOR Joanna Bryson, MIT

TITLE Making Modularity Work: Combining Memory Systems
    and Intelligent Processes in a Dialog Agent


One of the greatest obstacles to designing a mind is the
integration of different process types, time frames and
representational structures.  This paper describes a dialog
system being developed under a complex agent architecture. We are
using a behavior oriented design strategy to combine modules
exploiting statistical natural language aquisition with modules
containing explicitely programmed behavioral competences.  We
anticipate this project should be interesting to the Design a
Mind symposium for two reasons: first, it demonstrates a
methodology for creating complex, scalable, agents; and second,
it presents a framework for exploring the relation between
semantics and embodiment, a common concern for researchers into
mind. We will now address each of these issues in further depth.

We have elsewhere presented an analysis of the evolution of four
well-established complex agent design methodologies:
behavior-based architectures, layered architectures, BDI
architectures (particularly PRS) and Soar / ACT-R.  [Aaron -- I'm
not certain how much "overlap" you want here, but I could also
present a summary of that review if you are very interested.  But
I'd rather do the paper described here.] Our analysis suggests
two main lessons.  First, facilitating design is the most
critical role of an agent architecture.  This presumably reflects
the fact that intelligence is an extremely difficult problem,
which takes significant search cost to solve.  Consequently,
translating explicit / cultural knowledge to a program is far
more effective than any machine learning or constructive planning
strategy yet established.

Second, there are three architectural features that have
repeatedly proven useful:

-) Modularity, which simplifies the design task.
-) An action selection mechanism that exploits pre-programmed
  ``plans'' which provide focus to behavior through hierarchical and
   sequential structures.
-) A specialised attention mechanism for noticing unexpected priorty
   information and altering the focus of the action selection
   mechanism appropriately.

We have developed a methodology based on object-oriented design
for facilitationg the construction of a system along these lines.
 First, we have developed an action selection system which
provides for both the hierarchical and sequential structure
needed to combat the combinatorial complexity of action
selection, and the reactive attention shifting.  This reduces the
problem to developing two elements: a library of behavior modules
and an action selection hierarchy.  Behaviour decomposition for
the library is performed on the basis of plasticity requirements:
the categorisation, discrimination and recall that is at the
heart of a perception / action system.  Methods on these behavior
objects form the primitive elements at the leafs of the action
selection hierarchy.  The hierarchies and behaviors are
effectively coevolved by the designer.

As an example of this methodology, we consider the problem of
developing an agent capable of dialog.  In this agent, we are
attempting to exploit the fact that human-like semantics can be
aquired by statistical methods through simple exposure to
language, independent of the agent's own action or grounding.
One module of the agent is a semantic lexicon derived this way,
another is a simple grammatic model derived via n-grams.  The
agent's motivations and expertise in turn taking are developed
via relatively standard behavior-based programming methods.  We
are in the process of integrating these modules.  By January we
hope to additionaly have at least one topic that is actually
``contentful'' to the agent, in that it is motivated to create
and describe explicitly representated information.  The first
such competence module will probably be a simple database, such
as the political organization of the department, the second is
intended to be manipulation of the lexical entries themselves,
e.g. describe a ``goat''.



Joanna Bryson holds a BA from the U. of Chicago in Behavioral
Science, an MSc from Edinburgh in AI, and expects to receive an
MPhil from Edinburgh in Psychology in early 2000 [I'm vivaing 7
January] and a PhD in CS from MIT in 2001. Her thesis work is on
managing the complexity of distributed models of intelligence;
her main avocations are understanding  the brain and
understanding the spread of intelligence through culture.  She is
fairly well read in neuroscience and has taught "the biological
basis of behavior" at the undergraduate level.

Her MPhil work was conducted with Brendan McGonigle in the Center
for Neuroscience and Intelligent Systems.  Although her own work
concentrated on robot architectures, she was able to immerse
herself to some degree in the two main reseearch projects of that
center: work on hierarchical representations in monkeys and
understanding the rat hippocampus (Richard Morris).  Most of this
work was conducted in 1996-97, though the thesis was not
submitted until 1999.  In 1998, Joanna worked in the Wizard group
of LEGO Digital on the development of humanoid intelligence for
virtual reality, until the closure of the group.  In 1999, Joanna
has been working in the Human Communication Research Center in
the group formally known as Cognitive Science at the University
of Edinburgh with Johanna Moore.  She also began collaborating
with Andrew Whiten on modelling the underlying representations
for social learning in primates.

Her four most recent published papers are:

Hierarchy and Sequence vs. Full Parallelism in Reactive Action
Selection Architectures, Intelligent Virtual Agents 2 (1999)

Creativity by Design: A Character Based Approach to Creating
Creative Play, AISB Symposium on AI and Creativity in
Entertainment (1999)

Just Another Artifact: Ethics and the Empirical Experience of AI
(with Phil Kime), 15th International Congress on Cybernetics

Agent Architecture as Object Oriented Design (with Brendan
McGonigle), Agent Theories, Architectures and Languages 4 (1998)

She is currently working on a review paper for JETAI called:
"Evolutionary Trends in Autonomous Agent Architectures"

In the next year, she anticipates submitting papers on
implementing her architecture Edmund in PRS, a journal length
extension of the AISB paper (with Kris Thorisson), and a paper on
"Behavior Oriented Design".  She is currently still employed with
the HCRC, but anticipates re-registering full time for a final
year at MIT in the Fall of 2000, where she is working with Lynn
Stein and Gill Pratt.