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 Catriona M. Kennedy, The University of Birmingham

TITLE: Reflective Architectures for Survival:
    Bridging the Gap between Philosophy and Engineering

An essential feature of a mind is its ability to cope with
unusual events which deviate from unconscious expectations. Of
particular interest is when the mind recognises first that there
is something unusual about its own internal processing (e.g. a
sudden apprehension or increased alertness) due to a still
unidentified but novel feature of the environment. To design
artificial agents with this feature, we focus on the comparison
of whole architectures, instead of particular algorithms or

We may characterise the architecture of an agent as the pattern
of interrelationships (causal connections etc.) between entities
representing specialist functions (e.g. they may be layers,
individual components or sets of functionally similar
components). Each entity may be regarded as a "slot" into which a
specialist AI technique can be "plugged in" (e.g. a human
interface layer or a planning specialist).

Since our question involves interrelationships, the "slots" of
the architecture must be "shallow" (at least initially),
otherwise the problem becomes unmanageable. In practise, this
means that a slot contains only a minimal implementation of a
technique (e.g. a concept acquisition algorithm which solves only
one specific type of problem instead of a general class of

However, an architecture that is too "shallow" may tell us
nothing new. An extreme example is a synthetic "personality" that
only displays a surprised facial expression in response to an
external stimulus.

We attempt to find a middle ground by ensuring that mentalistic
concepts are given a concrete interpretation in an engineering
sense. Therefore, we use a bottom-up, low-level approach where an
agent must "survive" in an environment in which its software can
actually be damaged. I.e. the agent is "situated" in an
environment which can produce real faults (and not just
simulations of faults). The agent has no knowledge of the kinds
of damage that can occur; instead it must detect anomalies in its
software execution patterns (actually its rule firing patterns).
Then it must attempt to "repair" the damaged software or take
evasive action as necessary. To detect anomalies, it must have a
model of its own normal operation, which is acquired gradually by
self-observation during a "development phase".

Clearly this kind of anomaly-detection is extremely simple
initially and does not approach the complexity of surprised or
anxious states in humans. However, the long term aim is to
understand the kind of low-level processes that may collectively
give rise to such states.

To address the "survival" problem, we are exploring reflective
agent architectures: one is a layered architecture containing a
meta-level which checks the agent's rule firing patterns against
the model. This architecture has the weakness that it may not
detect anomalies in its own meta-level (e.g. if the meta-level is
deleted). An alternative is a distributed architecture where the
functionality of the first agent is produced collectively by two
or more agents. Each agent's meta-level not only monitors the
agent's own software but also the meta-level of at least one
other agent. In this way all meta-levels are "protected". We
investigate the practical feasibility of a distributed
architecture and whether it provides survival advantages over a
simple layered one.


Research Student,                   e-mail: C.M.Kennedy@cs.bham.ac.uk
School of Computer Science,         web:    http://www.cs.bham.ac.uk/~cmk/
University of Birmingham,
Edgbaston, B15 2TT


PhD thesis topic: ``Distributed Reflective Architectures for Anomaly

The aim is to explore architectures which allow an autonomous
system to detect and recover from anomalies without user
intervention. An anomaly is any event that deviates from the
model-predicted state of the world and may also occur in the
system's own software or hardware. This means that the system
must have a model of its own operation (reflection).

I am exploring forms of distributed reflection using a
multi-agent network, where each agent may specialise in a
particular aspect of the system's operation. The network is not
intended as a team of cooperating agents but instead as a
decentralised control system for a single autonomous agent.

The idea is inspired by various branches of philosophy and
biology, in particular autopoiesis theory, immune system models,
and Minsky's Society of Mind concept.


MSc Intelligent Systems, Brunel University, 1989
BSc (Hons. 2/1) Computer Science, Stirling University, 1982


1982-85:    Programmer, Civil Service, UK
1985-88:    Software designer, Logica, London, UK
1989-90:    English language assistant, Berlin, Germany
1990-93:    Scientific programmer, Berlin, Germany
1994-98:    Scientific assistant, University of Dresden, Germany
1998-now:   Research student, Birmingham, UK


1999: C. Kennedy. "Towards Self-Critical Agents",
      Journal of Intelligent Systems, Vol 9, No. 6, Dec 1999,
      Special Issue on Consciousness and Cognition: New Approaches

      C. Kennedy. "Collective Anomaly-Detection using Immune Systems
      Workshop on Collective Agent-Based Systems at the European
      Summer School on Logic Language and Information (ESSLLI99),
      Utrecht, The Netherlands, Aug 16-20, 1999

      C. Kennedy. "Distributed Reflective Architectures for Adjustable
      Workshop on Adjustable Autonomy at the International Joint
      Conference on Artificial Intelligence (IJCAI99),
      Stockholm, Sweden, July 30 - Aug 5, 1999

1998: C. Kennedy. "Evolution of Self-definition".
      Proceedings of the IEEE International Conference on Systems, Man and
      Cybernetics Invited Track on ``Immune Systems: Modelling and
      Simulation'', San Diego, October 1998

      C. Kennedy. "A Conceptual Foundation for Autonomous Learning in
      Unforeseen Situations",
      International Symposium on Intelligent Control (ISIC/CIRA/ISIS'98),
      Gaithersburg, Maryland, September 14-17, 1998.

      C. Kennedy. "Anomaly-Driven Concept Acquisition",
      14th German Conference of Artificial Intelligence (KI'98,)
      Workshop on Machine Learning and Concept Acquisition,
      Bremen, September 1998.

1996: E. von Goldammer, J. Paul and C. Kennedy.
      "DEAD AND LIVING SYSTEMS: Their Relation to Formal Logical
      Proceedings of the 13th European Meeting on Cybernetics and Systems
      Research, Vienna, April 1996.

      E. von Goldammer, C. Kennedy, J. Paul, H. Lerchner and R. Swik.
      "AUTONOMOUS SYSTEMS: Description and Construction",
      Proceedings of the 13th European Meeting on Cybernetics and Systems
      Research, Vienna, April 1996.