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: Jim Cunningham,
    Communicating Agents Group,
    Department of Computing,
    Imperial College, 180 Queens Gate, London SW7 2BZ
    http://comma.doc.ic.ac.uk, http://medlar.doc.ic.ac.uk/~rjc

POSTER TITLE: Towards an Axiomatic Theory of Consciousness

This is a step towards an axiomatic theory of consciousness as a quantified form of introspective awareness. A crucial step for presentation of the theory is use of an interval temporal logic to give formal expression to on-going conditions such as those represented by the progressive aspect in natural language. In this way we are able to enrich more stative mental models so that an agent's internal activities and its perception of external processes can be expressed more faithfully as current but durative and changeable logical properties.

Realisations of the intentional stance, notably variants of the Belief, Desire, Intention (BDI) paradigm, are popular as a philosophical basis for intelligent software agents (see bibliography). These are by no means merely toys of Artificial Intelligence laboratories. The idea of an agent as mobile light-weight software process with integrated intelligence has considerable attraction as a basis for modularity and local adaptability in a telecommunication network. Industrial standards already exist for agent host platforms and for communicative acts, although there is a considerable gap between research into intelligence and the industrial application of software agents.

One route to the realisation of abstract axiomatic theories of agent intelligence is through the computational mechanisation of the intensional logics in which they are typically expressed. But from the perspective of an agent designer, extant intensional theories of rational agents focus on stative concepts like knowledge, commitment, and the BDI concepts themselves, all of which can be regarded as computational data states. Activity, or process states, which are equally important in a computational model, have been ignored, or buried in naive computational models. But activity states like thinking, planning, learning and sleeping, the sensing and perceiving of external conditions, and the doing of actions, are equally important for a computational model of a rationality. We can express these through use of the progressive form of the verb in natural language, which we shall formally capture the by modifying the singular verb predicate by the modal operator prog, so that a formal rendering of j is sensing cbecomes prog sensesj c. To define the prog operator we use the interval temporal logic of Halpern and Shoham. This incorporates as modalities the interval relations first identified by Allen. The usefulness of the Halpern and Shoham logic for representing and modelling both tense and aspect is demonstrated in a thesis by Leith.

Once we have the ability to express temporal relations between interval based activities as logical properties, the interactions between activities and between activities and other mental states can be expressed by axioms. We may for instance consider that the formal axiom

prog perceivesj p <<--> prog sensesj c and believesj (c -->p)

expresses the idea that sensory perception amounts to an ongoing inferential interaction between sense and belief. Of course we have the ability to express other tenses, and also atemporal conditions, such as the idea that to sense a condition is coextensive with sensing by some subset of available senses, including in the human case the haptic sense as part of the sense of feel. This latter stipulation may be important if we are to allow a perception of self to enter into a property of awareness. To be aware is more tantamount to a mental activity than a data state, awareness includes both perception and belief, but differs by degree of mental focus - we switch awareness of a condition on and off by paying attention to it, either in response to change in perception through the senses, or through deliberate selection and volition, primitive mental processes whereby mental activity and ultimately action are controlled. In positing these processes we follow Carl Ginet by arguing not only for such philosophical abstractions themselves but in appreciation that they relate to notions of free will and causality through neurological elements such as the motor cortex.

We remark at this stage that despite being merely elements an axiomatic theory we do not consider that the properties of sensory and control activites we have so far discussed are distinct from those which a sophisticated but unconscious automaton could deploy in pursuit of a preset goal, and that enhancing capacity to plan and to learn does not change this, except in level of autonomy. To introduce consciousness, and critically, a consciousness of responsibility, we first allow the activity of being aware to be positively introspective to some degree, so that when an agent is aware it is also, at least for some conditions, aware this fact; and if it learns or itself posits a causal relationship, an agent with adequate introspective power can become aware that it is aware of this relationship, and of its consequences. Because the scope and degree of introspection can be graded there seems to be no evolutionary argument against the acquisition of introspective awareness, indeed it seems necessary for a sense of social responsibility. However the degree of introspection will always be limited, for full positive and negative introspection will each lead to forms of omniscience.

Finally we claim that once an agent has mental activities of positive introspective awareness it also has a form of consciousness, that in its weakest form consciousness is just an activity of being introspectively aware of something. We also note that by providing both a progressive temporal model of durative mental activity, leading to activities of focused awareness, and by indicating that there are graded degrees of conciousness we have plausible evolutionary justification to answer some of the substantive criticism from sceptics of mechanisation such as Tallis.

J. Allen, "Towards a General Theory of Action and Time", Artificial Intelligence 23(2), 123-154.
M. Bratman, Intention, Claims and Practical Reason, Harvard University Press 1997
P.R. Cohen and H.J. Leveques, "Intention is choice with commitment", Artificial Intelligence 1990, 42, pp213-261
C. Ginet, On Action, Cambridge University Press 1990
M.F. Leith, "Modelling Linguistic Events", PhD Thesis, Imperial College, University of London 1997
A.S. Rao and M.P Georgeff, "BDI agents: from theory to practice", Proc. Internat'l Conf. on Multi-agent systems, (ICMAS-95), San Francisco, CA, 1995 pp312-319
R.Tallis, The Explicit Animal: A defence of Human Consciousness, Macmillan Press Ltd. 1991 (reprinted 1999).

Short CV:

Formally I am a Senior Lecturer in the Department of Computing at
Imperial College, and lead an interdepartmental research group on
Communicating Agents. The group  embraces pure and applied
research on software agents. The group has been involved in six
recent ACTS projects on agents, embracing telecomms, e-commerce,
and human computer interaction. The group has interests in agent
software, cognitive robotics, automated reasoning, speech,
language and human face & guesture recognition.

As an academic I have always been been research active, but an
under achieving eclectic, despite some technical and managerial
skills. I left control systems for programming language theory in
'66, introduced object oriented progamming (Simula 67) 72-78
before tiring of pressure for Pascal and functional languages,
was an early user of Prolog but insufficiently doctrinal, left
formal methods despite award winning papers once the evidence
pointed to poor Human Computer Interaction as the main deficiency
in safety critical systems. I have had a long interest in making
automated reasoning more applicable and accessible for human
cognition, introduced term Knuth-Bendix rewriting methods to UK,
but tired of less productive detail, then decided intensional
rather than extensional logics were more appropriate for
expressing human language and contextual reasoning, and that
Agent technology is an appropriate experimental vehicle.)

At present I teach courses in Multi-Agent Systems, Modal and
Temporal Logic and Human Computer Interaction, also supervise
research students in automated reasoning, electronic commerce,
agent theories, and human communication. I have spent much of the
last 20 years involved in some wide ranging collaborative
research projects, mostly Esprit (7 in addition to the 3 of the
ACTs projects), including leading the two "Medlar" basic research
projects on mechanising the logics of practical reasoning. I
enjoy productive relations with some international industrial

Most recent completed papers are with recent research students,
but are not yet on the web:

M.Leith, J Cunningham "Aspect and Interval Tense Logic" a 50 page
paper reviewed and provisionally accepted for publication th the
Journal of Linguistics and Philosophy - (final revision in
March). Earlier related work reported in two conference papers on
the use of Halpern and Shoham's interval temporal logic for
representing verb aspect.

S.Paurobally, J. Cunningham, "Specifying the States and Processes
of Negotiation", accepted to appear in an Agent link collection
on Electronic Commerce (Ed., C Sierra). This work is an
application of dynamic logic which clarifies some ambiguities in
software modelling of tasks and states. Not related to e-commerce
paper on my web page.

I am scheduled to give an informal seminar on consciousness to
Igor Alexander's research group in EEE department on Feb. 15 and
to present material on multi-processing Active Agents to the
Software Methodology Agentlink Workshop in Saarbruecken, Feb 21.
(I think this work on multi-process agent architectures is the
most relevant work for mind construction).

Arranging FAPR 2000 conference on practical reasoning, Sept 18-20