A little while ago on this list, I asked whether anyone knew of any
agent specifications as formalisms -- logics, cellular automata or other
mathematical representations -- that had ever been implemented in a
problem space or to perform some function on a useful scale.  Only one
suggestion was offered: Rafael Bordini suggested the implementation of
an air traffic control system in Australia using a programming
architecture (PRS which became dMARS and then, I believe, AgentSpeak).
PRS and dMARS were developed by Rao and Georgeff then at the Australian 
AI Institute (AAII).  In looking at that literature, I have the
impression that both PRS and now dMARS use the semantics of BDI logics. 
It does not seem as if applications developed in these architectures are
necessarily sound and consistent relative to the BDI logic reported by
Rao and Georgeff in, for example, their ICMAS '95 paper.  Indeed, I have
not found any such claim by anyone working on agent architectures using
BDI semantics.  To offer two examples:

Martin, Cheyer and Moran at the Artificial Intelligence Center of SRI
International wrote in October '98:

     Another influential approach, which makes stronger assumptions
     about the knowledge and processing used within individual
     agents, is based on the structuring of agents' activities
     around the concepts of Belief, Desire and Intention (BDI) [Rao
     and Georgeff1995].  While BDI's emphasis on a higher level of
     abstraction has been extremely important in giving direction
     to work on agent based systems, its applicability may be
     limited by the structural requirements imposed on individual
     agents, and by difficulties in interoperating with legacy
     systems.

In fact, if we go back to the Rao-Georgeff paper, they seem to have kept
the theoretical and implementation perspectives quite distinct in
reporting their air traffic management application. To say that they are
related is not to say that they are identical.

There is no doubt that the work at AAII has shown the BDI semantics to
be a useful and insightful approach to the development of large-scale,
real-time systems.  What is in doubt is whether those systems are
implementations of the BDI formalism alone and, to the extent that other
elements enter into those systems, whether their success is due to the
BDI formalism or to the other elements of the programming environment or
some aspect of the combination.

I presume that the interest of both the computer science and social
simulation communities in multi-agent systems is in the systems rather
than the agents alone.  That is to say, the agents are a programming
device in computer science intended as components of reliable systems
with predictable behaviour and known applicability. Social scientists
are interested in the interactions among individuals and how these give
rise to and are mediated by institutional arrangements. Moreover, the
virtue of the formalisms is that they enable us to specify properties of
the phenomena they are used to represent.  It would seem, therefore, to
make some sense to ensure that our representations of systems are
consistent and sound relative to an appropriate formalism and that our
agents are specified pragmatically.

There is, however, a further reason for taking a systems perspective
here.  The successful sciences -- paradigmatically, physics, chemistry
and biology -- develop by explaining and predicting observations.
Special relativity theory arose from the need by physicists to bring
together the Galilean principle of relativity which conformed both to
common observation and to such fundamental laws of physics as the law of
inertia with the (then) more recent observational and theoretical truth
that the speed of light is constant and independent of the relative
velocity and direction of motion of the light source.  I don't know much
(i.e. anything) about particle physics but I do read that observation,
experimentation and the consequences for theory lead physicists to
postulate new concepts of particles ***together with the evidence that
the existence of such particles would manifest***.  The double helix was
identified on the basis of x-ray crystallographic observation and also
offered a physical explanation for geneticists' observations and,
ultimately, Darwinian evolution.  Important elements in chemistry
emerged in large measure from observation and production requirements in
the French bleaching industry.

In the social sciences, business historians (Chandler and Penrose) and
historians of technology (Rosenberg and David) have documented the ways
in which the need to solve problems gives rise to analytical,
organizational and technological developments that not only solve the
problems they were developed to address but enhance the  capacities of
scientists, managers, technologists, et al to engage in new activities
and develop new science and technology.

In all of these cases, successful developments follow from starting with
the problem and developing the techniques to resolve those problems. The
solution techniques also provide new opportunities that themselves
frequently require other, related problems to be solved resulting in yet
further capacities for activity or understanding.

Consider, for a moment, disciplines that do not include in their
theoretical research agenda the solution of empirical problems.  My own
favourite example, of course, is economics.  The commitment of
economists to the representation of agents as utility-maximizers is
bound up with the focus of analysis on equilibrium.  The inventors of
the economic equilibrium concept (e.g. John Bates Clark, 1893 and Alfred
Marshall, 1895) put it forward as a transitional step to a fully dynamic
analysis of economic systems.  Since no such system has every been
observed and, we now know, that unlimited computational capacity is a
necessary condition for equilibrium (Radner, 1968) in dynamic economies
where not everyone is identical, it has not been an objective of
economic theory to represent anything we actually observe.  What is true
is that the semantics of equilibrium theory are frequently used to
justify policy prescriptions intended to influence real socio-economic
systems.  Without the link between observation and problem solving on
the one hand and the development of agent representations on the other
hand, the creative tension that has driven the successful sciences,
technology and the development of social institutions has been absent. 
Even economists who consider themselves heterodox, assert some features
of economies to be more plausible than those assumed by conventional
economists, then invent out of their own heads some representation of
agent cognition and then define an environment that the invented agent
representation can act on.  There is still no demonstrable link between
such a model and the world we observe.

The answers, or lack thereof, to my original question about the
existence of practical applications of agents represented as formalisms
supports my conjecture that a successful research programme is unlikely
to start with such formalisms as agents in the hope that eventually
useful social science will emerge.  By useful, I mean social science
that explains a wide range of observations in an integrated manner and,
as a result, supports model based policy analysis. A successful social
science seems, on the historical record, more likely to emerge from a
usefully scaled, problem oriented approach than from an approach in
which the problem space is driven by agent representations.

The proposal is a programme for the social simulation community in which
we identify some key social problems to address or observations to
explain and model the systems or represent the relevant environment in a
manner that will capture the problem area or observation independently
of any agent specification.  Agent based means should be developed to
assess the confidence we can have in our model based analyses.  That is,
I propose that we identify on the basis of domain expertise who are the
important real actors and then develop agent representations of those
actors that are as detailed and descriptively accurate as our
computational resources allow.  This connection between observation and
agent representation will constitute the conditions of application of
the models.

Clearly, it is not possible to use such fine-grained presentations at
the useful scale required.  Techniques for abstracting from the 
entailed representations of agents are also required.  Such abstractions
should be specified in such a way that we can ascertain by statistical
and qualitative means that the abstractions do not entail agent
behaviours that distort the behaviour of the more fine grained agent
representations.  This in itself seems a fertile field for exploration.
There might well be cascades of such models so that the finest grained
representations are abstracted for use in (say) models of a region and
these in turn are abstracted to model interactions among regions and so
on up to a global scale.  There might also turn out to be some useful
canonical forms of models such as suggested in
<http://www.cpm.mmu.ac.uk/cpmrep49.html> that could be used in a variety
of applications.

There are some agent representations already in the literature that
attempt to capture real agents, their behaviour in social systems and
the interactions between the behaviour of those individuals and relevant
representations of their social systems. Examples include Rouchier's
model of potlatch reported at MABS98 or my own model of critical
incident management in JASSS.  But these are both free-standing models
that have not been used to support models of larger or higher level
systems.

To support aggregation and scaling up in general, the most natural set
of abstract representations to start with will be those we already know:
deontic and bdi logics, finite cellular automata, simulated annealing ,
genetic programming, and so on.

It will be clear from the above that I am not suggesting a problem
oriented, model based, usefully scaled social science as an alternative
to the representation of agents by logical and mathematical formalisms. 
What I am suggesting is that such formalisms do not easily connect with
observation and social problems in any direct way either in their
specification or in the problem spaces they support.  As such, they
might well provide links between agent representations that are as far
as possible descriptively accurate and scaled up problem and policy
analysis.

I turn finally to the properties of the systems within which the agents
are represented.

For purposes of policy analysis and generally for understanding social
systems, it will always be useful to identify any properties of those
systems that are or are not sensitive to particular patterns or
representations of individual behaviour.  My argument against building
policy oriented models entirely on agents represented by formalisms was
entirely pragmatic and based on observation: such formalisms do not in
practice provide conditions of application of the models and they have
not led to usefully scaled models.  There is a cost here in that the
absence of a formal basis for the agent representation allows
ambiguities and contradictions to survive unnoticed.

The methodology proposed here would yield models at a useful scale for
policy analysis with sufficient abstraction from detail to meet the
limits of computational capacity and the cognitive capacities of policy
analysts. If such models were, in addition, known to be sound and
consistent relative to some logical formalism, the virtues of the
avoidance of ambiguity and self-contradiction could be had without the
costs associated with the use of such formalisms as the lowest level
agent representations.

This seems to me to be a virtue of strictly declarative languages such
as SDML.  Because SDML conforms to a known formal logic (a fragment of
strongly grounded autoepistemic logics -- FOSGAL), any model that runs
in SDML is consistent and sound relative to FOSGAL.  Different logical
formalisms can still be used to represent individual agents of any
grain.  But the representation of the system itself has the properties
of a FOSGAL formalism with the attendant possibilities (in principle
though typically difficult in practice) for proving theorems about
relationships identified in simulation experiments.



--
Scott Moss
Director
Centre for Policy Modelling
Manchester Metropolitan University
Aytoun Building
Manchester M1 3GH
UNITED KINGDOM

telephone: +44 (0)161 247 3886
fax: +44 (0)161 247 6802

http://www.cpm.mmu.ac.uk/~scott


%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%