CPHC Working Group on Agent Based and Multi-Agent Systems
AGENT BASED AND MULTI-AGENT SYSTEMS
CPHC Research Strategy Working Group
This document has been produced by a working group convened as part of
the Conference of Professors and Heads of Computing
(CPHC)
Research Strategy Workshop organised by
Chris Hankin
and
Ian Watson, and held
in Manchester in January 2000. This group comprised leading UK
researchers in the areas of agent-based and multi-agent systems. Its
aims were to assess the current state of UK research in these areas,
to identify the key challenges for their development and to make broad
proposals concerning future funding priorities. As such, this document
provides a considered view of the future of research in these,
increasingly important, areas.
Agent-Based systems are beginning to be used in a significant number
of areas, and are suggested as providing appropriate solutions for an
even wider range of problems. Although there is still considerable debate
concerning the detail of what exactly constitues an agent, there is
general agreement that an agent-based approach often provides an
appropriate abstraction for modelling and implementing complex
systems. This is particularly important as such systems are becoming
more widespread, and will surely continue to do so in the
future. Although popularised relatively recently, the notion of an
`agent' can be seen as a natural evolution from work in distributed
systems (for example, via coordination and mobility), object-based
systems (for example, via concurrency and autonomy) and artificial
intelligence (for example, via rationality and emotion); indeed, many
researchers in those areas are tackling similar problems to those
considered by the agent-based systems research community.
Thus, while activity concerning agent-based systems is very broad,
being truly interdisciplinary, most researchers in this area share a
common vision of the type of systems that they are either using, or
would like to use. This vision involves open, dynamic systems that
comprise massive numbers of autonomous, interacting and evolving
components. These components exhibit a wide spectrum of capabilities,
from very functional `object-like' behaviours, to components
comprising more `intelligent' aspects such as learning, perception and
deliberation. Indeed, in this vision of complex systems, some of the
components may be `real' human elements. In developing individual
agents, the designer must be able to construct an appropriate
combination of techniques and must be able to analyse such
combinations to assess their feasibility/utility.
While individual components can exhibit a variety of behaviours, the
components themselves can be grouped together into an enourmous number
of different organisational structures, often reflecting the different
problem classes being tackled. These organisational structures not only
provide finer structure within the space of components, but can
themselves dynamically evolve. In particular, collections of
such components can be seen as `societies' that are able to adapt
themselves to the required situation. Components within such
structures interact in order to provide more complex behaviour,
typically via coordination, negotiation, cooperation or competition.
Thus, this vision of (future) complex systems is ambitious. In
addition, it has significant industrial relevance. For example,
systems conforming (at least partly) to this vision are beginning to
be used in electronic commerce (Internet auctions), virtual
environments, simulation/modelling of complex systems, organisation of
distributed components, etc. Such a vision also provides many future
exploitation possibilities. For example, work on the representation
and understanding of emotion within individual agents could lead to
enhanced `believable' agents and hence to even more applications in
the entertainment industry.
By viewing complex systems as being comprised of agents that are
themselves autonomous, interacting and evolving, we provide a very
natural abstraction for understanding and developing them. Indeed,
many researchers working in the areas of software engineering,
artificial intelligence and simulation and modelling are tackling
similar problems to the agent-based systems community. Thus,
regardless of whether or not components of this form are called
`agents', the vision outlined above is not only compelling, but
presents many fundamental research questions; these are the questions
that the agent-based systems research community is tackling.
While agents may be seen as an appealing abstraction for understanding
complex systems of the above form, there are many fundamental research
questions that remain to be answered. The UK contains a number of
strong, internationally recognised, research groups that have
contributed to developing prototypical examples of many of the
individual elements of the vision. In particular, significant research
has been carried out concerning agent communication languages, formal
methods, agent architectures, negotitation and cooperation.
Thus, we are now in a position where
- agents combining a few (AI) elements selected from neural
networks, rule-based systems, predictive planning, learning,
deliberation, perception, etc., can now be built (for example,
[11])
- agents within a multi-agent system can carry out simple
interactions concerning negotiation, auctions, task delegation,
markets, etc. (for example, electronic commerce [6])
- (relatively simple) agent communication languages are being used
(see, for example, the FIPA standardisation effort [5])
- the size of organisational structures comprising agents has now
reached tens of agents (for example, [7])
- there are standard toolkits, based upon conventional programming
languages, that allow the user to develop non-trivial agent
applications (for example, Aglets [1] and JATLite [9]), while
novel approaches to the programming of specifically agent-based
systems (for example, [3]) have been proposed
- there are now simple theories of organisational aspects of
multi-agent systems, for example cooperation, coordination and
negotiation between agents, for example [12]
- formal approaches to the specification (and, in some cases,
verification) of some of the above elements of agent-based systems
are beginning to become available (for example, [4])
- there are beginning to be a significant range of applications that
utilise agent technology, for example in areas such as Internet
search agents, air-traffic control, real-time fault monitoring,
Internet auctions, believable agents and entertainment, urban
traffic contol, interface agents and user profiling, resource
management, business process modelling, industrial process control
and telecommunications [8]
The strength of the UK in agent-based systems research, for example
exhibited through the large part played in the AgentLink EU Network of
Excellence [2], together with advances in underlying
software and hardware technology (for example, [10]), mean
that researchers in this area now have both the opportunity, and
(viable) infrastructure in which, to develop significant agent-based
systems.
In spite of the advances that have been made in agent-based systems
research, there is clearly a long way to go before the vision outlined
above can be realised. Key research questions that must be answered
involve a number of areas, including the following.
- Improved theories concerning massive numbers of interacting
agents, together with the practice of utilising such systems. This
involves considering aspects relating to
- communication - how can inter-agent communication be
managed efficiently and appropriately (e.g. what level of
language do the agents need to use?)
- management and organisation - what type of structures are
needed in order to ensure that massive numbers of agents can
together achieve some useful effect (e.g. what societal
structures are most appropriate in a given situation?)
- dynamic agent creation and open systems - as agents evolve
and spawn new agents, how can we control the organisational
structures we wish to use (e.g. how do we balance the
flexibility of adaption/evolution/cloning, with the
efficiency of fixed organisational structures?)
- complexity/tractability - what expectation can we have that
any algorithms for small multi-agent systems will be at all
feasible for massive systems (e.g. can we develop abstraction
mechanisms in order to approximate large-scale systems?)
- Refined techniques for organising multi-agent activity, for example
- cooperation/coordination - how can we move from individual
rational agents to a more coordinated effort whereby a group
of agents can excahnge information and share activities
- teams - how can we ensure that members of teams that have
common goals actually respect those goals, and how do the
teams evolve from a set of individuals in the first place
- organisations/roles - on a larger scale, how can we
designate roles to agents within complex organisations, and
how do these roles evolve
- evolution/adaption/emergence of organisational structure -
given that any fixed organisation structure within a
multi-agent system is certain to be very inefficient in
certain scenarios, the potential to evolve and adapt the
agent organisation is clearly important in complex
multi-agent systems, but how is this evolution controlled?
- Improved theories of individual agents, capturing the range of
possible components, for example learning, reflection, reactivity,
perception. For example, what are appropriate thoeries of
learning, or of deliberation?
- Techniques for putting together individual components to produce
an coherent and useful agent, c.f. agent architectures and evolution.
In particular, given specific techniques for each behavioural
aspect of the agent, how do we combine these techniques to ensure
that the overall agent provides the sum of the required techniques?
- Improved theories concerning how agents decide what to
do, for example balancing deliberation and reactivity. Even if
individual agents have a range of capabilities, the most important aspect
concerns how the agents decide which techniques to use at a given
time, when to suspend certain activity, etc. Such decision
processes are complex and difficult to analyse, but lie at the
heart of agent development.
- Techniques for interacting with and representing the `real world',
for example interacting with humans, interacting with legacy
systems, and interacting with devices.
- Design (including validation/verification) and implementation
techniques/methodologies for all of the above, including aspects
such as performance requirements. A range of enhanced techniques are
required, including logical methods, economic modelling,
statistical approximations, etc.
- Ontologies for communication/understanding between agents. How
should such ontologies be represented and analysed, and how should
they be used by the agents temselves?
Below we suggest several items which we believe can stimulate further
the research in agent-based systems within the UK.
- UK would benefit from a cooperative network of researchers
tackling the interdisciplinary aspects across the whole of agent
research
- Both understanding the techniques required in, and actually
undertaking, the construction of large scale multi-agent systems
is a difficult system engineering endeavour and the UK would
benefit from some form of collaborative programme in this area
with specific goals. For example, a programme tackling some
identified challenges in agent research relevant to the UK.
- There are many fundamental research questions that still require
investigation. Targeted funding is needed for foundational
research into, for example, models of cooperation and its genesis,
methods of managing emergence, non-standard agent architectures,
and new agent programming languages.
- UK needs specific (interdisciplinary) programmes to train
researchers/practitioners in agent development techniques
There are a number of risks associated with research into agent-based
systems that need to be avoided, namely:
- `overhype', and a tendency to `re-invent wheels' within the agent
community; lack of awareness of practice/methods used in the rest
of the world
- a tendency to jump on the `agent bandwagon' by the research
community in general
- fragmentation of agent research community, partly due to
interdisciplinary nature of research area
- no standardisation or, even worse, premature standardisation
- pressure to make research `applicable' may lead to `short termism'
and shallow research
- perceived failure, on the part of the agent research community, to
explain what the research area is all about
We are confident that the agent research community has the experience
to avoid these hazards in well-managed programmes.
- 1.
- Aglets Software Development Kit,
http://www.trl.ibm.co.jp/aglets
- 2.
- AgentLink (EU Network of Excellence),
http://www.agentlink.org
- 3.
- Programing Resoure-Bounded Deliberative
Agents. M. Fisher and C. Ghidini. In Proc. Int. Joint
Conference on Artificial Intelligence, 1999
- 4.
- On the Formal Specification and Verification of
Multi-Agent Systems. M. Fisher and M. Wooldridge. Int.
Journal of Cooperative Information Systems 6(1), 1997.
- 5.
- FIPA,
http://www.fipa.org
- 6.
- Agents that Buy and Sell: Transforming Commerce as
we Know It. P. Maes, R. Guttman and
A. Moukas. Communications of the ACM 42(3), 1999.
- 7.
- Agent-Based Business Process Manangement.
N. Jennings, P. Faratin, M. Johnson, T. Norman, P. O'Brien
and M. Wiegand. Int. Journal of Cooperative Information
Systems 5(2), 1996.
- 8.
- Agent Technology: Foundations, Applications and
Markets. N. Jennings and M. Wooldridge
(eds). Springer-Verlag, 1998.
- 9.
- JATLite,
http://java.stanford.edu/
- 10.
- JINI connection technology,
http://www.sun.com/jini
- 11.
- Building Cognitively Rich Agents using the
SIM_AGENT Toolkit. A. Sloman and
B. Logan. Communications of the ACM 42(3), 1999.
- 12.
- The Cooperative Problem-Solving Process.
M. Wooldridge and N. Jennings.
Journal of Logic and Computation 9(4), 1999.
AGENT BASED AND MULTI-AGENT SYSTEMS
CPHC Research Strategy Working Group
This document was generated using the
LaTeX2HTML translator Version 98.2 beta6 (August 14th, 1998)
Copyright © 1993, 1994, 1995, 1996,
Nikos Drakos,
Computer Based Learning Unit, University of Leeds.
Copyright © 1997, 1998,
Ross Moore,
Mathematics Department, Macquarie University, Sydney.
The command line arguments were:
latex2html -no_navigation -split 0 -address M.Fisher@csc.liv.ac.uk cphc00rep
Footnotes
- ...http://www.csc.liv.ac.uk/~michael1
- Chair of working group and editor of document.
M.Fisher@csc.liv.ac.uk