Multi-Agent Systems and Simulation: A Survey from an Application Perspective

1,633 words) Setting: Simulation in the Sciences of Complex Systems As social and economic systems are among the most complex systems in our world, the chapter will mainly deal with applications of simulation in general and agent-based simulation in particular in economics and the social sciences. Thus it will start with a discussion of the predecessors and origins of agent-based simulation mainly, but not only, in these sciences from the time when the first simulation models were created that used, or rather should have used, multi-agent systems. If one accepts that multi-agent systems have object-oriented languages as their prerequisites, one has also to accept that multi-agent systems proper could only be implemented after the early 1980s, but much earlier, namely in the 1960s first simulations, for instance in political science, were built that can be described as forerunners of multi-agent systems. At the same time, ingredients were developed that nowadays are a defining part of the agents in multi-agent systems, such as fact and rule bases [Abelson/Carroll 1965] in which early “agents” stored information that they communicated among each other, although they lacked the defining feature of autonomy. But for a long time, simulation approaches prevailed that did not address the fact that in social and economic systems there are actors who are endowed with a very high degree of autonomy and with the capability to deliberate. Although not for all purposes of the sciences dealing with these systems, autonomy and deliberation are necessary ingredients of theory and models, one would not content oneself with humans being modelled as deterministic or stochastic automata but prefer models that reflect some typically human capability. Subject: Predecessors and Alternatives This chapter will deal with the similarities and differences between agent-based simulation in multiagent systems and all the earlier approaches to simulation: continuous and discrete, event-oriented and oriented at equidistant time steps, microsimulation and system dynamics, cellular automata, genetic algorithms and learning algorithms. Many of these earlier approaches were developed for applications in biology, ecology, production planning and other management issues as well as in economics and the social sciences, and at the same time physicists exported some of their mathematical and computer based methods to disciplines such as economics and sociology, forming interdisciplinary approaches nowadays known as econophysics and sociophysics, and although they call their simulations often agent-based, these agents are often not much more than interacting particles whose interactions are based on “forces” that are described in the same manner as gravitational or electrostatic and electromagnetic forces. Agent-based computational demography is a child of both classical microsimulation and multi-agent systems, and the respective ancestry also holds true for many other interdisciplinary approaches, such as systems biology, evolutionary economics and evolutionary game theory, to name just a few. Socionics is another field where multi-agent systems, simulation and classical theory building methodologies of an empirical science come together. Among the early forerunners of multi-agent systems in the social sciences at least two can be named where processes of voter attitude changes are modelled and simulated. Although the poor computer languages of the late 1950s and early 1960s did not allow for agents in the sense of our days, any reimplementation would nowadays be a multi-agent system with several classes of agents (representing voters, candidates, media channels, as in [Abelson/Bernstein 1963] or in the Simulmatics project supporting John F. Kennedy’s election campaign [de SolaPool 1962]) as they dealt with the communications among citizens, between citizens and candidates as well as between citizens and media channels and modelled their behaviour and actions in a rule-based manner — which is to be extended in the long version of the contribution. Microsimulation [Orcutt et al. 1961] is another early forerunner of agent-based systems, as here, too, agents had to be modelled that changed their attributes according to certain stochastic rules — although up to now most microsimulation models do not include interaction between agents (except perhaps for some kind of marriage market), but agent-based computational demography [Billari/Fürnkranz-Prskawetz 2003] makes heavy use of inter-agent processes, for instance as in some recent papers the propensity to bear children is modelled as dependent on the perceived attitudes of a friendship network [Aparicio Diaz 2007]. Many modern tools for multi-agent simulation use the technique of cellular automata [Wolfram 1984] to give agents an environment with a topography that is sufficiently similar to the environment the simulated animals [Drogoul et al. 1995] or humans [Schelling 1971] live in, an approach that was extended into a full grown multi-agent model which reconstructs “social science from the bottom up” (where one has to take into account that the bottom-up approach is not always useful, see the discussion in [NZZ 22/08/2007, to be detailed]). Other early approaches to reconstructing the emergence of complex systems used continuous spaces [Doran/Palmer 1995], as is also the case in the ambitious NEWTIES project [Gilbert 2007]. Discrete-event simulation, too, has extended into the field of multi-agent systems, as in the traditional approach customers could never be modelled as human members of waiting queues often behave: They would move from one queue to look for another queue that seems to be served more quickly, or they leave the system before being served at all, they would negotiate with the server — behavioural features that are difficult to model with classical toolboxes and call for agent-based models. Contribution: Unfolding, Nesting, Coping with Complexity Multi-agent systems also lend themselves to coupling models of different types and to unfold models in a top-down way, starting with a macro model of the top level of the system and then breaking it off, replacing part of the rules of the macro system with autonomous software entities representing realworld elements of the modelled overall system, as for instance exemplified in [Möhring/Troitzsch 2001] and [Brassel et al. 2000]. An approach like this allows researchers to start with a macro view on a complex real-world system — given that all interesting real-world systems are complex. Whereas the complexity of many models derived from some of the existing system theories is restricted to complexity of the interactions between state variables of the system as such ([Forrester 1968]), we usually observe that systems are decomposable into interacting system elements which in turn might by systems of another “natural kind” [Bunge 1979]. On all the levels of such a nested system, agents can be used for representation, although not on all levels the respective agents would need to have all the features that are commonly attributed to them [Wooldridge and Jennings 1995]. Moreover, in such a view, not only the complexity of the domain can be mapped into a simulation model, but also the complexity of time — different time scales for the different levels of a nested system — as for every kind of agents different mechanisms of representing time can be used. In a way, “agents cover all the world” [Brassel et al. 1997] in that multi-agent systems can be used for all simulation purposes, as agents can always be programmed in a way that they behave as continuous or discrete models, can be activated according to event scheduling or synchronously or in a round-robin manner, can use rule bases as well as stochastic state transitions, and all these kinds of agents can even be nested into each other, thus supporting a wider range of applications than any of the classical simulation approaches. This leads to a third aspect of complexity (after the complexity of domains and the complexity of time): agent-based models can encompass several different approaches, both from a technical and implementation point of view, but also from the disciplines making use of simulation (for instance, disciplines such as neurophysiology, cognitive psychology, social psychology and sociology can combine their contributions into a deeply structured simulation model. It goes too far to say that all this would not be possible in a non-agents world (as everything is programmable in Assembler), but examples (to be detailed in the full version) make clear that models combining aspects from even neighbouring disciplines as the ones enumerated in the paragraph above are only understandable and communicable when they come in a modular form that is typical for agent-based models — and the same applies to ecological models where disciplines from physics, via biochemistry to population biology would play their co-operative roles. Issues for Future Research: The Emergence of Communication Although agent communication languages (see the special issue of Autonomous Agents and MultiAgent Systems, vol. 14 no. 2) such as KQML [Labrou and Finin 1997] have been developed for a long time (back to 1993), it is still an open question how agents in a simulation model could develop a communication means on their own and/or extend their communication tool to be able to refer to a changing environment (see the special issue of Autonomous Agents and Multi-Agent Systems, vol. 14 no. 1). Hutchins and Hazlehurst [1995] made a first step into the field of the emergence of a lexicon, but their agents were only able to agree on names of things (patterns) they saw. The NEWTIES project [Gilbert 2007], ambitious as it is, aims at creating an artificial society that develops its own culture and will also need to define agent capabilities that allow them to develop something like a language although it is still questiona