Abstract
The effects of distinct agent interaction and activation structures are compared and contrasted in several multi-agent models of social phenomena. Random graphs and lattices represent two limiting kinds of agent interaction networks studied, with so-called ‘small-world’ networks being an intermediate form between these two extremes. A model of retirement behavior is studied with each network type, resulting in important differences in key model outputs. Then, in the context of a model of firm formation, in which multi-agent structures (firms) are emergent, it is demonstrated that the medium of interaction—whether through individual agents or through firms—affects the qualitative character of the results. Finally, alternative agent activation ‘schedules’ are studied. In particular, two activation modes are compared: (1) all agents being active exactly once each period, and (2) each agent having a random number of activations in every period with mean 1. In many circumstances these two regimes produce indistinguishable results at the aggregate level, but in certain cases the differences between them are significant.
This paper was published in Multi-Agent Based Simulation, Springer Verlag Lecture Notes in Artificial Intelligence, 2000.
Introduction
One class of multi-agent systems (MAS) consists of a relatively small number of agents, each of whom has relatively sophisticated behavior (e.g., a rich cognitive model, perhaps for dealing with a complex task environment [18]). A different type of MAS involves relatively large numbers of behaviorally simple agents. This second family of multi-agent systems is of significant interest as the basis for empirically-relevant models of human social and economic phenomena. Such models typically involve the use of aggregate social or economic data to estimate parameters of a MAS in which agents have heterogeneous internal states (e.g., preferences) but a common repertoire of behaviors (e.g., economic exchange).
One reason for the elevated attention given to simple agents is that the prevailing norm in the mathematical social sciences is to build models that abstract from the details of cognition.2 Stated differently, the focus of economists and other quantitative social scientists on behaviorally simple models is a symptom of the lack today of anything like a universal model of cognition. A second reason for differential interest in models composed of moderate or large numbers of simple agents is that such systems are quite capable of complex aggregate behavior, involving, for example, the spontaneous emergence of behavioral norms (e.g., [19]) or the self-organization of multi-agent coalitions (e.g., [3]). Understanding the origin of these complex patterns of emerged behavior is often a significant challenge, and would be even more difficult if individual agents were complex in their own right—if individual decisions were also emergent.