Agent-based model

Agent-based modeling (ABM) is a computational methodology in which a system is represented as a collection of autonomous agents that interact according to local rules, producing global patterns that are not programmed into any single agent. ABM is the systems-theoretic approach to mathematical modeling: it does not assume that the system can be described by aggregate equations, but rather builds the system from the bottom up and observes what emerges.

The core premise of agent-based modeling is that many important phenomena — emergence, self-organization, phase transitions, cascading failure — cannot be captured by models that average over agents. A market of identical rational traders cannot produce bubbles. An ecosystem of identical predators cannot produce biodiversity. A city of identical commuters cannot produce traffic jams. These phenomena require heterogeneity: agents that differ in their rules, their information, their strategies, and their network positions. ABM builds this heterogeneity in from the start.

The methodology has three distinctive features. First, agents are autonomous: they make decisions based on local information and local rules, not on a global plan. Second, interactions are local: agents influence only their neighbors, and the network of interactions is often explicit and dynamic. Third, the system is emergent: global patterns arise from the accumulation of local interactions, and the modeler does not know in advance what patterns will emerge. This makes ABM an exploratory methodology, not a confirmatory one. It is used to discover what kinds of local rules produce what kinds of global patterns, and to test whether a hypothesized set of local mechanisms is sufficient to explain an observed global phenomenon.

ABM has been applied across domains: economics (market dynamics, herding behavior, financial contagion), biology (ecosystem dynamics, epidemiology, immune response), sociology (opinion formation, social contagion, institutional evolution), and computer science (distributed systems, peer-to-peer networks, swarm robotics). The unifying feature is that in all these domains, the system is composed of heterogeneous, interacting agents whose collective behavior is the object of study.

The primary critique of agent-based modeling is that it is underdetermined: many different sets of local rules can produce the same global pattern, and the modeler must choose among them without decisive empirical guidance. This is the inverse problem of ABM: given a global pattern, infer the local rules. The problem is ill-posed, and ABM practitioners must supplement their models with empirical data, experimental validation, and theoretical constraints to avoid overfitting. The response to this critique is that ABM is not intended to produce unique explanations. It is intended to produce possible explanations — to show that a set of local mechanisms is sufficient to produce an observed phenomenon, and to explore the space of alternative mechanisms that could do the same.