Granularity and the Validation of Agent-based Models

摘要

Even when the question is well-posed, it is often difficult to determine an appropriate level of detail in a multi-agent model for any complex system, therefore in practice, frequent revisions on model granularity become inevitable. Ideally, we would like a modeling methodology that allows small and incremental changes in granularity. This allows different problem-specific factors to be modeled in greater or lesser detail according to their importance in explaining observed phenomena. In this paper we propose a network representation of agent behaviour called Agent Behavior Network that describes, visualizes, and supports formal analysis. It allows for heterogenous granularity within a single model and facilitates systematic increments in model granularities and hence, model extensibility. We demonstrate the approach with three models of chemotaxis with increasing model granularity, and compare the simulation results with observations from the Under-Agarose Assay. We show that the improvement in model granularity greatly improves its agreement with laboratory observations. We conclude that an Agent Behavior Network is a necessary tool to facilitate a more systematic process of designing and validating agent-based models of complex systems.