Robust reasoning with agent-based modeling

Agent-based modeling (ABM) is a powerful representational formalism that has wide utility for modeling nonlinear systems. For ABM to achieve its potential as a scientific tool, our ability to build models that embody our knowledge must be complemented by rigorous means for making inferences using such models. Due to nonlinearity, this rigor cannot in general be based solely on demonstrating that a model reliably predicts the outcomes of available physical measurements. In this paper we describe an alternative approach to robust reasoning based on the concept of ensembles of alternative models. Ensembles of models can be defined that plausibly span classes of systems including the system of interest. Research methodologies for searching and sampling from such ensembles can be used to support plausible conclusions about invariant properties of ensembles of ABMs and hence of the classes of systems they represent. Notable among these are approaches that implement a competition between ensembles of problem formulations or challenges and conclusions robust to these challenges. This approach is demonstrated using examples drawn from our research.