Using Time Series Motifs to Explore the Parameter Space of Agent-Based Models: A Pilot Study

Agent-based modeling (ABM) has seen a resurgence of interest in recent years with the continued expansion of computational capabilities, allowing researchers to study the systemic effects of individual-level behavior on increasingly larger and more complex systems. As ABMs become mainstream tools in a variety of disciplines, a need has emerged for better techniques to assess the sensitivity of ABMs and identify the critical regions of the input space where small changes in the input variables result in large qualitative changes in the model's behavior. This work attempts to address this need by characterizing the behavior of ABMs in a two-step process. First, we use time series clustering to identify the behavioral motifs present in a model. Second, we use classification to find and describe the relationship between the motifs and the input space. This feasibility study successfully mapped the input space to the time series motifs for the Schelling model and identified further directions of inquiry towards a generalized method for more complex ABM characterizations.