On generating and exploring the behavior space of complex models

This paper introduces a novel approach for (i) exploring the spectrum of behaviors system dynamics models could generate and (ii) identifying and selecting exemplar simulation runs that are representative in terms of their dynamics and their origin in the input space. Our overall approach consists of two phases: a first phase of iterative behavior‐based sampling and simulation; and a second phase of behavior‐based classification or clustering, input‐space identification and selection of representative exemplars. For this approach, we introduce new ways to characterize the dynamic complexity of simulation runs, develop a new iterative output‐oriented sampling approach, and combine classification, clustering, data visualization and machine learning techniques. Using two well‐known system dynamics models, we show how our approach enables one to generate a wide spectrum of behaviors and group runs based on their type of behavior, and identify distinct areas in the input space based on the particular type of behavior they generate. Copyright 2016 System Dynamics Society