Disentangling Drivers of Emergent Behavior from Agent Based Models: Application to a Proof-of-Concept Model of Environmentally Driven Human Migration

Abstract:

The flexibility and ability of Agent Based Models (ABM) to show emergent behaviors that can be compared to real systems complexity, as well as improvements in computer power, has led to increasing use of the results of these models by managers and policy makers to inform decisions. However, as ABM complexity increases the model becomes more of a black box, being difficult to analyze the source of its emerging behavior from the agent interactions. This is especially important in high-stake situations involving emergency management or many risk analyses where we have no empirical data to compare to. Global Sensitivity Analysis (GSA) is a robust method able to identify the importance of the model input variables and factors and their interactions based on high-dimensional output variance decomposition. However, ABM intrinsic stochasticity requires additional treatment within GSA to separate the deterministic effects of the ABM input factors from that introduced by the stochastic interactions of the agents. Failing to formally consider this stochasticity in GSA leads to misinterpretation of the importance of ABM drivers. As a proof-of-concept we use an environmentally driven human migration ABM, which stablishes a gradient in water availability and cultural affinity among interfacing locations that results in “resistance to migration” patterns. The application of the extended GSA method proposed herein illustrates how in some instances most of the variability of the ABM output is stochastic, where in other instances there is a strong deterministic control by the input factors.