An empirical nonlinear dynamics approach to analyzing emergent behavior of agent based models

Abstract:

Our ability to construct complex agent based models (ABMs) exceeds our current capacity to evaluate their emergent dynamics. This creates an ‘explainability’ problem in convincing policymakers that ABM results can be trusted to correspond to the real-world that they are charged with regulating. One must go outside of the ABM to analyze its output since the maze of randomized interactions cannot be reverse-engineered to tie emergent behavior to the agents that gave rise to it. Since ABM output is often recognized to have nonlinear dynamics, we propose a deterministic alternative that embeds ABM dynamics in a nonlinear state space reconstructed from ABM output time-series records. Reconstructed ABM dynamics can be analyzed with a collection of recent empirical nonlinear dynamic (END) methods that detect the presence of dissipative (dimensional-reducing) deterministic nonlinear dynamics in ABM output, detect causal interactions among ABM covariates, and extract a phenomenological system of ordinary differential equations (ODEs) that serves as a meta-model of the ABM emergent dynamics. As a proof-of-concept we use an environmentally driven human migration ABM, which establishes a gradient in water availability and cultural affinity among interfacing locations that results in “resistance to migration” patterns. Another abstract in this session applies Global Sensitivity Analysis (GSA) to analyze the output of the human-migration ABM. Our presentation will compare END and GSA analyses to determine if the probability-based GSA method ‘fingerprints’ in some way deterministic structure uncovered by END.