Towards Emergent Design: Analysis, Fitness and Heterogeneity of Agent Based Models Using Geometry of Behavioral Spaces Framework

Detection and analysis of collective behavior in natural and artificial systems is a difficult task which is commonly delegated to a human observer. We present a statistical framework to automatically detect emergent, collective behavior of agents in agent based simulations which exhibit swarming and flocking behavior. Our tunable, transitional-, rotational, and scaleinvariant framework – geometry of behavioral spaces – identifies common behaviors among agents and translates these behaviors into a system’s behavioral primitives, along with the agent transitions from one behavioral primitive to another. Finally, we use complex network analysis to detect collectives of agents that gravitate into a common cluster of behavioral primitives as the system’s emergent behavior condenses or decays. We apply complex network theory to the analysis of collective behavior dynamics in the simulations of flocking and swarming to validate our analysis. Our framework does not use the knowledge of the parameter space that drive the models, and only relies on the temporal agent trajectories of exhibited behavior. The utility of detecting emergence from exhibited behavior makes this technique suitable as a fitness function for stochastic search algorithms, analyzing evolutionary dynamics of systems with collective behaviors, detecting structures in artificial chemistry experiments, or analyzing physical system such as bacterial formations.