Reverse-engineering emergent collective behaviors in an evolved swarm system

This study examines collective behaviors in large numbers of simple, identical software agents. The motivation for studying this architecture is the possibility of future “intelligent” technologies based on the same mechanisms of adaptive self organization observed in social insect colonies. In this study, a “swarm” is applied to a simulated spatial task loosely based on marine oil spill containment. The swarm consists of hundreds of identical individual software agents acting together without a system of communication. Each agent bases its actions only on information from its local surroundings. Despite the homogeneity and the restriction to local information, the swarms develop strategic, globally-coordinated spatiotemporal structures. We refer to these structures as emergent collective behaviors. “Emergent” refers to the fact that the intricate behavioral patterns observed in the swarms are neither centrally controlled, nor directly attributable to the internal mechanisms of the agents alone; they arise only through a combination of internal processes and interactions.; There is no general design theory for such swarms. This study introduces a method for developing swarm solutions by using a genetic algorithm to evolve the weight matrix for a recurrent neural network. Identical copies of this network control each individual agent in the swarm. The evolved swarms exhibit a variety of structured collective behaviors on the oil spill task. These solutions are reverse-engineered through multi-level analyses that correlate the collective behaviors, the individual behaviors, and the activity of the individual agents' artificial neural mechanisms. We test the hypothesized mechanisms exposed in the analyses by using them to hand-build state-based agent designs that replicate the behavior of the evolved solutions.; The analyses of the swarm activity uncover the use of stigmergy: instances in which the swarm's effects on the environment serve to coordinate and guide its future behaviors. We find that organized behavior emerges from a dense web of interactions in which each agent's behavior constantly shapes, and is shaped by, its environment. This reflects one of the basic tenets of the Distributed Cognition approach: that real cognitive work takes place both inside and outside the individual.