On the Optimal Interaction Range for Multi-Agent Systems Under Adversarial Attack (2004.06562v2)

Abstract: Consider a consensus-driven multi-agent dynamic system. The interaction range, which defines the set of neighbors for each agent, plays a key role in influencing connectivity of the underlying network. In this paper, we assume the system is under attack by a predator and explore the question of finding the optimal interaction range that facilitates the most-efficient escape trajectories for the group of agents. We find that for many cases of interest the optimal interaction range is one that forces the network to break up into a handful of disconnected graphs, each containing a subset of agents, thus outperforming the two extreme cases corresponding to fully-connected and fully-disconnected networks. In other words, the results indicate that some connectivity among the agents is helpful because information is effectively transmitted from the agents closest to the predator to others slightly farther away, but also that too much connectivity can be detrimental to the agility of the group, thus hampering efficient and rapid escape.