Appraisal of Autonomous Swarms through Analysis of Observed Behavior

Abstract

Swarms of autonomous vehicles are capable of performing complex missions in a variety of applications. Functions inherent to these missions include obstacle avoidance and collaboration with other swarm members. The logic for guiding autonomous agents through these functions can result in unanticipated emergent behaviors. Commanders of complex autonomous missions need a way to gain confidence in a swarm's behavior and detect adversarial behavior at runtime without inhibiting operations. The research described in this paper explores using measurements and analysis of external, observable characteristics, such as location data, to detect adversarial behavior in a simulated homogeneous swarm for a set of well-defined use cases. Initial results using directional and positional entropy of individual agents and the DBSCAN clustering algorithm demonstrate that measurements of external characteristics are a promising addition to a commander's toolset. Further research should be performed to determine the applicability to a broader set of use cases.