Toward Reinforcement Learning Identification for Swarms Engaged in Cooperative Pursuit

Space vehicle technology and space exploration has benefited from an uptick in research related to swarm robotics. Swarms in the aerospace arena offer measurement capabilities that are not possible via a single vehicle or satellite. Many algorithms have been implemented to achieve complex behavior where each agent in a swarm acts based on local perception and local coordination with each neighboring agent to complete complex tasks. Optimization and artificial intelligence (AI) based models, specifically reinforcement learning based networks, have been researched as key methods for control implementation. From a defense perspective, it may be important to determine what kind of control implementation a swarm is employing to better understand the mechanics of an unknown swarm and identify possible weaknesses. This research constructs a multi-agent pursuit evasion game simulation and experimental implementation using off board control models that employ either reinforcement learning (AI) or optimization based approaches. A simulation environment is developed that will be similar to the future hardware implementation. From the simulation, each agents' state data is used to train a machine learning algorithm to identify the control approach. Future work will include the pursuit evasion game implemented on real hardware using multiple drones engaged in pursuit, switching out the reinforcement learning and optimization based approach to test the algorithm. Understanding the control method employed in a swarm of agents is important in determining system vulnerability and developing successful counter-attack approaches. This research serves to provide insight toward the broader problem of whether a system's underlying control is artificial intelligence-based.