Multi-agent systems arise in diverse fields, including power systems, robotics, cyber-physical systems, and the Internet of Things. Coordinating these systems is often done using decentralized interactions, in which each agent only communicates with a small number of others. Decentralized algorithms offer several benefits, though they may have difficulty accommodating some performance demands, such as user privacy requirements. Toward addressing such challenges, I will present recent work on mixed centralized/decentralized decision protocols for multi-agent systems. Motivated by the availability of cloud computing, a centralized cloud computer is added to networks of agents in order to gather global information, perform centralized computations, and broadcast the results. As this happens, the agents continue to execute a decentralized behavior. The centralized nature of the cloud means it will be slower than the agents, though its slow, occasional transmissions do indeed enable multi-agent systems to handle various practical challenges. To this end, I will present mixed centralized/decentralized coordination algorithms that tolerate asynchronous information sharing and user privacy requirements, while still enabling strong theoretical guarantees of performance. In the asynchronous case, I will present an algorithm that allows each agent to perform useful work even if the agents have conflicting information about the network. For privacy, the framework of differential privacy is used, giving rise to a novel stochastic optimization algorithm. These algorithms draw from primal-dual optimization techniques and the theory of stochastic variational inequalities, and solve coordination tasks that are stated as convex optimization problems. The end result is a flexible coordination framework that tolerates an array of practical challenges, all while solving constrained coordination problems for teams of agents, regardless of whether an agent is a robot, a self-driving car, or any other physical entity. In addition to theoretical results, I will present robotic implementations of this work to demonstrate its applicability in practice.
Matthew Hale is a Ph.D. Candidate in Electrical and Computer Engineering at the Georgia Institute of Technology. In 2012, he received his B.S.E. in Electrical Engineering from the University of Pennsylvania, where he was a member of the GRASP Lab. He received his M.S. in Electrical and Computer Engineering from Georgia Tech in 2015, and was awarded the Colonel Oscar P. Cleaver Outstanding Graduate Student Award by the same department in 2013. His research interests include optimization and control for multi-agent systems, differential privacy, and hybrid systems. His work applies methods from these areas to cyber-physical systems and teams of robots.