Analysis and design of multi-agent systems under communication and privacy constraints

This dissertation presents techniques for design and analysis of multi-agent distributed systems with control oriented objectives. We study two problems: one related to networked estimation in Networked Control Systems and the other related to privacy in Cyber-physical systems.;In the first problem, we focus on congestion control in a communication network that is supporting remote estimation of multiple processes. A stochastic rate control protocol is developed using the network utility maximization framework. This decentralized protocol avoids congestion by regulating the transmission probabilities of the sources. The presence of estimation costs poses new challenges; however, for low congestion levels, the form of rate controller resembles that of the standard TCP rate controller. Stability of the protocol is analyzed in the presence of fixed network delays.;In the second problem, we address the issue of privacy of agents in a multi-agent LTI system which is monitored by a control center via the measurements sent to it by the agents. We show that such architecture is prone to privacy breaches in which an intruder can gain access to agents' sensitive parameters that govern their dynamics. To prevent this, we employ the differential privacy framework and develop a noise adding privacy mechanism in which the agents add synthetic noise while sending their measurements to the control center. We design the privacy noise by characterizing the sensitivity of the system. We substantiate our framework by studying two concrete examples of second-order consensus and LQR control. Our numerical results show that in an asymptotic regime of low privacy and high SNR, the privacy noise results in marginal performance degradation at the control center, when compared to the error suffered by the intruder in identifying the sensitive parameters.;We study another related privacy problem for a scenario where multiple agents cooperatively solve a quadratic optimization problem. To maintain privacy of their states over time, agents implement a noise-adding mechanism according to the classic differential privacy framework. We characterize how the noise due to the privacy mechanism degrades the performance of the multi-agent system. Interestingly, we show that depending on the desired level of privacy (and thus noise), the system performance is optimized by reducing the level of cooperation among the agents. The notion of cooperation level models the trust of an agent towards the information received from neighboring agents. For the prototypical examples of consensus and centroidal Voronoi tessellations, we are able to characterize the optimum cooperation level that maximizes the system performance while ensuring a desired privacy level. Our results suggest that for the class of problems we study, and in fact for a broad class of multi-agent systems, it is always beneficial for the agents to reduce their cooperation level when the privacy level increases.