Communication-efficient and privacy-aware distributed optimization in multi-agent systems

Abstract

The rapid development of control, sensing, and communication technologies has facilitated the emergence of multi-agent systems. However, concerns regarding communication overload have arisen due to the limited bandwidth and energy resources of networked systems. Additionally, preserving privacy during information transmission while guaranteeing convergence has become an intriguing topic. This thesis investigates the trade-off among communication, privacy, and accuracy in distributed optimization for multi-agent systems and proposes algorithms that are communication-efficient and privacy-aware. To enhance communication efficiency in networked competitive agents, we propose a novel stochastic event-triggered distributed Nash equilibrium seeking algorithm for non-cooperative games. The designed stochastic event-triggering law activates agents’ transmission with a probability dependent on certain events that reflect the importance of current information, thereby reducing the number of communication rounds and ensuring convergence of agents’ decisions to an accurate Nash equilibrium in a mean square sense. The proposed algorithm extends existing deterministic event-triggered mechanism and improves the trade-off between communication efficiency and convergence. To preserve privacy in distributed resource allocation problems over directed networks, we develop a di↵erentially private dual gradient tracking algorithm. This algorithm utilizes Laplacian noise to obscure transmitted messages in networks and leverages the robust push-pull technique to reduce noise accumulations and improve the convergence accuracy. Through careful design of step sizes and noise parameters, the designed algorithm ensures privacy preservation even when the total number of iterations approaches infinity. Finally, we address the joint concerns of communication efficiency and privacy preservation in distributed optimization. We explore the interplay between efficient communication and privacy, exploiting the random activation of agents and sparsification to amplify privacy guarantees. By leveraging the inherent characteristic of sparsification, our algorithm reduces the noise perturbation intensity without sacrificing differential privacy, thereby achieving a better balance among communication, privacy, and accuracy trade-offs.

Date of Award	2024
Original language	English
Awarding Institution	The Hong Kong University of Science and Technology
Supervisor	Ling SHI (Supervisor)