Distributed Multi-Agent Learning/Optimization with Delayed and Compressed Communication

Distributed multi-agent learning through optimization methods in networked cyberphysical systems has been a hot research topic in recent years due to its advantages, such as robustness, modularity, and scalability. It has been recognized that networked cyber-physical systems will become increasingly data-driven. This proposal aims to explore some fundamental research issues that arise at the intersection of these areas. Unlike conventional centralized learning/optimization methods, multi-agents conduct local training steps and exchange information with neighboring agents or a parameter server in the distributed learning/optimization methods. Such a distributed framework reduces centralized schemes' computation and communication burden and preserves owners' data privacy. The necessity of efficient distributed learning algorithms through optimization techniques with good convergence performance and low computational complexity for practical applications is not just a goal but a pressing need that should be addressed urgently.This proposal primarily addresses unbounded communication delays and compression as sources of imperfect communication. In complex networks, the transmitted data may be imperfect due to communication channels' technical or economic constraints. Hence,developing a new approach is necessary to guarantee the convergence of distributed learning/optimization algorithms under such imperfect communication conditions. This proposal primarily addresses unbounded communication delays and compression assources of imperfect communication. Hence, a two-time-scale approach is introduced to guarantee the convergence of distributed learning algorithms under these challenging conditions. The proposed approach shall provide significant theoretical and practicalimplications. Furthermore, the proposal also utilizes the two-time-scale approach to design effective time-triggered schemes that balance communication cost and convergence performance of distributed learning algorithms. If successfully implemented, the theoretical advancements in this proposal will serve as a cornerstone for the future of distributed learning /optimization algorithms in practical applications, demonstrating the potential impact of this research on the field.