UC Irvine

In an era where timely and accurate information is paramount across various domains---including wireless sensor networks, environmental monitoring, health care, vehicular networks, and federated learning systems---the \textit{Age of Information} (AoI) has emerged as a critical metric for quantifying information freshness in communication networks. This thesis investigates the quantification and minimization of AoI in two key areas: sensing networks and federated learning systems.

In the first part of the thesis, we focus on sensing networks, where timely data acquisition and dissemination are essential. We analyze AoI in multiple-sensing networks comprising multiple sources and multiple servers. Each source represents an independent piece of information, and its AoI is individually measured. We consider both homogeneous and heterogeneous scenarios in terms of arrival and service rates. Utilizing stochastic hybrid system (SHS) models, we derive closed-form expressions and develop efficient algorithms for computing the average AoI under various queuing policies, such as Last-Come-First-Serve with preemption (LCFS). Our analysis reveals insights into optimal update scheduling strategies and highlights the impact of network heterogeneity on information freshness.

Additionally, we study AoI in sensing networks over erasure channels with feedback. We consider a status updating system where updates are transmitted over an erasure channel, and an error-free feedback channel informs the source about successful packet deliveries. Each update consists of multiple packets, and a new update is available at each channel use. The key challenge lies in deciding whether to continue transmitting the current update or discard it in favor of a fresher one, based on the history of successful and failed transmissions. We derive bounds on the average and peak AoI and propose optimal transmission policies that minimize AoI. Our findings demonstrate how leveraging feedback can significantly enhance information freshness in unreliable communication environments.

In the second part of the thesis, we apply the concept of AoI to federated learning (FL), a decentralized machine learning framework that enables collaborative model training without sharing raw data. We address challenges such as load balancing, resource utilization, scalability, and convergence in FL systems. We propose a novel client selection policy based on AoI, defining a load metric as the number of rounds between subsequent selections of a client. By minimizing the variance of this load metric across clients, we ensure a balanced workload distribution and consistent participation. We introduce a decentralized Markov scheduling policy that reduces management overhead and adapts to dynamic network conditions. Our theoretical analysis and simulations show that this approach not only improves fairness and operational efficiency but also enhances the convergence rate of learning models.

Collectively, the contributions of this thesis advance the understanding of AoI in both sensing networks and federated learning systems. By providing practical strategies for optimizing information freshness, we address critical challenges in modern communication and learning systems where timely information is paramount.