User- and Data-Centric Optimization of Distributed Networks

Distributed networks supporting sensing, learning, and control applications must balance timely information delivery with limited communication and energy resources. This challenge naturally involves both data-centric considerations related to the relevance and freshness of information at the receiver, and user-centric considerations arising from decentralized decision-making and strategic behavior of networked agents. This thesis studies user- and data-centric optimization of distributed networks from a receiver-oriented perspective, in which freshness is interpreted as a property of the evolving knowledge of the receiver's state. Within this framework, the thesis progressively introduces richer models of data relevance, uncertainty, and cooperation, and studies their interaction with decentralized control and incentives. First, we develop adaptive retraining policies for edge-assisted learning under data drift, showing that receiver-aware update control can significantly reduce communication costs while maintaining learning performance. Then, we investigate data freshness, which is often quantified through age of information (AoI) or related metrics. In particular, random access networks are analyzed under both average and peak AoI, revealing trade-offs between efficiency and stability when users autonomously adjust their transmission strategies. To capture the quality of information beyond timestamps, receiver uncertainty is modeled through conditional entropy, leading to structural results that characterize threshold-based transmission policies under resource constraints. The notion of freshness is further generalized through the proposed age of federated information (AoFI), which accounts for quorum-based update requirements motivated by applications such as federated learning. In this setting, both centralized optimization and decentralized strategic behavior are studied, showing that selfish user actions can substantially delay the acquisition of fresh information. Finally, we analyze incentive mechanisms and repeated interactions to quantify inefficiency in repeated interactions, highlighting the temporal nature of performance loss in decentralized systems. Across the considered models, a unifying insight emerges: despite increasing semantic richness of data-centric freshness metrics, simple and structured policies often achieve near-optimal performance, while user-centric decentralization fundamentally reshapes both efficiency and timeliness of information delivery.