Goal-Oriented communication: theoretical analysis and practical algorithms

The increasing deployment of distributed intelligent systems composed of connected autonomous agents calls for a rethinking of communication architectures beyond the traditional objective of reliable data transmission. In many emerging applications, decision-making is sequential, interactive, and task-driven, requiring communication strategies that explicitly account for the goals, autonomy, and behavior of the agents involved. This perspective aligns with the concept of Goal-Oriented Communication (GoC), where communication, estimation, and control are jointly designed to optimize task performance rather than information fidelity alone. This thesis develops a comprehensive theoretical and algorithmic framework for GoC in networked decision-making systems, with a particular focus on cyber-physical systems. The first part establishes fundamental performance limits and structural properties of optimal communication and control strategies, characterizing the trade-offs between communication resources and task effectiveness. By leveraging modified modified policy optimization formulations, this part provides theoretical guidance on joint policy discovery when communication actions influence learning and decision-making dynamics. The second part addresses practical scenarios in which accurate system models are unavailable or intractable. It proposes distributed learning-based solutions for remote estimation and control under high-dimensional observations, unreliable wireless channels, and stringent bandwidth and memory constraints. In this setting, encoders, decoders, and control policies are parameterized as machine learning models and jointly optimized to maximize GoC performance. Finally, the third part investigates the security implications of joint communication and control strategies. It formalizes eavesdropping and side-channel attacks that exploit GoC mechanisms to infer system states, and it analyzes mitigation techniques that balance task performance, communication efficiency, and secrecy. Overall, this thesis provides a unified treatment of the theoretical foundations, learning-based implementations, and security considerations of GoC, contributing essential insights for the design of intelligent, efficient, and robust distributed autonomous systems.