Data-driven control of terrestrial and satellite communication networks

Artificial Intelligence and the pervasive presence of Big Data have vigorously become the technological protagonists of the new millennium. In the interconnected world in which we live, millions of virtual interactions take place, thanks to the development of increasingly sophisticated information, electronic and communications technologies, which allow us to enjoy experiences that were unthinkable up until twenty years ago. The availability of an enormous amount of information is also pushing the scientific community relating to automation and control science to move towards a data-driven paradigm, which is opposed to class methods based on the knowledge of the mathematical model of the process to be controlled. This thesis delves into the realm of data-driven control methods in the domains of terrestrial and satellite communication networks, aiming to prove the capability of model--free techniques to optimize network performance, rethink the network selection paradigm, tune dynamically transmission power, and improve signal quality, reliability and availability, allowing for continuous and ubiquitous connectivity. In order to reach these objectives, this essay presents empirical and simulated evidence demonstrating the effectiveness of data-driven control methods in improving the performance and reliability of both terrestrial and satellite networks. The findings have significant implications for the communication landscape, including (i) improved network performance and efficiency within terrestrial networks in relation with critical applications like autonomous driving and Mobile Augmented Reality, (ii) improved adaptability and dynamic decision-making capabilities, and (iii) signal degradation mitigation and uninterrupted connectivity even under challenging atmospheric conditions in satellite networks. The thesis is organized in three Parts. Part 1 discusses about the generalities of data--driven control methods relying on Artificial Intelligence and, in particular, Reinforcement Learning. The dissertation starts from the difference between methods based on knowledge of the model and those that rely solely on data coming from sensors, highlighting pros and cons of each one of the two paradigms. Then, the mathematical foundations of the Reinforcement Learning are provided, with the characterization of Markov Decision Processes (the single--agent domain) and Markov Games (the multi--agent scenario). Eventually, the discussion is shifted from discrete spaces to continuous states and actions, introducing the challenging concept of Deep Reinforcement Learning, which exploits a combination of neural networks to build, train, and test in real--time intelligent agents. Part 2 focuses on the generalities of terrestrial network, with emphasis on the role of the new generations of cellular networks (5G and beyond) and their critical applications, including Virtual and Augmented Reality in the cultural heritage sector. This part of the thesis presents three different control strategies. The first one is a decision framework for the solution of the network selection and traffic steering problems in downlink--only mobile connections. The second one instead considers the uplink plane, proposing a continuous control of transmitting power and image resolution for Mobile Augmented Reality applications. The third and last one poses the attention on another critical application domain of terrestrial networks, i.e., self--driving vehicles. It is shown how it is possible to control vehicle platoons even under the assumption of a complete communication fault. Part 3 spotlights the sphere of satellite communications, with a debate on the dualism between radio frequency and free space optics, showing how the latter constitute a disruptive technology for high--throughput and secure communication between ground stations and satellite assets. Later on, two Machine Learning--based control laws for site diversity implementation are exhibited, the first one using a single geostationary satellite, and the second one operating with a low Earth orbit satellite constellation. Eventually, Chapter 11 will draw conclusions on the work carried out, its scientific impact and practical implications, also showing all possible limitations and blind spots. The path will therefore be traced on possible strategies by which these limits can be overcome in the future. The author of this work hopes that future work and research in applied control science can draw innovative ideas and insights starting from the results exhibited in this doctoral thesis.