On fundamental trade-offs and architecture design in Networked Control Systems

It can be legitimately said that Networked Control Systems represent one of the biggest breakthroughs in engineering over the latest decades. Stemming from the intertwining among control, computer engineering, and telecommunications, these powerful systems received the legacy of classical communication and computer networks, but leveled it up by virtue of autonomy of each involved unit. Nowadays, examples of Networked Control Systems are smart power grids, smart homes and buildings, Industry 4.0 and Industrial Internet of Things, and smart agriculture, to mention a few. Even more futuristic applications, such as networks of autonomous vehicles or search-and-rescue robotic teams, are predicted to be available on the market in a matter of time. Despite the exponential growth of such systems both in industrial applications and in research, one main reason why the current development is somewhat refrained on several aspects is that designing a Networked Control System is challenging in nature. In fact, not only blending different engineering fields raises novel issues, but also the interdependence of individual subsystems makes it hard to design control and, in general, decision-making procedures at local level, whereas design at global level is not only undesired but sometimes even unfeasible. To mention just one example of design complexity, while it is well known that the optimal Linear Quadratic controller for a single system can be found by solving a (relatively simple) algebraic matrix equation, it is also known that solving the same problem for a distributed controller is NP hard. Because engineered systems must work in real life, the lack of strong theoretical results is typically replaced with ad-hoc and heuristic methods, that try to take the best of both worlds of human experience and available mathematical tools. While such solutions have already yielded impressive applications, relying on intuition might not always be the best strategy, and theoretical advancement is needed to unleash the full potential of Networked Control Systems. For example, recently introduced Multi-Agent Reinforcement Learning, even though it has proved powerful in some scenarios, still leaves open room for improvement before it can be safely deployed in the real world. This thesis investigates, and possibly questions, the role that conventional wisdom plays in design of Networked Control Systems. Specifically, the aim is to explore situations where common design beliefs might not match the real nature of the system to be designed, possibly causing loss in performance. Three conventions will be examined: more sensors improve estimation; more communication links increase control performance; more collaboration enhances cooperative tasks. While such conventions seem indeed reasonable, results exposed in this thesis will show that it is not always so: in fact, more sensors may hinder estimation under computational delays; more communication links may degrade control performance under communication delays; more collaboration may be dangerous under misbehaving agents. Even though most results are limited to analysis, and practical design indications are still preliminary, the hope with this piece of research is to offer high-level guidelines and insights that can improve classical conventions and possibly pave the way to novel research directions, to be exploited in the design of high-performing Networked Control Systems.