Dottorato in AUTOMATICA, BIOINGEGNERIA E RICERCA OPERATIVA -
ABRO

Titolo della tesi: Decentralised Learning for Intelligent Control Systems

This manuscript represents a collection of the most important research activities carried out by the candidate during his three years of PhD studies. This work leverages on the concept of Intelligent Control Systems, defined as a framework where control methods attempt to emulate important characteristics of human intelligence to generate control actions. Being referred to as a point of contact between the scientific fields of control theory and artificial intelligence, they aim to combine the mathematical rigor of the former with the representativeness of the latter in order to exploit the potential of both of them. While classical control theory model-based approaches commonly used to examine the characteristics of a given system in terms of its stability, safety and optimality, may fail to include environmental uncertainties and are subject to modelling errors, data-driven controller design techniques aim to capture such stochasticities and nonlinearities. This idea is behind the development of the first work which develops neural-based control solution which envisages the use of deep neural networks within the model predictive control framework with the aim to derive the optimal control law in a distributed fashion by means of a cascading combination of one-step predictors. The second work focuses on the learning processes of data-driven methodologies, with particular attention to neural networks, whose approximation capabilities make them of one of the most important tools in the approximation of system dynamics. The research activity, carried out by the candidate in the scope of the POR FESR FedMedAI project, develops a decentralised framework based on consensus-theory aimed at allowing the training of a neural networks over decentralised scenarios, namely on data belonging to multiple actors who communicate with each other and collaborate for the learning of a data-driven model aimed at solving an approximation task. The specifications of this framework were defined during the course of the project through interactions with the Italian Istituto Superiore di Sanità, allowing the realization of a platform aimed at enhancing a privacy-preserving collaboration among clinical institutions, without any exchange of clinical data. The investigation of the mechanisms underpinning the interaction between different actors is examined within the third work in the context of multi-agent systems. Since communication is one of the tools used by agents to collaborate, a learning-based strategy allowing agents to limit their communication while still achieving their objective is proposed leveraging on the multi-agent reinforcement learning framework. The proposed approach allows to cope with real-world scenarios where communication-related costs cannot be neglected. The fourth work discusses multi-agent scenarios where each agent attempts to accomplish its own objective independently of other agents' cooperation. Numerous settings find use for these non-cooperative scenarios, one of which being telecommunications. In this context, the convergence properties of a class of load-balancing strategies towards a set of approximate non-cooperative equilibria are examined. The candidate also explores non-cooperative approaches in the domains of mobile edge computing and automotive, whereby decentralised policy broadcasting mechanisms and decision-making processes based on reinforcement learning are proposed. All the studies incorporated into this work addresses various issues and challenges that may arise when intelligent control systems are employed in multi-agent context. In particular, control systems of this type find application in the control of complex systems, such as health-related ones, in which the interaction with the human being constitutes the most critical aspect. With respect to this issue, the high-level architecture of the PON CADUCEO, POR FESR FedMedAI and Allenamente project is described. Every study under consideration is predicated on the use of various control theory arguments and data-driven approaches, whose choice and combination is justified and validated over different scenarios.