Filippo Fabiani - Beyond the black box Certified machine learning for control and decision making

Speaker: Filippo Fabiani, DYSCO Research Unit at the IMT School for Advanced Studies Lucca (IT) Title: Beyond the black-box: Certified machine learning for control and decision-making Bio: Filippo Fabiani is an Associate Professor in the DYSCO Research Unit at the IMT School for Advanced Studies Lucca (IT). He received the B.Sc. degree in Bioengineering, the M.Sc. degree in Automatic Control Engineering, and the Ph.D. degree (cum laude) in Automatic Control from the University of Pisa (IT), in 2012, 2015, and 2019 respectively. In 2017-2018, he visited the Delft Center for Systems and Control at TU Delft (NL), where in 2018-2019 he spent one year as post-doctoral Research Fellow. Before joining IMT Lucca as an Assistant Professor, he was a Post-Doctoral Research Assistant in the Control Group at the Department of Engineering Science, University of Oxford (UK), from 2019 to 2022. Since 2025, he is Associate Editor for the IEEE Control Systems Letters. His research to date has been at the intersection of machine learning, game theory and control engineering, and concentrates on developing both data-driven theoretical tools and algorithmic methods for modern control design. These include optimal control of multi-agent and uncertain systems such as power and traffic networks, and emerging control applications in smart grids and smart cities. Abstract: The renewed interest in data-driven learning techniques is reshaping conventional paradigms in control and decision-making, where the convergence of accessible, large datasets and advanced machine learning offers transformative opportunities. On the other hand, accompanying these learning-based methodologies with rigorous performance certificates is an essential requirement to ensure safe and proficient operations in modern engineering systems. In the first part of the seminar, we will introduce recently developed neural network-based approaches to accomplish traditional control and multi-agent decision-making problems. First, we will consider the problem of designing a neural network surrogate of an unknown dynamical system from a finite number of data points such that the model obtained is also suitable for optimal control design [1]. To this end we propose a specific, yet easy-to-train, neural network architecture that, under a careful choice of its weights, produces a hybrid system model with structural properties that are highly favourable when used as part of a finite horizon optimal control problem. Successively, we will discuss the design of stabilizing neural network controllers for both deterministic [2] and uncertain systems [3]. Specifically, we focus on the problem of certifying the approximation quality of a neural network with rectified linear units in replicating the action of traditional optimization-based controllers, such as MPC. We develop an offline methodology requiring the construction and solution of mixed-integer linear programs so that, in case the resulting optimal solution meets certain problem-dependent value, the stability of the closed-loop system with a neural network controller is guaranteed. In the second part of the seminar, instead, we will focus on active learning-based policies for decision-making and control. First, we will consider the case in which an agnostic entity aims at learning a possible outcome of the multi-agent decision process observed [4]. We design an iterative scheme where such an external observer, endowed with a learning procedure, is allowed to make smart queries and observe the agents’ reactions through private action-reaction mappings, whose collective fixed point corresponds to a stationary profile. We establish sufficient conditions to assess the asymptotic properties of the proposed active learning-based approach. Finally, we will consider the design of control Lyapunov functions and associated state-feedback controllers for uncertain systems [5]. To this end, we design a counter-example guided inductive synthesis scheme, where the joint – yet “adversarial” – action of a learner (which computes tentative candidates) and a verifier (which generates suitable counter-examples) produces a solution with a finite number of samples. References - see eventpage: https://www.digitalfutures.kth.se/eve...