Abstract: We explore several theoretical and applied questions related to the use of data-driven methods in Control.

These methods are very useful to study/control complex, black-box systems or systems evolving in challenging, uncertain environments.

Many such control problems---such as verifying black-box systems or controlling systems a stochastic environments---can be framed as (chance-constrained) optimization problems.

Monte Carlo methods, such as scenario optimization, provide powerful tools for obtaining approximate solutions to these problems, but their generalization (i.e., correctness and safety guarantees) remains critical for deployment in safety-critical systems.

This leads to fundamental questions at the intersection of Machine Learning and Control.

In this talk, we first study what makes a scenario algorithm generalize, presenting recent insights on the necessity and sufficiency of existing generalization conditions, along with new such conditions, bridging scenario optimization and VC theory.

We then demonstrate the applications of these techniques on two control problems of great importance: the verification of black-box hybrid linear systems, and the safe optimal robot motion planning in stochastic environments.

Finally, we present new directions for improving scenario optimization techniques in control tasks, thriving on the approaches of Active Learning and Reinforcement Learning---where agents learn continuously during the operation---, with applications in MPC and online motion planning.

Bio: Guillaume Berger is a Postdoctoral Researcher at UCLouvain, Belgium, in the team of Raphaël Jungers, since 2023.

Previously, he was a Postdoctoral Researcher at the University of Colorado Boulder in the team of Sriram Sankaranarayanan.

He obtained his Ph.D. in the study of Cyber-Physical Systems from UCLouvain in 2021.

His area of research is the analysis, control, and optimization of complex systems, like cyber-physical systems and hybrid systems.

He is particularly interested in formal verification and data-driven analysis of these systems, using tools from Control, Optimization, Computer Science and Machine Learning.