Pietro Gori

University of Pisa

Tuesday, July 14, 2026, 11:00 - 11:59

SR 01-012

Model Predictive Control (MPC) is a powerful framework for constrained optimal control, but its computational complexity often limits its applicability in real-time settings. Learning-based approximations have emerged as a promising alternative, yet they typically rely on large datasets generated by repeatedly solving expensive MPC problems. In this talk, I will present two complementary approaches that make learning MPC policies significantly more efficient. First, I will discuss an informative sampling strategy that improves dataset generation by focusing on the most relevant regions of the state space and active constraint sets. Second, I will introduce an unsupervised learning framework that formulates policy training as a single large-scale nonlinear optimization problem, eliminating the need to solve hundreds of thousands of individual MPC instances. Numerical examples illustrate the potential of these methods for scalable and efficient MPC policy approximation.