Machine Learning Methods for Model Predictive Control

Semi-plenary lecture by Alberto Bemporad at the European Control Conference 2021, July 2, 2021.
Slides: cse.lab.imtlucc...
ABSTRACT:
Machine learning is a set of techniques to extract mathematical models from data that has recently become extremely popular and very successful in many fields, including control. In my talk, I will present several approaches in which machine learning can help design and calibrate model predictive control (MPC) laws and simplify the associated online computations. I will focus on techniques for learning prediction models tailored to nonlinear and hybrid MPC, global and preference-based optimization methods using surrogate functions for actively learning optimal MPC parameters from calibration experiments, and unsupervised and supervised learning techniques for reducing the number of optimization variables in MPC.