Datum predavanja: ponedjeljak, 24. 5. 2021.
Vrijeme predavanja: 11h

Uroš, Kalabić, Mitsubishi Electric Research Laboratories in Cambridge, Massachusetts, USA
Title: A constraint-separation principle in model predictive control

Abstract:
In 2002, Kouvaritakis, Cannon and Rossiter published the paper “Who needs QP for linear MPC anyway?” where they popularized the fact that, in the case of linear time-invariant systems, an MPC control can be decomposed into a feedback component with LQR gain, and a constraint-enforcing component obtained by solving an optimization problem with a cost function that is independent of the state.
In this talk, I will show that this decomposition can be performed for all locally linear MPC problems, i.e., constrained optimization problems for linearizable, potentially time-varying, systems with a quadratic cost function, and I will also provide a generalization of the result to nonlinear systems. It turns out that a locally linear MPC problem can always be decomposed into an inner-loop LQR problem, which has a convenient closed-form solution, and an outer-loop minimum-norm projection problem, which projects the control sequence onto a constraint set.
I propose to call this a constraint-separation principle since it locally separates stabilization from constraint enforcement in MPC. In particular, I am interested in the practical applications of the result and will cover praxis during the second part of the talk. The literature thus far has focused on the numerical benefits of decomposition. Other practical considerations include: What does it mean to use MPC for stabilization and when is it useful? Does decomposition help in design or certification of constraint-enforcing control schemes? Are linear tracking MPC and extended command governors equivalent?

Bio:
Uroš Kalabić is a Principal Research Scientist at Mitsubishi Electric Research Laboratories in Cambridge, Massachusetts, USA. He received a PhD in aerospace engineering in 2015 from the University of Michigan, Ann Arbor. He also received an MS in mathematics in 2014 and an MSE in aerospace engineering in 2011 from the University of Michigan, and a BASc with honours in engineering science from the University of Toronto in 2010. His main research interest is the control of systems subject to state constraints, and he has authored over 40 peer-reviewed papers and is named an inventor on over ten patent applications in the area.