(193e) Advances on a Solver for Multiparametric Model Predictive Control

Multiparametric programming is the mathematical optimization methodology of solving an optimization problem explicitly and offline to reduce real-time optimization's computational burden. Multiparametric Model Predictive Controllers (mpMPCs) is a class of Model Predictive Controllers (MPC) where the entire MPC optimization problem is solved offline. This type of controller has published laboratory and pilot-scale applications in the energy, biomedical, robotics, power electronics, and automotive sectors. Multiparametric programming based methods have been formulated to solve integration of design, control and scheduling, multilevel optimization, robust optimization and more. However, current state of the art solvers do not leverage parallel algorithms and does not scale with hardware innovations that have to lead to a parallel computing paradigm.

In this presentation, we describe the PPOPT (Python Parametric OPTimization) package. It is a general-purpose multiparametric programming package that features: A) Efficient and parallel implementations of multiparametric programming algorithms for all common problem types, B) Problem reformulations to reduce the computational overhead and increase numerical stability, C) Export explicit solutions in C, JavaScript, Python, MATLAB, and Rust. The speed and scaling behavior of PPOPT are explored with computational studies on solving mpMPC controller problems and synthetic problem sets. Additionally, the exported solutions are benchmarked on a wide array of platforms, from micro-controllers to desktop computers.