Note: This content is accessible to all versions of every browser. However, this browser does not seem to support current Web standards, preventing the display of our site's design details.


Methods and Tools for Embedded Model Predictive Control


A. Domahidi

University of California Berkeley, CA, USA

This talk deals with methods and tools for the implementation model predictive control (MPC) on embedded platforms. Particularly, two approaches will be presented that aim at bridging the gap between problem formulations with high performance and the limited computing power of low-cost embedded control platforms. The first part of the talk deals with learning decision rules from simulation data to mimic the binary decisions of a hybrid MPC building controller as closely as possible while maintaining high performance. To this end, we employ AdaBoost, a well known concept from machine learning that combines a number of weak learners into a strong classifier. The resulting controller maps directly to a majority voting system using simple if-then-else rules; a structure that building operators are familiar with. We discuss feature selection, i.e. which measurements are actually important for the decision to be made? Closed-loop simulation results show that the learned controller, while offering a significant complexity reduction, is very close in performance to hybrid MPC. In the second part of the talk, we present efficient interior point methods tailored to convex multistage problems, a class that most linear MPC problems can be cast in. We discuss aspects of structure exploitation when solving the underlying KKT system and give examples of MPC problems that are significantly faster to solve than the standard quadratic program (QP) with polytopic constraints, e.g. a quadratically constrained QP. We introduce the FORCES code generator, available at, which allows one to generate tailored second-order solvers that are orders of magnitude faster and smaller than existing interior point codes.


Type of Publication:


File Download:

Request a copy of this publication.
(Uses JavaScript)
% No recipe for automatically generating a BibTex entry for (06)Talk
Permanent link