Authors:
Edward Kikken
;
Jeroen Willems
;
Branimir Mrak
and
Bruno Depraetere
Affiliation:
Flanders Make, Lommel, Belgium
Keyword(s):
Optimal Control, Motion Control, Trajectory Planning, Artificial Neural Networks, Mechatronics.
Abstract:
Model Predictive Control is an advanced control technique that can yield high performance, but it is often challenging to implement. Especially for systems with dynamics that are complex to model, have strong nonlinearities, and/or have small time constants, it is often not possible to complete the needed online optimizations fast and reliable enough. In this work we look at approximating the MPC solutions using black-box models i.e. deep neural networks, so that the computational load at runtime is strongly reduced. We use a supervised learning approach to train these models to yield outputs similar to those of an example dataset of offline pre-computed MPC solutions. We illustrate this approach on three realistic (active-suspension system, parallel robot, and a truck-trailer), illustrating the typical workflow and how the approach has to be set up to address the varying challenges. We show that the approximate MPC solutions yield a high level of performance, reaching nearly the lev
el of the original MPC, yet at a strongly reduced computational load.
(More)