Safe continual learning in MPC with prescribed bounds on the tracking error

arXiv2304.10910MaRDI QIDQ6433879FDOQ6433879

Authors: Lukas Lanza, Dario Dennstädt, Thomas Berger, Karl Worthmann 

Publication date: 21 April 2023

Abstract: Recently, a two component MPC scheme was introduced, consisting of pure feedback control (funnel control) and model-based predictive control (funnel MPC). It achieves output tracking of a given reference signal with prescribed performance of the tracking error for a class of unknown nonlinear systems. Relying on the feedback controller's ability to compensate for tracking errors even in the presence of noise and uncertainties, this control structure is robust with respect to model-plant mismatches and bounded disturbances. In the present article, we extend this control structure by a learning component in order to adapt the underlying model to the system data and hence to improve the contribution of MPC. Since the combined control scheme robust funnel MPC is inherently robust with respect to model-plant mismatches and the evolution of the tracking error in the prescribed performance funnel is always guaranteed, the additional learning component is able to perform the learning task online without an initial model or offline training.

This page was built for publication: Safe continual learning in MPC with prescribed bounds on the tracking error

Report a bug (only for logged in users!)Click here to report a bug for this page (MaRDI item Q6433879)