To path indicator Subpages of “Nachrichtendetail“ To navigation To quick access To footer with other services
To content
Logo of the Control and Cyberphysical Systems Group
Department of Mechanical Engineering
Studierende sitzt am Arbeitsplatz und schaut auf den Laptop. © Aliona Kardash​/​TU Dortmund
at - Automatisierungstechnik

Paper in at - Automatisierungstechnik

A diagram illustrating two approaches to data-driven control design: 1) An indirect approach involving "Model Identification" from "Data" followed by "Model-Based Design" to "Control". 2) A "Direct Data-Driven Design" path bypassing explicit model identification. The concept of "Implicit Predictors" provides a model-based interpretation of predictions made by direct data-driven control schemes. © MKL ​/​ RCS
Implicit predictors provide an indirect (i.e., model-based) view on the predictions made by direct data-driven predictive control schemes.
Manuel Klädtke and Moritz Schulze Darup have published a paper on explainability in data-driven predictive control (DPC) with regularizations. The paper proposes two tools designed to enhance explainability: implicit predictors and the trajectory-specific effects of regularizations. It demonstrates their application and explores further extensions for broader applicability.

Data-driven predictive control (DPC) has emerged as a promising alternative to traditional model predictive control (MPC), utilizing linear combinations of recorded trajectory data without relying on explicitly defined prediction models. However, this absence of a clear model can lead to opacity regarding the effects of commonly used regularization terms, resulting in various challenges and limited interpretability.

In this paper, we address these challenges by proposing two analysis tools focused on enhancing explainability: implicit predictors and the trajectory-specific effects of regularization. Building upon these previously introduced tools (see our earlier L-CSS paper ), we provide concise explanations, interpretations, and demonstrations using commonly employed quadratic regularization. Following this overview, we analyze further extensions such as affine DPC, offset regularizations, slack variables, and terminal constraints to broaden the applicability of our methods.

This work aims to enhance both the explainability and reliability of regularized DPC by providing a deeper understanding of these mechanisms, ultimately benefiting practical applications. Our article is published in the special issue "Data-Driven and Learning-Based Control – Perspectives and Prospects" of the journal "at - Automatisierungstechnik.". Contact Manuel for further details or have a look at the paper:

M. Klädtke and M. Schulze Darup, Towards explainable data-driven predictive control with regularizations, in at - Automatisierungstechnik, DOI: 10.1515/auto-2024-0161.

To top of page