Title | ||
---|---|---|
Stability Certificates for Neural Network Learning-based Controllers using Robust Control Theory |
Abstract | ||
---|---|---|
Providing stability guarantees for controllers that use neural networks can be challenging. Robust control theoretic tools are used to derive a framework for providing nominal stability guarantees - stability guarantees for a known nominal system - controlled by a learning-based neural network controller. The neural network controller is trained using data from an existing baseline controller that achieves desirable closed-loop performance which might, however, not provide provable properties such as stability. Examples of possible applications are human-driver-data-based controllers for autonomous driving, or the learning of control strategies for chemical plants based on the control actions of human operators. To provide stability guarantees for the learning-based controller, the controller is reformulated in form of diagonal nonlinear differential form. This representation exploits the fact that the neural network activation functions are sector-bounded and that their slopes are globally bounded. Based on this representation, sufficient closed-loop stability conditions are established in form of Linear Matrix Inequalities for the nominal system, as well as for the disturbed system controlled by the learning-based controller. For nonlinear activation functions that do not satisfy the necessary conditions, a loop transformation is outlined that allows the application of the presented stability certificate. |
Year | DOI | Venue |
---|---|---|
2021 | 10.23919/ACC50511.2021.9482637 | 2021 AMERICAN CONTROL CONFERENCE (ACC) |
Keywords | DocType | ISSN |
Neural network-based control, deep learning, trustable AI, stability analysis | Conference | 0743-1619 |
Citations | PageRank | References |
0 | 0.34 | 0 |
Authors | ||
6 |
Name | Order | Citations | PageRank |
---|---|---|---|
Hoang Hai Nguyen | 1 | 0 | 0.68 |
Tim Zieger | 2 | 0 | 0.68 |
Sandra C. Wells | 3 | 0 | 0.34 |
Anastasia Nikolakopoulou | 4 | 0 | 2.03 |
Richard D. Braatz | 5 | 417 | 108.65 |
Rolf Findeisen | 6 | 0 | 0.68 |