Start of funding 01.07.2020

Safety Assurances for Control based on Bayesian Neural Networks

Prof. Dr. Sandra Hirche
Technische Universität München
Informationoriented Control (ITR)

Alexandre Capone
Technische Universität München
Informationoriented Control (ITR)

Prof. Dr. Claire Tomlin
University of California, Berkeley
EECS - Electrical Engineering and Computer Science



Machine learning inspired data-driven approaches have already led to very promising results in the control area in the last years. Especially, neural networks require only a minimal prior knowledge for the modeling of complex dynamics. However, the major drawback of neural networks manifests as unpredictable outcomes and the absence of guarantees about the stability of the control loop and performance limitations, which is translated as compromised safety in control systems. Therefore, the current application of neural network approaches in control is often limited to non-critical and/or low performance systems.

The aim of this research project is to establish safety in control of complex systems. The project makes use of Bayesian neural networks (BNNs) to derive stability and performance guarantees. For this purpose, concepts for learning with respect to the physical structure of the systems, exploiting the Bayesian methodology, will be developed. In particular, the project focuses on the identification of theoretical properties of BNNs and the integration of physically intelligent control structures. As a result, we will develop data-driven based control approaches with safety assurances.