An Approach to Build and Verify Stable Neural Network Controllers for Cyber Physical Systems with Non-Linear Dynamics
Conference
He, X. (2023). An Approach to Build and Verify Stable Neural Network Controllers for Cyber Physical Systems with Non-Linear Dynamics
. 638-649. 10.1109/QRS60937.2023.00068
He, X. (2023). An Approach to Build and Verify Stable Neural Network Controllers for Cyber Physical Systems with Non-Linear Dynamics
. 638-649. 10.1109/QRS60937.2023.00068
Cyber-physical systems (CPS) have become increasingly important in the functioning of our society. In recent years, machine learning (ML) approaches have started to become an attractive choice to design CPS controllers for better performance and adaptability. Assuring the correctness of this type of controller is extremely difficulty and remains a grand research challenge. Stability is a critical correctness property of any CPS. Designing stable conventional controllers for linear system dynamics is well-understood based on the control theory from the past half century. Designing stable deep neural net (DNN) controllers for linear system dynamics is new while designing stable DNN controllers for non-linear system dynamics is a major research challenge. Although demonstrating the stability of a controller using simulation is a widely accepted practice, this approach does not provide required assurance for CPS involved in safety and mission-critical applications. A provable stability region of attraction (RoA) provides a strong assurance guarantee. In this paper, we have developed an approach to build and verify stable DNN controllers for nonlinear systems, which includes using deep reinforcement learning to design and train a DNN controller based on nonlinear system dynamics, approximating the non-linear system dynamics with several well-known linearization techniques, and leveraging a recent method in deriving the RoA on linearized system dynamics. We have applied this approach to one well-known benchmark system with non-linear dynamics and have obtained their approximated stability RoAs based on several linearization techniques.
