Towards Automated Controller Parameter Design in Cyber-Physical Systems: Improving Computational Cost

Fully-automated optimal design of Cyber-Physical Systems is a challenging task that involves multiple components. In this paper, we are interested in the automated design of one of these components: the system controller. State-of-the-art con-trol parameter automated tuning techniques are optimization-based. However, for high-dimension control parameter spaces, computational costs can be high. We present a multistage controller tuning framework that decomposes controller tuning into sub-tasks, each with a reduced-dimension search space. We show formally that this framework reduces the sample complexity of the control-tuning task. We empirically validate this result by applying a Bayesian optimization approach to tuning multiple PID controllers in an unmanned underwater vehicle benchmark system. We demonstrate an 86 % decrease in computational time and a 36 % decrease in sample complex-ity. Furthermore, the proposed framework highlights existing challenges in fully automated control parameter tuning.