Performance-Driven Safe Bayesian Optimization for Intelligent Tuning of High-Order Cascade Systems

Automatic parameter tuning of high-order cascade controllers suffers from sampling inefficiency and strong couplings. This work presents a performance-driven, systematic, and safe intelligent parameter-tuning framework for high-order cascade systems. To achieve data-efficient and noise-robust hyperparameter calibration, an intelligent tuning framework based on Bayesian optimization is proposed to calibrate the control parameters from the buffer of performance-metric measurements. Furthermore, we improve the Bayesian-optimization-based framework in three aspects, involving effectiveness, security, and sampling efficiency. First, a comprehensive control performance assessment combining the error-integral and statistical performance criteria is designed to evaluate the cost of a sampling point in terms of final precision, response rapidity, and vibration. Meanwhile, the security of sampling exploration is heightened by imposing composite hard parametric and soft performance-metric (maximum input, overshoot, etc.) constraints on the acquisition points. Additionally, a hierarchical optimization strategy is proposed to further boost sampling exploitation by alternatively tuning the control parameters of subsystems and refining the constraint on the sampling space. The proposed framework is applied to automatic parameter tuning of high-order dynamic-surface-based backstepping control combined with antidisturbance rejection control. The comparative simulation results demonstrate that the proposed intelligent tuning possesses superiorities in sampling efficiency and security for parameter calibration problems.

I. Introduction

Designing a controller with a better performance relies on the selection of tuning parameters of the controller that may strongly impact the closed-loop stability (precision, robustness, etc.) and transient performance (maximum control input, overshoot, rapidity, etc.). In practice, these parameters are usually selected via trial-and-error experiments or heuristic-based strategies that rely on expensive closed-loop simulations or experiments, which can become prohibitive when system uncertainties have an effect on control performance [1]. Autotuner, as an intelligent system, seeks the optimum parameters of the controller by applying a data-driven performance optimization method [2], [3]. Automatic tuning (or autotuning) relieves the pain of manual tuning and has been successfully applied in numerous practical fields, such as industrial process control [2], [4], flight vehicles [5], and weak grids [6].