Data-driven control law optimization via Kriging surrogate model with adaptive domain reconstruction

Abstract

Data-driven control parameter design methods rely on an appropriate initial design domain, which is challenging to define for complex systems with poorly understood dynamics. This reliance creates a dilemma: overly large domains risk instability and high computational costs, while conservative domains may exclude global optimal solutions. To address this issue, a new data-driven control law design method is proposed, combining Kriging surrogate optimization with a dual-mode design domain adaptive reconstruction (DAR) strategy. Taking Active Disturbance Rejection Control (ADRC) as an example, a data-driven Kriging surrogate-based design framework is constructed with control parameters as inputs and control performance index as output. The proposed method dynamically relocates and resizes the search space through stability-constrained boundary adjustments, eliminating dependence on empirical domain settings. Experimental validation on several numerical benchmark problems and two control system applications reveals that the proposed method offers enhanced optimization efficiency and superior global convergence. Its robust adaptability to diverse extreme initial domains effectively lowers the barriers to engineering applications of control law design. This work provides a new reference for future control system design with high-dimensional nonlinear dynamics by bridging the gap between data-driven exploration and deterministic control approaches.