Safety-Critical Randomized Event-Triggered Learning of Gaussian Process With Applications to Data-Driven Predictive Control

• Published in: IEEE transactions on automatic control Vol. 70; no. 6; pp. 3920 - 3935
• Main Authors: Zheng, Kaikai, Shi, Dawei, Shi, Yang
• Format: Journal Article
• Language: English
• Published: New York IEEE 01.06.2025; The Institute of Electrical and Electronics Engineers, Inc. (IEEE)
• Subjects: Accuracy; Adaptation models; Algorithms; Computational modeling; Confidence intervals; Control systems; Data models; Data-driven control; Dynamical systems; Event detection; event-triggered learning; Gaussian process; Machine learning; Nonlinear control; Nonlinear systems; Prediction models; Predictive control; Predictive models; randomized learning; Safety; Safety critical; safety-critical system; Simulation; Statistical analysis; Training; Trajectory; Uncertainty
• ISSN: 0018-9286; 1558-2523
• DOI: 10.1109/TAC.2024.3523682

Safety and data efficiency are important concerns in data-driven control, especially for nonlinear systems with unknown dynamics and subject to disturbances. In this work, we consider a class of control-affine nonlinear systems with partially unknown dynamics and aim to introduce an event-triggered learning-based control approach with guaranteed safety and improved data utilization efficiency. Specifically, a randomized learning approach is employed to evaluate the safety of state trajectories by defining and estimating its confidence interval, with data from a multisample of randomly generated state trajectories. Using the proposed randomized learning algorithm, a nominal trajectory with a high probability safety guarantee is designed, thus ensuring the disturbed system states to remain within a prespecified range around the nominal trajectory with a sufficiently high probability. Through removing irrelevant data, a local prediction model around the nominal trajectory is learned with satisfactory precision, and is updated online using an event-triggered learning strategy. Based on the learned model, an efficient data-driven predictive controller is designed to force the system states to evolve within the vicinity of the designed safety nominal trajectory. The effectiveness of the proposed event-triggered learning and data-driven control approaches is validated through comprehensive simulation studies.