Disturbance observer-based adaptive reinforcement learning for perturbed uncertain surface vessels

• Published in: ISA transactions Vol. 130; pp. 277 - 292
• Main Authors: Tu Vu, Van, Pham, Thanh Loc, Dao, Phuong Nam
• Format: Journal Article
• Language: English
• Published: United States Elsevier Ltd 01.11.2022
• Subjects: Adaptive/approximate reinforcement learning (ARL); Disturbance observer (DO); Lyapunov stability theory; Optimal control; Surface vessels (SVs); Adaptive/approximate reinforcement learning (ARL); Lyapunov stability theory; Optimal control; Surface vessels (SVs); Disturbance observer (DO)
• ISSN: 0019-0578; 1879-2022
• DOI: 10.1016/j.isatra.2022.03.027

This article considers a problem of tracking, convergence of disturbance observer (DO) based optimal control design for uncertain surface vessels (SVs) with external disturbance. The advantage of proposed optimal control using adaptive/approximate reinforcement learning (ARL) is that consideration for whole SVs with only one dynamic equation and without conventional separation technique. Additionally, thanks to appropriate disturbance observer, the attraction region of tracking error is remarkably reduced. On the other hand, the particular case of optimal control problem is presented by directly solving for the purpose of choosing the suitable activation functions of ARL. Furthermore, the proposed ARL based optimal control also deals with non-autonomous property of closed tracking error SV model by considering the equivalent system. Based on the Lyapunov function candidate using optimal function and quadratic form of estimated error of actor/critic weight, the stability and convergence of the closed system are proven. Some examples are given to verify and demonstrate the effectiveness of the new control strategy. [Display omitted] [Display omitted] •The proposed control scheme is implemented by a framework of a DO and a RL based optimal control strategy without conventional separation technique.•The proposed controller develops the simultaneous tuning of RL strategy for guaranteeing the tracking problem of non-autonomous closed systems.•The stability and convergence of learning method are strictly proven from optimality principle and Lyapunov theorem on stability. Moreover, the method of choosing activation functions of RL is approached and verified based on the solution of a special case of optimal control problem.