Reinforcement Learning-Based Predefined-Time Tracking Control for Input-Saturated Nonlinear Systems With Performance Guarantees

• Published in: IEEE transactions on automation science and engineering Vol. 22; pp. 19876 - 19888
• Main Authors: Zhao, Wei, Wang, Jing, Lv, Shaoyu, Su, Lei, Shen, Hao
• Format: Journal Article
• Language: English
• Published: IEEE 2025
• Subjects: Automation; Control systems; Convergence; Costs; input saturation; Nonlinear systems; Performance analysis; Predefined-time control; reinforcement learning; Simulation; Stability criteria; Thermal stability; Training
• ISSN: 1545-5955; 1558-3783
• DOI: 10.1109/TASE.2025.3598545

This paper addresses the predefined-time tracking control problem for nonlinear systems subject to predetermined performance metrics and input saturation. With the assistance of the reinforcement learning method, the optimization control objectives of nonlinear systems are achieved. Therein, the execution costs and control behaviors are trained effectively under the actor-critic neural network framework. In addition, the unknown nonlinearities and external disturbances are delicately compensated through the neural approximation rule and an adaptive updated law, respectively. To meet the performance requirements, a performance function is designated to acquire the satisfaction of performance metrics. By incorporating an auxiliary control signal, the saturation constraint on control input is satisfied. In particular, under the action of the developed scheme, the boundness of each system state can be ensured in a predefined time interval. Finally, the effectiveness and superiority of the developed scheme are demonstrated via simulation results. Note to Practitioners-This paper aims to achieve the optimized predefined-time tracking control for nonlinear systems with input saturation and performance guarantees. With the improvement of automation, predefined time convergence becomes one significant performance index in many time-critical practical projects, such as teleoperation systems, missile guidance systems, and so on. As a matter of fact, since actuator saturation and external disturbances exist extensively in practice, they constantly impair the operation performance of engineering systems. To deal with the aforementioned challenges, a reinforcement learning-based predefined-time anti-saturation controller is designed in this paper, which ensures that the system converges within a user-specific time while minimizing the cost. Preliminary simulation comparisons demonstrate that this developed scheme applies to practical servo systems with satisfactory control results.