Adaptive Finite‐Time Optimal Control of Stochastic Nonlinear Systems via an Identifier‐Critic‐Actor Reinforcement Learning Algorithm: Application to Intelligent Ship Autopilot

ABSTRACT In this article, an adaptive finite‐time neural optimal control technique is proposed for stochastic nonlinear systems (SNSs) by applying the reinforcement learning (RL) technique of the identifier‐critic‐actor architecture. Radial basis function neural networks (RBFNNs) are used to estimate the uncertain functions in the considered system. Unlike a deterministic system, SNS control must account for stochastic disturbance in the process of system control. Therefore, the finite‐time update rules for the critic and actor in SNSs are formulated based on the negative gradient of a positive function, which is derived from the partial derivatives of the Hamilton–Jacobi–Bellman (HJB) equation discussed herein. Using the discussed approach, the RL algorithm is simplified, and two preconditions are released: Persistent excitation and known dynamics. The controller design method proposed in this article not only simplifies the control scheme but also solves the singularity problem that occurs in the traditional backstepping optimal control techniques. Stability analysis reveals that the tracking error tends toward a vicinity of zero, while the boundedness of all system signals is ensured within a finite time. Finally, a simulation example involving an intelligent ship autopilot was provided to verify the effectiveness of the theoretical results. Meanwhile, it can also demonstrate the effectiveness of the academic achievements.