Physics-Informed Neural Networks-Based Adaptive Optimized Control and Its Application to Automated Surface Vessels

The abundant knowledge of data and physics models can be simultaneously utilized in learning-based modeling, prediction, and control methods, which makes the balance between model efficiency, accuracy, and complexity. Thus, this work investigates the physics-informed neural networks (PINNs)-based adaptive optimized control method with essential learning designs for the whole learning framework. More specifically, the proposed method efficiently realizes the adaptive learning performance with PINNs modeled system dynamics via continuous learning with online data and system physics. Meanwhile, the PINNs model with autodifferentiation is employed by the adaptive dynamic programming approach to iteratively approximate the solution of the continuous-time Hamilton–Jacobi–Bellman equation with neural networks, providing more accuracy and efficiency over approaches that solely utilize either data-driven or physics-based models. As an outcome, the proposed method enables the PINNs-based learning control method to have superior performance in model transfer adaptation and learning efficiency. The proposed method is applied to automated vessel control problems, and its effectiveness and practical applicability are demonstrated through comparative simulations and hardware-in-the-loop tests.