Data‐driven control based on simultaneous perturbation stochastic approximation with adaptive weighted gradient estimation

Abstract

This study proposes one novel data-driven control strategy based upon the simultaneous perturbation stochastic approximation method with adaptive weighted gradient estimation for general discrete non-linear systems. A function approximator is used to construct the controller, and here, it is fixed as a neural network (NN), whose structure is fixed previously, while allowing its connecting weights to be updated. The control parameters are then the connecting weights of the NN controller. The biggest advantage of this data-driven control approach is that it can generate a control signal to affect system's future performance without establishing the plant's mathematical model first. In this novel approach, to improve the control ability and accuracy, an adaptive weighted gradient estimation method is designed to do the parametric estimation with convergence analysis. Non-linear tracking problems for typical discrete-time non-linear plants are introduced for simulation comparison tests, and the convergence and feasibility of this newly proposed data-driven control strategy are well demonstrated through the simulation results. Finally, empirical study on a simulated wastewater treatment system is carried out to further illustrate the effectiveness of this newly proposed approach.