Adaptive data driven control based upon simultaneous perturbation stochastic approximation and neural network

Abstract

A novel data-driven control strategy is proposed based upon the Simultaneous Perturbation Stochastic Approximation (SPSA) method with adaptive weighted gradient estimation for general discrete nonlinear systems. The controller is constructed through use of a Function Approximator (FA), which is fixed as a neural network here. The number of layers and notes in each layer is fixed previously, while the connecting weights, which are then the control parameters, are allowed to be updated. This proposed data-driven control approach just need to take as input past and current system information and produce as output a control value to affect future system performance, and it does not need to establish the mathematical model of controlled plant previously. In this novel approach, the parametric estimation is designed to be adaptive. On the other hand, in order to accelerate the convergence property, a PI compensator is accommodated into the proposed data-driven control strategy. The control ability has been greatly improved and the proposed control strategy is finally applied to solve nonlinear tracking problems for discrete-time nonlinear systems. Simulation comparison tests were conducted on typical non-linear plants, through which, the convergence and feasibility of the proposed data-driven control strategy is well demonstrated.