关键词: 同时扰动随机逼近, 数据驱动控制, 梯度逼近, 单边SPSA, 数据利用, 非线性系统

Abstract:

Using simultaneous-perturbation stochastic approximation ( SPSA) algorithm to establish a data-driven control method is feasible when the mathematical models of controlled systems are totally unknown. In this paper,the principle and effectiveness of both the SPSA algorithm and its improved algorithms are investigated,and the shortcomings of the SPSA algorithm in the control-oriented situation,such as lower gradient approximation efficiency,are analyzed. Then,a one-sided SPSA algorithm is proposed for the control problem,which saves one-third data
compared with the SPSA algorithm in the actual control,thus obtaining enough convergence rate and accuracy of the optimization procedure for the control problem. Then,by combining the one-sided SPSA algorithm and the datadriven control method,a data utilization scheme is designed based on dynamic deviations,and the connection weights in a neural network function approximator are adjusted directly and adaptively. Simulatied results show that the proposed data-driven control method secures excellent control over nonlinear systems without the help of any model.

Key words: simultaneous-perturbation stochastic approximation, data-driven control, gradient approximation, onesided SPSA, data utilization, nonlinear systems