Parameter estimation in neural networks by improved version of simultaneous perturbation stochastic approximation algorithm

This paper describes the parameter estimation and update in neural networks (NN) using a modified version of simultaneous perturbation stochastic approximation (SPSA) algorithm in order to obtain a low computational cost and better performance in the proposed system here. Also, this SPSA is used as learning rule applied to a neuro-controller (NC). In this paper, we apply a direct inverse control scheme by a NN. The NN must learn an inverse system of the objective plant. When using a type of gradient method as a learning rule of the NN, the Jacobian of the plant is required. On the other hand, this control scheme described here does not require any information about the plant Jacobian, because the modified version of SPSA estimates the gradient using only values of the error defined by output of the plant and its desired one. We propose to reduce the oscillation in the single flexible link used as plant in this paper in order to confirm the feasibility of the proposed method.