抄録
The Robbins-Monro stochastic approximation procedure is a useful method for finding the root θ of the regression function f(x) = α recursively based on the noisy observations Y(x) of f(x). This procedure updates the estimate based on only the information of the current estimate, and hence the convergence to θ is generally slow. In this paper, a multi-stage stochastic approximation procedure is proposed to accelerate the convergence. The proposed procedure updates the estimate by using the information of the past estimates in addition to current one. This idea is also applied to nonlinear state estimation problem. Numerical examples illustrate its applicability.
