An Efficient Multistep Stochastic Approximation Algorithm

This paper presents an efficient and easily implemented stochastic approximation algorithm. The procedure is based on the reprocessing of input data for several steps (iterations) of the algorithm and is especially suited to the case where the input data are expensive to obtain. We show that this multistep algorithm has the usual a.s. convergence property of the standard Robbins-Monro algorithm. We also present some preliminary results on choosing the optimal number of steps for use in such data reprocessing and ilustrate the results with two numerical studies.