Quasi-efficient stochastic approximation on the basis of neural networks
Alexander V. Nazin · 2002
A three-step recursive algorithm of a stochastic approximation type is proposed. The estimates are generated by the Polyak-Ruppert averaging technique with a neural network-based transformation of observations. The algorithm includes a procedure for neural network tuning to approximate the optimal transformation function. Theorems on convergence and asymptotic normality which demonstrate quasi-efficiency of the estimates are formulated.>