A Randomized Stochastic Approximation Algorithm for Self-Learning

A new stochastic approximation algorithm with input perturbation for self-learning is designed with test perturbations and has certain useful properties, such as consistency of estimates under almost arbitrary perturbations and preservation of simplicity and performance with the growing size of the state space and increasing number of classes. An example on computer-aided modeling of learning is given to illustrate the performance of the algorithm.__________Translated from Avtomatika i Telemekhanika, No. 8, 2005, pp. 52–63.Original Russian Text Copyright © 2005 by Granichin, Izmakova.