ABSTRACT
Based on the Projection Gradient Method [2], let us consider the following recurrent procedure:
Notice that this procedure can be implemented only in the case where the information concerning the considered Markov chain is complete. We consider the case where only the realizations of loss function and the state trajectories are available. In this situation, the use of Stochastic Approximation Techniques [3-4) seems to be very appropriate. In fact, stochastic approximation techniques have been used to solve many engineering problems [5], and in a resurgence of interest have been considered as learning algorithms for neural networks and neuro-fuzzy systems synthesis [6-8).
