Structure adaptation of stochastic neural nets using learning automata technique

Abstract

Abstract The selection of a number of nodes in artificial neural nets containing stochastic noise perturbations in the outputs and inputs of each node is examined. The suggested approach is based on a reinforcement learning technique. To solve this optimization problem we introduce a special performance index in such a way that the best number of nodes corresponds to the minimum point of the suggested criterion. This criterion presents a linear combination of a residual minimization functional and some ‘generalized variance’ of the involved disturbances of random nature. A large value of the noise variance leads to a different optimal number of neurons in a neural networks because of the ‘interference’ effect. The optimal point is obtained by the learning procedure based on the Bush-Mosteller reinforcement scheme. This numerical method is commonly used in Intelligent Control Theory. Simulation modelling results are presented to illustrate the effectiveness of the suggested approach