An Aproach of Design and Training of Artificial Neural Networks By Applying Stochastic Search Method

Although vast research works have been paid (throughout some 20 years back) regarding formal synthesis of an ANN it is somehow still open issue. This paper does not consider the mentioned formal synthesis aspects but intends to introduce an original engineering approach offering some advantages whenever training and designing of artificial neurel networks are under consideration. In that sense the author uses powerful combined method of stochastic direct search, the univerzal approximator and simulation method. Author has created specific stochastic serch algorithm wich in some cases having advantages over numerous known methods wich are based on the application of the gradient. The said algorithm incorporates universal approximation and simulation during designing and training neurel networks. The offered approach is applicable for wide range of artificial neural networks types icluding recurrent ones in real time. The presented numerical examples illustrate applicability of the offered advanced approach.