An approach to Synthesis and Analysis of Complex Recurrent Neural Network

This paper presents an approach to optimization of recurrent artificial neural networks (RNN) that leans on the appliance of stochastic search (SS). Favor algorithm SS with information accumulation (SSAI) is simple in numerical sense, and does not require a lot of computing time in the optimization process i.e. RNN training, and gives suboptimal results in comparison to gradient methods. In certain sense, suggested approach more appropriate for engineering practice than back propagation error (BPE) method, because it does not condition the differentiability of activation neuron function, as well as transformation of RNN in corresponding multi-layered network with forward propagation signal, and after that gave the problem with a great deal of dimensions. Behind the corresponding theoretical analysis, SSAI is applied on optimization of structure and RNN parameters (supervised learning algorithm), for creation of predictive model which serves for content of useful component in input raw material in technological process of flotation in real time.