Optimization of a Recurrent Neural Network Based on the Structure of Stochastic Automata

The neural network approach is an effective method for solving problems of diagnostics, pattern recognition, classification and clustering of patterns, as well as for solving forecasting problems. In cases where the input data are members of a time sequence, or other context-oriented information in which the order of the elements is important, in this case it is appropriate to use recurrent neural networks. The presence of information about the context, or local environment, allows such networks to quickly adapt to control tasks when the receptor state vector is input, and a control command is formed at the output, an indirect assessment of the effectiveness of which will be fed to the input of such a network along with data on the state of the receptors at the output. next step of operation. Another method for solving such a problem is certain designs of automata, characterized by expedient behavior. In this paper, we propose a method for constructing a recurrent neural network based on the characteristics of the corresponding probabilistic automaton with a variable structure. The proposed method allows you to optimize the set of connections of the original recurrent neural network and the required number of neurons by removing both redundant neurons and redundant connections between the remaining neurons.