SYNTHESIS AND USAGE OF NEURAL NETWORK MODELS WITH PROBABILISTIC STRUCTURE CODING

Abstract

Context. The problem of encoding information of models based on artificial neural networks for further transmission and use of such models is considered. The object of research is the process of coding artificial neural networks using probabilistic data structures. Objective of this work is to develop a method for coding neural networks to reduce the resource intensity of the process of neuroevolutionary model synthesis. Method. A method for encoding neural networks based on probabilistic data structures is proposed. At the beginning, the method uses the basic principles of the approach of direct encoding of network information and, based on sequencing, encodes a matrix of interneuronal connections in the form of biopolymers. Then, probabilistic data structures are used to represent the original matrix more compactly. For this purpose, hash functions are used, the initial matrix goes through the hashing process, which significantly reduces the requirements for memory resources. The method allows to reduce memory costs when sending artificial neural networks, which significantly expands the practical use of such models, preventing a sharp decrease in the accuracy of their operation. Results. The developed method is implemented and investigated in solving the problem of classification of the state of South German creditors. The use of the developed method allowed increasing the rate of neuromodel synthesis by 15–17.6%, depending on the computing resources used. The method also reduced the share of information transfers by 8%, which also indicates faster and more efficient use of resources. Conclusions. The conducted experiments confirmed the efficiency of the proposed mathematical software and allow us to recommend it for use in practice, when encoding models based on artificial neural networks, for further solving problems of diagnostics, forecasting, evaluation and pattern recognition. Prospects for further research may consist in pre-processing data for more strict control of the encoding process in order to minimize the loss of quality of models based on neural networks.