Development of a method for structural optimization of a neural network based on the criterion of resource utilization efficiency

At present, mathematical models in the form of artificial neural networks (ANNs) are widely used to solve problems on approximation. Application of this technology involves a two-stage approach that implies determining the structure for a model of ANN and the implementation of its training. Completion of the learning process makes it possible to derive a result of the approximation whose accuracy is defined by the complexity of ANN structure. In other words, increasing the ANN complexity allows obtaining a more precise result of training. In this case, obtaining the model of ANN that implements approximation at the assigned accuracy is defined as the process of optimization. However, an increase in the ANN complexity leads not only to the improved accuracy, but prolongs the time of computation as well. Thus, the indicator «assigned accuracy» cannot be used in the problems on determining the optimum neural network architecture. This relates to that the result of the model structure selection and the process of its training, based on the required accuracy of approximation, might be obtained over a period of time unacceptable for the user. To solve the task on structural identification of a neural network, the approach is used in which the model’s configuration is determined based on a criterion of efficiency. The process of implementation of the constructed method implies adjusting a time factor related to solving the problem and the accuracy of approximation. The proposed approach makes it possible to substantiate the principle of choosing the structure and parameters of a neural network based on the maximum value for the indicator of effective use of resources