ИССЛЕДОВАНИЕ ПРИМЕНЕНИЯ АНСАМБЛЕЙ НЕЙРОННЫХ СЕТЕЙ ДЛЯ ПОВЫШЕНИЯ КАЧЕСТВА РЕШЕНИЯ ЗАДАЧ РЕГРЕССИИ

Abstract

Artificial neural networks (ANN) are one of the information and statistical methods of analysis. ANN models base on the principle of black box.Itmeans that the search of functions describing the dependence of the output variables from the input is going during running of the available data sample. Modeled functions depend on the structure of the neural network, i.e. the number of input and output variables, the presence and number of hidden layers, type of activation function, etc. Theoretically proved that there is not any universal selection algorithm of a neural network structure and it can`t be proved that one structure better than another one. This creates uncertainty of choice. Another problem associated with training on a black box is theoverfitting, i.e. precise adjustment of the modeled function to the available data. As the result suchnetwork loses predictive ability. Thus, in addition to the correct formulation of objectives, the choice and normalization of data for network training,the task solving using neural networks has difficulty of choosing the most optimal network structure specifically for the task and control of network training to prevent it from overfitting. The perspective way to improve the ANN method is combining of elementary neural networks set into a single system that can allow solve rather difficult forecast tasks at asatisfactory level. Neural networkscombined into the system (ensemble) theoretically have higher predictive ability. This article presents the experiment results of usingneural networks ensembles in order to improve the quality of solving regression taskson example of the task of bottom-hole acidizing (BHA) forecasting efficiency (increases of oil production and liquid flow rate) at the productive reservoir JH 1 of Urevskoe oil field. The conclusion based on these results about the effectiveness of neural networks ensembles for predicting in comparison with single network with composite outputs has been made.