Özet:

Abstract The problem of training sample forming is solved to automate the construction of neural network models on precedents. The sampling method is proposed. It automatically selects the training and test samples from the original sample without the need for downloading the entire original sample to the computer memory. It processes an initial sample for each one instance with hashing transformation to a onedimensional axis, forming cluster templates on the generalized axis, minimizing their number. This allows to increase the speed of sampling, to reduce the requirements to computing resources and to computer memory and to provide an acceptable level of accuracy of the synthesized models. The developed method does not require multiple passes through the sample, being limited by only three viewing. At the same time the method keeps in a random access memory only the current instance and the generated set of one-dimensional templates, which is minimized by volume. Unlike the methods based on random sampling and cluster analysis the proposed method automatically determines the size of the formed training and test samples without the need for human intervention. Software realizing proposed method is developed. On its basis the practical task of decision-making model building to predict the individual state of the patient with hypertension is resolved. References Субботін С. О. Нейронні мережі : навчальний посібник / С. О. Субботін, А. О. Олійник ; під заг. ред. проф. С. О. Субботіна. – Запоріжжя : ЗНТУ, 2014. – 132 с. 2. Computational intelligence: a methodological introduction / [R. Kruse, C. Borgelt, F.Klawonn et. al.]. – London : Springer-Verlag, 2013. – 488 p. DOI: 10.1007/978-1-4471-5013-8_1 3. Олешко Д. Н. Построение качественной обучающей выборки для прогнозирующих нейросетевых моделей / Д. Н. Олешко, В. А. Крисилов, А. А. Блажко // Штучний інтелект. – 2004. – № 3. – С. 567–573. 4. Subbotin S. A. The training set quality measures for neural network learning / S. A. Subbotin // Optical memory and neural networks (information optics). – 2010. – Vol. 19. – № 2. – P. 126–139. DOI: 10.3103/s1060992x10020037 5. Субботин С. А. Критерии индивидуальной информативности и методы отбора экземпляров для построения диагностических и распознающих моделей / С.А. Субботин // Біоніка інтелекту. – 2010. – № 1. – С. 38–42. 6. Encyclopedia of survey research methods / ed. P. J. Lavrakas. – Thousand Oaks: Sage Publications, 2008. – Vol. 1–2. – 968 p. DOI: 10.1108/09504121011011879 7. Hansen M.H. Sample survey methods and theory / M. H. Hansen, W. N. Hurtz, W. G. Madow. – Vol. 1 : Methods and applications. – New York: John Wiley & Sons, 1953. – 638 p. 8. Кокрен У. Методы выборочного исследования / У. Кокрен ; пер. с англ. И. М. Сонина ; под ред. А. Г. Волкова, Н. К. Дружинина. – М. : Статистика, 1976. – 440 с. 9. Multivariate analysis, design of experiments, and survey sampling / ed. S. Ghosh. – New York: Marcel Dekker Inc., 1999. – 698 p. 10. Smith G. A deterministic approach to partitioning neural network training data for the classification problem : dissertation ... doctor of philosophy in business / Smith Gregory. – Blacksburg: Virginia Polytechnic Institute & State University, 2006. – 110 p.

Anahtar Kelimeler: