|
ИСТИНА |
Войти в систему Регистрация |
Интеллектуальная Система Тематического Исследования НАукометрических данных |
||
Feature Selection in Neural Network Solution of Inverse Problem Based on Integration of Optical Spectroscopic Methodsстатья Исследовательская статья
Информация о цитировании статьи получена из
Scopus
Дата последнего поиска статьи во внешних источниках: 24 февраля 2021 г.
- Авторы: Isaev Igor, Sarmanova Olga, Burikov Sergey, Dolenko Tatiana, Laptinskiy Kirill, Dolenko Sergey
- Сборник: Advances in Neural Computation, Machine Learning, and Cognitive Research IV
- Серия: Studies in Computational Intelligence
- Том: 925
- Год издания: 2021
- Первая страница: 234
- Последняя страница: 241
- DOI: 10.1007/978-3-030-60577-3_27
- Аннотация: This study considers a neural network solution to the inverse problem based on the integration of optical spectroscopy methods for determining ion concentrations in aqueous solutions. The effect of integration of physical methods is studied using the selection of significant input features. The previously formulated thesis is confirmed that if the integrated methods differ much by their accuracy, then the integration of these methods is ineffective. Use of joint selection of significant input features may only slightly improve the quality of solution in a limited number of situations. Among the tested methods of feature selection, embedded method based on the analysis of weights of an already trained neural network (multi-layer perceptron) gives better results than filter methods with selection based on standard deviation, cross-correlation, or cross-entropy, at the expense of some additional computation. Selection of the significant input features allows improving the quality of the ANN solution relative to the solution obtained on the full sets of features. The improvement is the greater, the worse is the initial solution.
- Добавил в систему: Доленко Сергей Анатольевич