Significant Feature Selection in Neural Network Solution of an Inverse Problem in Spectroscopy*

Abstract

Abstract Raman spectra are a valuable source of information about the studied object widely used in laser spectroscopy of liquids. Estimating the values of object parameters from Raman spectra is an incorrect inverse problem with high input dimensionality, which may be successfully solved by artificial neural networks. One of the ways to reduce this dimensionality is selection of significant input features, which can both reduce the error of parameter determination and bring some information about relationship of the determined parameters and the inputs of the problem. In this study, various methods of significant feature selection are considered for the problem of identification and determination of ionic composition of a multi-component water solution of inorganic salts.