Frequency selection using genetic algorithms
Sandia National Labs., Albuquerque, NM (United States), Edward V. Thomas, USDOE · 1993
Modern spectroscopic instruments are used to make optical measurements of various materials at a number of frequencies in one or more regions of the optical spectrum.Coupled with empirically determined calibration models, these measurements allow for quantitative analysis of various material properties over a wide range of applications.Calibration methods capable of using a large number of frequencies, such as partial least squares regression (PLS), are often used to construct these calibration models.Because spectral measurements are so easy to obtain, the number of potential spectral frequencies to use is often very large, perhaps thou~nds.However, in many applications, measurements from a large number of the available spectral frequencies are irrelevant or difficult to model.Use of the measuremen-ts from these irrelevant (or difficult) spectral frequencies can have significant negative consequences on calibration method performance (measured by the ability to predict the chemical/physical property of interest), so it is important to select an appropriate subset.Often, due to the complexity of materials to be analyzed, subject-matter knowledge is insufficient to select an appropriate subset.Genetic algorithms have been proposed as a mechanism for performing frequency selection (see Kateman andLucasius 1991 andLi et al. 1992).Unlike traditional methods, frequency selection based on genetic algorithms can utilize the synergy among measurements from different frequencies.This paper investigates some of the statistical and computational issues associated with using genetic algorithms for frequency selection.Based on an analysis of these issues, a specific strategy for selecting frequencies is proposed.The proposed strategy is illustrated by using near-infrared spectral data to estimate the concentration of glucose in multicomponent aqueous solutions.Patent Ca•tion -Thu doctament maJ reveal patenta6le n6jed matter.The informatioa mut not 6e dinlged o•uide Sandia National Ldoratoriu toitho•t the approval of the Pateat aad