Study on evolution methods for the optimization of machine learning models based on FT-NIR spectroscopy

Abstract

Least square support vector machine (LSSVM) is a prevalent machine learning method for near infrared calibration. The chemometric models were established by LSSVM for Fourier transform near infrared (FT-NIR) spectral analysis to determine the content of berberine in Coptidis Rhizoma samples. In order to improve the model prediction ability, LSSVM parameters need to be optimized by using some evolution methods. In this paper, genetic algorithm (GA) and differential evolution (DE) were applied for searching the optimal LSSVM parameters. In comparison with grid search, the GA- and DE-optimized models were validated to have appreciate enhancement for the FT-NIR calibrations. The model performance was systematically assessed by the correlation coefficient of test (RT), the root mean square error of test (RMSET), and the relative standard deviation of test (RSDT). For the GA-optimized LSSVM model, the RT, RMSET and RSDT were obtained as 0.900, 0.218 (%) and 9.112, respectively; while for the DE-optimized model, they were observed as 0.904, 0.215 (%) and 8.859, respectively. These results are more preferable than those of the grid searched LSSVM model. Furthermore, the experiments show that DE has fast convergence speed for parameter optimization. Thus we concluded that the evolutional methods (GA and DE) are feasible to optimize the LSSVM parameters in FT-NIR prediction on berberine content. It is expected that GA and DE can be utilized for model improvement in other near infrared applications.