Evolutionary algorithm applied for improving the accuracy of the automated detection of psychedelic amphetamines

Abstract

We are presenting a comparative study regarding the improvement of the correct classification rate of an artificial intelligence application designed to recognize the class identity of psychedelic amphetamines based on the similarity of their ATR-FTIR spectra. For this purpose, the most relevant absorptions were first selected by using a metaheuristic, i.e. a genetic algorithm (GA). The latter is a type of evolutionary algorithm (EA) that mimics the natural selection process and which is recommended especially for improving classification models and optimizing searching procedures. Regression models were built with the original spectral dataset, representing the absorptions measured at 1869 wavenumbers, and with the dataset formed by the absorptions measured at the 187 wavenumbers selected as being the most significant by the genetic algorithm. Both models were tested by using the K-Nearest Neighbors and the Random Forest procedures. Several classification figures of merit were determined and compared for these two cases. The results indicate that the GA wavenumber selection leads to a significant improvement of the classification accuracy.