A new genetic algorithm is presented for sensor selection, in order to predict the octane number and the content of benzene, toluene and total aromatics in gasolines, based on partial least squares multivariate calibration of near infrared spectral data. The algorithm separately labels each of the selected wavenumber ranges with a relative inclusion ranking. Relative prediction errors (% of root mean square error with respect to the mean calibration value) achieved after sensor selection are: octane number, 0.5%, benzene, 0.6%, toluene, 0.6%, and total aromatics, 1.1%. The results are compared with those provided by the spark-ignition engine fuel research method for octane number and by gas chromatography for the remaining parameters.