Color technology needs specifications to which extent physical differences of stimuli correspond to differences in perception. Generalized linear models (GLMs) have proved successful to provide such specifications from choice-based experiments. However, the use of GLMs imposes practical restrictions on the experiment and stimulus parameters. We propose an alternative analytic approach based on machine learning and demonstrate its use in designing and analyzing choice-based experiments with multiple stimulus dimensions. (c) 2012 Wiley Periodicals, Inc. Col Res Appl, 38, 334-343, 2012