Kullback-Leibler distance optimization for artificial chemo-sensors

A gas sensor optimization method for odor discrimination is introduced in this paper. The method deals with a performance index widely used in the information theory, namely the Kullback-Leibler distance (KL-distance), which gives a quantitative measure of mutual difference between two probability distributions. We argue that optimizing this index over the controllable operating parameter namely ß (i.e., the operating temperature) of a single sensor will allow maximizing the spread of the odor-class prototypes (i.e., the class centers) in the feature space so that a better discrimination of odorants will be possible. We demonstrate on a sample dataset that finely tuning the operating temperature of a metal oxide sensor based on the suggested criterion not only yields a substantial improvement in classification performance but also warns about the existence of temperatures that cause a total confusion in the odor discrimination.