Classification of electronic nose data with support vector machines

We investigate a new pattern recognition technique, called support vector machines (SVM), by applying it to the classification of e-nose data. SVM have the advantage of relying on a well-developed theory and have already proved to be successful in a number of practical applications. We analyze the test error of SVM as a function of (a) the number of principal components (on which the data are projected), (b) the kernel parameter value, for both the polynomial and the RBF kernel, and (c) the regularization parameter. This permits to explore the insurgence of underfitting and overfitting effects, which are the principal limitations of non-parametric learning techniques. In particular, we found out that the regularization parameter, often set a priori to C = 1, strongly influences SVM performance. SVM were trained on two electronic nose dataset of different hardness, collected with the Pico electronic nose developed at the Brescia University.