In this work we analyze the optimization of tapped delay support vector machines (TD-SVMs) for analyzing quantitative e-nose data. Here, an array of nanostructured and polymer based sensors is exposed to several NO2-NH3-RH mixtures in order to built a suitable data set for testing its real time concentration estimation capabilities. TD-SVM performance depends on both SVM and TD lines parameters. The partial knowledge about their mutual relationships and availability of a GRID infrastructure made a brute force approach on performance optimization feasible. Results indicate that while it is not advisable to optimize SVM and TD lines parameters separately, for this problem a region of quasi optimality is detectable for SVM parameters.
展开▼
