Two alternative schemes to update SVM approximations for the identification of explicit decision functions

This paper presents a new adaptive sampling technique for the construction of explicit decision functions using support vector machine (SVM). Two approaches are used for the adaptive selection of the training samples. The first approach involves the construction of a single initial decision function using a limited number of samples, which is then up- dated by adding subsequent samples on the decision function by maximizing the minimum distance. The second approach involves the generation of competing decision functions using di®erent SVM parameters. The difference between the competing approximations provides an idea about the regions of design space with possible need for improvement. Several examples are presented to show the construction of decision functions using the proposed methods. The update schemes are validated by comparing the predicted explicit functions to actual analytical decision functions. Also, the results obtained using the two methods are compared to each other.