Interpolation in Som: Improved Generalization by Iterative Methods

Improved Generalization by Iterative Methods Josef G oppert, Wolfgang Rosenstiel University of T ubingen, Sand 13, D-72076 T ubingen Abstract. Interpolation and self-organizing maps have proved to be powerful in several applications. These methods are able to combine the topology preservation of the self-organizing map (SOM) and real valued data evaluation as required in sensoring and monitoring. The basic idea of the kind of interpolation in this paper is the extraction of geometric information in the input space in order to use these relations in the output space of a counter-propagation network. In this paper some important investigations about the in uence of the interpolation parameters will be presented and di erent strategies for nding these parameters will be compared and improved. Di erent types of data are used to show empirical results.