A novel method for the design of radial-basis-function-networks and its implication for knowledge extraction

The focus of this work is on the design of radial-basis-function networks for classification in pattern recognition problems, e.g. visual industrial quality control. Based on the review of current design techniques, novel schemes for an incremental network design are presented. Deviating from the scheme of fixed kernel functions, which possess the property of radial symmetry, a novel method is presented to adaptively determine arbitrary kernel functions according to the sample data. The developed methods are demonstrated and verified with problem data from visual industrial quality control. The basis functions can be exploited to determine membership functions of each feature element for every category of the classification problem. The implication of such an approach for feature selection as well as knowledge aquisition is pointed out