Neural Network based Construction ofFuzzy

Neural Network based Construction of Fuzzy Graphs Michael R. Berthold and Klaus{Peter Huber Institute for Computer Design and Fault Tolerance (Prof. D. Schmid), University of Karlsruhe, Germany Abstract| Function approximation using example data has gained considerable interest in the past. One interesting application is the approximationof the behaviour of simulationmodels, called metamodelling. The goal is to approximate the behaviour as well as to extract some understandable knowledge about the simulation model. In this paper a combination of a special type of Neural Network (Rectangular Basis Function Network) with a (de{)fuzzication module is used. The resulting system approximates real valued functions with an adjustable precision. A constructive algorithm builds the network from scratch, resulting in a structure where each hidden unit represents a rectangular area with a corresponding membership function (or a fuzzy point). The underlying knowledge can be extracted from the network in form of a Fuzzy Graph.