On neural approximation of fuzzy systems

It is noted that, by utilizing the function approximation capability, a dynamical system which is represented by fuzzy logic can be approximated by a feedforward sigmoidal network. An example of such approximation is described. The approximate neural system is obtained by simulating a manifold of the input-output product space of the fuzzy system, where the network parameters are adjusted by a numerical optimization algorithm. It is shown that, where such an approximation takes place, the resultant neural system and the fuzzy system share similar dynamical characteristics in which an input space is transformed into an output space by clustering the input space and interpolating the regions between the clusters. The issue of network intermediate nodes necessary to approximate a given fuzzy system is discussed