From Circuits to Neurofuzzy Networks: Synthesis by Numerical and Linguistic Information

Neurofuzzy networks allow to directly manipulate the information concerning the realization of a required input-output mapping. They constitute an important sector of the emerging field of Intelligent Signal Processing. In the paper a comparison is preliminarily carried out between the neurofuzzy and the circuit approaches, taking account that traditionally the latter plays the same role as the former with respect to signal processing. In the case of neurofuzzy networks the mapping of interest is described by numerical examples and by linguistic sentences regarding its properties, as given by experts on the basis of their experience. This information is manipulated by neurofuzzy networks on the basis of fuzzy logic. After a short survey of the basic ingredients of a neurofuzzy network, two representative architectures and several synthesis procedures are proposed. Traditional synthesis methods cannot be applied for pursuing numerical information and are consequently replaced by clustering algorithms. The linguistic information, on the contrary, can be directly incorporated in the network architecture, as it is given by the experts. As a consequence, the neurofuzzy networks partially mimic humans in facing the task to be accomplished. Detailed examples are presented for illustrating the proposed architectures and synthesis procedures.