An improved neural network for fuzzy reasoning implementation

Neural networks or connectionist models are massively parallel interconnections of simple neurons that work as a collective system, can emulate human performance and provide high computation rates. On the other hand, fuzzy systems are capable to model uncertain or ambiguous situations that are so often encountered in real life. One way for implementing fuzzy systems is through utilizations of the expert system architecture. Recently, many attempts have been made to “fuse” fuzzy systems and neural nets in order to achieve better performance in reasoning and decision making processes. The systems that result from such a fusion are called neuro-fuzzy inference systems and possess combined features. The purpose of the present paper is to propose such a neuro-fuzzy system by extending and improving the system of Keller et al. (1992). The present system makes use of Hamacher's fuzzy intersection function and Sugeno's complement function. After a brief outline of the operation of the system its features are established with the aid of four theorems which are fully proved. The capabilities of the system are shown by a set of simulation results derived for the case of trapezoidal fuzzy sets. These results are shown to be better than the ones obtained with the original neuro-fuzzy system of Keller et al.