On Integrity Problem of Inference Systems in Adaptive Fuzzy-Neural Computing

This paper explores the learning fuzzy inference systems implemented as adaptive fuzzy-neural networks. The research into application of learning techniques to fuzzy inference systems (FIS) has matured into a family of adaptive fuzzy inference systems (AFIS). In most cases, the learning FIS and AFIS families can be interpreted as a partially connected multilayer feedforward neural network with Gaussian activation function for the hidden neurons. The connection can be interpreted in terms of rules. Often, these rules are designed a priori implying the connections are a priori fixed, and their strengths can be adapted from input and output data. However, the strengths of the rules and membershipfunction parameters are adapted in the learning process from an input-output training data set, such that the error function is minimized. The latter as well as information granulation gave rise to integrity problem which must be observed in applications.