Comparison of network-based inference mechanisms for fuzzy logic

The authors consider criteria on which to compare neural network structures for fuzzy logic inference associated more with expert systems than with control situations. They examine three types of fuzzy inference neural networks with respect to these criteria. The networks include first the standard feed-forward multilayer perceptron. The inference process can be viewed as a form of functional approximation, and therefore feed-forward neural networks can be utilized. Since inference deals with fuzzy sets, fixed architecture networks whose nodes implement particular fuzzy set theoretic connectives and whose weights are hand-crafted have been introduced to provide a firm theoretical framework in which to cast the inference activity. Finally, trainable evidence aggregation networks whose nodes compute parametrized families of fuzzy set theoretic operators have recently been proposed to combine the best parts of the former approaches. These various neural network-like models for fuzzy logic inference are compared