Fuzzy neural network model with three-layered structure

In this paper, a three-layered fuzzy neural network model is developed to execute parallel fuzzy inference with linguistic knowledge representation. Each linguistic variable and its linguistic term set is encapsulated into a single linguistic neuron, which may operate in normal mode or reverse mode. In normal mode, it has the functions of fuzzification and matching degree calculation. In reverse mode, it has the functions of evidence combination, conclusion making and defuzzification. In the three-layered model, the input (premise) layer is composed of a set of linguistic neurons operating in normal mode, while the output (conclusion) layer contains a set of linguistic neurons operating in reverse mode during inferencing but operating in normal mode during learning. Between the input layer and the output layer, a rule layer composed of rule neurons constitutes the truth-value flow channel from input layer to output layer in fuzzy inference. Each rule neuron represents a fuzzy rule. Such a three-layered structure makes a natural representation for fuzzy expert systems, and has faster inferencing and learning speed. This paper further developes a learning algorithm with the advantage of quick convergence. The learning algorithm includes a clustering phase before rule construction, whose results can provide useful information to construct rules by only building necessary links.