A New Fuzzy Inference System Based on Artificial Neural Network and Its Applications
Jacek M. Łęski, E. Czogała · Studies in fuzziness and soft computing · 1999
In this paper a new artificial neural network based fuzzy inference system (ANNBFIS) has been described. The novelty of the system consists in the moving fuzzy consequent in if-then rules. The location of this fuzzy set is determined by a linear combination of system inputs. This system also automatically generates rules from numerical data. The proposed system operates with Gaussian membership functions in premise part. Parameter estimation has been made by connection of both gradient and least squares methods. For initialization of unknown parameter values of premises, a preliminary fuzzy c-means clustering method has been employed. For cluster validity Xie-Beni, Fukujama-Sugeno and our new indexes have been applied. The applications to the design of a classifier are considered in this paper as well. The method of classifier construction for two classes and an extension for a greater number of classes has been presented. The selection method of target values for classifier outputs minimizing number of false classifications is also presented. The applications to prediction of chaotic time series, pattern recognition and system identification are considered in this paper. The tests of the ANNBFIS are carried out on the basis of the data bases known from literature: Mackey-Glass chaotic time series, Anderson’s iris and MONKS classification problems, Box-Jenkins data from the gas oven.