Composition Methods and Learning Algorithms of Fuzzy Neural Networks(Journal of Japan Society for Fuzzy Theory and Systems)

概要

論文の詳細を見る
There have been many researches on application of neural networks to fuzzy inference. The authors call the neural networks Fuzzy Neural Networks (FNNs). The authors have proposed two kinds of FNNs. The FNNS realize fuzzy reasoning whose consequences are expressed by constants/first-order linear equations. These FNNs can automatically identify the fuzzy rules and tune the membership functions. But the FNNs have some problems in the identified fuzzy rules and the setting of the learning rates of their connection weights. This paper presents new composition methods of the FNNs using the center of gravity method with normalizing units in the premise parts. Using the center of gravity method, the fuzzy rules which express exactly the input-output relationships of the networks and the characteristics of a system can be identified. This paper also proposes a new FNN of which consequences are expressed by fuzzy variables, and a method for making the setting of the learning rates easier. The capability of the FNNs are examined using simple numerical data. The membership functions in the premises of the FNNs have a distinguishing feature in the learning and they can be tuned appropriately using back-propagation algorithm. The identified fuzzy rules by the FNNs express the characteristics of the system well.
日本知能情報ファジィ学会の論文
1992-10-15