Nonparametric Tuning of Fuzzy Modeling by Neural Network Technology

• Main Authors: Chen, Huang-Chi, 陳煌琦
• Other Authors: Hwang, Rey-Chue
• Format: Others
• Language: zh-TW
• Published: 1998
• Online Access: http://ndltd.ncl.edu.tw/handle/29699213095666205029

碩士 === 義守大學 === 電機工程研究所 === 86 === 　　In this thesis, a non-paramertic fuzzy model tuning technique by using poylnomial neural network is presented. The associations between input and output of fuzzy set will be clearly described based on the adapiive capability of neural network. 　　Fuzzy theory has been applied into a wide variety of engineering and business areas. Fuzzifier, rule base and defuzzifier are three main cores in fuzzy mechanism Rule base could be built generally by designer's expertise based on the characteristics of the system approached. It consists of a set of fuzzy IF-THEN rules. It is not very difficult to construct such a rough fuzzy rule base, if the designer has a great mass of professional knowledge about the system needed to proceed. In fact, a good rule base designed by an expert still can not cover all possible situations that may happen in the system. Therefore, the original designed fuzzy rule base may need to be modified or extended as a higher accuracy of the fuzzy system performance is desired. 　　In the whole fuzzy system, membership function plays an important role in fuzzifier and defuzzifier. The performance of the fuzzy system is affected significantly by that the designed membership function is appropriate of not, On typical fuzzy models, the antecedent and consequent fuzzy sets are characterized by monotonic membership functions, such as bell shape function, triangular shape function and trapezoid shape function. An interesting question appears immediately.Do these functions effectively express the relationships between fuzzy sets of the utilized fuzzy model? Especially, the fuzzy model is used for a complex, nonlinear system. In order to improve the performance of a fuzzy model, several methods based on neural network have been proposed for tuning parameters of fuzzy model. Most of these methods were tuning the specified parameters of both antecedent and consequent terms. 　　In this study, a roughly fuzzy model is constructed by using traditional procedure with a prior expert knowledge. Each step of fuzzy model constructing is simplifed as possible as we can. The monotonic membership functions are still utilized in the beginning of the rough model we developed. Of course, the designed fuzzy model has to be made sure that it can work at least for the system required to be proceeded. We assume that the complex and non-monotonic relationships may exits between the input and output variables of fuzzy sets. Then, each membership function of antecedent terms of fuzzy model is replacey by an individual neural network that has completely learned the characteristics of corresponding function, respectively. Fuzzy rules are unchanged in the entire process of experiments. During the operation period of fuzzy model, the input/output relationships of membership function of fuzzy sets will be adjusted based on error of the system output. The finally real mapping information of input/output of fuzzy sets will be stored in the weights of neural network after the whole process of experiment is finished. Then, the real shape of each membership function of antecedent terms can be recalculated and redrawn by these trained neural networks. The specified parameters of consequent terms are also tuned on-line based on error signal. 　　Form the experimental results, we can conclude that the technique we proposed is successful. System performance has a significant improvement on modeling cases. The developed technique can also be easily used to another area applications. It has a great potential for fuzzy application in very complex system.