New Methods for Generating Fuzzy Rules from Numerical Data

• Main Authors: Chi-Hao Chang, 張志豪
• Other Authors: Shyi-Ming Chen
• Format: Others
• Language: en_US
• Published: 2001
• Online Access: http://ndltd.ncl.edu.tw/handle/62727720069770369894

碩士 === 國立臺灣科技大學 === 電子工程系 === 89 === The fuzzy classification system is an important application of the fuzzy set theory. Fuzzy classification systems can deal with perceptual uncertainties in classification problems. In order to design a fuzzy classification system, it is an important task to construct the membership function for each attribute and generate fuzzy rules from training instances for handling a specific classification problem. There are two approaches to construct the membership function for each attribute and generate fuzzy rules from training instances. One approach is based on human experts’ assistance, and the other approach is by applying machine learning techniques, such that the fuzzy classification system can construct membership functions and generate fuzzy rules from the training instances automatically. In recent years, many researchers have proposed different methods to construct membership functions and to generate fuzzy rules for handling fuzzy classification problems. However, there are some drawbacks in the existing methods: (1) Some existing methods need human experts to predefine initial membership functions, i.e., these methods can not construct membership functions from the training data set fully automatically. (2) Some existing methods are too complicated and need a lot of computation time. (3) Some existing methods generate too many fuzzy rules. In this thesis, we proposed two methods to construct the membership function for each attribute and to generate fuzzy rules automatically from training instances for handling fuzzy classification problems. The first method is based on the exclusion of attribute terms that can achieve a higher average classification accuracy rate and generate less fuzzy rules than the existing methods. The second method generates weighted fuzzy rules from training instances that can construct membership functions automatically without any human experts’ interaction and can generate less fuzzy rules than the existing methods.