Learning function-based classifiers based on rough set and genetic programming

• Main Author: 楊瑞祥
• Other Authors: 錢炳全
• Format: Others
• Language: en_US
• Published: 2003
• Online Access: http://ndltd.ncl.edu.tw/handle/94051120435574304583

碩士 === 義守大學 === 資訊工程學系 === 91 === Classification is one of the important research topics in knowledge discovery and machine learning. Many classifier learning approaches have been proposed including decision trees, Bayesian networks, neural networks and genetic algorithms. A new classifier learning method using genetic programming has been developed for classifying numerical data recently. However, it is difficult for a function-based classifier to classify general nominal data, because nominal data may contain possible large distinct values without order. In this thesis, we present a new scheme based on rough set theory and genetic programming to learn a function-based classifier from the data set containing both nominal and numerical attributes. The proposed scheme first transforms the nominal data into numerical values by rough membership functions. The technique of genetic programming is then used to generate the classification functions. We also develop a new fitness function with adaptable dynamic range intervals for solving the problem of how to specify the ranges of interval for fitness functions in previous researches. The new fitness function can support the design of conflict resolutions and improve the recognition rates of learned classification functions. Several data sets from UCI Machine Learning repository are selected to demonstrate the feasibility and performance of the proposed schemes. The selected data sets are grouped into subsets according to their characteristics. The experiments show that the rough membership function can transforms nominal data into numerical data effectively for genetic programming to learn high accurate function based classifiers. We compare our results with the previous famous methods. The results of comparisons also show that the proposed learning algorithm can obtain higher classification accuracy than the previous methods. Finally, the thesis also show that genetic programming embeds the capability of features selection without using extra mechanisms of features selection of data preprocessing.