In this thesis, fuzzy clustering-based RBFNN pattern classifier is proposed. The objective of this study is focused on the comparison of preprocessing algorithms such as Principal Component Analysis(PCA), Linear Discriminant Analysis(LDA) etc. Because each algorithm extracts the features of data different way respectively. By comparing the result of feature extraction utilizing formulas and figures, the originality of this thesis is shown. In the thesis, two types of data is used. Type 1 is machine learning data and type 2 is spectrum data. Data is acquired from UCI website and Laser-Induced Breakdown Spectroscopy(LIBS) respectively.
Features which are extracted from preprocessing algorithms are used as the inputs of RBFNN pattern classifier. The proposed classifier consists of four parts ? preprocessing part, condition part, conclusion part and inference part. In the condition part of the proposed classifier, fuzzy clustering is used as activation function instead of gaussian function. FCM clustering also known as fuzzy clustering trains the weights between input layer and hidden layer. The bigger the weights, the greater the degree of belonging to the cluster. Hence, weights which are expressed as numeric help to reflect the characteristics of data. In the conclusion part of the proposed classifier, connection weights between hidden layer and output layer are trained by Least Square Estimation(LSE). LSE is a method of estimating the coefficient toward minimizing the sum of squared errors. In the inference part of the proposed classifier, the weights trained by fuzzy clustering from the condition part and the connection weights estimated by LSE from the conclusion part are calculated by using centroid method each class. The largest value which is computed results is selected as final output of the proposed classifier. Not only RBFNN classifier, But weka data mining software are used for comparing the classification performance. The proposed classifier is demonstrated by showing the excellence from the viewpoint of classification performance.