การจัดกลุ่มอย่างเหมาะสมที่สุดสำหรับฟังก์ชันความเป็นสมาชิกในระบบนิวโรฟัซซีเพื่อการจำแนกประเภท

Abstract

This thesis proposes an optimal clustering approach for the membership function initialization in the neuro-fuzzy system for classification. Three experiments of the neural fuzzy system were conducted for determining the initialization of the membership function. First, the original method that uses the mean and the standard deviation. Secondly, the k-means clustering uses the mean and the standard deviation of the resulting cluster. Finally, the quartile method uses the quartile values that divide data into quarters. Every method was tested with 2 to 5 of membership functions that are easy to understand and suitable for rule-based classification. The experiment consisted of 3 steps. The first step determines the initialization of the membership function with the specific number of membership functions. Second step calculates the membership value to convert the input data to fuzzy value and convert the resulting value to bipolar value for the next step calculation. Third step deals with neural network learning. It is a combination of the products of the weights with the bipolar value and the bias. Then the resulting value was put into the activation function. The last step was to adjust the parameters, which are the mean and standard deviation of the membership function, i.e., weights and bias of neural network step. The above classification was tests against the 12 datasets from the UCI database, which is a dataset with difference class numbers, number of attributes and the numbers of samples. Datasets were divided into learning datasets and testing datasets with the 10 folds cross-validation test to calculate the average accuracy from the classification results. The results showed that grouping for membership functions by initializing the quartile method and assigning the number of membership functions to 3 resulted in greater efficiency compared to other methods tested. The quartile method yields higher accuracy than the original neuro-fuzzy system from 8 out of 12 datasets.