洪欽銘Chin-Ming Hong陳瑄易Shyuan-Yi Chen2019-09-042006-7-32019-09-042006http://etds.lib.ntnu.edu.tw/cgi-bin/gs32/gsweb.cgi?o=dstdcdr&s=id=%22GN0693700346%22.&%22.id.&http://rportal.lib.ntnu.edu.tw:80/handle/20.500.12235/98992ㄧ個可以解釋因果關係的分類系統，必須具備容易解釋語意的表示方式以及使用方便等特性。因此本研究建構於類神經-模糊網路之上，利用其模糊理論之可解釋性與類神經網路之學習能力對樣本進行訓練以期獲得一個優越的模糊分類器。其中對特徵選取以及歸屬函數最佳化均分別用了不同之方法進行試驗，同時提出如何利用K-means演算法獲得歸屬函數初始參數之方法，並提供一特徵選取方式，使所設計之模糊分類器可以用最低特徵考量達至最高推論精確度。 為克服類神經-模糊網路分類器繁瑣之最佳化過程，本研究提出一個垂直合併型歸屬函數概念之快速圖形式模糊分類器，將原本模糊集合之歸屬度由值之水平移動進行歸屬度之計算方式，改由由各值之垂直移動對應模糊區間之計算方式，同樣可以藉由圖形化之概念進行可讀性之解釋，同時步驟簡易且具可解釋性，更重要的是大量減低歸屬函數之使用，且仍可保有相當高之準確率。而本研究所提出之演算法均經由實驗證明其可行性，並進行分析與討論。An efficient and simple decision support system must have the characteristics such as interpretable, easy understanding, convenient, et al. For this reason, the designed classifier in this study was based on a neuro-fuzzy network to combine the transparent characteristic of fuzzy system and learning ability of neural network. First, this study proposes a refined K-means algorithm and a gradient-based learning algorithm to logically determine and adaptively tuned the fuzzy membership functions for the employed neuro-fuzzy network. Moreover, this study also uses grey relational algorithm to perform feature selection and proposes a novel feature reduction algorithm to overcome the drawbacks of grey relational algorithm. Because optimized processes contain complex and long steps, this study proposes a Fast Graph Fuzzy Classifier (FGFC) which has a novel determining scheme of the membership function degree and can prevent to confront an abstruse classifier algorithm as well as keep the advantages of the traditional fuzzy systems. All of the above-mentioned methods were implemented and analyzed in this study.類神經-模糊網路分類特徵選取K-means演算法歸屬函數Neuro-Fuzzy Network ClassifierClassificationFeature SelectionK-means AlgorithmMembership Function