A Method of Fuzzy-Neural Network to Pattern Classification

利用模糊類神經網路解型態分類的方法

張介仁(Jieh-Ren Chang) ； 王乃堅(Nai-Jian Wang)

《宜蘭技術學報》 9期電機資訊專輯 (2002/12) Pp. 147-157

https://doi.org/10.6179/nit.eec.2002.09.25

摘要

在資訊或生物科技工程上，分類問題一直是一個重要的研究方向，更是資料探勘領域上一們基礎的研究。類神經模糊規則架構常被連用在分類系統的研究上，本論文是以模糊量測（Fuzzy measures）為基礎，從資料中自動建立模糊規則庫及可調變的類神經結構並應用於分類系統上，並以範例加以實驗測試以證明此演算法的正確性。本文中所提出的方法是利用輸入所有的訓練資料，並記錄下各類別每一個輸入變數的最小和最大值來建立包圍類別區域的超立方體作為類神經結構的第一層。接著測試這些超立方體之間重壘情形，我們分別定義了模糊量測、資訊提供度及資料分離度並建立一結構調變參數，用以判別此重疊區域是否需要再被更細的分割，而分割後的模糊規則庫即成為類神經結構的第二層。依此類推，可建立多迴圈的類神經模糊架構，直到結構調變參數太小。此方法可以合理的使用訓練資料以建立有效的分類器，並可依照資料的不同來簡化計算的複雜度，已達快速分類之目的。

關鍵字

分類器 ； 模糊量測 ； 模糊類神經網路 ； 資料探勘

並列摘要

The pattern classification is an important issue on Information technology and biological engineering. It is also a key element to data mining research. Recently, Fuzzy-Neural network system is used in many pattern classifiers. In this paper, a new method is proposed for setting a variable Fuzzy-Neural network structure directly from numerical data. It provides several examples of operation that demonstrate the strong qualities of this method. We propose a variable fuzzy-neural structure network, which constructs by 3 layers for pattern classification. The first layer is for data input, and third layer is for output decision. We make the unit of second layer of network, which is set by each activation fuzzy hyper-box for each class. The fuzzy hyper-box is an n-dimensional box defined by a min point and max point with a corresponding membership function. Then, we test the condition of overlap of these hypercube by defined tuning-structure parameter, which is made by fuzzy measure, information support degree and data separation degree. We decide how many loops at most in the second layer of the network should be rebuild again. We also create a feedback node in third layer to decide the parameter value of each unit in second layer. We can generate a high efficiency classifier by this dynamic neural-fuzzy network structure using sufficient information of all training data. We also decrease the complexity of classification computation according to different test data.

並列關鍵字

pattern classification ； variable neural-fuzzy network structure ； data mining ； fuzzy measure