Please use this identifier to cite or link to this item: http://hdl.handle.net/11455/7066
標題: 以分段法建構模糊規則為基礎之倒傳遞神經網路
Fuzzy Rule Based Backpropagation Neural Network Constructed by Divide-and-Conquer Technique
作者: 陳舜賢
S.H.Chen
關鍵字: 類神經模糊網路;倒傳遞神經網路;TSK(Takagi-Sugeno-Kang);模糊群集;架構學習;參數學習
出版社: 電機工程學系
摘要: 
本論文提出一個全新的類神經模糊網路架構,即以模糊規則為基礎之倒傳遞神經網路(FRBPNN),其是以分段克服為基礎的方式建構出來。在模糊規則部分FRBPNN採TSK(Takagi-Sugeno-Kang)形式而其推論結果是BPNN型態。FRBPNN採用模糊群集概念設計,他將輸入之訓練資料分割成不同的群集,而每一個群集的輸入輸出對應關係之學習是採用BPNN架構。在FRBPNN架構中,一旦一則新的模糊規則被建立,則對應的BPNN也將對應產生。
FRBPNN的架構與參數學習是同時進行的,網路初期架構中無任何規則,所有規則皆由網路依輸入資料經演算法而即時建立。在建立規則的同時會產生一個倒傳遞神經網路,並自動搜尋適當的前件部;現存中若無適當者再依演算法建立新的模糊集合,此方法可以減少模糊集合之數量。BPNN之架構隨著每一個新建立的模糊規則產生時,可以被規劃成固定模式或者自動模式;在自動架構模式部分,初始BPNN只有包含一個隱藏節點,當一段時間的學習後,誤差下降率無法滿足時,則會在BPNN架構之中加入一個新的節點。在自動參數學習部分,在BPNN與模糊規則中預先設定部分的所有自由參數都同樣採用最陡坡降法倒傳遞方式調整。
我們使用了四組範例進行模擬,來評估FRBPNN的效能。此效能評估方式是與BPNN和其他型態的模糊類神經網路作為比較,來證明其優越性。

This thesis proposes a novel neural fuzzy network, the Fuzzy Rule Based Back-Propagation Neural Network (FRBPNN), constructed based on the divide-and-conquer technique. In fuzzy rule form, the FRBPNN is of Takagi-Sugeno-Kang (TSK)-type, where the consequence is a BPNN. The architecture design of FRBPNN employs the concept of fuzzy clustering that divides the input training data to different clusters, and the input-output mapping of each cluster is learned by a BPNN. In FRBPNN, once a new fuzzy rule is built, a corresponding BPNN will be built successively.
Learning of FRBPNN is based on simultaneous structure and parameter learning. Initially, the rule base is blank. All of the rules are constructed on-line by fuzzy clustering. For a newly generated rule, a criterion is proposed to determine whether a new fuzzy set should be generated on each input variables. This way we can reduce the number of fuzzy sets. The structure of BPNN at the consequence of each new fuzzy rule can be assigned in advance or automatically built. For automatic building, the initial BPNN contains only one hidden node. When the learning error decreasing rate is not satisfied over a period of time, a new node is added to the BPNN in the rule with the largest distributed error. For parameter learning, all of the free parameters in the BPNN and precondition part of the fuzzy rules are learned by gradient descent.
To measure the performance of FRBPNN, four examples are simulated. Performance of FRBPNN is compared with BPNN and other type of neural fuzzy network to verify its superiority.
URI: http://hdl.handle.net/11455/7066
Appears in Collections:電機工程學系所

Show full item record
 
TAIR Related Article

Google ScholarTM

Check


Items in DSpace are protected by copyright, with all rights reserved, unless otherwise indicated.