跳到主要內容

臺灣博碩士論文加值系統

(44.222.189.51) 您好!臺灣時間:2024/05/18 17:53
字體大小: 字級放大   字級縮小   預設字形  
回查詢結果 :::

詳目顯示

我願授權國圖
: 
twitterline
研究生:陳國寅
研究生(外文):Chen Guo-Yin
論文名稱:類神經網路及模糊類神經網路的學習性能之研究
論文名稱(外文):On the Study of the Learning Performance for Neural Networks and Neural Fuzzy Networks
指導教授:蘇順豐
指導教授(外文):Shun-Feng Su
學位類別:碩士
校院名稱:國立臺灣科技大學
系所名稱:電機工程技術研究所
學門:工程學門
學類:電資工程學類
論文種類:學術論文
論文出版年:1998
畢業學年度:86
語文別:中文
中文關鍵詞:類神經網路模糊類神經網路
相關次數:
  • 被引用被引用:0
  • 點閱點閱:268
  • 評分評分:
  • 下載下載:0
  • 收藏至我的研究室書目清單書目收藏:1
類神經網路與模糊系統可以從輸入-輸出的資料組中預估函數以及具有聯想式記憶的行為,因為這兩系統是model-free estimators, 所以是直接地從所給的訓練資料中,學習輸入-輸出的關係而不需其他的知識,事實上,這兩種方法在某些環境下已被證明是universal approximates,而這性質超過了一般的特性,但由於不好的學習能力便無法被滿足於實際的情況,在這篇的研究中,我們目標是探討這兩個universal approximates在一般的學習觀念及基本原理的不同而不是針對追求其改善學習的方法。在這篇論文中,將討論三種數值型的學習系統,它們是類神經網路、模糊類神經網路以及具有學習其網路架構的模糊類神經網路。在我們的研究中,使用三種系統,它們是模糊車系統、sinc 函數的近似以及地形定位的系統,這些系統分別代表不同學習的問題,模糊車系統是學習具有雜訊及輸入-輸出關係的不確定性之訓練資料,sinc 函數是從一個明確而已知的系統來學習,因此,這系統的誤差以及所加入雜訊的大小可以正確的定義進而評估學習網路的學習性能,而地形定位的系統所代表的是一個非常複雜的學習目標,除此之外,在學習具有雜訊及不確定性的大量之訓練資料中而導致許多訓練上的問題,在我們所實現的地形定位的系統中,發現了幾個象現,在我們研究中,第一個象現稱做是假收斂,在這篇論文中,提出了階層式的模糊系統的方法來解決此問題,而具有階層式的模糊系統之架構,其收斂的速度快而且訓練的誤差也明顯的下降,另外我們對於地形定位的系統在模糊類神經網路中加入了一些知識,而這些知識使用在具有TSK型式的模糊法則之模糊類神經網路上,加入了知識後,其學習性能獲得大大的改善,最後,提出並討論一種不同於傳統的卡門濾波器型式而具有知識的網路架構之資料融合的新方法,而結果已證明了我們所用的方法優於傳統的卡門濾波器。在這篇論文中,在具有學習其網路架構的模糊類神經網路中可以找到兩種調整模糊法則的後件部之線性函數的參數的方法,它們是傳統的逆傳遞法以及遞回最小平方法,從所實行的結果中,我們說明了遞回最小平方法的使用並不是完全可以取代逆傳遞法的一種想法。

Neural networks and fuzzy systems can be used to estimate functions frominput-output data pairs and behave as associative memories. Since both approaches are model-free estimators, the resultant systems can be said to directly model the input-output relationship from the given training patterns without requiring other knowledge. As a matter of fact, those two approaches have been proven to be universal approximators under certain circumstances. It is more than often that such a universal property cannot be satisfied in the actual cases due to poor learning capability. In this research, instead in pursuit of approaches to improve the learning schemes, we were aimed at studying the general learning concept and the fundamental differences between those two universal approximators. In this thesis, three kinds of numerical learning systems are discussed. They are neural networks, neural fuzzy systems, and neural fuzzy systems with structure learning. Three systems are used in our study; they are the fuzzy car system, the sinc function approximation, and the terrain location identification system. Those systems represent different kinds of learning problems. The fuzzy car system is to learn from the training data that are very noisy and with non-deterministic input-output relationship. The sinc function system is to learn form an exactly know system, and therefore, the system errors and the added noise magnitude can be defined exactly to evaluate the learning performance of the employed learning mechanisms. The terrain location identification problem on the other hand represents a very complicated learning target. Beside of noisy and non-deterministic training data, the training task must learn from a very large size of training data, which may cause lots of learning problems. In our implementation of the terrain location identification system, several phenomena have been discovered. The first phenomenon is called the fake convergence in our research. In this thesis, a fuzzy hierarchical approach is proposed to resolve the problem. With this fuzzy hierarchical structure, the learning process can become fast and the training errors are also significantly reduced. Another issue in the terrain location identification problem is regarding about embedding domain knowledge into the learning structure of neural fuzzy networks. The domain knowledge is used in a neural fuzzy network in which the TSK fuzzy rule model is equipped. With such inclusion of knowledge, the learning performance is dramatically improved. Finally, with the structure of using domain knowledge, a new way of fusing data other than the traditional Kalman filter type of data fusion is proposed and discussed. Our results have also demonstrate the superiority of our approach to the traditional Kalman filter. In our study of neural fuzzy networks with structure learning, two approaches of tuning the parameters in the linear functions of the consequent part of fuzzy rules can be found in the literature. They are the traditional backpropagation algorithm (BP) and the recursive least square method (RLS). From our implementation, we may say that it is not always a good idea to use the RLS training to replace the BP training.

QRCODE
 
 
 
 
 
                                                                                                                                                                                                                                                                                                                                                                                                               
第一頁 上一頁 下一頁 最後一頁 top
無相關期刊