系統識別號 U0026-0812200911525571
論文名稱(中文) 在機器學習中適當樣本量之決定
論文名稱(英文) Determination of the meaningful size of learning data sets
校院名稱 成功大學
系所名稱(中) 工業與資訊管理學系碩博士班
系所名稱(英) Department of Industrial and Information Management
學年度 94
學期 2
出版年 95
研究生(中文) 張耀仁
研究生(英文) Yao-Jen Chang
電子信箱 ppapa0511@hotmail.com
學號 r3693117
學位類別 碩士
語文別 中文
論文頁數 40頁
口試委員 口試委員-李賢德
中文關鍵字 分類  樣本數  準確度 
英文關鍵字 accuracy  Sample number  classification 
中文摘要   在工業製造上,常會運用許多不同的分類方法將產品做分類。大部分的分類方法在建立其統計模型時,會假設樣本均來自同ㄧ母體分配,然後透過統計上所使用的最小足夠樣本量 來做為最少所需要的樣本量,但是通常求出的 都蠻大的,所以我們思考是否 這個指標適合用在工業的製造上。首先在前提假設上,我們認為假設所有的樣本均來自於同ㄧ母體分配是不合理的,因為在工業生產中雖然前後的樣本可能會有高度的相關,但是由於時間性和外在環境上一定都會有些微的不同,所以是否能夠將所有的樣本均假設為同ㄧ個母體是我們所懷疑的。

英文摘要 In this modern generation, there are quite a few classification methods to classify the products in the field of manufacturing. The most common method used by most manufactories is to establish statistical models that are assumed all samples come from the same population. Then the smallest sample amount m0 needed from statistical measurements would be used for the sample estimation; However, m0 is quite large. Therefore we have to consider whether m0 is suitable or not for the field of manufacturing.

We doubt the assumption that all samples come from the identical population is unreasonable. In industrial production, there could be some kind of correlation between samples that their appearances are sequential because the factors of time and external environmental conditions can cause the slightly difference concerning the amount of samples, therefore we substitute the assumption all samples are derived from the same population.

This research explores to establish an estimation model. After analyzing the data characteristics and then putting them into this model, we can find out two suitable sample quantities, and divide samples collected into three parts according to the two critical values:(1) When the sample quantity has not reached the first critical value, we regard it as the small sample, and deal the data with small sample processing methods for classification. (2) Between the first and the second critical values, we regard it as the suitable sample quantity. (3) If the amount of sample exceeds the second critical value, the amount of sample is oversized. The factors of time and external environment conditions have played important roles on affecting the amount of samples. The initial samples also have a side effect on the accuracy of later classification. This leads into worse results for classification accuracies; therefore it would be appropriately to remove the initial samples in order to improve the accuracy of classification.

By way of this model establishment, we can put the collected data into this model to calculate the appropriate two samples quantity target (that is two critical values), and choose the suitable classification methods according to the two critical values in order to improve the efficiency and the accuracy of classification. On the other hand, researchers can determine the necessary sample quantity based on the proposed model to avoid the wastes of manpower and material resources.

論文目次 目錄
摘要 I
Abstract II
誌謝 IV
目錄 V
表目錄 VI
圖目錄 VII
第一章、緒論 1
第一節、研究背景和動機 1
第二節、研究目的 2
第三節、論文架構 3
第二章、文獻探討 4
第一節、計算小樣本的方法 4
一、透過PAC-Learnable找出足夠的樣本量 (sufficient sample length) 4
二、透過Vapnik-Chervonenkis dimension來定義小樣本的範圍 5
第二節、倒傳遞類神經網路(Back Propagation Network) 7
二、定義 7
三、網路架構 8
四、倒傳遞類神經網路所使用的非線性轉換函數 10
五、倒傳遞類神經網路演算法 11
六、倒傳遞類神經網路的優缺點 15
第三章、研究方法 16
第一節、模式架構 16
第二節、模擬樣本 17
第三節、倒傳遞類神經網路做分類時的變數設定 19
第四節、倒傳遞類神經網路做預測時的變數設定 21
第四章、模擬分析及實證研究 23
第一節、運用倒傳遞類神經網路做分類的模擬分析 23
第二節、運用倒傳遞類神經網路做預測的正確率驗證 34
第五章、結論與建議 37
Reference 39

參考文獻 Bengio, Y. (1996). Neural Networks for Speech and Sequence Recognition. International Thomson Publishing Inc.

Bo K. Wong , Thomas A. Bodnovich , Yakup Selvi , (1997) , “Neural network applications in business : A review and analysis of the literature(1988-95)” , Decision Support systems 19 , pp.301-320

Brad Warner & Manavendra Misa, (Nov.,1996), “Understanding Neural Network as Statistical Tools” , The American Statistician , Vol.50 , No.4 , pp.284-293

Braga-Neto,U.M. , Dougherty,E.R. , (2004)“Is cross-validation valid for small-sample microarray classification?” , Bioinformatics , 20 , pp.374-380

Dana Angluin , (1988) , “Queries and Concept Learning” , Machine Learning , 2 , pp.319-342

David Haussler , Michael Kearns & Robert E. Schapire , (1994) , “Bounds on the Sample Complexity of Bayesian Learning Using Information Theory and the VC Dimension” , Machine Learning , 14 , pp.83-113

Kidong Lee & David Booth & Pervaiz Alam , (2005) , “A comparison of supervised and unsupervised neural networks in predicting bankruptcy of Korean firms” , Expert Systems with Applications 29 , pp.1-16

Mohammed Waleed Kadous , Claude Sammut , (2005) , “Classification of Multivariate Time Series and Structured Data Using Constructive Induction” , Machine Learning , 58 , pp. 179–216

Sally Floyd & Manfred Warmuth , (1995) , “Sample Compression, Learnability, and the Vapnik-Chervonenkis Dimension” , Machine Learning , 21 , pp. 269–304

S. Gunasekaran. , B. Venkatesh, B.S.D. Sagar, (2003), “Convergence index for BPN training” , Neurocomputing 55 , pp.711-719

Tom M. Mitchell , (1997) , “Machine learning” , McGraw-Hill International editions

Tsong-Lin Lee ,(2004), “Back-propagation neural network for long-term tidal predictions” , Ocean Engineering 31 , pp.225-238

Vladimir N. Vapnik , (1999) , “The Nature of Statistical learning Theory ” , Springer , Second edition

W-C.Chiang Tl Urban & GW Baldridge, (1996), “A Neural Network Approach to Mutual Fund Net Asset Value Forecasting” , Omega ,Vol. 24 , No.24 , pp.205-215 ,

Wenjiang J. Fu ,Edward R. Dougherty , Bani Mallick , Raymond J. Carroll , (2005) , “How many samples are needed to build a classifier : a general sequenrial approach ” , Bioinformatics , Vol. 21 , pp.63-70



  • 同意授權校內瀏覽/列印電子全文服務,於2006-06-26起公開。
  • 同意授權校外瀏覽/列印電子全文服務,於2006-06-26起公開。

  • 如您有疑問,請聯絡圖書館