資訊管理學報
葉怡成;陳重志;黃冠傑;
頁: 135-154
日期: 2008/04
摘要: 徑向基底函數網路(RBFN)常用於分類問題，它的核有形心與半徑二種參數，這二種參數可用監督式或無監督式學習來決定。但它有一個缺點是視所有自變數有同等地位，故分類邊界是圓形，但事實上每一個自變數對分類的影響力不同，分類邊界是應該是橢圓形較合理。為克服此一缺點，本文提出具自適應核形狀參數的徑向基底函數網路，並以監督式學習推導出其學習規則。為證明此一架構優於傳統的徑向基底函數網路，本研究以五個人為的與七個真實的分類例題進行比較。結果顯示，此一架構確實比倒傳遞網路及傳統的徑向基底函數網路更為準確，狀參數值的大小確實能表現出自變數對分類的影響力高低。
關鍵字: 半徑基神經網路;監督式學習;核函數;分類;
Abstract: Radial Basis Function Network (RBFN) is usually employed for classification problems, whose kernel has centroid and radius parameters determined with supervised or unsupervised learning. However, it has a shortcoming that it regards each independent variable as the same position; hence, the boundary of classification is circle. But in fact, each independent variable has different influence to the classification, it is more reasonable that the boundary of classification is ellipse. To overcome the shortcoming, we proposed the RBFN with adaptive kernel shape parameters and deduced its learning rule, using supervised learning. To verify whether the architecture is more accurate than conventional RBFN, experiments with five human-made problems and seven real-world problems were conducted. The results showed that this architecture is really more accurate than Back-Propagation Network and conventional RBFN, and the shape parameters can represent the influence of independent variable to classification.
Keywords: Radial basis function network;supervised learning;kernel function;classification;
瀏覽次數: 12687 下載次數: 123
引用 導入Endnote
頁: 135-154
日期: 2008/04
摘要: 徑向基底函數網路(RBFN)常用於分類問題，它的核有形心與半徑二種參數，這二種參數可用監督式或無監督式學習來決定。但它有一個缺點是視所有自變數有同等地位，故分類邊界是圓形，但事實上每一個自變數對分類的影響力不同，分類邊界是應該是橢圓形較合理。為克服此一缺點，本文提出具自適應核形狀參數的徑向基底函數網路，並以監督式學習推導出其學習規則。為證明此一架構優於傳統的徑向基底函數網路，本研究以五個人為的與七個真實的分類例題進行比較。結果顯示，此一架構確實比倒傳遞網路及傳統的徑向基底函數網路更為準確，狀參數值的大小確實能表現出自變數對分類的影響力高低。
關鍵字: 半徑基神經網路;監督式學習;核函數;分類;
Radial Basis Function Networks with Adaptive Kernel Shape Parameter
Abstract: Radial Basis Function Network (RBFN) is usually employed for classification problems, whose kernel has centroid and radius parameters determined with supervised or unsupervised learning. However, it has a shortcoming that it regards each independent variable as the same position; hence, the boundary of classification is circle. But in fact, each independent variable has different influence to the classification, it is more reasonable that the boundary of classification is ellipse. To overcome the shortcoming, we proposed the RBFN with adaptive kernel shape parameters and deduced its learning rule, using supervised learning. To verify whether the architecture is more accurate than conventional RBFN, experiments with five human-made problems and seven real-world problems were conducted. The results showed that this architecture is really more accurate than Back-Propagation Network and conventional RBFN, and the shape parameters can represent the influence of independent variable to classification.
Keywords: Radial basis function network;supervised learning;kernel function;classification;
瀏覽次數: 12687 下載次數: 123
引用 導入Endnote