• 作者： 王景弘; 洪宗貝; 曾憲雄
• 作者服務機構： 中華電信研究所; 義守大學資訊管理學系; 國立交通大學資訊科學研究所
• 中文摘要： 於本篇論文中，我們提出一個混合型學習演算法，它結合傳統樹型學習法及Back-propagation類神經網路學習法，去建立一個樹型網路分類器。在整個學習過程中，我們不需要頂先設定學習網路的架構，網路最後的架構會隨著學習樹的成長自動建立。最後，我們將此學習演算法應用於腦瘤疾病診斷領域，並和其它三種學習法比較，於準確度比較，我們發現本演算法高於傳統樹型學習法及Back-propagation學習法，於學習速度比較，我們發現本演算法快於Back-propagation學習法。
• 英文摘要： In this paper, we propose a tree net classifier (TNC) that integrates the learning strategies of decisiontrees and modified back-propagation neural networks. The TNC uses a tree-pruning mechanism to avoidoverfitting problems．It has small tree sizes and low error rates, especially for complex classification.Furthermore, users need not lay out the structure of a tree net in advance；the structure is automaticallyconstructed by the tree-growing process. Finally, results of experiments on diagnosing brain tumors aregiven to compare the proposed algorithm with three other learning methods in terms of accuracy, complexityof the knowledge structure, and learning speed. The experimental results show that the proposed classifierhas very high accuracy rates and learns much faster than the back-propagation neural network.
• 中文關鍵字： back-propagation; decision tree, ID3; neural network; non-linear boundary; tree net
• 英文關鍵字： --