Dynamic Induction Network and its Application

• Main Authors: Chwng, Wen-Cheung, 鄭文泉
• Other Authors: Kuo Yau-Hwang
• Format: Others
• Language: zh-TW
• Published: 1998
• Online Access: http://ndltd.ncl.edu.tw/handle/43867538848706209304

碩士 === 國立成功大學 === 資訊工程學系 === 86 === Knowledge acquisition may be the most critical issue in developing an expert system. In fact, how to acquire accurate and compact knowledgefrom imprecise or incomplete training samples is still an open problemon most application fields. Traditionally, knowledge acquisition is realized by interviewing with one or more experts. But, such a told-by-expert approach suffers from low rule generation rate and knowledge inconsistency produced by different experts. Therefore, many machine learning algorithms, such as ID3 and neural network, had been developed to solve these problems. But these methods also have some disadvantages.In this thesis, a dynamic induction (DI) approach based on a four-layered network structure is proposed to solve the knowledge acquisition problems mentioned above. The proposed DI model provides an incremental learning ability on induction, which can acquire partial conclusions from some incomplete training data. Besides, it adopts a structural knowledge representation scheme that is different from the distributed knowledge representation scheme used in conventional neural network models. The principle of self-organization is also applied in the DI model to increase the processing speed and system flexibility.Two real problems are adopted to verify the DI model. Several experiments have been made in various aspects including learning efficiency and accuracy, learning speed, etc. Fortunately, all experimental results affirmatively confirm the superiority of the dynamic induction approach.