Structure adaptation of hierarchical knowledge-based classifiers

This paper introduces a new method to identify the qualified rule-relevant nodes to construct hierarchical neuro-fuzzy systems (HNFSs). After learning, the proposed method analyzes the entire history of activities and behaviors of all rule nodes, which reflects their levels of involvement or contribution during the process. The less qualified rule-relevant nodes can then be identified and removed, reducing the size and complexity of the HNFS. Upon the repetitive learning process, the method may be repetitively applied until a satisfactory result is obtained, simultaneously improving the performance and reducing the size and complexity. Incorporated with the method is a new HNFS architecture which addresses both the scalability problem experienced in rule based systems and the restriction of the "overcrowded defuzzification" problem found in hierarchical designs. In order to verify the performance, the proposed method has been successfully tested against five well-known classification problems whose results are provided and then discussed in the concluding remarks.