應用類神經方法建構物件導向軟體可維護性預測模式

Abstract

近年來以物件導向技術開發的軟體，逐步地取代傳統軟體技術，成為軟體系統開發的主流技術，因此有越來越多物件導向軟體系統出現在市面上。由於軟體開發過程所需投入資源相當龐大，因此透過維護作業以保持軟體系統高效率地持續運作是相當重要的，而軟體可維護性預測模式可幫助管理維護階段資源及對系統採取防護性設計(defensive design)，能降低未來維護系統時所需花費心力。軟體度量為軟體可維護性預測模式建立的基礎，但傳統的度量方法已不適合物件導向特殊設計方法，因此發展出許多物件導向度量，其中Chidamber 與 Kemerer[11]制定的度量(C&K度量)與 Li 與 Henry[18]制定的度量與軟體可維護性間具有高度相關，提升了度量方法在物件導向軟體可維護性預測上的實用性。然而度量與軟體品質特性之關係複雜且非線性，傳統統計方法來預測品質特性會使預測準確性受限，由於類神經網路對於變數間具非線性關係有不錯預測能力，因此本研究使用自組性演算法與廣義迴歸類神經網路建構軟體可維護性預測模式，再透過比較上述兩種類神經方法的準確性，提出一套準確性高且方便業界使用之軟體可維護性預測模式。本研究最後利用國外文獻所提供之商業軟體系統QUES(Quality Evaluation System)三年之維護資料，驗證本研究所提出之軟體可維護性預測模式具準確性與再現性。

In recent years, object-oriented software development techniques have gradually replaced traditional techniques of software development, and became the mainstream technique. As a result, more and more object-oriented software systems appear on the market. Due to the vast amount of resource required in the software development process, it is crucial to maintain highly efficient and persistent operations of software systems by means of maintenance procedures. The maintainability prediction model of a system can help managing system’s maintenance resource and adopt a defensive design mode towards system development in order to decrease the time and energy needed in future maintenance processes. Software metrics is the basis for the maintainability prediction model of a system, but the usual recommended traditional metrics is not suitable for object-oriented design. Hence, many object oriented-based metrics have been developed, in which that established by Chidamber and Kemerer (the C&K metrics) and that established by Li and Henry are highly related to system maintainability, and improves the practicality of metrics on the maintainability prediction model of a system. However, the relationship between metrics and software quality is complicated and nonlinear. Thus, the accuracy of the prediction model will be limited by traditional statistical approaches. Earlier studies have suggested that artificial neural approaches have performed quite well on nonlinear relationship predictions. Therefore, this study applied the Group Method of Data Handling(GMDH) and General Regression Neural Network(GRNN) to build the maintainability prediction model of a system, and address a highly accurate and easily accessible prediction model via comparing the accuracy of the two artificial neural approaches.Lastly, three years of maintenance data from the commercial software – QUES (Quality Evaluation System) were utilized to verify that the proposed maintainability prediction model possesses both accuracy and reproducibility.