- 學位論文
應用倒傳遞類神經網路和多因子軟體專案分群預估軟體工作量
Applying Back Propagation Neural Network to Estimate Software Effort by Multiple Factors Software Project Clustering
摘要
在科技產業中，在每個專案的軟體開發上常常會遇到的問題是如何評估一個軟體開發成本、專案所需人力及時程規劃等，這些常常會從以前的經驗來去評估一個專案所需的工作量及相關成本，一旦評估錯誤可能會導致一個專案的虧損或失敗，所以精確的評估每個專案的工作量是很重要的。而本研究將以倒傳遞類神經網路(Back Propagation Network)來對專案進行分析和預估軟體工作量，並以皮爾遜相關積差係數分析( Pearson product-moment correlation coefficient ) 以及one-way ANOVA分析並選出數個因子，並帶入不同的分群演算法去分群(K-mean method分群演算法, Ward’s method 分群演算法)，利用平均誤差率(MMRE)及評估等級(PRED)做比較，本研究利用COCOMO 中的63筆歷史專案來進行測試，並由實驗結果得知經由分群過的專案跟COCOMO原本的3種分類模式在預估軟體工作量相較下是比較準確的。
關鍵字
軟體工作量 ； 皮爾遜積差相關係數分析 ； one-way ANOVA分析 ； K-mean method分群演算法 ； Ward’s method 分群演算法 ； 倒傳遞類神經網路 ； 評估等級 ； 平均誤差率並列摘要
In the technology industry, the problem often encountered in each project's software development is how to estimate the cost of a software development schedule planning and project the necessary manpower, these often come from previous experience to estimate a project required effort and associated costs, once the estimated false may lead to the loss or failure of a project, so an accurate estimate of the effort of each project is very important. The study will be Back Propagation Network software effort analysis and estimate of the project, and use the Pearson product-moment correlation and one-way ANOVA analysis to select a number of factors, and through a different clustering algorithms to the clustering method ( K-mean clustering algorithm and Ward's method clustering algorithm), the MMRE and prediction level (PRED) to compare project, the study of 63 COCOMO in the history of the project to be tested by experimental results, through the clustering project and multiple factors analysis that was compared to originally with the COCOMO three kinds of classification model is more accurate to estimate software effort.