在資訊軟體開發過程中,成本的估計是一個重要的議題,足以影響整個專案的成功與否。若開發者能正確的預估成本,將可為此開發專案作完整的規劃,以分配適當的資源;再經由管理者的領導,使發展過程控制得宜,減少突發事件的產生,使專案在預定時程及預算內達成目標。在過去的二十年有許多的估計模型被提出,如統計估計的方法,COCOMO方法等,然而這些方法的對成本的預測的準確率卻很低。有鑑於此,本研究發展出兩種新的預估模型:模式一應用類神經網路於軟體開發成本的預估,以COCOMO資料庫所提供的63筆專案資料,採用倒傳遞神經網路,建立該模式得以預估軟體開發成本;模式二應用蒙地卡羅方法於軟體開發成本的預估,同樣以COCOMO資料庫所提供的63筆專案資料為基礎,以蒙地卡羅抽樣模擬大量資料再利用蒙地卡羅方法,建立該模式得以估計軟體開發成本。本研究的實驗結果顯示,此兩種模式都有不錯的預估能力。
In a software development its cost estimation is an important issue. Many cost estimation models have been proposed, such as statistical methods and COCOMO model; yet, they have a main drawback of the low accuracy of prediction. To overcome this disadvantage, this thesis developed two cost approaches for software estimation : neural network and Mote Carlo mothods. A back-propagation neural network (BPN) was used to learning the pattern of the COCOMO dataset and to compare with the results of COCOMO itself. Experimental results showed that the BPN was superior to the COCOMO model. Also, Mote Carlo mothod built a cost estimation model using the COCOMO dataset. Experimental results indicate that the Mote Carlo approach was useful for measuring development cost of software.