带优化参数的SVR模型及在软件成本估算中的应用  

Hybrid Support Vector Regression Model with Optimized Parameter for Software Effort Estimation

作　　者：吴登生[1,2] 宋浩[1,2,3] 李建平[1] 蔡晨[1] 

机构地区：[1]中国科学院科技政策与管理科学研究所,北京100190 [2]中国科学院研究生院,北京100039 [3]山东经济学院统计与数学学院,山东济南250014

出　　处：《运筹与管理》2011年第1期143-149,共7页Operations Research and Management Science

基　　金：国家自然科学基金重大研究计划重点项目(90718042);国家自然科学基金重点项目(70531040)

摘　　要：随着社会各行各业对软件开发投资的日益增长,产业界和学术界越来越关注可靠的软件成本估算,以有效控制软件开发过程中相关风险。为了能更准确地估算软件成本,提出一种带遗传算法优化参数的支持向量回归机模型,用遗传算法来优化支持向量回归机模型中的参数集(C,γ,ε),可以避免参数选择的盲目性,能显著提高支持向量回归机模型的预测能力。分别用IBM DP、Kemerer和Hallmark三个数据库来验证模型的有效性,并与常用的线性回归模型进行对比,结果显示采用遗传算法优化的支持向量回归机模型具有很好的学习精度和推广能力,在MMRE和Pred(0.25)两个标准上都优于线性回归模型。With software development becoming an essential investment for many organizations recently,both the software industry and academic communities have become more and more concerned about reliable and accurate estimates of software development effort.This paper examines the potentials of the software effort estimation model by integrating a genetic algorithm（GA）to the support vector regression（SVR）model with RBF kernels.The genetic algorithm（GA）method is adapted to optimize the parameters（Cγ,ε）in support vector regression.The experiments are carried out using three datasets of software projects from IBM DP,Kemerer and Hallmark.The effectiveness of the GA-SVR model is compared with other techniques such as linear regression in terms of the error measure which is mean magnitude relative error（MMRE）and Pred（0.25）.The results have shown that GA-SVR model has superior learning accuracy and generalization performance and significantly outperforms linear regression in three datasets.

关 键 词：可信软件 软件成本估算 参数优化 支持向量回归机 遗传算法 

分 类 号：TP311.5[自动化与计算机技术—计算机软件与理论]