论文研究 基于神经网络的软件模块风险性预测模型.pdf

采用学习矢量量化神经网络对软件质量进行预测,提出基于学习矢量量化神经网络的软件模块风险性预测模型,与BP神经网络预测模型相比,实验结果表明提出的模型获得更精确的预测效果。1082007,43(18)Computer Engineering and Applications计算机工程与应用型错误率(错淏判定的个数测试样本的个数)和总体的错误率(两和BP神经网络模型。LVQ神经网络在有监督的状态下将自组种类型错误判定的总个数测试样本的总个数)。织性和竞争性学习思想进行有机结合,具有更好的性能。本文3.3预测结果及分析提出基」学习矢量量化神经网络的软件模块风险性预测模型,与在实验中,分别选择30个高风险性模块的尺度数据与70BP神经网络预测模型相比,实验结果表明基于LVQ神经网络的个低凤险性的尺度数据对VO神经网络进行训练,隐含层的预测模型得更精确的预测结果。(收稿日期:2006年11月)神经儿个数为20,学习速率为0.05,训练次数为300次,训练的结果如图2所示。参考文献I1 LeGall C, Adam M F, Derriennic H, et al. Studies on measuringi I. asoftware [ JIEEE Journal of Selected Areas in Communications1990,8(2):234-245[2 Karunanithi N, Whitley D, Malaiya Y K Using neural networks inreliability prediction[J]IEEE Software, 1992, 9(4): 53-59[3 Munson J C, Khoshgoftaar T MThe detection of fault-prone pro-grams[.IEEE Transactions on Software Engineering, 1992, 18(5)423-433[4] Porter AA, Selby R WEmpirically guided software developmentusing metric-based classification trees J-IEEE Software, 19907(2):46-54图2基于LVQ的神经网络训练结果5 Khoshgoflaar T M, Lanning D L, Pandya A S A neural network mod经过300次训练后,网络误差达到0.21这个比较理想的eling methodology for detection of high-risk programs[Cy/proc 4th Int数值。在LVQ神经网络测试中,分别选用没有参加训练的23Symp Software Reliabilily Eng, Denver, CO, Nov 1993: 302-309个高风险性模块、31个低风险性模块和90个中等风险模块数6 Khoshgoftaar T M., Pandya A S, Lanning d L.A comparative study据进行验证,模型的输出如果为1则表示为低风险性模块,如of pattern recognition techniques for quality evaluation of telecom果输出结果为2则表示为高风险性。基于LVQ模型的预测结munications softwre [J].IFFE. Journal on Selested in Communica-ions,194,12(2):279-291.果以及B神经网络预测模型的预测结果见附表1-附表3,(7 Khoshgoftaar T M,AenE.B. Hudepohl J P,etal. pplication of附表1是高风险性模块预测结果,附長2是低风险性模块预测neural networks to software quality modeling of a very large结果,附表3为中等风险性模块预测结果。telecommunications system [J.IEEE Transactions on Neural Net-从3个表的结果可以看出,在第一种类型中,基」BP神works,1997,8(4):902-909经网络预测模型的错误预测数为2个,测试用例数为23个,错81 Khoshgoftaar T M, Allen e.B. Predicting testability of program淏率为2123=8.69%;而基于LⅤQ神经网络预測模型的错误预modules using a neural network [C]/Proceedings of Symposium of测数为1个,测试用例数为23个,错误率为123=4.34%。在第Application-Speeific Systems and Software Engineering Technolo-二种类型中,基于BP神经网络预测模型的错误预测数为5gv,2000.个,测试用例数为31个,错误率为5/31=1613%,而基于IVQ9 Wei zhi, Khoshgoflaar T M. Software quality prediction f神经网络的预测模型的错误预测数为1个,测试用例数为31surance network telecommunications systems computer Journal个,错误率为1/31=3.23%。基于BP神经网络的预测模型的两2001,44(6):557-568种错误类型共有7个错误预测,测试用例共54个,错误率为7/0 Neumann D h. An enhanced neural network technique for soft54=12.96%;而基于LVQ神经网络的预测模型的两种错误类型ware risk analysis[JJ.IEEE Transactions on Sollware Engineerin共有2个错误预测,测试用例共54个,错误率为2/54=3.70%。002,28(9):904-912对j中等风险性模块预测介这两种类型之间,不会严重影响1 I Miemie Thel Thwin, Tong-Seng qui4 oplication of neural net-work for predicting software development faults using object-ori软件产品的质量,没有特别重要的价值。可见基于LⅤQ神经网ented design metrics Cp/Preceeding of 9th Internatinal Confer络的预测模型的三个错淏率都比基于B神经网络预测模型ence on Neural Information Processing(ICONIP'07), 2002, 5要低,基于LVQ神经网络的预测模型的预測精确度要高。比较2312-2316.结果如表1所示。[12] Quah Jon T S, Mie Mie Thet Thwin. Prediction of software readi表1基于LVQ神经网络的预测模型与基于BP神经网络的ness using neural network CV/Proceedings of lst International预测模型的预测结果比较Conference on Information Technology &Applications( ICITA错误类型2002), Australia, 25-28 Nov 200模型第一类型第二类型总体[13 Mie Mie Thet Thwin Tong-Seng Quah Application of neural net-个数锴误率/%个数错误率/%个数错误率works for software quality prediction using objiect -oriented metBP神经网络预测模型28.69516.1371296rics[J]Journal of Systems and Software, 2005, 76(2): 147-156LVQ神经冈络预测模型14.3413.2323.70[14] Kanmani S, Rhymend Uthariaraj V, Sankaranarayanan V, et al. Object oriented soflware quality prediction using generregressIon4结语neural networks [CP/ACM SIGSOFT Software Engineering Notes2004,29(5):1-6软件质量直接影响到软件系统的可维护性、有效性、可靠[15] Mahaweerawat A, Sophatsathit P, Lursinsap C, et al. Fault predic性。在软件开发初期通过对软件模块风险的判定,可以帮助软tion in object-oriented software using neural network techniques件管理人员合理地分配人力和物力资源来提高软件质量。现阶[C]/Proceedings of the In Tech Conference, Houston, 2-4Dec段的软件模块风险性预测模型有判别式分析模型、决策树模型2004:27-34.與求索,钟诚:基于神经网络的软件模块风险性预测模型2007,43(18)109附表1高风险性模块预测结果模块数VG11 G LOC HEAIlYQ预测类型HP预测类型4703852824269530215420609228146.958131220811268219158814.45173031724013365207171112411011432241627910486105156523013368262191332953105.314.571851584494764523.387113286197159918.382243314910261725899.361522175935516l73656227.531(19)317761954463292955.456802.1367632037058附表2低风险性模块预测结果模块数VG1 DG ELOC HEE2m12LVQ预测类型BP预测类型26()671631610.12731432622132127211212100.280.05131340.01923330150.4510219813163280.01660.614618034.3220.051111210.016().691090.917374100034(.18I812).45l(01ll2821102007,43(18)omputer engneeringdations计算机工程与应用表3中等风险性模块预测结果模坎数vVQ预测类型HP预测类型33288.6152345678901236000005附7789928780977666Ol05060500000004321912l12232544515199c01104350040807224330241111811121112281344154841121678185420099834442117632669679519122348210123456789032343678904234533345111114522215292522213266077)00000.000.0111584(3200655065708006545483172306033038102005311111821662801000.0003123282933129431222232245831000000003492802卯SI9c900011112ll067211111464444555503312843122323331332127516000000003615641445816116861111133823200906523686304008402626309m8969452465739902098394065935111210.6589662643121267)008671347131530.302033543007742522099188252420100009822594888900006421121121501211252222222238312223206111l12112355576118)3102111116951312223238160000u0u05537888888I164555171812169222719010111112-1111111412