神经网络在软件失效预测中的性能比较和研究  

Comparative Study of Software Fault Prediction Models Based on Different Neural Networks

作　　者：肖中元[1] 王琪[1] 于波 朱杰[1] 

机构地区：[1]上海交通大学电子工程系,上海200030 [2]朗讯科技光网络有限公司,上海200233

出　　处：《计算机仿真》2005年第10期179-182,共4页Computer Simulation

摘　　要：在软件开发的早期预测有失效倾向的软件模块,能够极大地提高软件的质量。软件失效预测中的一个普遍问题是数据中噪声的存在。神经网络具有鲁棒性而且对噪声有很强的抑制能力。不同结构的神经网络在训练算法和应用领域都有差异。该文主要就软件失效预测这个应用领域叙述几种适用的网络,并比较这几种网络在训练结果和性能上的差异。上述方法在SDH通信软件的失效预测中得到了成功的应用。试验结果显示虽然MLP、PNN、LVQ网络都能解决这类模式分类问题,但是只有MLP网络训练结果比较稳定,在不同的数据集上训练出的网络都有很好的预测效果。Predicting the fault - prone software module early in the software development can improve the software quality significantly. A common problem in software fault prediction is presence of the noise in the data. Neural networks are robust to have good tolerance to noise. Different neural networks differ both in training algorithm and in application field. This issue mainly discusses several types of neural networks in the field of software fault prediction, and compares these networks in the differences of training results and performances. The above approaches are applied to the fault prediction of SDH telecommunication software developed by Lucent Technology Optical Network Company. The results show that though MLP, PNN, LVQ could be used to solve this kind of pattern recognition problems, MLP is the most stable network. All the MLP networks trained by different data sets have good predicting accuracy.

关 键 词：软件质量预测 神经网络 主成分分析 

分 类 号：TP311.5[自动化与计算机技术—计算机软件与理论]