数据驱动的软件可靠性模型研究

Research on data-driven software reliability models

传统的软件可靠性增长模型通常基于一系列不符合实际情况的假设,因而其适用性和准确性受到了较大的影响。近年来,数据驱动的软件可靠性楚模方法受到了广泛的重视,一些文献提出了基于人工神经网络和基于支持向量机的软件可靠性模型。由于数据驱动的软件可靠性模型不对软件的内部错误及失效过程做任何假设,因而其适用范围较传统的软件可靠性增长模型更广。对于数据驱动的软件可靠性模型,其使用的软件失效数据对模型的预测精度有很大的影响,但现有的文献未对此进行研究。本文提出了一种基于支持向量机的软件可靠性模型,并指出对数据驱动的软件可靠性模型,应使用累计失效数据而不应使用失效问隔数据,应使用最近的部分失效数据而不应使用所有观测到的失效数据。本文同时还提出了一种基于遗传算法的模型参数优化方法。应用三组现有文献中的实际软件测试数据,对所提出的软件可靠性模型与现有的数据驱动的软件可靠性模型进行了对比研究。结果表明,本文所提出的软件可靠性模型具有较好的预测精度。

Traditional software reliability growth models（SRGMs）are generally based on several impractical assumptions,which to a large extend limits their applicability and accuracy.In recent years,data-driven approach to software reliability modeling has attracted a lot of attention,and several artificial neural network（ANN）based and support vector machine（SVM）based software reliability models（SRMs）have been proposed in the literature. Data-driven SRMs require no assumptions on the properties of software faults and software failure process,thus they appear to have wider applicability compared with SRGMs.For data-driven SRMs,software failure data used have great impact on model prediction accuracy;however,to the best of our knowledge,this issue has not been studied in the literature.In this paper,an SVM-based SRM is proposed.It is also demonstrated that for data-driven SRMs accumulative software failure data rather than inter-failure data should be used,and recent failure data rather than all historical failure data should be used.A genetic algorithm（GA）based algorithm for optimizing model parameters is proposed.Based on three failure data sets published in the literature which are taken from real-life software projects,comparative studies of the proposed SRM and existing data-driven SRMs are conducted.Results show that the proposed SRM seems to have the highest prediction accuracy.