静态多目标软件缺陷预测策略研究

发布时间：2018-08-04 18:30

【摘要】：软件缺陷预测是根据历史数据以及已经发现的缺陷等软件度量数据预测未来开发软件缺陷性的技术。本文针对软件缺陷预测做了以下工作:首先,针对缺陷检出率和误报率这两个目标,已有的研究主要集中于多目标微粒群算法,但该算法生成规则的分类器包含多条规则,每条规则分布不均匀,规则之间有重复区域覆盖,且生成的规则需要进行组合才能对软件缺陷进行预测。这些问题在很大程度上影响算法性能。为此,论文引入多目标定向布谷鸟算法用以优化预测模型的参数,布谷鸟算法的个体当前位置即为其个体历史最优位置,这一特性使得该多目标算法生成的规则分布较为均匀,从而改善了算法性能。为了验证该算法,论文选择了NASA数据库的八个数据集,并与八个比较算法进行比较,实验结果表明,基于MOOCS的多目标软件缺陷预测技术的平均性能较优。其次,我们对NASA数据库的六个数据集中的模块代码行数分布不均衡性分布进行了分析,结果发现大部分软件模块的代码行数较少(称为小模块),仅有少量模块的代码行数较长(称为大模块)。根据Arisholm的研究结论:软件模块的测试资源大小和软件模块的代码行数成正比,大模块在缺陷预测过程中产生的误报率会严重浪费测试资源。为了降低这种由于大模块误报而导致的资源浪费,我们将一个数据集按照代码长度分为大模块和小模块,并各自分配一个支持向量分类器,以达到降低总体测试资源浪费的目标。为了进一步验证技术性能,论文将该算法与九个算法进行比较,结果表明了基于多目标双支持向量机软件缺陷预测技术的有效性。最后,基于双支持向量机的多目标缺陷预测需要将一个数据集依据代码行数分割为大模块和小模块,这个比例的选择对缺陷预测的效果影响很大,因此我们采用黄金分割法对五个数据集的分割比例做了选择,实验结果表明,该比例在40%-80%之间,效果较优。
[Abstract]:Software defect prediction is a technology to predict future development software defects based on historical data and the detected defects, such as defects. In this paper, the following work is done for software defect prediction: firstly, the research focuses on the two targets of defect detection rate and false alarm rate, which are mainly focused on the multi-objective particle swarm optimization algorithm, but this calculation is calculated. The classifier of the rule generating rules contains many rules, each rule is not evenly distributed, and the rules have repeated area coverage, and the rules that are generated need to be combined to predict the software defects. These problems affect the performance of the algorithm to a great extent. The individual current position of the cuckoo algorithm is the optimal location of individual history, which makes the rule distribution of the multi-objective algorithm more uniform and thus improves the performance of the algorithm. In order to verify the algorithm, eight data sets of the NASA database are selected and compared with the eight comparison algorithms. The experimental results are compared. It shows that the average performance of the MOOCS based multi target software defect prediction technology is better. Secondly, we analyze the distribution of the unbalance distribution of the number of code lines in the module of the six data sets of the NASA database. The results show that most of the software modules have less code lines (called small modules), and only a small number of modules have a long number of code lines. According to the research conclusion of Arisholm: the test resource size of the software module is proportional to the number of code lines in the software module, the false alarm rate generated by the large module in the defect prediction process is a serious waste of test resources. In order to reduce the resource waste caused by the large module misinformation, we put a data set in accordance with the code. The length is divided into large modules and small modules, and a support vector classifier is allocated each to reduce the waste of the overall test resources. In order to further verify the technical performance, the paper compares the algorithm with the nine algorithms. The results show the effectiveness of the software defect prediction technology based on the multi target dual support vector machine software. Finally, The multi target defect prediction based on dual support vector machine (SVM) needs to divide a data set into large modules and small modules based on the number of code lines. The ratio selection has a great influence on the effect of defect prediction. Therefore, we choose the proportion of five data sets by the golden segmentation method. The experimental results show that the ratio is in 40%-80%. The effect is better.
【学位授予单位】：太原科技大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP311.5

【参考文献】