摘要
分析用于软件质量评价的神经网络在训练时只有小样本的原因,及由此导致的网络泛化能力差和泛化能力无法定量估算等问题,提出输出一致假设,用“分析评价结果和表决结果”作为解决问题的途径,以已经训练出对小样本准确率高且未过拟合的少量相对合格网络为前提,依据蒙特卡洛模拟方法生成的大量随机数,采用表决成功率(VSR)和评价相同率均值(SR)反推出泛化能力最优的网络组合,用t分布定量估算最优网络组合泛化能力。实例结果表明,该方法基于小样本,能够挑选出对新软件项目泛化能力较强的最优网络组合,为从软件度量直接评价软件综合质量的工作和集成学习提供新思路。
After analyzing the reason why the neural network used for software quality evaluation has only small samples in trai-ning and the two resulting problems including the poor generalization ability and the inability,the generalization ability was estimated quantitatively,and the output consistency hypothesis was proposed,the analyzing evaluation results and voting results method was used to solve the problems.Taking a small number of relatively qualified neural networks that had been trained with high accuracy and had not been over-fitted for small samples as a premise,based on a large number of random numbers generated by Monte Carlo simulation,vote success ratio(VSR)and evaluation sameness ratio mean value(SR)were used to deduce the optimal network combination with best generalization ability in the opposite direction,and t distribution was used to estimate the generalization ability quantitatively.An example shows that the above method can select the optimal network combination with stronger generalization ability for new software projects based on small samples,an idea is provided for evaluating the comprehensive quality of software directly from software metrics and ensemble learning.
作者
陈鑫
CHEN Xin(Xi’an Aeronautics Computing Technique Research Institute,Aviation Industry Corporation of China,Xi’an 710068,China;Xi’an Kai-Xiang Computer Software Limited Company,Xi’an 710068,China)
出处
《计算机工程与设计》
北大核心
2023年第11期3380-3386,共7页
Computer Engineering and Design
关键词
小样本
神经网络
软件质量评价
输出一致假设
蒙特卡洛模拟
最优组合挑选
表决成功率
评价相同率均值
small sample
neural network
software quality evaluation
output consistency hypothesis
Monte Carlo simulation
optimal combination selection
vote success ratio(VSR)
evaluation sameness ratio mean value(SR)
