SHF: Small: Causal Foundations of Statistical Fault Localization

这项研究的目标是提高自动化技术的有效性，这些技术寻求在软件中定位导致测试或操作使用期间发生的观察到的故障（故障）的故障，以便可以修复故障。这个目标很重要，因为正常运行的软件在商业、通信、国家安全、交通、科学和许多其他活动中都是至关重要的。所期望的改进是通过采用最近跨多个学科开发的方法来实现的，以便能够准确和无偏见地估计各种治疗、暴露或干预（例如，医学治疗）对感兴趣的结果（例如，疾病）的因果影响。如果成功的话，这项研究有可能帮助软件开发人员有效地定位和修复他们产品中的故障，从而防止经济损失、伤害甚至死亡等危害。该研究还将有助于在软件工程社区传播合理的因果推理方法。具体而言，本研究将探讨和完善因果统计故障定位（CSFL）的基础，包括因果模型的形式、因果状态的抽象和迭代的处理。将开发一种基于值的CSFL方法，包括分析和分析程序变量的值，并将其与基于谓词的CSFL相结合，以便更准确地估计程序元素导致失败的影响。我们将探索一种新的CSFL方法，该方法采用多层统计模型来集成不同类型和粒度级别的执行数据，既适用于给定的程序版本，也适用于跨版本。为了考虑分数分布的特征和预测个人分数可信度的因素，我们将对用CSFL获得的怀疑分数集应用元分析技术。还需要研究的是选择偏差问题如何影响不同环境下的SFL技术，以及如何减轻它。最后，本研究将探讨病例对照方法在软件故障不频繁的情况下提高SFL的成本效益的潜在价值。

The goal of this research is to improve the effectiveness of automated techniques that seek to locate the faults in software that caused observed failures (malfunctions) to occur during testing or operational use, so the faults can be repaired. This goal is important because properly functioning software is critical in business, communications, national security, transportation, science, and many other activities. The desired improvements are to be achieved by employing methodology that has been developed recently, across several disciplines, to enable the causal effects of various kinds of treatments, exposures, or interventions (e.g., medical treatments) upon outcomes of interest (e.g., diseases) to be estimated accurately and without bias. If successful, the proposed research has the potential to help software developers to efficiently localize and repair faults in their products, thereby preventing harms such as economic loss, injury, and even death. The research will also help to disseminate sound causal inference methodology in the software engineering community. More specifically, the research will investigate and improve the foundations of causal statistical fault localization (CSFL), including the form of causal models, the abstraction of causal states, and the handling of iteration. A value-based approach to CSFL will be developed, which involves profiling and analyzing the values of program variables, and this will be integrated with predicate-based CSFL, in order to more accurately estimate the failure-causing effects of program elements. A new approach to CSFL will be explored that employs multilevel statistical models to integrate execution data of different types and granularity levels, both for a given program version and across versions. Meta-analysis techniques will be applied to the set of suspiciousness scores obtained with CSFL, in order to take account of features of the score distribution and of factors that predict the credibility if individual scores.Also to be investigated is how the problem of selection bias affects SFL techniques in different settings and how it can be mitigated. Finally, the research will explore the potential value of case-control methodology for improving the cost-effectiveness of SFL in scenarios where software failures are infrequent.