SHF: Small: Automatically Localizing Functional Faults In Deployed Software Applications

尽管大多数软件应用程序在发布给客户之前都经过测试，但这些应用程序仍然包含导致现场故障的生产（或现场）功能故障，这些故障具有昂贵的后果并且修复成本高昂。由于其局限性，现有的自动调试方法不能充分隔离和识别现场故障的生产故障。 之前对测试经理的采访和对错误存储库的研究表明，程序员平均花费近50%的时间来定位生产故障，这是软件系统和软件项目失败的主要因素。该项目的教育创新是通过将概率图模型应用于软件工程问题来开发一种综合教学方法。该提案的目标是创建一个新的理论基础，允许利益相关者仅使用症状（例如，输出值的符号是不正确的），并且不需要装备部署的应用程序来收集运行时数据，从而避免了部署运行时开销，并且不需要使用Oracle进行任何测试来发现故障，不需要执行对比成功和失败的运行，并且不需要从现场故障收集运行时数据。有了这个理论基础，研究人员可以更紧密地合作，规划未来的故障定位的概率图形模型的基础上扩展彼此的结果作为共同的抽象。 仅基于在给定应用程序的部署期间发生的故障症状，将确定源代码中的故障位置，以及从可能的故障到可以修复这些故障的代码的导航路径。该项目将创建、评估和部署：（1）用于自动获得近似软件应用程序特定故障模型的概率图形模型的新理论、算法和技术;（2）使用基于模型的鉴别诊断来执行溯因推理以在给定现场故障的症状的情况下定位生产故障的新颖方式，以及（3）用于评估用于生产故障定位的算法的有效性的综合实验框架。除了本地化生产功能故障之外，该实现还可以用作广泛的实验平台，用于创建和测试各种软件调试和测试想法的假设，例如，用于指导测试选择和优先级排序。

Even though most software applications are tested before they are released to customers, these applications still contain production (or field) functional faults that result in field failures, which have costly consequences and are expensive to fix. Due to their limitations, existing automatic debugging approaches do not adequately isolate and identify production faults for field failures. Prior interviews of test managers and studies of bug repositories revealed that programmers spent close to 50% of their time on average to localize production faults, which is a major factor in software system and software project failures. The educational innovation of this project is in developing an integrated approach to teaching by applying probabilistic graphical models to software engineering problems. The goal of this proposal is to create a novel theoretical foundation that allows stakeholders to predict and localize functional faults for field failures automatically with a high degree of precision using symptoms only (e.g., the sign of the output value is incorrect) and without instrumenting deployed applications to collect runtime data, thus avoiding the deployment runtime overhead, and without having any tests with oracles to uncover the fault, without performing contrasting successful and failed runs, and without collecting runtime data from field failures. With this theoretical foundation, researchers can collaborate more closely in planning the future of fault localization by expanding each other's results based on probabilistic graphical models as common abstractions. Based only on failure symptoms occurring during deployment of a given application, the location of faults in the source code will be determined, as well as navigation paths from likely faults to the code that can fix these faults. The project will create, evaluate and deploy: (1) new theories, algorithms and techniques for automatically obtaining probabilistic graphical models that approximate specific fault models for software applications; (2) a novel way in which model-based differential diagnoses are used to perform abductive reasoning to localize production faults given symptoms for field failures, and (3) a comprehensive experimentation framework for evaluating the effectiveness of the algorithms for localizing production faults. In addition to localizing production functional faults, the implementation can be used as a broad experimental platform for creating and testing hypotheses for various software debugging and testing ideas, e.g., for guiding test selection and prioritization.