Continuous Exploration of Infinitely Configurable Cyber-Physical Systems for Sample-based Testing (Co-InCyTe)

当今的软件包含多达数千种配置选项，可根据不同的要求、环境和平台进行调整。每个（布尔）选项使配置空间的大小加倍，这使得每个单独配置的质量保证几乎不可能。采样策略通过定义用于选择优选的小的但具有代表性的配置子集的标准和算法来绕过这个问题。不幸的是，最近的采样方法还没有为即将到来的安全关键网络物理系统（CPS）时代做好准备：（1）它们限于有限的配置空间，而CPS实际上是无限可配置的，（2）它们需要明确指定有效配置空间的配置模型，该配置模型实际上不可用于CPS，（3）它们以一次性方式生成样本，需要预先对配置空间有完全的了解，这对于CPS也是不可行的，（4）它们应用黑盒选择标准，而CPS的有效采样在没有附加的领域知识的情况下是不可能的，例如，非功能属性。拟议的项目提供了一种新的抽样方法，全面应对挑战（1）-（4）。对于（1），我们提出了一种新的配置模型，它集成了两个概念：扩展到无限配置空间的特征模型和训练的分类器，以处理特别复杂的配置约束。关于（2），我们采用从已知配置集合中提取配置模型的技术。为了解决（3），我们交错配置模型提取和样本选择，以不断探索未知的配置。最后，关于（4），我们以行为模型的形式使用进一步的解决方案空间知识，并应用基于族的分析来识别关键配置。

Today's software comprises up to thousands of configuration options to adjust to diverse requirements, contexts and platforms. Each (Boolean) option doubles the size of the configuration space which makes quality assurance of every individual configuration practically impossible. Sampling strategies bypass this issue by defining criteria and algorithms for selecting preferable small, yet representative subsets of configurations. Unfortunately, recent sampling approaches are not ready for the upcoming era of safety-critical cyber-physical systems (CPS): (1) they are limited to finite configuration spaces, whereas CPS are literally infinitely configurable, (2) they require a configuration model explicitly specifying the valid configuration space which is not available for CPS in practice, (3) they generate samples in a one-shot manner requiring complete knowledge of the configuration space in advance, %which is also infeasible for CPS, (4) they apply black-box selection criteria, whereas effective sampling of CPS is impossible without additional domain knowledge, e.g., non-functional properties. The proposed project contributes a novel sampling methodology comprehensively tackling challenges (1) -- (4). For (1), we propose a novel configuration model integrating two concepts: feature models extended to infinite configuration spaces and trained classifiers to handle particularly complicated configuration constraints. Concerning (2), we adapt techniques for extracting configuration models from sets of known configurations. To tackle (3), we interleave configuration-model extraction and sample selection for continuously exploring unknown configurations. Finally, concerning (4), we use further solution-space knowledge in form of behavioral models and apply family-based analysis to identify critical configurations.