From power grids to transportation and logistics systems, autonomous systems play a central, and often safety-critical, role in modern life. Even as these systems grow more complex and ubiquitous, we have already observed failures in autonomous systems like autonomous vehicles and power networks resulting in the loss of human life [1]. Given this context, it is important that we be able to verify the safety of autonomous systems prior to deployment; for instance, by understanding the different ways in which a system might fail and proposing repair strategies. Human designers often use their knowledge of likely failure modes to guide the design process; indeed, systematically assessing the risks of different failures and developing repair strategies is an important part of the systems engineering process [2]. However, as autonomous systems grow more complex, it becomes increasingly difficult for human engineers to manually predict likely failures. In this paper, we propose an automated framework for predicting, and then repairing, failure modes in complex autonomous systems. Our effort builds on a large body of work on testing and verification of autonomous systems, many of which focus on identifying failure modes or adversarial examples [3, 4, 5, 6, 7, 8], but we identify two major gaps in the state of the art. First, many existing methods [4, 5, 9, 7] use techniques like gradient descent to search locally for failure modes; however, in practice we are more interested in characterizing the distribution of potential failures, which requires a global perspective. Some methods exist that address this issue by taking a probabilistic approach to sample from an (unknown) distribution of failure modes [6, 10].

Support Vector Data Description (SVDD) is a popular one-class classifiers for anomaly and novelty detection. But despite its effectiveness, SVDD does not scale well with data size. To avoid prohibitive training times, sampling methods select small subsets of the training data on which SVDD trains a decision boundary hopefully equivalent to the one obtained on the full data set. According to the literature, a good sample should therefore contain so-called boundary observations that SVDD would select as support vectors on the full data set. However, non-boundary observations also are essential to not fragment contiguous inlier regions and avoid poor classification accuracy. Other aspects, such as selecting a sufficiently representative sample, are important as well. But existing sampling methods largely overlook them, resulting in poor classification accuracy. In this article, we study how to select a sample considering these points. Our approach is to frame SVDD sampling as an optimization problem, where constraints guarantee that sampling indeed approximates the original decision boundary. We then propose RAPID, an efficient algorithm to solve this optimization problem. RAPID does not require any tuning of parameters, is easy to implement and scales well to large data sets. We evaluate our approach on real-world and synthetic data. Our evaluation is the most comprehensive one for SVDD sampling so far. Our results show that RAPID outperforms its competitors in classification accuracy, in sample size, and in runtime.