| dc.description.abstract | Machine learning components such as Deep Neural Networks (DNNs) are used extensively in Cyber-Physical Systems (CPS). However, such components may introduce new types of hazards that can have disastrous consequences and need to be addressed for engineering trustworthy systems. Although DNNs offer advanced capabilities, they must be complemented by engineering methods and practices that allow effective integration in CPS. This dissertation presents approaches for assurance monitoring of learning-enabled CPS as well as decision making. We present an approach for assurance monitoring of learning-enabled CPS based on the Inductive Conformal Prediction (ICP) framework. In order to allow real-time assurance monitoring, the approach employs distance learning to transform high-dimensional inputs into lower size embedding representations. Then, we discuss the use of sequential data on assurance monitoring and present a feedback loop that is used to query the sensors for a new input to further refine the predictions and increase the classification accuracy. Further, we present a selective classification approach that evaluates the trustworthiness of predictions and rejects those that cannot be trusted. We quantify the credibility and confidence of each prediction by computing aggregate statistical p-values from multiple subsequent inputs. We examine different multiple hypothesis testing approaches for combining p-values computed using ICP focusing on their ability to produce valid p-values for sequential data. Last, we use the Inductive Venn Predictors framework for computing probability intervals regarding the correctness of each prediction in real-time. We propose taxonomies based on distance metric learning to compute informative probability intervals in applications involving high-dimensional inputs. An adaptive approach allows for improving the quality of computed the probability intervals during runtime by pseudo-labeling input data. The experimental results demonstrate that all the presented approaches can form well-calibrated assurance monitors, take decisions that minimize the number of alarms and are computationally efficient allowing their use in real-time. | |
