NSF Award
2427265
This project concerns the development of efficient and reliable algorithms for detection of anomalies in data. This topic is of current in interest in many industries such as computer systems management, healthcare, and power systems. For these three examples, anomalies may represent a serious threat, such as a data breach, a medical emergency, or collapse of the power grid. Research will focus primarily on online settings for which speed of response is critical, such as early detection of events that may lead to disaster. The goal is not only to identify a sudden change but also anticipate the onset of undesirable behavior. Dissemination will be strengthened through local workshops, for which the PI has been active since 2012. In particular, the annual “workshop on cognition and control” organized by the PI has attracted speakers from all over the globe.<br/><br/>There is a long history of research on algorithms for threat detection, with most of the research focused on algorithm design based on statistical models of threatening behavior. While such approaches may bring both complexity and fragility, the large literature on quickest change detection offers enormous insights on appropriate architectures for change detection algorithms in a model-free setting. Several approaches will be explored based on recent advances in machine learning, notably reinforcement learning and gradient free optimization. In addition, it is believed that development of these approaches within the domain of threat detection will help advance these general approaches to machine learning.<br/><br/>This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.
