Sponsor: National Science Foundation
In modern management of critical infrastructures, such as transportation networks and power grids, machine intelligence is expected to perceive and understand system situations, track and infer system dynamics, control and mitigate system threats. For example, in power grid management, machine intelligence can assist in characterizing electrical distribution (what happened) and forecasting future electrical demand (what will happen); in traffic light control, machine intelligence can simulate traffic evolution under traffic light control policies (how it changes) and generate better control policies (how to change it) for safe and efficient transportation. This project will develop novel machine learning techniques to equip machines with the perception intelligence to understand what happened and what will happen, and the prescription intelligence to understand how it changes and how to change it. The research outcomes can help computers to better identify situations, extract semantics, forecast trends, detect anomalies, discover causality, simulate system behaviors, and, moreover, prevent, mitigate, and eliminate threats, which are important for the operations and defense of critical infrastructures. Furthermore, this research provides new courses, research, and internship opportunities for undergraduate, graduate, and underrepresented students.
The interconnected critical infrastructures can be viewed as dynamic network systems that generate big graph sequence data. Such data is an essential source of the perception and prescription intelligence. This project will develop a transformative framework that generalizes and unifies perception and prescription into a joint and interactive learning architecture. This project will address three fundamental research challenges: (1) How can a unified learning paradigm be designed to simultaneously perform perception and prescription in graph sequences? (2) Can the new learning paradigm be used to develop precise representation and reliable projection capabilities of graph sequences? (3) Can the new learning paradigm be used to develop system simulators and intervention planners with interaction and feedback capabilities? This project will result in new algorithms, including reinforced graph imitative embedding, adversarial confidence training, prescriptive intervention, and interactive learning with external and causal knowledge. The project will be complemented by a comprehensive evaluation plan with transportation networks, power grids, social networks data. This research effort will provide new exploration and insights into distillation, transfer, and feedback mechanism between predictive knowledge in perception and actionable knowledge in prescription.