Learning on Graphs for Resilience Decision-Support in Real-World Networks
Abstract
Project Abstract: Approved for Public ReleaseLearning on Graphs for Resilience Decision-Support in Real-World NetworksPrincipal Investigator: Jie Zhang, The University of Texas at Dallas (UTD)Team Members: Yulia R. Gel (UTD), Souma Chowdhury (University at Buffalo)Motivation: Massive networks of complex interdependent systems, such as critical infrastructure and emerging distributed systems, are known to be widely vulnerable to natural and anthropogenic hazards. Resilience of these networks depends on their ability to recover quicklyafter disruption, through operational reconfiguration of the network and restoration of impaired elements of the ne Most existing optimization or heuristics based decision-support frameworks for post-event resilience are not able to adapt to the large scale and complex topology of networks, and time-critically handle the high-dimensionality and stochastic characteristics of the underlying combinatorial optimization problems. Machine learning techniques, which can learn to solve combinatorial problems overgraphs, in general provide promising opportunities to fill this gap. Reported success with such techniques however remain limited to scenarios where the graph structure is simple (e.g., homogeneous and mono-layered), labeled samples are abundant or inexpensive toevaluate, and the problem is deterministic. In contrast, real-world networks tend to present large, heterogeneous, multi-layered structures and uncertain behavior, and are usuallyexpensive to simulate.Objective: The overall goal of our proposed project can be stated as: To develop a unified learning on graphs framework that enables generation of robust, scalable, and generalizable combinatorial policies over graphs, to support real-time decision-making for recovery of disrupted real-world networks. In addition, we aim to use networked microgrid resilience as the demonstrative application to evaluate the effectiveness and potential impact of our proposed methods, at unprecedented levels of complexity.Technical Approaches: To accomplish our goal, we will pursue a four-year projectwith integrative research activities that can be divided into three major technical thrusts: 1) abstraction of topological descriptors of graphs; 2) integrated Artificial Intelligence (AI) framework with graphconvolutional networks, reinforcement learning and topological inputs/outputs, to generate network reconfiguration and restoration policies; and 3) training and evaluation setup/process on grid resiliency problems. The central intellectual merit of our proposed methods for learningon graphs lies in the ability to tackle head-on the complex combinatorial optimization problems underlying operational resilience in real-world networks.Expected Impact on DoD Capabilities: Tractable learning over graphs provides a unique ability to move the burden of planning across a wide range of possible extreme scenarios to offline computations, while providing fast schemes to support time-critical online decisions. Whileour approaches are to be evaluated over microgrid resiliency problems, the proposed contributions could be translated to solving many other combinatorial problems of societal importance e.g., recovery of transportation networks and deployment of multi-robotic disaster response solutions. The developed approach can assist decision making associated with mission planning, re-planning and execution of Naval operational missions at the tactical and operational command levels. Immediate impact of the project will be the enhanced reliability and resiliency of microgrids in the presence of natural disasters, cyber attacks, and contingencies.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Jun 09, 2021
- Source ID
- N000142112530
Entities
People
- Jie Zhang
Organizations
- Office of Naval Research
- United States Navy
- University of Texas at Dallas