Applications of graph neural networks in modeling and decision-making of dynamic networks

Abstract

A dynamic network is a system whose components are interrelated in time and space. Dynamic networks play is an important tool for modeling various man-made and natural systems. However, modeling and decision-making of dynamic networks via analytic methods are demanding tasks due to their inherent complexity. Recent advances in hardware and theoretical advances in deep learning techniques have enabled the modeling and decision-making of complex systems using data-driven techniques instead of using analytical techniques. This dissertation studies the applications of graph representation learning to model, plan/control, design, and meta-learn (dynamic) networks. The first case introduces a graph representation learning method that models networks with fixed points, the second case deals with sequential decision-making (planning) on dynamic networks to solve combinatorial optimization problems, the third case models wind farm networks and utilizes the model to solve network design problems and the fourth case studies an efficient mete-learning method and applies it to the non-stationary networks. Through the four cases, we experimentally confirmed that the graph expression learning method is more effective than the conventional analytical method and other deep learning methods on various dynamic networks.