ABSTRACT:

Due to their dynamic nature and the multitude of involved parameters, network optimization problems are typically complex. The complexity of these problems is further increased by the distributed nature of these networks, coupled with the need to simultaneously take into account the variety of objectives and constraints, mainly related to the supported services. Conventional optimization methods generally fail to address these multiobjective problems, as they fail to provide scalable and robust solutions capable of dynamically adapting to changing network conditions. To overcome these obstacles, several machine learning-based approaches, and particularly reinforcement learning, have emerged in recent years. Although effective, these approaches are generally unable to grasp the graph topology of the networks and the richness of information, they also pain to generalize the learning, in addition to being unexplainable black boxes, which limits the practical utility of these approaches. In this context, Graph Neural Networks (GNNs) have emerged as a promising approach, leveraging the power of deep learning to process graph-structured data effectively, which is typically needed in several networking problems.

This tutorial will explore the applications of GNNs in network optimization problems such as virtual services’ placement for network slicing. We will discuss the ability of GNNs to capture graph structures and apprehend representations directly from network data to enable more efficient and intelligent resource allocation strategies. By exploring both the promises and limitations of GNNs, this talk aims to provide a comprehensive understanding of their applicability in network management, and opens the door for future research directions in this rapidly evolving field.