Challenging the generalization capabilities of Graph Neural Networks for network modeling

Abstract

Today, network operators still lack functional network models able to make accurate predictions of end-to-end Key Performance Indicators (e.g., delay or jitter) at limited cost. Recently, a novel Graph Neural Network (GNN) model called RouteNet was proposed as a cost-effective alternative to estimate the per-source/destination pair mean delay and jitter in networks. Thanks to its GNN architecture that operates over graph-structured data, RouteNet revealed an unprecedented ability to learn and model the complex relationships among topology, routing and input traffic in networks. As a result, it was able to make performance predictions with similar accuracy than resource-hungry packet-level simulators even in network scenarios unseen during training. In this demo, we will challenge the generalization capabilities of RouteNet with more complex scenarios, including larger topologies.

Is Machine Learning Ready for Traffic Engineering Optimization?

“…The TE scenario described in Section II is implemented through a MARL architecture that, thanks to the use of GNN, yields good generalization properties over networks [19], [20]. We will first introduce our generic MARL+GNN framework, which is especially designed for distributed networking tasks, and then we will provide details on how it is adapted to the intradomain TE use case addressed in this paper.…”

“…Thus, after training, each agent v ∈ V can be interpreted as a replica of this universal agent that behaves based on its local environment. This so-called parameter sharing feature provides compelling generalization and scalability properties, which can be beneficial to effectively deploy the solution in networks with topologies of different size and structure, not necessarily seen during the training phase [20], [42].…”

“…Unlike previous ML-based proposals for TE (e.g., [11], [12], [18]), the combination of MARL and GNN allows us to handle topologies of various sizes and structures in a distributed fashion; and more importantly, to achieve combinatorial generalization over the information exchanged by agents in the network, which is naturally represented as a graph [19], [20].…”

Results and Achievements of the ALLIANCE Project: New Network Solutions for 5G and Beyond

“…In this context, we have applied RouteNet for automatic routing optimisation in several QoS-aware optimisation use cases (e.g., minimise end-to-end delay, jitter). Our experimental results show that unlike previous ML-based proposals, this GNN model can operate successfully in network scenarios with different topologies, routing configurations, and traffic never seen during the training phase [28]. Likewise, we have recently pioneered the first network optimisation architecture that combines DRL and GNN for routing optimisation [25], and have applied it to a classic routing optimisation problem in optical networks, as well as to traffic engineering in IP networks [29].…”

Learning Combinatorial Optimization on Graphs: A Survey With Applications to Networking

“…A limitation of the model is that it supports only topologies with identical link capacities. The model has been further explored by Suárez-Varela et al [91] and extended by Badia-Sampera et al [92] to model nodes (their queue size) too.…”