Graph Neural Network for Building Prediction Agents in Intent-Based Zero-Touch Networks

Intent-based network operation is an essential paradigm to create autonomous networks as it eliminates the complexity of manual configurations and replaces it with an abstracted, automated method. To enable intent-based networks that can handle and resolve conflicts under shared and limited network resources, prediction of the impact of proposed configuration and topology changes is crucial. In this respect, Graph Neural Networks (GNN)s are gaining attention in the networking domain thanks to their ability to process data based on close to real modeling of networks as a graph. In this paper, we develop a GNN framework of intent-based networks that takes the network information as inputs and predicts the specific key performance indicators (KPIs) of intents. We propose to make use of GNNs for accurate prediction of the impacts of proposed actions on all of the active KPIs of an intent-based network. Experiments with different configurations are conducted and the empirical evaluations demonstrate that a GNN based prediction agent outperforms a baseline neural-network based agent in both prediction performance and ability to generalize to previously unseen configurations.