Graph-Based Machine Learning Estimation Methods for Backbone Optical Network Optimization

The development of new technologies, causing an immense increase in the amount of data transmitted through the backbone infrastructure, triggers the growing need for new, effective optimization methods. Routing and spectrum allocation (rsa) algorithms are the basis of network management, and new solutions aided by machine learning (ml) techniques are gaining popularity within the research community. However, broad testing is essential to validate the effectiveness of proposed methods across diverse traffic conditions, and to set expectations on how the network will operate in the upcoming days when using the chosen algorithm. On the contrary, the recently-surging solutions based on reinforcement learning (rl) fail to provide overwhelmingly better quality while being more complex than traditional methods. In this context, we address the problem of predicting the performance of heuristic-operated dynamic resource allocation algorithms in multilayer optical networks. We show how the massive scale of the dynamic rsa problem can be coped with using various aggregation methods to create a regression representation of the employed algorithm. Through broad experimental evaluation, we demonstrate the benefits of using graph representations with various meta-features to create versatile predictors independent of the number of connection requests and the physical topology. The proposed methodology allows for fast prototyping of new algorithms and quick estimation of the operation of existing ones with changing traffic conditions. The developed graph-based models achieve great prediction quality, and statistically outperform the baselines.