α-OMC: Cost-Aware Deep Learning for Mobile Network Resource Orchestration

Abstract

Orchestrating resources in 5G and beyond-5G systems will be substantially more complex than it used to be in previous generations of mobile networks. In order to take full advantage of the unprecedented possibilities for dynamic reconfiguration offered by network softwarization and virtualization technologies, operators have to embed intelligence in network resource orchestrators. We advocate that the automated, data-driven decisions taken by orchestrators must be guided by considerations on the cost that such decisions involve for the operator. We show that such a strategy can be implemented via a deep learning architecture that forecasts capacity rather than plain traffic, thanks to a novel loss function named α-OMC. We investigate the convergence properties of α-OMC, and provide preliminary results on the performance of the learning process in case studies with real-world mobile network traffic.