Machine-learning-aided resource allocation in 5G metro networks

The 5G is emerging to address very stringent and heterogeneous requirements of numerous new services and applications. In particular, urban areas introduce challenging requirements in terms of throughput, latency, and reliability to enhance the performance of the new use cases related to eMBB, uRLLC and mMTC. In addition, mobile data traffic exhibits repetitive patterns with spatiotemporal variations thanks to the highly predictable daily movements of large populations of citizens in urban areas. Hence, the upcoming applications require a dynamic and flexible optical metro network capable of integrating network and IT resources to carry the network demands ensuring service performance and network efficiency. Such flexibility can be achieved thanks to software-based technologies as Network Function Virtualization (NFV) and Software-Defined Networks (SDN); however, the network reconfiguration is not immediate because of the time to compute and assign the required resources. Machine learning techniques can therefore be used to help the allocation of resources by predicting the traffic expected ahead of time. This chapter analyzes two use cases that exploit machine-learning-aided optimization to allocate resources. The first use case is an optical metro network used as the backbone for the mobile service, we present a dynamic resource allocation exploiting the programmability leveraged by SDN. We use the predicted traffic variation to solve o✏ine mixedinteger linear programming instances of an optical routing and wavelength assignment optimization problem. Results demonstrate the e↵ectiveness of the method, such that the prediction-based optical routing reconfiguration optimization matches almost perfectly the behavior with an oracle-like traffic prediction. The second use case is the allocation of Virtual Network Functions (VNFs) that implement the RAN baseband functions over a metro network. We propose a mixed-integer linear programming optimization that uses the hourly predicted traffic to place the VNFs with the goal of minimizing the network operators’ costs. Results show that the proposed machine-learning-based optimization is able to efficiently compute the resource assignment in advance without significant losses for the operators in terms of costs and performance degradation.