Timely Admission Control for Network Slicing in 5G with Machine Learning

For guaranteeing the strict requirements foreseen for 5G, network slicing has been proposed as a dynamic and scalable mechanism for the allocation of customized resources to service providers. Many solutions have been proposed in the literature for the scenario where multiple service providers share the same pool of resources, while the exclusive allocation to different providers is still an open issue due to the associated complexity. In this work, we define a policy-based admission mechanism for exclusive intra-service slice allocation, at fine and adaptable timescales. In particular, we consider the case where optimal admission strategies are pre-computed offline for network state conditions that are representative of typical traffic loads and resource availability. This offline phase is also used to train a Machine learning algorithm; a neural network (NN) learns the best admission policies from a more computationally expensive mechanism in previously studied network conditions. Thus, the NN is used for providing near-optimal admission decisions at runtime under network conditions for which no optimal policy has been computed. The potential of the 5G marketplace in terms of revenue and quality of service is demonstrated for the particular case of services with strict latency constraints by means of a proof of concept tested over network traces from a real network operator. Different strategies are compared for the computation of the admission strategies and results are provided in terms of efficiency in resource utilization, fairness to the service providers, network owners’ revenue and complexity. This study confirms the feasibility of a policy-based approach for exclusive intra-service resource allocation, especially if computationally-efficient mechanisms are adopted in the case of missing information about network states.