A deep map from performance to resources - intelligent resource allocation for network slicing

Network slicing allows flexible and automatic provisioning of resources to satisfy contractually defined Key Performance Indicator (KPI) targets. When admitting new slices and during the slice run-time, the mapping between a KPI target and the resources required to fulfill it should be as accurate as possible. However, in mobile wireless networks, the number of resources to achieve a given network performance varies with the users' connection quality over time. An incorrect mapping may result in additional costs due to excessive over-provisioning or fines paid due to a contract breach. This work introduces a framework that predicts a cost efficient KPI target to resource mapping for a future time window. It does so by learning with network condition history and the slice's contractual agreement. Different operators can use the framework in both slice admission and run-time. Simulations show that the framework successively decreases costs related to an inefficient KPI target to resource mapping.