MUESLI: Multi-objective Radio Resource Slice Management via Reinforcement Learning

5G Radio Access Network (RAN) slicing concerns strategies to share radio resources while guaranteeing differentiated service requirements. Current state of the art approaches make use of strict isolation or dedicated RAN physical resource block (PRB) partitioning among slices to ensure differentiated services. However, spectrum multiplexing may be rendered suboptimal due to isolation of resources; it further cannot handle variations in traffic patterns or intents in a dynamic way. In this paper, we propose a flexible multi-service partitioning strategy that can balance functional isolation and optimal sharing of resources. This system, called Muesli: Multi-objective Radio Resource Slice Management, makes use of model-based reinforcement learning techniques to dynamically modify PRB partitions. The reinforcement learning reward structure ensures that the system is trained to meet multiple objectives such as network slice Service Level Agreement (SLA) compliance, spectrum usage efficiency and fairness among customer classes. On a real use case from Ericsson, the throughput levels for individual services are shown to be optimized with accurate PRB partitioning.