DRL for Network Slicing: Resource Efficiency Optimisation under Reconfiguration Avoidance

Network Slicing (NS) can satisfy the differentiated service demands of vertical industries by tailoring a common infrastructure to multiple isolated logical networks. Such slices may need to be reconfigured given the varying traffic patterns, service requirements, device mobility and resource constraints. Frequent slice reconfigurations lead to service quality degradation for slice consumers. However, avoiding reconfigurations of existing slices leads to fewer deployed slices, in comparison to allowing reconfigurations. To mitigate this problem in the avoidance of reconfiguration, this paper proposes using Deep Reinforcement Learning (DRL) to optimise successful slice deployments over time, while avoiding reconfigurations. The DRL agent instructs a Linear Programming (LP) algorithm to deploy slices to maximise the chance of successful slice deployments over time. The results show that more slices are deployed compared to avoiding reconfiguration with LP-only algorithms. Our DRL-based slice orchestrator provides approximately 11% more slices, compared to existing slice orchestrations without such DRL guidance, demonstrating the potential advantage of DRL in NS applications.