User and resource allocation in latency constrained Xhaul via reinforcement learning

Abstract

The Flexible Ethernet (FlexE) is envisioned for the provisioning of different services and hard slicing of the Xhaul in 5G and beyond networks. For efficient bandwidth utilization in the Xhaul, traffic prediction for slot allocation in FlexE calendars is required. Further, if coordinated multipoint (CoMP) is used, the allocation of users to remote units (RUs) with an Xhaul path of lower latency to the distributed unit/central unit will increase the achievable user bit rate. In this paper, the use of multi-agent deep reinforcement learning (DRL) for optimal slot allocations in a FlexE-enabled Xhaul, for traffic generated through CoMP, and for offloading users among different RUs is explored. In simulation results, the DRL agent can learn to predict input traffic patterns and allocate slots with the necessary granularity of 5 Gbps in the FlexE calendar. The resulting gains are expressed in terms of the reduction of mean over-allocation of slots in the FlexE calendar in comparison to the prediction obtained from an autoregressive integrated moving average (ARIMA) model. Simulations indicate that DRL outperforms ARIMA-based prediction by up to 11.6%