Machine Intelligence in Supervising Bandwidth Allocation for Low-latency Communications

This paper presents the exploitation of an artificial neural network (ANN) to facilitate insights into existing bandwidth allocation schemes in optical access networks and supervise bandwidth allocation decisions that reduce the latency. Specifically, based on the classic and predictive dynamic bandwidth allocation (DBA) schemes, we train a multi-layered ANN at the central office (CO) to learn the uplink latency corresponding to varying bandwidth allocation decisions. Multiple network feature knowledge, such as network load, traffic/packet statistics, fiber link distances and the number of optical network units (ONUs), is for the first time considered and utilized in the training process. Then, with the dependency between bandwidth allocation and the resultant latency learned by the ANN, we numerically analyze the latency performance of existing DBA schemes and show the optimal bandwidth decisions supervised by the ANN in achieving low latency. With extensive simulations, we show that exploiting the ANN to supervise bandwidth allocation at the CO, termed as ANN-DBA scheme, effective improvement in latency performance is realized.