Dynamic bandwidth allocation based on adaptive predictive for low latency communications in changing passive optical networks environment

Abstract

The industry is considering redesigning the architecture of optical access networks (PONs) to address the challenges of 5G developments. One of the major challenges is to achieve low communication delay. To better address this challenge, PONs not only need to be improved from the perspective of underlying facilities, but also need to combine with effective resource allocation strategies. This paper proposes a dynamic bandwidth allocation strategy based on adaptive predictive algorithm. First, the amount of arriving packets at each ONU during each polling cycle is predicted using the XGBoost ensemble learning algorithm. As such, arriving packets can be allocated bandwidth in advance, so as to achieve low communication delay. Simulations verify that the DBA algorithm based on XGBoost is basically the same as the DBA algorithm based on neural network algorithm in terms of delay performance, but requires less computing time. At the same time, for increasing the robustness of the algorithm, dealing with the changing PONs environment (e.g., the number of ONUs online in PONs changes over time caused by sleep mode), we use dynamic perceptual algorithm which uses the ideas of reinforcement learning to adjust the result of the predictive algorithm, this perceptual algorithm helps to reduce the predictive error caused by environmental changes, so as to reduce the error impact on achieving low delay. Through simulation experiments with different PON environments, the effectiveness of the proposed algorithm is verified and the delay effect of the original predictive algorithm is improved by nearly 10% with the adaptive predictive algorithm.