View article

Time-division-multiplexing passive optical networks (TDM-PONs), with their massive deployment worldwide, are considered a fundamental technology for supporting not only traditional Internet broadband services, but also for new emerging 5G latency-sensitive services, such as Ultra-Reliable and Low Latency Communications (URLLC). Traditional dynamic bandwidth allocation (DBA) mechanisms, currently used to allocate network resources in TDM-PONs, are not suited to meet the requirements of these new services with strict latency requirements, as they use a polling mechanism which can result in a high queuing delay and ultimately violate URLLC latency requirements. In this work, we propose a new predictive-based DBA mechanism for Gigabit Symmetrical PON (XGS-PON) that allows to reduce the latency to fulfill requirements of emerging latency-sensitive services. Our solution employs reinforcement learning (RL) to predict the ingress buffer occupancy of ONUs in the next DBA cycle. Results show that the proposed RL method outperforms traditional DBA approaches in terms of upstream delay while maintaining similar frame loss ratio.