Addressing Concept Drift of Dynamic Traffic Environments through Rapid and Self-Adaptive Bandwidth Allocation

Passive optical networks are envisioned to become increasingly complex as they support more and more diverse and immersive services that have different capacity, latency, and reliability needs. In the near term, they are expected to support the delivery of a diverse and immersive set of services including mixed reality, holographic communication, human-to-machine/robot communications, Tactile Internet, and digital sensing. However, in supporting these diverse and immersive services, traffic on the network will become increasingly dynamic across a range of different time scales. The upstream bandwidth in a passive optical network is typically shared by a group of end users, meaning that the uplink latency performance as experienced by each end user is thus highly dependent on the amount and when bandwidth to that end user is allocated. Machine learning enhanced bandwidth allocation algorithms have been proposed but are typically stationary, primarily-designed or pre-trained based on certain network configurations. In dynamic network conditions where traffic can evolve over time, concept drift, a phenomenon whereby the underlying distribution of the training data will no longer be representative of that in deployment, may occur. In view of future dynamic network conditions, we present a novel online reinforcement learning based bandwidth allocation scheme to address concept drift in machine learning enhanced passive optical network. The scheme facilitates self-adaptive decisions in real-time to accommodate dynamic network environments with varying traffic types and network loads. Results from comprehensive performance evaluation of the scheme show that rapid and self-adaptive bandwidth decisions can be achieved, yielding ~ 60% latency improvement in dynamic traffic environments.