Reinforcement Learning-Based Dynamic Wavelength Allocation in Elastic Optical Networks
Abstract
Revolution in Internet services, 5G/6G apps and cloud computing has led to unparalleled growth of traffic on optical backbones. Elastic Optical Networks (EONs) are projected to bypass the shortcomings of the fixed-grid Wavelength Division Multiplexing (WDM) by providing a flexible yet granular spectrum. Nevertheless, in case of very variable traffic, the issue of dynamic wavelength allocation (DWA) is still open. Conventional heuristic and optimization algorithms are disadvantaged because of the inability to respond to realtime traffic conditions. To cope with this, we present an RLbased dynamic wavelength allocation framework in EONs. The problem of allocation is cast in an MDP where, by Q-learning, an agent can learn without a prior mind of the traffic patterns and learn optimal allocation policies. Thorough simulations of the NSFNET and COST239 topologies show that the RL-based DWA can considerably minimize the likelihood of blocking, improve spectrum utilization, and optimization of fairness of the requests requests compared to First-Fit and Random-Fit heuristics. Also, both scalability and robustness testing ensures that the proposed approach is applicable to larger EON deployments.