Deep Reinforcement Learning for Proactive Spectrum Defragmentation in Elastic Optical Networks

Spectrum fragmentation is a critical challenge in elastic optical networks (EONs), where inefficient spectrum utilization can lead to increased blocking probability (BP).This paper presents DRL-Defrag, an online Deep Reinforcement Learning (DRL) framework for proactive spectrum defragmentation (SD), designed to reduce service blocking ratio (SBR) .Unlike conventional heuristics, which typically handle only parts of the defragmentation process and offer limited adaptability, DRL-Defrag jointly addresses the three key aspects of SD: deciding the defragmentation starting point in both time and connections position and determining the new allocation spectrum slot .By incorporating service attributes, fragmentation metric to capture the spectrum occupancy state, and the reconfiguration cost, DRL-Defrag consistently selects suitable actions dynamically adapting modification to traffic conditions.Three DRL agents (DQN, A2C, and PPO) are employed to learn optimal reallocation strategies without pre-collected training data.Extensive simulations on the NSFNET, USNET, and EURO28 topologies demonstrate substantial improvements, reducing the SBR by 59.2%, 49.49%, and 61.63%, respectively, compared with heuristic methods.