Enhanced resource allocation in elastic optical networks using VAE-GNN and deep reinforcement learning
Year: 2026
Type: article
Abstract: Elastic optical networks (EON) have emerged as a promising solution for supporting heterogeneous and time-varying bandwidth demands through flexible spectrum allocation. Nevertheless, efficient resource management in EONs remains challenging due to dynamic traffic conditions, spectrum fragmentation, and the need to jointly optimize multiple objectives such as blocking probability, transmission delay, energy consumption, and spectrum utilization. In this paper, we propose VAE-GNN-RL, an integrated framework that combines a Variational Autoencoder (VAE) for stable traffic data augmentation, a Graph Attention Network (GAT) for topology-aware prediction of spectrum requirements and network performance indicators, and a Deep Reinforcement Learning (DRL) stage for adaptive real-time spectrum allocation. The proposed framework is evaluated on the NSFNET topology under dynamic traffic scenarios with Poisson-distributed arrivals and exponentially distributed holding times. Experimental results show that the proposed approach achieves very high predictive performance, reaching 99.93% classification accuracy and an average mean squared error of 0.0037, while consistently outperforming the GAN-ESN baseline and several conventional deep learning models. In addition, the adaptive optimization stage improves allocation efficiency and network performance under dynamic operating conditions. These findings demonstrate that the proposed VAE-GNN-RL framework provides an effective and intelligent solution for topology-aware and adaptive resource allocation in elastic optical networks.
Source: Optical Fiber Technology
Institutions: University of Batna 2, Badji Mokhtar-Annaba University
Language: English
Subfield: Electrical and Electronic Engineering
Field: Engineering
Domain: Physical Sciences
Open Access status: hybrid
Estimated APC paid: $2,360
OpenAlex ID: W7221050795