Routing in Elastic Optical Networks Based on Deep Reinforcement Learning

Abstract

Modern optical network schemes consist of eighty channels involving fixed grids and offering speeds of up to 400 Gbps per channel rendering the allocated bandwidths insufficient during certain timeframes and underutilized during others. These speeds will become inefficient when compared to the accelerated growth of traditional traffic, which threatens to be tripled by 2025. In order to exploit the capacity of elastic optical networks, efficient algorithms are essential. This article assesses Deep reinforcement learning RMCSA algorithms, focused on providing high availability to the allocation of routing, spectrum and modulation resources over a given network topology, in order to determine how to allocate physical resources in optical transport networks and increase capacity while accounting for traffic variability and demands. Reinforcement learning is currently used in many scientific fields for the solution of different problems. In optical networks, this work will use this method to determine the best path and resources for each type of traffic-related demand. Simulations are performed by increasing volumes of variable traffic over a fourteen-node topology. The results show a significant 40% improvement in blocking probability for 22 links with 100 frequency slots compared to current transmission systems. They also reflect low latency, higher speeds, and high availability, which can deliver better quality of service for the end user.