Reinforcement learning-driven multicast routing in Elastic Optical Networks: A multi-objective cost optimization framework

Abstract

Data traffic demand is increasing every year, and networks must respond to these increasing demands. On the other hand, Wavelength Division Multiplexing (WDM) networks, due to their static allocation, waste resources and are unable to meet the necessary requirements. To solve this problem, Elastic Optical Networks (EONs) with dynamic spectrum allocation have been developed, and it has been shown that EONs can enhance resource utilization relative to WDM-based networks. The main challenge in EONs is efficiently allocating resources to optimize network capacity utilization. On the one hand, Dynamic spectrum assignment and diverse available schemes for demands lead to higher spectrum utilization. Multicasting, also known as multi-destination data transfer, is one of the most cost-effective and efficient techniques for providing flow in computer networks, a technology attracting considerable attention due to the widespread adoption of internet-based services. This paper introduces a novel RL-driven framework for multi-objective cost optimization in dynamic multicast routing. We propose two algorithms: MIN (Multicasting through Intermediate Nodes) and MIN2P (via 2 Paths), which integrate hop count, physical distance, and frequency slot congestion into a unified cost function. MIN leverages intermediate nodes (destination and non-destination) via SPT algorithm for logical tree construction, while MIN2P enhances load balancing by splitting traffic across two paths with first-fit and last-fit spectrum allocation. For the first time, a Q-learning-based reinforcement learning (RL) agent adaptively tunes the cost function weights in real-time, responding to network conditions to minimize blocking probability. Simulations on NSFNET and JPN12 topologies, with varying traffic loads (100–800 Gbps), demonstrate that hop count is the most influential metric, reducing blocking rates by up to 30% compared to benchmarks like OL-M-SFMOR and MT3A. MIN excels in low-congestion scenarios, while MIN2P outperforms in high-load, large topologies due to effective load distribution. This framework provides actionable insights for designing adaptive, efficient EONs, advancing multicast capabilities in next-generation optical networks.