A deep reinforcement learning approach for flexible multipath routing using graph neural networks

Conventional dynamic routing protocols such as RIP and OSPF predominantly adopt singlepath transmission strategies, which often result in inefficient bandwidth utilization and network congestion, especially under dynamic traffic demands. To solve these problems, this paper proposes a flexible multipath routing optimization framework that integrates Deep Reinforcement Learning (DRL) with Graph Neural Networks (GNN). Specifically, the proposed approach introduces a multi-path decision-making mechanism that enables each traffic demand to be distributed across one or two selected paths combinations, with the goal of maximizing long-term network throughput. Then, a DRL agent is designed wherein the Q-value function leverages GNN-based state representations, capturing topological dependencies and link states. The action space of proposed agent is defined over combinations of candidate paths, allowing adaptive multi-path selection. Furthermore, a dynamic traffic allocation strategy is employed to proportionally split traffic based on the real-time residual capacities of selected paths which enhances load balancing and alleviates congestion. Simulation experiments conducted on realistic network topologies demonstrate that the proposed method achieves a 9.76% improvement in total allocated bandwidth compared to existing single-path DRL+GNN methods, validating its effectiveness in complex and dynamic network environments.