Deep reinforcement learning for SDN routing optimization enhanced by graph multi-head attention

Emerging applications impose stringent quality of service (QoS) requirements on the Internet. Advances in traffic classification, software-defined networking (SDN), and programmable network devices have facilitated the rapid identification of user requirements and the implementation of fine-grained routing control. Meanwhile, the heterogeneity of diverse service flows necessitates the collaborative optimization of multi-objective QoS. However, the existing solutions still lack sufficient capability to optimize traffic forwarding paths to meet such demands in online scenarios. To address this challenge, this paper presents GMHA-DRL, a deep reinforcement learning (DRL)-based routing algorithm incorporating a Graph Multi-Head Attention (GMHA) network. Built upon the actor-critic framework, the proposed algorithm tackles inherent limitations through two core architectural designs. First, a customized graph multi-head attention module is devised to extract features from network topology and traffic types efficiently, substantially enhancing the model's capability to capture complex network states and improving its generalization. Second, a service-oriented multi-objective reward function is constructed to support iterative routing policy updates through sustained interaction with the network environment. This approach satisfies multi-dimensional QoS requirements and equips the routing algorithm with strong dynamic adaptability. We implemented GMHA-DRL under the SDN architecture and evaluated it on the Mininet platform using real network topologies and traffic traces. Experimental results demonstrate that the algorithm effectively adapts to the QoS requirements of various service types and exhibits excellent adaptability and generalization across different network topologies and load scenarios.