基于分层策略强化学习的多类型流量差异化路由优化

Routing is an important method for optimizing network resource allocation. However, traditional routing algorithms rely on static strategies to optimize single quality of service metrics, making it difficult to address the differentiated requirements of explosive growth in multi-type traffic. Although deep reinforcement learning has provided new ideas for routing optimization in dynamic network environments, existing methods still lack fine-grained perception of traffic types and cannot flexibly adjust routing strategies. To this end, this paper designs a traffic-aware routing algorithm based on hierarchical policy reinforcement learning for the differentiated routing requirements of different traffic types. First, a traffic classification module is introduced to achieve fine-grained perception of the differentiated service requirements of different traffic. Second, graph convolutional networks are used to efficiently model the network topology, based on which a hierarchical decision network and a differentiated reward function are designed to guide the agent to generate adaptive routing decisions and realize dynamic adjustment of routing strategies for each traffic category. Meanwhile, a global attention mechanism is introduced into the actor-critic framework to enhance the agent＇s ability to model the spatio-temporal dependency of network states, and the training efficiency and stability are improved through generalized advantage estimation and proximal policy optimization algorithms. Finally, the effectiveness of the proposed algorithm is verified on various network topologies.