基于GAT的分布式联合路由与频谱接入机制OA

To address the limited situational awareness and high decision coupling of traditional routing and spectrum ac-cess methods in dynamic network topologies,a joint optimization framework that integrates graph attention networks(GAT)with deep reinforcement learning(DRL)was proposed.The distributed path establishment process was formu-lated as a partially observable Markov decision process(POMDP),enabling hop-by-hop decentralized decisions via DRL.GAT was implemented to aggregate local observations to capture irregular topologies and inter-node interference,improving adaptability to complex environments.During training,prioritized experience replay enhances sample effi-ciency.Extensive simulations under random,clustered,and multi-flow scenarios demonstrate the method's effective-ness:in random topologies,it achieves approximately 10%higher bottleneck throughput while reducing both channel switching frequency and path hop count.In clustered topologies,it reduces channel switches by about 10%and hop count by about 13%,and in multi-flow scenarios,its performance is comparable to baseline approaches.