Abstract: Aiming at the problems of weak adaptability of traditional flying ad hoc network protocols and poor effect in large-scale network application scenarios, a multi-mode routing algorithm based on deep reinforcement learning is proposed. The algorithm constructs the value function by comprehensively using the parameters such as system throughput, packet delivery rate and average end-to-end delay. The agent automatically adjusts the routing mode of each UAV, decomposes the large network into the main network and several small heterogeneous networks connected with it, reduces the system complexity, optimizes the local performance, and improves the performance of the whole network. The agent automatically adjusts the routing mode of each UAV, decomposes the large network into the main network and several small heterogeneous networks connected with it, reduces the system complexity, optimizes the local performance, and improves the performance of the whole network. Simulation results show that the algorithm is significantly better than the traditional protocol, and the larger the network scale and the higher the load, the more obvious the advantage is. The average throughput is increased by 55.46%, the packet delivery rate is increased by 39.85%, and the average end-to-end delay is reduced by 60.94%.

Key words: flying ad hoc network（FANET）, deep reinforcement learning（DRL）, adaptive routing algorithm, hybrid routing

摘要： 针对传统飞行自组网协议自适应能力不强、大规模网络应用场景效果不佳的问题，提出了一种基于深度强化学习的多模式路由算法。该算法综合利用系统吞吐量、分组递交率和平均端到端时延等参数构建价值函数，通过智能体自动调节各个无人机的路由工作模式，将大型网络分解为主体网络和数个与之相连的小型异构网络，降低了系统复杂度，局部性能达到最优，提升了整个网络的性能。使用NS3仿真平台测试了算法和传统协议AODV、DSDV的性能指标。仿真结果表明，算法显著优于传统协议，且网络规模越大、负载越高则优势越明显，平均吞吐量提升了55.46%，分组递交率提升了39.85%，平均端到端时延降低了60.94%。

关键词: 飞行自组网, 深度强化学习, 自适应路由算法, 混合路由