可扩展的无人机多机飞行冲突解脱算法研究  

Study on the Scalable Intelligent Algorithm for UAV Multi- Actor Conflict Resolution

作　　者：张凯[1] 隋东[1] 邹国良[2] ZHANG Kai;SUI Dong;ZOU Guo-liang(Nanjing University of Aeronautics and Astronautics,Nanjing 211000,China;Civil Aviation Administration of China,Beijing 100000,China)

机构地区：[1]南京航空航天大学,江苏南京211000 [2]中国民用航空局,北京100000

出　　处：《航空计算技术》2022年第2期62-66,共5页Aeronautical Computing Technique

基　　金：中国民用航空局安全能力建设项目资助(TM 2018-5-1/2)。

摘　　要：随着无人机数量的增长,无人机间飞行冲突的自动解脱研究成为热点。针对冲突无人机数量变化的多机冲突解脱问题,采用一个集中式的深度多智能体强化学习算法——BiCNet算法。BiCNet算法包含参数共享机制和双向循环神经网络,使得冲突解脱模型可以支持动态扩展,即能够使用不同数量UAV的多机冲突场景进行训练,并理论上可以解脱任意数量冲突无人机的多机冲突,进而提高冲突解脱模型的训练效率。此外,基于微软面向无人机的开源仿真环境AirSim,设计了大量高密度的无人机多机冲突场景,并对冲突解脱模型进行了训练和测试。从实验的结果来看,训练曲线和测试结果表明解脱模型在求解时间和解脱率方面有很好的表现。With the increasing number of UAVs, the research on the automatic resolution of flight conflict between UAVs has become a hot spot.In this paper, a centralized deep multi-agent reinforcement learning algorithm, BiCNet, is adopted to solve the problem of multi-actor conflict resolution with the change of the number of conflicting UAVs.BiCNet algorithm includes parameter sharing mechanism and bidirectional recurrent neural network that can support the conflict resolution model to achieve the function of dynamic expansion, that is, the multi-actor conflict scenario of different numbers of UAVs can be theoretically used to train and test the resolution model.Moreover, BiCNet also can improve the efficiency of training together with all agents.Finally, based on Microsoft′s open-source UAV simulation environment AirSim, this paper designs a large number of high-density UAV multi-actor conflict scenarios, and tests the conflict resolution model.The experimental results show that the resolution model has a good performance in computing time and success rate.

关 键 词：无人机 空中交通管制 飞行冲突解脱 动态扩展 多智能体强化学习算法 

分 类 号：V279[航空宇航科学与技术—飞行器设计] V355.1