Computer Engineering and Applications ›› 2024, Vol. 60 ›› Issue (11): 336-345.DOI: 10.3778/j.issn.1002-8331.2302-0383

• Engineering and Applications • Previous Articles     Next Articles

Joint Optimization of UAV Trajectory and Resource Allocation for Federal Learning

YAO Xiancai, ZHENG Jianchao, ZHENG Xin, YANG Xiaolong   

  1. 1.College of National Defense Science and Technology, Academy of Military Science, Beijing 100071, China
    2.School of Information and Communication Engineering, Beijing Information Science and Technology University, Beijing 100101, China
  • Online:2024-06-01 Published:2024-05-31

面向联邦学习的无人机轨迹与资源联合优化

姚献财,郑建超,郑鑫,杨小龙   

  1. 1.军事科学院 国防科技创新研究院,北京 100071
    2.北京信息科技大学 信息与通信工程学院,北京 100101

Abstract: Unmanned aerial vehicles (UAVs), due to their mobility and flexibility, are widely used in tasks such as search and tracking. The large amount of data generated during task execution can be significantly leveraged by machine learning (ML) algorithms to improve the intelligence of UAV clusters. Federated learning (FL), as a distributed machine learning architecture, is more suitable for UAV networks with limited bandwidth and energy budget, as it only requires model parameter transmission during training. To fully exploit the advantages of FL in UAV networks, this paper models factors affecting training energy consumption, such as bandwidth, computational frequency, FL accuracy, and training delay. A joint optimization algorithm is proposed to minimize the overall training energy consumption of FL by jointly optimizing training parameter settings, UAV trajectories, resource allocation of communication and computation in UAV networks. The proposed joint optimization algorithm decomposes the mixed integer nonlinear integer programming problem (MINLP) into three sub-problems, and transforms the non-convex sub-problems into convex sub-problems through methods such as successive convex approximation (SCA) and block coordinate descent to obtain suboptimal solutions. Simulation results show that, based on the constraints of training accuracy, UAV movement, and FL delay, the proposed algorithm reduces the overall training energy consumption of unmanned aerial vehicles by more than 15% compared to the existing joint training and resource optimization schemes.

Key words: federal learning, unmanned aerial vehicles (UAVs), reducing energy consumption, trajectory optimization, resource allocation

摘要: 无人机(unmanned aerial vehicle,UAV)因其机动、灵活的特点被广泛地应用于搜索、跟踪等任务。在任务执行过程中会产生的大量数据,通过机器学习(machine learning,ML)算法加以利用,可以显著提高无人机集群的智能。联邦学习(federated learning,FL)作为一种分布式机器学习架构,因训练过程中只需要进行模型参数传输,更加适用于带宽和能量预算有限的无人机网络。为充分发挥无人机网络中联邦学习的优势,从影响训练能耗的带宽、计算频率、联邦学习精度和训练延迟等因素进行建模,通过联合优化训练参数设置、无人机轨迹、无人机网络中通信以及计算资源分配来最小化联邦学习整体的训练能耗。提出的联合优化算法将所建模的混合整数非线性规划问题(mixed integer nonlinear integer programming problem,MINLP)分解为三个子问题,并对非凸的子问题进行转化,通过逐次凸逼近(successive convex approximation,SCA)和块坐标下降等方式求得问题的次优解。仿真结果表明,在满足训练精度、无人机运动和联邦学习延迟等约束的基础上,提出的算法与现有的联合训练与资源优化的方案相比,降低无人机的整体训练能耗15%以上。

关键词: 联邦学习, 无人机群, 降低能耗, 轨迹优化, 资源分配