群体智能算法作为一种模拟自然界中种群智能行为的启发式优化算法，已被广泛应用于解决无人机集群问题。然而，随着5G 等前沿无线通信技术的突破，无人机集群的规模不断扩大，内部关系逐渐多样，应用场景也日益复杂，传统启发式群体智能算法缺乏理论保障的缺点逐渐显现，且性能效果明显下降，为无人机集群的策略优化带来挑战。近年来，深度学习的突飞猛进为群体智能提供了新颖有效的范式和框架，逐步成为无人机集群的核心驱动算法。本文以无人机集群的通信博弈场景为研究对象，设计深度学习算法，分别解决了无人机集群的三维区域覆盖、连通性快速自愈以及层级结构检测三个子问题，具体贡献如下：（1）我们针对攻方无人机集群的三维区域覆盖的问题进行研究。其中，我们通过构建DEM 模型，将三维区域的覆盖问题转化为二维Patch 选择问题，以降低问题复杂度。此外，针对无人机集群固有的联合空间维度高、邻居动态变化和观测元素维度不一致等挑战，我们提出了群体强化学习算法SDQN 进行解决，以得到更优的Patch 选择规划方案。仿真结果表明，SDQN 算法相比已有算法具有更短的覆盖时间和更高的覆盖率，从而验证了其有效性。（2）我们研究了在不可预测的损毁下攻方无人机集群的快速连通性恢复问题。具体而言，我们首先提出了一个图卷积操作?(⋅)，在理论上保证了无人机集群恢复后的连通性。其次，我们将?(⋅) 扩展为图卷积网络RGCN，并根据优化问题的拉格朗日函数设计了损失函数，以最小化连通性的恢复时间。此外，我们还设计了元学习算法以降低不同无人机集群拓扑下RGCN 的训练时间，从而提高算法的在线运行效率。之后，我们合并了元学习和RGCN 两部分算法，构建了完整的M-RGCN框架。实验结果表明，相比于现有的连通性恢复算法，提出的M-RGCN 框架可以在保证无人机集群的连通性可以恢复的基础上有效减少恢复时间。（3）我们还从守方的角度研究了针对攻方无人机集群的层级结构检测问题。由于所有无人机外观一致不可辨认，守方只能通过观测其飞行行为挖掘内部层级结构关系。为此，我们提出了GASSL 算法，其核心思想是通过自监督的方式寻找无人机集群的注意力中心以确定其层级结构。我们利用多进程技术对GASSL 算法进行开发实现，以降低其在线运行的时间复杂度。仿真实验中大量数值结果验证了GASSL 算法的有效性，并表明了多进程方式可以大幅降低算法的时间复杂度。

Swarm intelligence algorithm, as a heuristic optimization algorithm simulating the intelligent behavior of natural populations, has been widely used to solve unmanned aerial vehicle (UAV) swarm problems. However, with the breakthroughs in cutting-edge wireless communication technologies, such as 5G, the scale of UAV swarms is constantly expanding, the internal relationships are becoming increasingly diverse, and the application scenarios are becoming more complex. The shortcomings of traditional heuristic swarm intelligence algorithms, which lack theoretical guarantees and exhibit a significant decline in performance, have gradually become apparent, posing challenges for the strategic optimization of UAV swarms. Nonetheless, in recent years, the rapid development of deep learning has provided novel and effective paradigms and frameworks for swarm intelligence, gradually becoming the core driving algorithm of UAV swarms. This paper focuses on the communication game scenarios of UAV swarms. Specifically, we design deep learning algorithms to solve three sub-problems: 3D area coverage problem, rapid self-healing connectivity problem, and hierarchical structure recognition problem. The contributions of this paper are summarized as follows:(1) We investigate the problem of 3D area coverage for UAV swarms of the attacking side. Specifically, we constructed a digital elevation model (DEM) to convert the coverage problem of a 3D area into a 2D patch selection problem, in order to reduce problem’s complexity. In addition, we proposed the SDQN swarm reinforcement learning algorithm to address challenges such as the high joint space dimensionality, dynamic neighbor changes, and inconsistent observation element dimensions inherent in UAV swarms. This algorithm is designed to obtain optimal patch selection plans for better area coverage. The simulation results demonstrate that the SDQN algorithm is more effective than existing algorithms in terms of shorter coverage time and higher coverage rate, thereby validatingits effectiveness.(2) We also investigated the problem of rapid connectivity recovery for attacking UAV swarms under unpredictable destruction. Specifically, we first proposed a graph convolution operation ?(⋅) that theoretically guarantees the connectivity of the swarm after recovery. Next, we extended ?(⋅) to a graph convolutional network called RGCN and designed a loss function based on the Lagrangian function of the optimization problem to minimize the recovery time of the swarm’s connectivity. In addition, we designed a meta-learning algorithm to reduce the training time of RGCN under different UAV swarm topologies, thereby improving the online running efficiency of the algorithm. We then combined the meta-learning and RGCN algorithms to construct the complete M-RGCN framework. Experimental results show that, compared with existing connectivity recovery algorithms, the proposed M-RGCN framework can effectively reduce the recovery time while ensuring that the connectivity of the UAV swarm can be restored.(3) In addition, we investigated the problem of hierarchical structure detection for attacking UAV swarms from the defender’s perspective. Since all UAVs have a uniform appearance and are indistinguishable, the defender can only explore the internal hierarchical relationships of the swarm by observing its flight behavior. To address this problem, we proposed the GASSL algorithm, which uses a self-supervised approach to find the attention center of the UAV swarm and determine its hierarchical structure. We developed and implemented the GASSL algorithm using multi-processing techniques to reduce its time complexity during online operation. Numerical results from extensive simulation experiments demonstrated the effectiveness of the GASSL algorithm and showed that the multi-processing approach can significantly reduce the time complexity of the algorithm.