Dynamic UAV Swarm Networking: A Two-Stage Adaptive Learning-Based Approach

The swarm of Unmanned Aerial Vehicles (UAVs) has garnered considerable attention, particularly in scenarios with critical situations or limited communication infrastructure. In such cases, Mission UAVs (MUs) often be deployed in clusters to provide communication services. However, the high mobility of MUs always leads to frequent changes in swarm network topology, posing challenges for the network performance. To tackle this issue, we deploy additional Relay UAVs (RUs) with a two-stage adaptive learning-based approach. In the first stage, we employ a Delaunay triangulation-based algorithm to optimize RUs’ position and construct the initial topology. In the second stage, we implement a centralized learning and decentralized execution (CTDE) reinforcement learning framework to ensure continuous network connectivity and optimize performance throughout the task cycle. To further enhance cooperation among RUs, we introduce a sequential update technique coupled with an entropy regularization term during the policy network updates. Finally, extensive simulation results demonstrate the effectiveness of our proposed algorithms.