Multi-Agent Reinforcement Learning for Resilient UAV Ad Hoc Backhaul Networks

In disaster scenarios, reliable emergency communication is essential for rapid medical response and public safety; however, terrestrial infrastructure is often damaged or congested. This article proposes a cooperative uncrewed aerial vehicle (UAV) ad hoc network architecture driven by multi-agent reinforcement learning (MARL) to jointly support wide-area exploration and resilient backhaul formation. We design a multiobjective reward that balances: 1) landmark discovery for expanding coverage in affected areas; 2) multihop connectivity to maintain end-to-end reachability to a ground base station (BS); and 3) mutual distance regularization to avoid excessive clustering and isolated UAVs. Using multi-agent proximal policy optimization (MAPPO), we evaluate the emergent UAV deployment behaviors under various reward-weight settings and clarify the characteristic tradeoffs between exploration and connectivity. Furthermore, after determining the UAV placements, we assess the interference-limited link quality by computing the signal-to-interference-plus-noise ratio (SINR) under multiple propagation models, including free-space, two-ray ground reflection, and 3GPP urban macro aerial vehicle (UMa-AV)/urban micro aerial vehicle (UMi-AV) LOS models. The simulation results demonstrate that appropriate reward balancing enables UAVs to reach distant targets while preserving multihop connectivity and that multipath-induced altitude sensitivity can cause severe SINR degradation in two-ray environments. These findings provide practical insights for designing learning-driven UAV-assisted emergency communication systems with robust connectivity and radio-aware deployments.