Intelligent UAV-RIS-Enabled D2D Networks for Disaster Recovery: A Reinforcement Learning Approach

Abstract

In disaster-stricken areas, establishing reliable wireless communication remains a critical challenge due to damaged infrastructure and dynamic user conditions. This article investigates a temporary device-to-device (D2D) communication network over a disaster zone, where a reconfigurable intelligent surface (RIS) is mounted on a hovering uncrewed aerial vehicle (UAV) and nonorthogonal multiple access (NOMA) is employed to serve multiple users simultaneously. A temporary base station (TBS) is deployed to initiate communication, while the UAV-RIS enhances coverage and mitigates blockage effects. An analytical framework is developed to characterize the outage probability and throughput performance under Rician fading, considering both static and mobile multiuser scenarios through a mobility-aware, time-correlated channel model. To maximize the network throughput, a model-free machine learning (ML) framework is proposed, termed exploration-attenuated Levenberg–Marquardt stochastic gradient (EA-LMSG). The proposed framework is lightweight, adaptive, and data-efficient and jointly optimizes the UAV-RIS position and power allocation by integrating reinforcement learning (RL) with a neural network trained using the Levenberg–Marquardt (LM) algorithm for fast convergence. The learning agent dynamically adapts UAV positioning and allocates transmission power on a block-by-block basis according to real-time channel conditions and user mobility. Simulation results demonstrate that the proposed EA-LMSG framework significantly improves throughput performance while maintaining robustness under user mobility and imperfect channel state information (CSI), thereby validating its effectiveness for adaptive and reliable disaster-resilient communications.