Deep Reinforcement Learning-Enabled Resource Allocation for UAV-Assisted Communications

Abstract

Unmanned Aerial Vehicles (UAVs) are increasingly employed in wireless networks to provide dynamic, on-demand connectivity, particularly in emergency and infrastructure-limited scenarios. This thesis presents a comprehensive AI-enabled framework that integrates user clustering, mobility modeling, and multi-agent reinforcement learning for optimizing UAV-assisted communications. The proposed system leverages a realistic user mobility model (STEP), silhouette-based K-Means clustering for UAV-UE association, and a hybrid reinforcement learning architecture combining Deep Q-Networks (DQN) and Multi-Agent Deep Deterministic Policy Gradient (MADDPG) to jointly optimize UAV placement, bandwidth allocation, and power control.

The research progresses through three stages: (1) joint resource allocation in a single-UAV static-user scenario; (2) power optimization in a multi-UAV static-user environment using user clustering and MADDPG; and (3) adaptive UAV deployment and resource scheduling in a dynamic-user setting. Simulation results demonstrate substantial improvements in data rate, UAV utility, and user coverage, with the hybrid DRL approach outperforming traditional baselines by up to 41%. The findings validate the potential of AI-driven, mobility-aware UAV coordination for scalable and intelligent next-generation wireless communication networks.