Multi – task reinforcement learning for UAV-enabled urban systems: Balancing trajectory planning and communication fairness

Abstract

Enhancing multi-task performance in unmanned aerial vehicle (UAV)-enabled urban communication systems remains challenging due to conflicting objectives, particularly the trade-off between trajectory-planning efficiency and communication fairness. This paper addresses this issue by jointly optimizing UAV path planning and hovering communication tasks. A UAV Network Fairness-Efficiency Model (UFEM) is proposed to quantify overall system performance by integrating an energy-efficiency metric for trajectory planning with a communication-fairness index for user scheduling, thereby capturing the inherent trade-off between these objectives. Building on this framework, a reinforcement learning (RL)-based urban downlink communication system is developed to track dynamic operating conditions through three-dimensional position observations and state-transition functions, while incorporating key environmental uncertainties such as a random wind model and a probabilistic line-of-sight (LoS) model. Task-specific reward functions are further designed to balance competing objectives and enable adaptive task switching. Based on these components, we introduce the Multi-Task Reinforcement Learning for UAV Maneuvers (MRLUM) algorithm, which jointly optimizes path planning and communication scheduling by fusing flight-state information and communication-channel data through an adaptive task-switching strategy. Simulation results demonstrate that MRLUM significantly improves both trajectory-planning efficiency and communication fairness under the UFEM metric, offering a promising solution for UAV-enabled urban communication systems facing multi-task conflicts and environmental uncertainties.