Dynamic Trajectory Planning for UAV-Assisted Visible Light Communications Using Deep Recurrent Q-Networks

Unmanned Aerial Vehicles (UAVs) assisted Visible Light Communication (VLC) systems offer a promising solution for high-speed data transmission and illumination in applications such as traffic offloading and post-disaster rescue. This work presents a framework for dynamic trajectory planning in a UAV-assisted VLC system. We cast this dynamic trajectory planning as a multi-objective Partially Observable Markov Decision Process (POMDP), in which each UAV must (i) maximize its cumulative VLC capacity at the service points, (ii) minimize the risk of collisions while navigating a 3D environment with stochastically placed obstacles and serving ground terminals, and (iii) minimize the energy consumption. To solve this POMDP, we introduce a Deep Recurrent Q-Network (DRQN) to aggregate temporal dependencies. The resulting policy enables UAVs to navigate complex 3-D environments more safely and with higher communication capacity than memory-less Deep Q-Network (DQN) or simulated annealing (SA) baselines in both single- and multi-agent settings.