Deep Meta Advisor-aided Exploration for UAV Trajectory Design in Vehicular Networks

Unmanned Aerial Vehicles (UAVs) deployed as aerial base stations (UABS) offer an adaptable solution for enhancing network performance, especially in vehicular networks. A key challenge is optimizing UABS trajectories in these dynamic environments. While Deep Reinforcement Learning (DRL) algorithms show the potential to solve this issue, their outcome wildly depends on the exploration phase carried out at the beginning of training. To address this, we introduce a deep meta advisor that, by applying efficient adaptation across similar tasks, learns an optimal exploration policy using augmented state inputs as additional context. Numerical results show that our approach improves agents learning efficiency across multiple tasks by enhancing the exploration phase, allowing to reach target performance with fewer training episodes compared to existing methods.