Bayesian Optimization Enhanced Deep Reinforcement Learning for Trajectory Planning and Network Formation in Multi-UAV Networks

원문 초록

In this paper, we employ multiple UAVs coordinated by a base station (BS) to help the ground users (GUs) to offload their sensing data. Different UAVs can adapt their trajectories and network formation to expedite data transmissions via multi-hop relaying. The trajectory planning aims to collect all GUs' data, while the UAVs' network formation optimizes the multi-hop UAV network topology to minimize the energy consumption and transmission delay. The joint network formation and trajectory optimization is solved by a two-step iterative approach. Firstly, we devise the adaptive network formation scheme by using a heuristic algorithm to balance the UAVs' energy consumption and data queue size. Then, with the fixed network formation, the UAVs' trajectories are further optimized by using multi-agent deep reinforcement learning without knowing the GUs' traffic demands and spatial distribution. To improve the learning efficiency, we further employ Bayesian optimization to estimate the UAVs' flying decisions based on historical trajectory points. This helps avoid inefficient action explorations and improves the convergence rate in the model training. The simulation results reveal close spatial-temporal couplings between the UAVs' trajectory planning and network formation. Compared with several baselines, our solution can better exploit the UAVs' cooperation in data offloading, thus improving energy efficiency and delay performance.