Towards using Deep Reinforcement Learning for Connection Steering in Cellular UAVs

This paper investigates the fundamental connection steering issue in cellular-enabled Unmanned Aerial Vehicles (UAVs), whereby a UAV steers the cellular connection across multiple Mobile Network Operators (MNOs) for ensuring enhanced Quality-of-Service (QoS). We first formulate the issue as an optimization problem for minimizing the maximum outage probability. This is a nonlinear and nonconvex problem that is generally difficult to be solved. To this end, we propose a new approach for solving the optimization problem based on Deep Reinforcement Learning (DRL), considering two important reinforcement learning algorithms (i.e., Deep Q-Learning (DQN) and Advantage Actor Critic (A2C)). Simulation results show that under the proposed approach, the UAVs can make optimal decisions to select the most suitable connection with MNOs for achieving the minimization of the maximum outage probability. Furthermore, the results also show that in our new approach, the A2C-based algorithm is better than the DQN-based one, especially when the number of MNOs increases, while the DQN-based algorithm can be executed in a shorter time.