Optimal Dynamic Trajectories for UAVs in Mobility-Enabled Relay Systems

We consider a UAV which acts as a mobile relay and must plan a trajectory to enable data transfer between multiple pairs of communication nodes. For each pair, successful data transfer is only possible in non-convex regions (relay regions) where a given quality of service requirement may be satisfied for both nodes. Trajectories consist of 1) locations where the UAV stops to relay (relay positions) and 2) a dynamic transition policy which determines the sequence in which the pairs are serviced. We are interested in minimizing the average time a bit waits at a source before being sent to a destination and first pose a general, non-convex problem that calls for optimization over both the relay positions and the dynamic transition policy. To find approximate solutions, we formulate an average cost semi-Markov decision process and propose a deep-reinforcement-learning-based algorithm to solve it. To validate our approach, we present the results of several simulation experiments, which show our approach significantly outperforms the state-of-the-art.