Deep Reinforcement Learning-Based Resource Allocation for Multi-UAV-Assisted Full-Duplex Wireless-Powered IoT Networks

In this paper, we investigate a resource allocation problem for a multi-unmanned aerial vehicle (UAV)-assisted full-duplex wireless-powered Internet-of-things (IoT) network, where the slot partition, power allocation, user association, and three dimensional (3D) UAV placement are jointly considered to maximize the sum bit rate of all IoT devices under the imperfect self-interference cancellation and generalized probabilistic air-ground channel model. To deal with the formulated mixed-integer non-convex problem, we propose a novel resource allocation strategy with three nested parts by integrating the model-based optimization theory with the data-based learning theory. Particularly, the data-based deep deterministic policy gradient algorithm is only explicitly used to train the 3D UAV placement policy, while the model-based Lagrange dual theory and matching theory are implicitly used to explore the hidden tractability of the rest two parts and design efficient algorithms, where the optimization results are passed onto the data-based part through reward values. Simulation results show that the proposed strategy greatly cuts down the execution time of the exhausting search-based genetic algorithm by 4 orders of magnitude at the cost of less than 5.1 percent performance loss.