Deep Reinforcement Learning-Based Intelligent Resource Management in Multi-UAVs-Assisted MEC Emergency Communication System

This paper investigates a multi unmanned aerial vehicles (UAVs) assisted mobile edge computing (MEC) emergency communication system in which each UAV acts as a mobile MEC server for computing tasks offloaded by ground sensor users. Considering the stochastic dynamic characteristics of multi-UAVs-assisted MEC systems and the precision of spectrum resources, the deep reinforcement learning (DRL) algorithm and the non-orthogonal multiple access (NOMA) techniques are introduced. Specifically, we design an offloading algorithm based on a multi-agent deep deterministic policy gradient that jointly optimizes the UAVs' flight trajectories, the sensors' offloading powers, and the dynamic spectrum access to maximize the number of successfully offloaded tasks. The algorithm employs the Gumbel-Softmax method to effectively control both the discrete sensor access action and the continuous offloading power action. Sufficient simulation results show that the proposed algorithm performs significantly better than other benchmark algorithms.