Research On Dynamic Deployment Of UAV For MEC

Battlefield information changes rapidly in military operations,so the processing of combat information must satisfy the criteria for ultra-low latency.Technology known as mobile edge computing(MEC)can successfully address the issues listed above,and it can support reconnaissance robots to achieve low-latency and low-energy computing requirements.However,there are high mobility requirements for reconnaissance robots on the battlefield,and fixed server base stations are easily damaged by the enemy.The capabilities of fixed edge server base stations in this scenario are limited,so we consider carrying edge servers on unmanned aerial vehicles(UAVs).Combining the high mobility of UAVs with the powerful computing capabilities of MEC provides the possibility for rapid combat.However,the battery capacity of UAVs is limited,which results in issues like inadequate service time and incomplete services in the service process of UAVs equipped with servers.Therefore,it is necessary to plan reasonably for their deployment trajectory.The battlefield environment changes in real time,and traditional optimization algorithms cannot achieve real-time planning.Deep reinforcement learning(DRL)has strong learning ability and adaptability.According to the service requirements of reconnaissance robots,DRL can support the real-time deployment of UAVs equipped with edge servers,so as to solve the problem that the computing power of reconnaissance robots cannot meet the computing needs.Based on this,the primary tasks are as follows:(1)In terms of the dynamic deployment of a single UAV.This paper suggests an improved DDQN algorithm that takes into account the existence of multiple obstacles on the ground and the constraints of UAV energy consumption and realizes the dynamic deployment of a single UAV edge server.The goal is to maximize the geographic location fairness of the reconnaissance robot and the obstacle avoidance of the UAV.First,formulate the problem of serving multiple ground terminals with a single UAV and propose optimization objectives.Second,in order to realize the training and learning of DRL,a Markov decision model must be established.Finally,to address the algorithm’s shortcoming of slow convergence speed,an action exploration strategy based on greedy-pseudo count(ε-pseudo count)is proposed based on the DDQN algorithm.The simulation results demonstrate that the UAV can smoothly avoid obstacles,and the improved double deep Q network algorithm has a higher average system reward and a faster convergence speed.In terms of geographical location fairness for reconnaissance robots,the improved DDQN algorithm outperforms Q-learning,DQN,and DDQN by 50%,20%,and 15.38%,respectively.(2)In terms of multi-UAV dynamic deployment.The DDQN algorithm cannot be applied due to the continuous action space considered in the multi-UAV scene,resulting in an action space dimension explosion.Therefore,for the complex scene of multiple obstacles and multiple UAVs,this paper implements the dynamic deployment of multi-UAV edge servers based on the DDPG.By planning the deployment trajectory of the multiple UAVs,the coverage of the multiple UAVs to the reconnaissance robot,the energy utilization of the multiple UAVs,and the obstacle avoidance of the multiple UAVs are maximized.In order to achieve this optimization goal,the mathematical model and Markov decision model of the system must be established.And to achieve the dynamic deployment of multiple UAVs,the DDPG algorithm is used.The outcomes of simulations demonstrate that multi-UAV can successfully avoid obstacles.In terms of reconnaissance robot coverage,the deployment algorithm based on DDPG outperforms AC,A3 C,and PG by 112.50%,21.43%,and 142.86%,respectively.In terms of energy utilization of the UAV cluster,the deployment algorithm based on DDPG is 28.57%,12.50% and 50% higher than AC,A3 C and PG,respectively.