Research On Resource Allocation Method Of Edge Computing Based On Deep Reinforcement Learning Mechanism

Edge computing is a computing paradigm that can bring practical value to most modern enterprises.When it is integrated into the Internet of Things system,it can improve the Quality of Service(Qo S)of mobile applications,and can realize real-time management of the generated big data.5G promotes the development of the edge computing paradigm further,and mobile users can obtain low-latency and high-speed access Qo S.In this thesis,we study the resource allocation method of edge servers and service migration criteria in the Mobile Edge Computing(MEC)environment.Resource allocation method absorbed in solving the problem of allocation of limited resources in the MEC system,such as computing resources and available bandwidth,with the goal of maximizing the average resource utilization and task processing capacity of edge servers in the MEC system,while considering some differential processing for delay-sensitive applications,describe this resource allocation problem as a Finite-state Markov Decision Process(FMDP),and consider the continuity of the user state,and proposed a new Edge Computing Resource Allocation Algorithm based on Deep Deterministic Policy Gradient(ECRAA-DDPG)to find the optimal strategy for resource allocation.One of the challenges in deploying MEC service in cellular networks is to support user mobility,especially when moving at high speeds,so that offloaded tasks can seamless migration between BSs(Base Stations)without affecting resource utilization efficiency and link reliability.Considering virtualization,I/O interference between virtual machines,and multi-user access interference,the problem is described as a multi-objective optimization problem,a new method based on relaxation and rounding is proposed to maximize overall Qo S and minimize migration costs.This method includes the optimal iterative method for integer relaxation problems and the proposed of restoring integer solutions.Finally,a great quantity of test are used to bear out the performance of the new algorithm.The test results show that the method can make the optimal decision in a real scene.