Research On Intelligent Offloading Strategy In D2D-Assisted Mobile Edge Computing Network

With the widespread adoption of intelligent devices and the rapid development of communication technology,the current communication platform architecture is becom-ing increasingly inadequate to meet the needs of users.Mobile Edge Computing(MEC)significantly reduces task latency and improves user’s quality of service by providing IT service environment and cloud computing capabilities at the edge of the wireless access network near users.Device-to-device(D2D)communication,as a key technology in the 5G era,can offload tasks to idle devices in the network and cooperate with MEC tech-nology to further improve system resource utilization.In the research of MEC system assisted by D2D,it is often necessary to consider the competition for resources among mobile devices and the collaborative task offloading of multiple agents under limited re-sources.How to design efficient task offloading strategies to meet the service needs of users in the scenario is one of the main challenges.The main work of this thesis is as follows:(1)To minimize the average task delay and drop ratio by decision-making task of-floading and bandwidth allocation in the D2D-MEC system under task deadline constraints,this thesis first dynamically divides active/idle devices based on device load to make them more flexible.Then,considering the fact that tasks may last for multiple time frames,a queue model is used to model the load.Next,according to the problem’s sequential rela-tionship,it is divided into two sub-problems: task offloading and bandwidth allocation.Faced with the massive action space of task offloading and the dynamic network envi-ronment,a multi-agent reinforcement learning algorithm based on Independent Proximal Policy Optimization(IPPO)algorithm is used to solve the problem.A dynamic program-ming algorithm based on group backpack is proposed to minimize the expected transmis-sion time of tasks according to the characteristic that devices only switch tasks in new time frames for bandwidth allocation problem.(2)To minimize the average task delay by decision-making task offloading and CPU frequency selection under battery capacity constraints in D2D-MEC system,this thesis in-troduces Energy Harvesting(EH)and Dynamic Voltage Frequency Scaling(DVFS)tech-nology to cope with energy constraints and enhance service reliability.Faced with a two-dimensional discrete action space and a dynamic network environment,a Multi-Agent Proximal Policy Optimization(MAPPO)multi-agent reinforcement learning algorithm is used under the centralized training + decentralized execution architecture to ensure the performance and timeliness of the algorithm,and dual-end Actor networks are used to output the offloading strategy and CPU frequency selection.(3)The performance of the proposed algorithms in this thesis is verified through sim-ulation experiments: ① Under the task deadline constraint,in the decision-making task offloading and bandwidth allocation in the D2D-MEC system,the proposed algorithm reduces the average delay and drop ratio by 11.2% and 22.7% respectively compared to the suboptimal Advantage Actor Critic(A2C)algorithm;② Under the battery capacity constraint,in the decision-making task offloading and CPU frequency selection in the D2D-MEC system,the decentralized execution architecture algorithm used in this thesis has no difference in performance compared with the centralized execution architecture algorithm in the average delay indicator,and it reduces 84.8% compared with the Proxi-mal Policy Optimization(PPO)algorithm,48.2% compared with the MEC mode,59.6%compared with the greedy algorithm,and 43.42% compared with the random algorithm.