Research On Multi-task Offloading Strategy Based On Deep Reinforcement Learning

With the rapid development of artificial intelligence and wireless communication,emerging applications represented by image recognition,natural language processing,and virtual reality have received extensive attention.However,current user equipment is limited by computing resources and energy resources,and it is difficult to meet the performance requirements of these applications.From Mobile Edge Computing to Multi-access Edge Computing(MEC),MEC technology can quickly respond to user service requests by deploying computing,storage and other resources close to the terminal side,effectively reducing task process delays and improving user experience.Computing offloading strategy is the core component of edge computing network architecture,and designing appropriate computing offloading strategy can effectively improve network performance and user experience.Existing computing offloading strategies mainly consider the optimization of a single computing task and a single performance objective,ignoring the cooperative relationship between multiple tasks in heterogeneous systems and the problem of differentiated performance requirements.Deep reinforcement learning has the characteristics of learning while interacting in a dynamic environment,and it is suitable for solving the computational offloading decision-making problem of dynamic environment changes in edge heterogeneous networks.Based on deep reinforcement learning,this paper studies the multi-task offloading decision-making problem in edge heterogeneous networks from different scenarios.The main research contents are summarized as follows.(1)Aiming at the multi-access multi-task edge-device hybrid computing offloading scenario,a multi-task hybrid offloading strategy based on deep reinforcement learning is proposed.Make global offloading decisions for multiple computing tasks,offload them to edge computing servers and adjacent terminals,and optimize the long-term total system delay.At the same time,the feature information of task and network state is extracted by using the recurrent neural network,which improves the convergence speed and stability of the deep reinforcement learning model.The simulation results show that the proposed algorithm can effectively reduce the total long-term system delay and terminal energy consumption.(2)On the basis of research content 1,a multi-objective computing offload optimization strategy based on multi-agent deep reinforcement learning is proposed,considering the scenarios where different users have different optimization objectives in the cloud-edge-deviceheterogeneous network.Two reinforcement learning agents are designed to make global offloading decisions for a set of delay-sensitive tasks and energy-sensitive tasks,respectively,and offload tasks to cloud centers,edge servers or nearby idle terminals,and implement differentiated performance requirements for users.optimization.The simulation results show that the proposed algorithm can meet the differentiated performance requirements of different users,and has better performance than the delay-energy weighted collaborative optimization algorithm.In this paper,for different edge heterogeneous network scenarios,and according to different optimization objectives,the corresponding computational unloading decision algorithms are designed respectively,which has profound practical significance for constructing mobile communication networks with low delay,low energy consumption and high performance.