Research On Edge Computing Task Scheduling Method Based On Reinforcement Learnin

Edge computing extends cloud computing services to the edge of the network,and is considered an effective solution for handling computationally intensive tasks due to its low latency and high computational performance.Developing a reasonable edge computing task offloading algorithm is crucial for ensuring quality of service in edge computing scenarios.Deep reinforcement learning has been widely used in task offloading problems due to its powerful capability in handling high-dimensional and non-convex problems.Existing edge computing task offloading methods based on deep reinforcement learning have three typical problems:Firstly,the vast majority of task offloading algorithm only considers the information of the devices and tasks themselves,and does not utilize the spatial information implied by the objective connection relationships between devices in the environment.Secondly,in the scenario of multi-edge server collaborative task offloading,the sharing of global rewards by agents will lead to a credit assignment problem,and each agent will find it difficult to evaluate its own contribution.Finally,the generalization ability of edge computing task offloading algorithms based on deep reinforcement learning is poor,and it is necessary to retrain network parameters to adapt to new environments when the environment changes.To solve the above problems,this paper proposes three edge computing task offloading algorithms,aiming to meet the time sensitive requirements of tasks.The main content includes:(1)To address the problem of underutilized potential spatial features among devices in edge computing environments,an edge computing task offloading algorithm based on the joint optimization of deep reinforcement learning and graph neural networks is proposed in this paper,which comprehensively considers the indivisibility of tasks,the transmission and computation queues of devices,and the load level on edge servers.In this algorithm,the end devices leverage the graph attentional agent and the recurrent neural network to respectively extract the spatial features among devices in the environment and the temporal characteristics of the load level on edge server,and the task offloading decision is subsequently made by the offloading agent.The experimental results demonstrate that the proposed algorithm outperforms other baseline algorithms in terms of low task latency and low task dropout rate performance under both single-task and multi-task settings.(2)To addresses the potential credit assignment problem in multi-agent collaborative edge computing scenarios,a multi-agent collaborative edge computing task offloading algorithm based on counterfactual inference is proposed in this paper,which comprehensively considers the severability of tasks,the energy consumption of devices,and the timevarying wireless channels.In this algorithm,each agent constructs a counterfactual action buffer to store better actions,and defines a new advantage function based on the counterfactual action buffer to evaluate the contribution of the agent to the entire system.The experimental results demonstrate that the proposed algorithm outperforms other baseline algorithms in terms of low task latency performance under multiple scenario settings,while also solving the potential credit assignment problem in the scenario.(3)To address the problem of poor generalization ability in task offloading algorithms based on deep reinforcement learning,a meta-reinforcement learning-based algorithm for fast adaptation of edge computing task offloading is proposed in this paper,which comprehensively considers the internal dependencies in tasks,computation and transmission capabilities of devices,and the availability of resources.In this algorithm,the graph neural network is utilized to extract potential spatial features between sub-tasks,and the inner and outer modules are constructed based on the concept of meta-reinforcement learning.Afterwards,the synergistic optimization of the two modules is utilized to improve the generalization ability of the proposed algorithm.The experimental results demonstrate that the proposed algorithm has faster convergence speed compared to other baseline algorithms in various environmental settings,and it also exhibits better generalization ability in different task attribute settings.