Abstract: To enhance the quality of user computing services in Industry 5.0 scenarios,a novel resource scheduling mechanism for computing power networks based on the deep reinforcement learning was proposed,which integrated cloud-edge-end collaboration by jointly addressing task scheduling and cooperative resource allocation across terminal devices,edge computing nodes,and centralized cloud infrastructures.Initially,a cloud-edge-end collaborative architecture that unified computing power networks with cloud-edge synergy was introduced.This framework reformulated the challenges of task scheduling and resource allocation into an optimization problem aimed at minimizing the overall system cost,while simultaneously satisfying constraints related to computing capacity,communication bandwidth,and energy consumption across network nodes.Subsequently,an enhanced Deep Q-Network(DQN) algorithm was developed with dynamically arriving tasks,and the innovative reward strategies,soft updates of network parameters,experience replay and a dual-network scheme were introduced to robustly learn optimal scheduling policies under varying system conditions.Experimental results demonstrated the feasibility of the proposed method,with effective performance in task scheduling and resource allocation.

Key words: cloud-edge-terminal collaborative architecture, computing power networks, task scheduling, resource allocation, deep reinforcement learning

摘要： 为了更好提升工业5.0场景下用户计算任务的服务质量,同时考虑任务调度决策以及终端设备、边缘算力服务节点、云计算中心的协同资源分配,提出一种基于深度强化学习的云-边-端协同的算力网络资源调度机制。首先提出一种融合算力网络与云边协同的云-边-端协同架构,将计算任务调度与资源分配问题转化为网络节点计算、通信、能源约束下的最小系统总成本优化问题;然后提出一种考虑任务动态到达的改进深度Q网络求解该问题,并在网络学习的过程中引入新的奖励方式、网络参数软更新、经验回放与双网络机制。通过实验验证了所提方法的可行性,在任务调度与资源分配方面性能良好。

关键词: 云边端协同架构, 算力网络, 任务调度, 资源分配, 深度强化学习