摘要： 在移动边缘计算中，移动设备通常受限于自身的处理性能和电源容量，需要其他设备协助进行任务处理。将移动设备上的一系列具有依赖关系的任务卸载到边缘服务器执行，以应对移动设备资源受限的问题，可提高任务计算和能源效率。针对信道状态动态变化的移动边缘计算环境下任务延迟和移动设备能耗优化问题，根据依赖任务卸载模型计算得出依赖任务调度顺序和优化目标，设计一种基于A3C的依赖任务卸载与资源分配（DTORA）算法。通过定义状态空间、动作空间和奖励函数，将依赖任务卸载问题转化为马尔可夫决策过程下的最优策略问题，采用异步并发求解得到高效的任务卸载和资源分配策略，并在具有标准多核CPU的单个机器上进行并行学习，降低神经网络参数更新的相关性，提升学习效果。实验结果表明，在信道状态动态变化场景下，对于多种不同依赖关系的任务，DTORA算法相比于4种基线算法任务延迟减少14%~61%，移动设备能耗降低8%~66%。

关键词: 移动边缘计算, 深度强化学习, 依赖任务卸载, 资源分配, 能耗优化

Abstract: In Mobile Edge Computing（MEC），Mobile Device（MD） are usually limited by their processing performance and power constraints and need other devices to assist them in task processing. Offloading a series of tasks with dependencies on MD to the edge server for execution to cope with the resource constraint of the MD can improve the computational and energy efficiencies of the tasks.To address the task delay and MD energy consumption optimization problems in MEC scenarios with time-varying channels，a Dependent Task Offloading and Resource Allocation（DTORA） algorithm based on Asynchronous Advantage Actor-Critic（A3C） is proposed based on the established dependent task offloading model to calculate the dependent task scheduling order and optimization objectives. By defining the state space，action space，and reward function，the dependent task offloading problem is transformed into an optimization policy problem under a Markov Decision Process（MDP），and an efficient task offloading and resource allocation policy is solved using asynchronous concurrency，which can be learned in parallel on a single machine with standard multi-core CPUs to reduce the correlation of neural network parameter updates and improve the learning effect.Experiments show that DTORA reduces the task delay by 14%-61% and reduces the energy consumption of MDs by 8%-66% compared with that of the four baseline algorithms for various tasks with different dependencies under the channel dynamic change scenarios.

Key words: Mobile Edge Computing（MEC）, Deep Reinforcement Learning（DRL）, dependent task offloading, resource allocation, energy consumption optimization

中图分类号: 

• TP391