基于图神经网络的具有依赖关系任务的计算卸载方法

Computational Offloading Method for Tasks with Dependency Based on Graph Neural Network

摘要:

计算卸载作为移动边缘计算的关键技术之一，通过将任务就近迁移至边缘服务器上执行大幅降低了用户的等待时延。针对具有依赖关系任务的计算卸载问题，为了解决以往文献在将表示任务依赖关系的有向无环图输入深度强化学习算法的神经网络时存在的丢失结构信息的问题，提出了一种有向无环图神经网络（Directed Acyclic Graph Neural Network，DAGNN），并将其与深度强化学习相结合，用以做卸载调度的决策。卸载决策的过程被描述为马尔科夫决策过程，用提出的DAGNN评估深度强化学习算法中每个卸载动作的Q值，进而做出卸载调度决策。仿真实验表明，所提出算法在各种条件下的表现均优于其它所有基线算法，并表现出较好的稳定性和通用性。

Abstract:

As one of the key technologies of mobile edge computing, computing offloading can greatly reduce the user's waiting time by moving the task to the edge server.For the problem of computing offloading for tasks with dependencies, in order to solve the problem of losing the structural information in the previous literature when the directed acyclic graph representing the task dependency is input into the neural network, came up with a directed acyclic graph neural network(DAGNN), and combine it with deep reinforcement learning to make decisions about offloading scheduling. The process of offloading decision is described as Markov decision process, and the Q value of each offloading action in the deep reinforcement learning algorithm is evaluated by the DAGNN proposed in this paper, and then the offloading scheduling decision is made. Simulation results show that the proposed algorithm performs better than all other baseline algorithms under various conditions, and shows good stability and versatility.