移动边缘计算中基于图到序列深度强化学习的复杂任务部署策略

Graph-to-sequence deep reinforcement learning based complex task deployment strategy in MEC

借助于移动边缘计算(MEC)和网络虚拟化技术,可使移动端将执行各类复杂应用所需的算力、存储和传输等资源需求就近卸载至边缘服务节点,从而获得更高效的服务体验。面向边缘服务商,研究其在进行复杂任务部署时所面临的能耗优化决策问题。首先将复杂任务部署于多个边缘服务节点的问题建模为混合整数规划(MIP)模型,然后提出了一种融合图到序列的深度强化学习(DRL)求解策略。该策略通过基于图的编码器设计提取并学习子任务间潜在的依赖关系,从而根据边缘服务节点的可用资源状态及使用率自动发现任务部署的通用模式,最终快速获得能耗优化的部署策略。在不同的网络规模中,将所提策略与具代表性的基准策略进行了全面对比。实验结果表明,所提策略在任务部署错误率、MEC系统总功耗和算法求解效率等方面均显著优于基准策略。

With the help of mobile edge computing(MEC)and network virtualization technology,the mobile terminals can offload the computing,storage,transmission and other resource required for executing various complex applications to the edge service nodes nearby,so as to obtain more efficient service experience.For edge service providers,the opti-mal energy consumption decision-making problem when deploying complex tasks was comprehensively investigated.Firstly,the problem of deploying complex tasks to multiple edge service nodes was modeled as a mixed integer pro-gramming(MIP)model,and then a deep reinforcement learning(DRL)solution strategy that integrated graph to se-quence was proposed.Potential dependencies between multiple subtasks through a graph-based encoder design were ex-tracted and learned,thereby automatically discovering common patterns of task deployment based on the available re-source status and utilization rate of edge service nodes,and ultimately quickly obtaining the deployment strategy with the optimal energy consumption.Compared with representative benchmark strategies in different network scales,the experi-mental results show that the proposed strategy is significantly superior to the benchmark strategies in terms of task de-ployment error ratio,total power consumption of MEC system,and algorithm solving efficiency.