SDCN中基于深度强化学习的移动边缘计算任务卸载算法研究

Research on task offloading algorithm of mobile edge computing based on deep reinforcement learning in SDCN

随着网络技术的不断发展,基于Fat-Tree的网络拓扑结构分布式网络控制模式逐渐显露出其局限性,软件定义数据中心网络(software-defined data center network,SDCN)技术作为Fat-Tree网络拓扑的改进技术,受到越来越多研究者的关注。首先搭建了一个SDCN中的边缘计算架构和基于移动边缘计算(mobileedge computing,MEC)平台三层服务架构的任务卸载模型,结合移动边缘计算平台的实际应用场景,利用同策略经验回放和熵正则改进传统的深度Q网络(deep Q-leaning network,DQN)算法,优化了MEC平台的任务卸载策略,并设计了实验对基于同策略经验回放和熵正则的改进深度Q网络算法(improved DQN algorithm based on same strategy empirical playback and entropy regularization,RSS2E-DQN)和其他3种算法在负载均衡、能耗、时延、网络使用量几个方面进行对比分析,验证了改进算法在上述4个方面具有更优越的性能。

With the continuous development of network technology,the network topology distributed network control mode based on Fat-Tree gradually reveals its limitations.Software-defined data center network(SDCN)technology,as an improved technology of Fat-Tree network topology,has attracted more and more researchers’attention.Firstly,an edge computing architecture in SDCN and a task offloading model based on the three-layer service architecture of the mobile edge computing(MEC)platform were built,combined with the actual application scenarios of the MEC platform.Through the same strategy experience playback and entropy regularization,the traditional deep Q-leaning network(DQN)algorithm was improved,and the task offloading strategy of MEC platform was optimized.An im-proved DQN algorithm based on same strategy empirical playback and entropy regularization(RSS2E-DQN)was compared with three other algorithms in load balancing,energy consumption,delay and network usage.It is verified that the improved algorithm has better performance in the above four aspects.