Deep Reinforcement Learning-based Edge Computing Offloading In SDN-IoT

With the massive growth in data generated by IoT(Internet of Things)devices,the limitations of computing power in end devices become apparent when processing massive computational tasks.This requires the offloading of computational tasks from resource-limited end devices to edge servers with stronger computing capabilities.However,when network conditions and task requirements change rapidly,data-driven intelligent algorithms struggle to obtain comprehensive statistical data for accurate predictions,leading to decreased performance and difficulty in adaptive adjustments.Improving environmental awareness and intelligent optimization in order to make computational offloading algorithms adapt to dynamic changes in network conditions and task requirements and achieve global multi-objective optimization such as latency and energy consumption is currently a challenge.Therefore,this article studies the optimization of collaborative task processing on the edge side through deep reinforcement learning in the context of edge IoT.The details are as follows:(1)A software-defined edge computing architecture,SDEC(Software Defined Edge Computing)is proposed to provide global status for the collaboration of edge servers.The control layer in the software-defined architecture is fused with the edge layer in edge computing,and multiple controllers share global network status information through east-west message exchange.(2)To solve the problem of server load imbalance in edge environments,a software-defined IoT edge computing offloading algorithm,ECO-SDIoT(Edge Computing Offloading Algorithm in Software-Defined IoT)based on deep reinforcement learning is proposed,allowing controllers to offload computational tasks to the most suitable edge server based on global network information obtained.Performance metrics of edge computing offloading are evaluated in terms of unit task processing latency,load balancing of edge servers,task processing energy consumption,and task completion rate.Simulation results show that ECO-SDIoT can effectively reduce task completion time and energy consumption compared to other strategies.(3)To achieve parallel offload processing of multiple edge tasks,an edge computing offloading algorithm based on multi-agent reinforcement learning,ECOMARL(Edge Computing Offloading Algorithms based on Multi-Agent Reinforcement Learning)is proposed,allowing controllers to asynchronously update model parameters based on global network information obtained by each agent.Each controller in ECO-MARL is treated as an intelligent agent,and each agent can allocate computational tasks to the most suitable edge server based on global network information to minimize task completion time and energy consumption,improving performance.Figure [35] Table [5] Reference [97]...