Research On Task Offloading And Resource Allocation Technology In Edge Intelligent Environment

With the rapid development of the Industrial Internet of Things,edge intelligence has become an important research direction.However,realtime neural network inference for resource-limited IoT devices remains a challenge for efficient machine learning task processing.Introducing edge computing on top of cloud computing can effectively improve the processing efficiency of machine learning tasks.However,traditional research work lacks optimization of the inference accuracy of machine learning tasks,as well as joint optimization of task offloading,data quality adjustment,and resource allocation.Therefore,this paper studies the technology of task offloading and resource allocation in the edge intelligence environment,and the specific research work is divided into two parts:(1)This paper proposes an machine learning task offloading scheme to minimize the total processing delay of tasks in the edge intelligent IoT system while guaranteeing the required inference accuracy of tasks.It considers multiple task attributes,task inference accuracy,and the impact of error inference on task processing delay.Wireless channel allocation and computing resource allocation problems are modeled together with task offloading.Given the high complexity of the optimization problem,this paper designs an algorithm that decomposes the problem into computing resource allocation sub-problem and task offloading and channel allocation sub-problem,and solves them separately.Through simulations and comparisons with four other schemes,the superiority of the proposed scheme is demonstrated.(2)This paper proposes a novel joint resource management scheme for Industrial Internet of Things scenarios consisting of multiple sensors and edge servers-equipped base stations.A time-slot system model is proposed,which includes task offloading,data quality adjustment,power control,wireless channel allocation,and computing resource allocation.Data quality adjustment balances the task inference accuracy and task transmission delay by changing the task data size.The optimization objective is to minimize the total cost,including the total processing delay of tasks,the total energy consumption of the system,and the total penalty for error inference,while guaranteeing the required inference accuracy of all tasks.This paper designs a multi-agent reinforcement learning-based algorithm to provide near-optimal solutions to this problem.A fast numerical method is also proposed to reduce the training complexity of the multi-agent reinforcement learning model for data quality adjustment.Extensive experiments are conducted by varying key parameters and comparing with three other schemes,and the superiority of the proposed scheme is demonstrated.