Research On Task Offloading Based On Reinforcement Learning In Edge Computing

As the quantity and variety of application data continue to increase,there is a growing demand from users for high-quality services.Edge computing,as a new computing model of the Internet of Things(Io T),has become a highly virtualized platform that provides computing,storage,and networking services between end devices and traditional cloud data centers.Edge node is the important infrastructure of edge computing network,including switch,router and embedded server.With the continuous development of internet terminal devices,image-based mobile platforms in automated driving,smart cities and other fields have been widely used.In the field of edge computing,as the variety and quantity of image processing task generated by Io T devices continue to increase,there is also an increasing demand from end users for high-quality mobile services.Furthermore,the increasing number of devices connected to edge nodes and the insufficient supply of resources will lead to high costs and severe load imbalances between edge nodes.Therefore,a comprehensive and effective image processing task offloading strategy is crucial for the development of edge computing networks and the improvement of application performance.This paper focuses on the specific task requirement of image processing in automated driving,and research on reinforcement learning-based optimization strategies for edge computing task offloading.It constructs new optimization models based on different constraints and proposes corresponding solutions.Subsequently,through simulation experiments and comparisons with existing methods,the feasibility of the theoretical models and the correctness of the proposed solutions are validated.The specific research work includes the following three aspects:(1)Firstly,we introduce a task offloading method for application-driven tasks in edge computing based on reinforcement learning is proposed.This method aims to design a reasonable image processing task offloading scheme to achieve resource allocation among multiple users with dependency relationships,taking into account the application-driven image processing task requests in mobile edge networks and automated driving scenarios.In addressing this problem,this paper investigates application-driven image processing tasks in multi-user scenarios in edge networks.It presents an application-driven task offloading strategy based on the joint optimization of latency and energy consumption for the overall cost of users.Firstly,the paper models the problem of application-driven image processing task offloading in edge computing automated driving scenarios and introduces a novel priority factor to enable the topological sorting of parallel subtasks based on analyzing the Vdependencies of application-driven image processing tasks.Secondly,building upon this,the paper proposes a task offloading strategy based on a deep Q-learning network.This strategy trains a fully connected neural network to achieve the joint optimization of latency and energy consumption for multiple users in the edge network.Finally,in order to validate the effectiveness of the proposed solution,the paper constructs simulation experiment platforms based on the proposed strategies were constructed separately,and tests were conducted using obstacle detection as a specific task type in autonomous driving.The experimental results demonstrate that the reinforcement learning-based task offloading method for application-driven tasks in edge computing,incorporating the sorting mechanism,yields favorable results in terms of convergence and overall cost.(2)Secondly,we propose task offloading method for edge-cloud architecture based on reinforcement learning.This method focuses on designing a reasonable image processing task offloading scheme to achieve collaborative task offloading in the edge-cloud system under edge resource constraints.considering the distribution of multiple users and their offline task requests in the given edge-cloud architecture.For different characteristics of offline independent task scenarios with multiple mobile automated driving users,this paper investigates image processing task offloading methods for multiple automated driving users in the edge-cloud collaborative architecture.The overall cost optimization is the objective,taking into account the trade-off between edge network resource constraints,communication latency,and energy consumption.Firstly,this paper models the problem based on the image processing task offloading scenario for multiple automated driving users in the edge-cloud architecture and proves that jointly optimizing the total latency and energy consumption of multiple users under capacity constraints is an NP-hard problem.Secondly,a novel joint update mechanism is introduced to optimize latency and energy consumption.Specifically,this paper considers the avoidance of high complexity caused by adjacent edge nodes and proposes a heuristic algorithm to filter out feasible nodes.Based on this,an image processing task offloading strategy based on deep reinforcement learning in the edge-cloud collaborative system is proposed.Finally,to validate the effectiveness of the proposed solution,a simulation experimental platform based on the edge-cloud network architecture is designed.The proposed method is compared with other algorithms,and extensive simulation experiments demonstrate that the image processing task offloading solution proposed in this paper performs well under different user request scenarios.(3)Finally,we study an edge collaborative online task offloading method based on reinforcement learning.This method focuses on designing a reasonable online task offloading scheme to achieve resource scheduling for multiple automated driving users in an active area,considering the given mobile edge network and online task requests from multiple automated driving users.In the study of this problem,this paper investigates task offloading methods for multiple automated driving mobile users in the online image processing task scenario,research on task offloading methods for multiple mobile users aimed at edge collaborative operations,aiming to jointly optimize the latency and energy consumption for mobile users under constraints of capacity and computing capability.It explores continuous image processing task offloading methods based on the online scenario of mobile users.Firstly,this paper models the collaborative image processing task offloading problem of edge nodes in mobile edge computing based on the online task scenario,and designs a trajectory prediction method based on the mobility characteristics of automated driving to reduce latency caused by user mobility.Secondly,centered around image processing task-requesting vehicle,a task optimization and scheduling mechanism based on the multi-dimensional features of tasks and the edge resource situation is proposed to avoid invalid states in the decision-making module.An approximate strategy is introduced in the decision-making module based on the feasibility mechanism.Building upon this,a reward evaluation method based on deep reinforcement learning is proposed to achieve online task offloading for multiple automated driving vehicle,jointly optimizing the overall latency and energy consumption under constraints of physical edge resources.Finally,to validate the effectiveness of the proposed solution,simulation experiments are conducted on synthetic and real datasets.Our joint optimization method is compared with several state-of-the-art methods,and the experimental results are analyzed and evaluated from different perspectives to provide corresponding conclusions.The simulation results demonstrate that the proposed solution can effectively improve the overall revenue of task requests from the image processing of automated driving in the edge collaborative task offloading area and guarantee the service quality for tasks with higher priorities.