Research On Complex Task Deployment Combining Graph Neural Network And Deep Reinforcement Learning In Edge Computing

With the advent of the Internet of everything era,the rapid growth of mobile devices makes the traditional cloud computing architecture unable to meet the real-time requirements of computing tasks.To solve this problem,edge computing came into being.Edge computing uses computation offloading technology to reasonably allocate computing tasks between devices and edge servers,so as to prolong the battery life of devices,improve the computing efficiency of tasks,reduce the network delay of task transmission,and improve the privacy protection of user data.This paper mainly studies the sub-task deployment problem of complex tasks in the problem of computation offloading,that is,a complex task can be divided into multiple sub-tasks and deployed to different edge servers for execution.In this paper,we first mathematically model the problem and formulate it as a combinatorial optimization problem with constraints.Then,using the graph neural network method and the deep reinforcement learning framework,we propose a graphto sequence structure neural network model to approximate the deployment strategy for solving the subtasks.The proposed algorithm is superior to the existing comparison algorithms under a number of performance indicators,which proves the effectiveness and practicability of the proposed method.The main research contents of this paper are as follows:(1)In this study,the deployment problem of complex computing tasks in edge computing is deeply modeled,and the dependency between subtasks is fully considered.By building a specialized task model,the dependencies between subtasks are represented as a directed acyclic graph.On this basis,this paper further considers the impact of resource requirements and communication delay in the process of sub-task deployment,and establishes a mathematical model with the optimization goal of minimizing the energy consumption of edge servers.The problem is actually a mixed integer linear programming problem,which is NP-Hard.(2)Aiming at the dynamic problem in the edge computing environment,this study proposes a dynamic construction strategy of subtask dependency graph.Since the dependencies between subtasks in complex tasks may be different in different operating environments,this paper adopts the self-attention mechanism to find the internal relevance between subtasks.Then,the k-nearest neighbor algorithm is used to filter out the most closely related adjacent subtasks,and a dependency graph with a clear structure is constructed.(3)In this study,deep reinforcement learning methods,specifically the graph-tosequence neural network model,are used to solve the constructed mathematical problem.Firstly,the Markov decision process is used to model the state space,action space and reward in the whole problem.Then,the graph-to-sequence neural network is used to approximate the policy function,and the policy gradient optimization algorithm with baseline is used for continuous iterative learning,and finally an optimal subtask deployment policy is found.