Research On Edge Computing Offloading Based On Deep Reinforcement Learning And Multi-Objective Optimization

With the rapid development of 5G wireless technology and the advent of the digital Io T era,people’s expectations for global intelligent interconnection and high-reliability low-latency communication are increasing.Mobile edge computing can provide users with localized services at the network edge close to the user,effectively solving the problem of limited performance of mobile devices when executing delay-sensitive and computeintensive tasks.The problem of task offloading is one of the core issues in the field of mobile edge computing.How to make reasonable decisions on task offloading and resource allocation directly affects the user’s usage experience.This thesis focuses on the problem of task offloading and resource allocation in the edge computing process,with the following specific work:In the context of independent unit mobile edge computing scenarios,this thesis focuses on the problem of computation offloading and resource allocation during the edge computing process,with device energy consumption and task execution latency as optimization objectives.The study introduces the channel parameter estimation error in the data uploading process and constructs a multi-objective optimization problem.A deep reinforcement learning-based multi-objective optimization algorithm was proposed for this problem.The original problem was decomposed into a group of scalar optimization subproblems based on the decomposition idea.Then,each subproblem was modeled as a neural network.By using the neighborhood-based parameter transfer strategy and the deep reinforcement learning training algorithm,the model parameters of all subproblems were collaboratively optimized.The Pareto front of the problem can be directly obtained by the trained model,and a dynamic weight factor was designed to assist in selecting the optimal solution.Experimental results show that the proposed algorithm can more efficiently obtain the Pareto front of the problem,and compared with existing computing offloading methods,it can achieve lower system total cost and more reasonable resource allocation.The scenario of task offloading for independent computing units has been extended to consider the problem of interference that may arise when different devices use the same communication channel in adjacent units.A communication and computation offloading model for cloud-edge collaboration across multiple units has been designed.The optimization problem is formulated as a multi-objective optimization problem with device energy consumption and task execution latency as the objectives,and a cloud-edge collaboration offloading strategy based on deep reinforcement learning is proposed.A pre-partitioning algorithm is designed to partition devices according to their offloading destinations.Due to the combinatorial nature of the problem,it is difficult to solve directly,so the problem is further decomposed into two sub-problems: the local optimal computation cost and the allocation of sub-channels and other resources,which are solved by convex optimization and a multi-objective optimization algorithm based on deep reinforcement learning,respectively.Simulation results show that the proposed algorithm can obtain better resource allocation schemes and smaller system total cost compared with conventional algorithms,and can still maintain good performance as the number of units and devices increases.