Research On Mobile Edge Computing Task Offloading Strategy Based On Deep Reinforcement Learning

Mobile Edge Computing(MEC)has received extensive attention as an emerging computing paradigm.MEC migrates computing tasks on user equipment to edge servers that are closer to it,thereby achieving lower latency,energy consumption,and higher reliability.Computation offloading is a key technology in the MEC system,which determines the processing method of computing tasks.An efficient computation offloading strategy can significantly improve the service experience quality of user equipment.With the development of big data and artificial intelligence,more and more computing-intensive applications with low latency requirements enrich people’s lives.The limited computing power and battery capacity of user equipment make it difficult to meet the latency and energy consumption requirements of these applications.On the one hand,considering that the existing computation offloading strategies are mainly binary computation offloading.However,edge server computing resources are also limited,and completely offloading computing tasks will also put pressure on network bandwidth.On the other hand,considering that computing task offloading is scheduled in the multiedge server scenario,the computation offloading decision and resource allocation decisions are decision spaces in different domains.The existing computation offloading strategy is mainly completed in two stages,which is difficult to adapt to dynamic the changing edge computing environment.Deep reinforcement learning combines the advantages of deep learning and reinforcement learning,which can effectively mine high-dimensional data features in dynamic environments and learn to control decision-making in the process of dynamic interaction.Therefore,this thesis considers introducing deep reinforcement learning into edge computing scenarios,digs deep into computing tasks and computing scene features,and studies mobile edge computing task offloading strategies to improve user service experience quality.The main work of this thesis is as follows:(1)Aiming at the Data Partitioned Oriented Applications(DPOA)with high parallelism,this thesis establishes a partial computation offloading framework based on deep reinforcement learning.The framework takes minimizing user equipment delay as the optimization goal,and defines its state space,action space and reward function in detail.Under this framework,this thesis proposes two partial computation offloading algorithms based on deep reinforcement learning,including a partial computation offloading algorithm based on Q-learning,and a partial computation offloading algorithm based on Deep Deterministic Policy Gradient(DDPG).Specifically,Q-learning computational offloading algorithm adopts discrete partial offloading decisions,and the DDPG computational offloading algorithm outputs partial offloading decisions on a continuous action space.The experimental results show that,compared with other baseline schemes,the two partial computation offloading algorithms proposed in this thesis are effective and can significantly reduce the delay of user equipment.(2)In addition,considering the mobility of users,computing tasks need to be scheduled under multiple edge servers.In this scenario,the computation offloading space is a hybrid action space,including the computation offloading decision,transmission power,and edge server resource allocation of the user equipment.However,existing computational offloading strategies for deep reinforcement learning cannot handle such mixed action spaces well.Based on the previous work,this thesis proposes an End-to-End Hybrid Computation Offloading(E2EHCO)framework based on deep reinforcement learning.With the goal of jointly optimizing the delay and energy consumption of mobile users,the computational offloading problem in the scenario of a dynamic multi-user multi-server mobile edge computing system is studied.In order to better study the impact of user mobility on computation offloading in real-world environment.In this thesis,E2 EHCO framework is applied to find the characteristics of user trajectory on the public user trajectory data set through data-driven method.Experimental results show that E2 EHCO framework can effectively reduce the delay and energy consumption of user equipment.In summary,this thesis studies the computation offloading strategy in mobile edge computing.This thesis makes some supplements to the existing research on computation offloading strategies in mobile edge computing,and discusses possible future research directions.