| In the past decade,with the advent of the information age,the number of intelligent devices has grown exponentially,and applications in the fields of automatic driving,intelligent medicine,and smart cities have also emerged endlessly.However,due to the limitations of intelligent devices’ storage capacity and computing capacity,they cannot handle many complex and centralized calculations.This cannot meet the needs of these applications for low latency and low energy consumption.At this time,intelligent devices can use edge servers to reduce latency and energy consumption by offloading computing-intensive tasks to edge servers deployed nearby.Therefore,in the edge offloading scenario composed of edge servers and intelligent devices,optimizing the allocation strategy for partial offloading is the main content of this dissertation.1.Research on computational task offloading based on deep reinforcement learning.Conventional edge computing scenarios all use edge servers with fixed base stations for resource allocation.In this dissertation,aiming at the computing scenarios where edge servers are mobile,an optimization scheme combining intelligent device offloading decision and mobile device mobility is proposed.DDPG(Deep Deterministic Policy Gradient)algorithm is used to solve the problem of continuity of intelligent device uninstallation decision,which has the advantages of less sampling data and high efficiency.Firstly,the optimization goal is to minimize the overall processing delay of partial offloading of computing tasks.Secondly,the optimization problem is transformed into a Markov decision-making process.Finally,the DDPG algorithm of the two-layer Actor-Critic framework is adopted to improve stability and convergence.In addition,the multi-agent MADDPG algorithm is extended to multiagent systems.Simulation experiments show that compared with other reinforcement learning algorithms,the proposed algorithm has better convergence performance and advantages in reducing users’ overall delay.2.Research on parallel task offloading with dependency.The current task offloading work often ignores the unique task topology on smart devices,resulting in insufficient resource utilization of edge servers.This dissertation focuses on offloading parallel tasks with dependencies and proposes a PDOA(Parallel Dependency Offload Algorithm)algorithm.Firstly,considers the dependency properties of subtasks in the task topology,then establishes the priority according to the dependencies of subtasks,pays full attention to the parallelism of subtasks,then optimizes the offloading strategy with the optimization goal of reducing the computing delay of the user devices,and finally realizes the offloading allocation of subtasks through the PDOA algorithm.Simulation comparison results show that compared with algorithms that do not consider task dependencies and parallel relationships,the proposed algorithm can effectively reduce the delay. |
