摘要
Abstract
The vigorous convergence of communication and internet of things enables the architecture of edge network become increasingly dense and heterogenous.Meanwhile,the differentiated tasks,highly dynamic network and decentralized deployment of computing-networking resource undermine the efficiency of service caching and resource allocation.To address the problems,this paper studies the joint optimization of task offloading,service caching and resource allocation in the decentralized MEC scenario.First,a service caching model and a computing offloading model are built.Then,a joint optimization of service caching 1 computing resource allocation and transmit power control is modeled and abstracted into a partially observable Markov decision process to minimize the task processing cost.Moreover,to tackle the privacy leakage in centralized model training,a distributed model training approach based on federated learning is designed,and a distributed service caching and resource allocation algorithm based on federated multiagent deep reinforcement learning is proposed to make service caching,computing resource allocation and transmit power control decisions independently.Considering the differences of local models,attention mechanism is employed to assign different weights for the local models when aggregating the parameters.Finally,our simulation results shows the proposed algorithm markedly decreases the task execution cost and improves the cache hit ratio.关键词
边缘智能/多智能体/资源分配/计算卸载/服务编排Key words
edge computing/multi-agent/resource allocation/computing offloading/service caching分类
信息技术与安全科学
