移动边缘计算中分布式智能服务缓存和资源分配联合优化

Joint optimization of distributed intelligent service caching andresource allocation in mobile edge computing

随着通信技术和物联网技术深度融合,边缘网络结构日益密集化、异构化。同时,在边缘网络环境中,广域差异化业务、去中心化算网资源部署以及高度动态化网络环境等特征,制约了网络服务缓存和资源分配的效率。针对上述问题,对去中心化的边缘网络场景下的任务卸载、服务缓存和资源联合优化展开研究。首先,研究并建立了服务缓存和任务卸载模型。其次,以最小化任务的处理成本为目标,将服务缓存、算力资源分配和传输功率控制的联合优化问题抽象为部分可观测马尔科夫决策过程。然后,针对模型集中训练存在的隐私泄露问题,设计了一种基于联邦学习的分布式模型训练方法,提出了一种基于联邦多智能体深度强化学习的分布式服务缓存和资源分配算法来自主决策服务缓存、算力资源分配和传输功率控制决策。此外,考虑到本地模型的差异性,在进行参数聚合时,使用注意力机制,为本地模型分配不同的权重。最后,仿真验证了所提算法在任务处理成本、缓存命中率方面性能明显提升。

ed 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.