联合联邦学习与深度强化学习的并行服务功能链部署算法

打开文本图片集

中图分类号：TP393.0 文献标志码：A

Abstract： To address the dynamic deployment of parallel service function chains （SFCs） in multidomain edge cloud networks，this paper constructs an optimized paralll SFC architecture and proposes a novel approach combining federated learning （FedAvg） and deep reinforcement learning（DRL）， termed FA-D3QN-PER . This method resolves the issues of imbalanced resource allcation and privacy leakage inherent in existing single-agent DRL and centralized decision-making frameworks when handling SFC partitioning and deployment. By enabling independent training of agents within each domain and leveraging FedAvg for model parameter sharing，it optimizes global strategies while preserving data privacy. The deployment phase involves the analysis and optimization of the hybrid SFC parallel structure;intelligent partitioning of the optimized hybrid SFC into sub-chains and their allocation to suitable edge domains; mapping of virtual network functions （VNFs） within each sub-chain to physical nodes in target domains. Simulation results demonstrate that the FA-D3QN-PER method exhibits strong stability and fast convergence， significantly improving SFC deployment acceptance rates while effectively reducing average latency and total costs. Compared to FA-DQN，DFSC，and MuL， the FA-D3QN-PER method increases acceptance rates by 11.6% ，while reducing average latency and total costs by 17% and 18.56% ，respectively.

Keywords： multi-domain edge cloud networks； paralel service function chain； dynamic deployment; federated averaging；deep reinforcement learning

随着第六代移动通信系统（6G）的引入和物联网（IoT）业务多样性的增长，电信服务提供商（TSP）面临着满足用户多样化需求的挑战。（剩余20804字）