光接入网低碳联邦边缘智能:理论与优化算法OA

[Objective]With the deep convergence of the 6th Generation Mobile Communication Technology(6G)communica-tion and Artificial Intelligence(AI)technologies,networks are evolving from simple connectivity pipelines into integrated infra-structures encompassing sensing,computing,and intelligence.As a distributed edge AI paradigm,federated edge intelligence aligns inherently with the native Point-to-MultiPoint(P2MP)architecture of Passive Optical Networks(PON),offering an ideal pathway for deploying intelligence at the network edge.Given this synergy,this paper introduces,for the first time,the concept of"PON-based Federated Edge Intelligence".However,with the rapid advancement and application of AI technologies,issues of high energy consumption and carbon emissions have become prominent,making it imperative to develop low-carbon and low-en-ergy optimization theories and methods for this federated edge intelligence system.[Methods]This paper systematically con-structs a theory and implementation framework for low-carbon federated edge intelligence in AI-native optical access networks un-der the Independent and Identically Distributed(IID)data scenario.The specific contributions include:① Establishing a carbon emission model that clearly quantifies the energy consumption and carbon emissions throughout the entire process,from data col-lection and local training to model aggregation;② Proposing a low-carbon participant mechanism for federated training that selects clients based on a low carbon emission coefficient;③Extending the existing federated learning convergence theory in IID scenarios by proving the unbiased expectation of model updates when averaging a fixed number of sampled clients,thereby establishing a convergence theory applicable to client sampling scenarios;④Revealing the biconvex structure of the optimization problem in this scenario and proposing an efficient optimization algorithm based on alternating convex search.[Results]A simulated network en-vironment was built on a Raspberry Pi-based hardware platform.The proposed theories and methods were validated using a Con-volutional Neural Network(CNN)model on the Modified National Institude of Standards and Technology(MNIST)dataset.Ex-perimental results show that under the IID data scenario,by selecting edge nodes for federated training based on low carbon emission coefficients compared to full client participation,the proposed approach achieves reductions of up to approximately 35%in energy consumption and 86%in carbon emissions under the experimental configuration of this study.[Conclusion]The simulation results verify the effectiveness of the proposed optimization theory and methods in optimizing energy consumption and carbon emissions.