面向算力网络的软掩码策略梯度联邦学习算法

Soft-masked Policy Gradient Federation Learning Algorithm for Computing Power Networks

针对算力网络中终端的异构性而导致的设备资源不均衡以及各终端动态加入或离开联邦学习任务导致协同体系不稳定、内生全局模型收敛性差等问题,文中提出了一种面向算力网络的软掩码策略梯度(Soft-Masked Policy Gra⁃dient,SMPG)联邦学习算法。该算法能通过联合计算优化机理,自适应地调整异构终端局部模型来更新频率与聚合时间,以合理分配终端计算及通信资源,提升联邦学习内生模型的收敛速度及终端的资源利用率。

Aiming at the imbalance of equipment resources caused by the heterogeneity of end points in computing power networks,the instability of the cooperative system caused by the dynamic addition or departure of each end point from federated learning tasks,and the poor convergence of the endogenous global model.This paper proposes a soft-mask policy layer(SMPG)federated learning algorithm for computing power networks.The algorithm can adaptively adjust the heterogeneous end point local model to update the frequency and aggregation time through the joint computing optimization mechanism,so as to rationally allocate end point computing and communication resources,improve the convergence speed of the federated learning endogenous model and the resource utilization rate of the end point.