Design And Research Of Cloud-Edge-Device Federated Online Optimization And Control Based On Two-Timescales Lyapunov Approach

Federated learning is an effective distributed computing method,which can use different data sources for collaborative model training,while maintaining users’ privacy and security.In this paper,we study the federated learning process in the cloud side collaboration system,in which the edge server collects data from the device and trains the local model,while the cloud is used to aggregate the global model.The purpose of the system is to minimize the cost of data acquisition and iterative model training.Because the devices collect data at an uncertain rate,and the resource cost for edge data acquisition changes with time and is unknown,it is impossible to make the best online data acquisition decision in a long time.Therefore,facing this challenge,the specific results of this paper are as follows:Firstly,we describe the online optimization problem of minimizing the total cost in federated learning system as a mixed integer problem.Our modeling controls data collection and other resource consumption,and maintains the stability of the federated learning queue and the quality of the higher federated learning global model by adjusting the lost convergence parameters.Then,in order to solve the above online optimization problem of Federated learning,we propose a new online control algorithm of Federated learning.We apply the traditional Lyapunov algorithm to twotimescale problems.At the same time,we can online optimize the data collection of the cooperative system of cloud side federated learning and the cost of iterative federated learning model training.Without knowing the future information,we can make data collection,transmission and unloading decisions only based on the current information,while maintaining the stability of Lyapunov queue and high model quality.Finally,the original Lyapunov queue control method takes a lot of time in optimization to get the optimal solution of the current system.In order to solve this problem,we use the historical Lyapunov queue information to design the empirical virtual queue and learn the empirical state information for optimization,so as to reduce the delay,speed up the convergence speed,and finally speed up the process of Federated learning.Moreover,we combine the Lyapunov optimization technology of twotimescales with NP-hard problem,carry out theoretical proof and experiments,and propose a weighted random rounding algorithm.The algorithm solves the NP-hard problem with low time complexity,and we also prove its performance.