Research On User Scheduling And Resource Allocation Algorithm For Federated Learning

With the continuous progress of information technology,Artificial Intelligence(AI)has become an indispensable part of our life and work.However,to make AI more accurate and intelligent,a large amount of data is needed to support it,among which there are often various kinds of private information,such as personal identity information,financial information,etc.,which may be used maliciously and therefore there is a serious privacy and security problem.To address this problem,Federated Learning(FL)is proposed as a distributed machine learning(ML)framework,in which the data set is distributed to different User Equipment(UE)for training,avoiding the privacy problem of storing the data set in one place.This avoids the privacy issues that may arise from storing data sets in one place.Since the data does not leave the UE locally during the training process,only the updated parameters of the model are transmitted to the server,thus protecting the data privacy to a great extent.However,with the advent of the Internet of Things(Io T)era,the number of UEs in wireless environments is rapidly increasing.Due to the limited resources of wireless networks,it is difficult for large-scale UEs to participate in the FL training process at the same time,resulting in low efficiency of FL training models.To address this problem,this thesis will propose corresponding optimization algorithms to improve the performance of wireless FL from two perspectives of user scheduling and resource allocation.The main research contents of this thesis are as follows:(1)This thesis considers that in the case of synchronous training,the training delay in each training cycle depends on the UE with the worst communication and computational capabilities,and some UEs that have completed training will keep waiting for the remaining UEs to complete training,which will generate a huge additional overhead of delay.To address this problem,an optimization problem with the objective of minimizing the total training delay of the wireless FL system is established,and a Multi-Armed Bandit(MAB)based user online scheduling algorithm is proposed,in which the Channel Status Information(CSI)between the Base Station(BS)and the UE is assumed to be unknown,the user scheduling problem is transformed into a MAB problem,and the reward function is a weighted sum of the delay consumed by the UE in each round and the significance of local updates,and then the tradeoff between MAB exploration and utilization is implemented based on the Upper Confidence Bound(UCB)algorithm.Simulation experiments show that the proposed method is more accurate and generates lower training latency compared with the traditional centralized FL policy and random user scheduling policy models.(2)This thesis considers that the user scheduling policy will incur a large amount of energy overhead during the execution of the joint UE per-round consumption delay and the significance of local updates.To address this problem,this thesis establishes the BS communication energy consumption model,the computational energy consumption model and the communication energy consumption model of UE,and proposes a FL-oriented multivariate resource allocation algorithm.The algorithm establishes the energy consumption optimization equation based on the user scheduling policy to filter valuable UEs,and uses the control variable method to transform the original problem into three sub-problems containing the energy consumption in the model download,model update,and model upload phases,optimizing the transmit power of BS,the computational resources of the central processing unit(CPU)of UE,and the energy consumption of sub-channel uplink of UE.The bandwidth solution of the uplink of the sub-channel of the UE is optimized.The numerical results are also analyzed by simulation experiments,which show that the proposed method can reduce the energy consumption of the wireless federation learning training process compared with the traditional user scheduling strategy and the bandwidth-only resource allocation scheme.