Wireless Resource Management And Model Aggregation Method For Multi-task Federated Learning

The integration of Federated Learning(FL)framework and Mobile Edge Computing(MEC)scenario can provide users with an efficient and secure task computing environment,which has a broad prospect.On the one hand,with its distributed machine learning paradigm,FL can break down the barriers between data sources on the computing side and improve the training effect on the premise of protecting the privacy and security of the client.On the other hand,MEC based on distributed wireless network deployment framework can provide highperformance,low-latency and high-bandwidth service environment on the communication side.However,the combination of them will face many challenges such as environment fluctuation,system heterogeneity and statistical heterogeneity at the same time,especially when multiple learning tasks coexist,model heterogeneity further increases the challenge of system optimization design.In view of the above challenges,this paper considers two situations of terminal communication and computing status known and unknown,and carries on the model establishment and algorithm design for the existence of multivariate heterogeneous scenarios,in order to improve the overall learning performance and efficiency.First of all,in view of the known information,combined with the common influence of communication environment and local training effect on the global model,a scheme of user selection and radio resource allocation is proposed to improve the performance and efficiency of the global model.Specifically,firstly,the optimization problem is determined by minimizing the weighted sum of the loss function of the global model.Then,for each round,considering the influence of wireless communication factors and local training parameters on the global convergence speed,a federated learning user task selection mechanism for multitask scenarios is designed to improve resource utilization on the premise of ensuring task accuracy.Next,the influence of wireless communication factors on the average global convergence speed is analyzed,the optimization problem is simplified,the factors affecting the training accuracy in each task are analyzed by multiple correlation,the task weight is determined,and a resource allocation algorithm is proposed based on KM optimal matching.Finally,the simulation results show that,compared with the comparison algorithm,the proposed algorithm can obtain a global model with higher accuracy,and the baseline accuracy is improved by 7.5%,and the global convergence speed is also improved accordingly.Secondly,in view of the unknown information,combined with the common influence of heterogeneous data and aggregation pattern on the global model,a hierarchical learning architecture is established,and task clustering algorithm and sliding window algorithm are designed to improve the performance and efficiency of the global model.Specifically,first of all,we integrate the idea of asynchronous federation learning to establish a hierarchical federation learning architecture FedHS,including task clustering layer and central aggregation layer.Aiming at the task clustering layer,on the one hand,based on Pearson Correlation Coefficient(PCC),a two-dimensional correlation clustering algorithm is designed to cluster the client training network model matrix to solve the problem of nonindependent and identical distribution of task sample data;on the other hand,a sliding window algorithm is designed to complete the task aggregation within the cluster and the Earth Mover’s Distance(EMD).Based on the consideration of EMD,aggregate turn and loss function,a task selection index is defined as"leading intensity" to design a sliding window opening drive mechanism for each round,and to evaluate the contribution of client tasks to global updates to formulate an intra-cluster aggregation strategy.Aiming at the central aggregation layer,based on the influence of the amount of data owned by the cluster and the number of rounds it participates in on the overall training effect,an asynchronous update strategy for global updating of cluster parameters is designed.Finally,the simulation results show that the proposed algorithm has a 9%improvement in the accuracy of global aggregation and a reduction of about 20%-33%in terms of convergence delay compared with pure synchronous federation learning or pure asynchronous federation learning.