Research On Efficient Resource Allocation Strategy For Wireless Edge Federated Learning

As the deep integration of machine learning(ML)and wireless networking,ubiquitous computing resources at network edge are interconnected to improve the learning and decision-making capabilities of artificial intelligence.However,the training of cognitive model needs massive data and powerful computing power.Traditional cloud computing needs to gather all terminal information to the central server.With huge bandwidth demand and high end-to-end delay,it can not meet the real-time cognitive needs of machine terminals in new application scenarios.Wireless edge federated learning can be considered promising solution.which makes full use of the distributed perceptual data and computing resources in the network to carry out efficient distributed training of cognitive models.In the process of wireless edge federated learning,the terminal nodes update the local models periodically,and frequently exchange model parameters with the edge server through wireless communication,so as to update the global model.Due to scarce wireless spectrum resources and limited training delay budget,only a limited number of devices are allowed to upload local models in each communication round.Therefore,user selection and system resource allocation are one of the bottlenecks of model training performance.To solve this problem,the main contributions of this thesis are as follows:(1)A time efficient wireless federated learning strategy based on age of parameter is proposed.The delay of model training in wireless federated learning is determined by the number of convergence rounds,the computation delay,and communication delay in each communication rounds.For non-ⅡD terminal data sets,aiming at the influence of the data diversity of the selected terminal set on the number of convergence rounds,this thesis proposes to use the age of terminal to measure the importance of its parameters,which can measure the influence of selection decision on the number of convergence rounds.In addition,in order to avoid the reduction of training accuracy caused by terminal drop-outs,energy harvesting technology is introduced to prolong the lifetime of terminals.By comprehensively considering the effects of heterogeneous computing power,dynamic energy level and time-varying channel state among multiple devices on training delay,as well as the influence of the importance of device parameters on convergence rounds,an optimization problem aiming at minimizing the model training delay is established.In order to solve this problem,Lyapunov optimization method is used to decouple the long-term energy constraint into each communication round,and a user scheduling and bandwidth allocation algorithm based on age of parameter is proposed.Simulation results show that compared with the traditional training strategy,proposed strategy can not only maintain the training accuracy,but also shorten the training delay and ensure the system energy stability.(2)A time efficient wireless federated learning strategy based on gradient quantization is proposed.Based on the above work,the random multi-level quantization technology is introduced to compress the gradient at terminal,so as to further reduce the communication delay in each communication round.The quantized gradient is used as an index to measure the influence of terminal selection decision on the number of convergence rounds.By comprehensively considering the effects of heterogeneous computing power at the terminal,time-varying channel state and quantization level on training delay,as well as the effects of age of parameter and quantized gradient on convergence rounds,an optimization problem aiming at minimizing model training delay is established.Afterwards,a user scheduling,bandwidth allocation and gradient quantization strategy based on deep reinforcement learning is proposed.The simulation results show that the proposed strategy can effectively reduce the training delay compared with traditional training strategy,and improve training accuracy compared with the fixed precision gradient quantization strategy.