Research On Resource Allocation Mechanism Of Wireless Federated Learning

In recent years,with the development and popularization of Internet of Things(Io T)technology,the demand for intelligent services such as speech recognition,augmented reality,and image processing is increasing day by day,which makes the application of machine learning in wireless networks more and more extensive.As an emerging distributed machine learning architecture,Federated Learning(FL)provides new ideas for the development of intelligent services in Io T with its features of making full use of users computing capacity and protecting data privacy.However,the limited communication,computation and energy supply capacity of wireless end-users constrains the performance of wireless Federated Learning.On the one hand,federated learning is a long-term training process consisting of several update rounds,and it is difficult for users’ limited energy to support their participation in the whole process.How to reasonably allocate the limited resources to each round of the learning to improve the performance of federated learning is a problem with little research but practical significance.On the other hand,with the recent emergence of wireless applications based on massive samples’ learning,hierarchical federated learning aggregating multi-cell user learning models has gradually become mainstream,but some devices lack the opportunity to participate in training due to resource constraints,leading to model performance degradation.To this end,the thesis focuses on the contradiction between the limited resources of users and the performance of wireless federated learning.From the perspective of the whole learning process,the limited energy of users is dynamically allocated,making the learning energy consumption adapt to the changes of the wireless environment.And digital twin is introduced to assist the training of resource-limited users,achieving the balance between the scale of users’ participating and the learning performance.Research contents of the thesis is as follows:1)To address the problem of how to use limited resources more efficiently for each update round of federated learning,the thesis proposes a long-term stable mechanism for joint optimization of user selection and heterogeneous resources,so that users can use energy dynamically according to the change of wireless channels in each learning round to achieve optimal learning results.On this basis,the thesis establishes an optimization model with the objective of minimizing the cumulative learning delay and maximizing training accuracy,and considers users constrained by the energy budget to jointly optimize the user selection and computational communication resource allocation.To solve the problem above,the thesis proposes the joint optimization algorithm URSLP(Userselection and Resource-allocation Subject to Long-term Performance)based on long-term perspective,which uses Lyapunov optimization to transform the long-term energy constraints into a single-round energy schedule,and then solves it with the help of convex optimization and greedy search methods.Simulation results show that the URSLP algorithm can not only effectively optimize the cumulative delay but also guarantee the training accuracy under the limited energy constraints.2)To address the problem of resource-constrained users’ difficulty in participating in learning in hierarchical federated learning,the thesis proposes a digital twin based training method selection and jointly optimizing mechanism of heterogeneous resources,in which users can choose to train the model locally or train with the help of the digital twin after synchronizing states with the edge server.On this basis,the thesis establishes an equilibrium model of time delay and energy consumption to characterize the overall system overhead,discusses the impact of communication resource scheduling on the users’ choice of training method,and jointly optimizes the learning accuracy and computational resource allocation to ensure the learning effect while reducing the learning cost of system.In order to solve the problem above,the thesis proposes the digital twin assisted training and resource allocation algorithm DTTRS(Digital Twin assisted Training and Resource Scheduling)based on genetic algorithm,and the superiority of DTTRS in the optimization of delay energy and learning accuracy control is finally verified by simulation experiments.