User Scheduling For Federated Learning In Wireless Networks Based On Reinforcement Learning And Matching Theory

Through the local model training of the network terminal and the convergence of the global model at the server,federated learning realizes the effective utilization of the data of the terminal users in the network on the premise of protecting the privacy of the terminal data.In the wireless network environment,the users who participate in the training process of federated learning tasks have a great impact on the training delay and accuracy of the model.The federated learning user scheduling for a wireless network is usually affected by the wireless channel state,the interference relationship between users,the computing power of users,and the training behavior of users.To solve the above problems,this thesis studies the federated learning user scheduling problem in wireless networks,and proposes two federated learning user scheduling methods,which are as follows:To solve the problem of multiple tasks federated learning user scheduling in the wireless network,when the server can’t know the prior information such as channel state information and computing power in advance,a multi-task federated learning user scheduling strategy based on multi-arm bandit and matching is proposed.First of all,the problem of minimizing user scheduling delay in multi-task federated learning is described as a multi-arm bandit problem.Secondly,due to the limited computing power of users,multi-task federated learning will produce scheduling conflicts in the process of user scheduling.To solve scheduling conflicts,matching theory is introduced to avoid scheduling conflicts through iterative matching of each task.Finally,a multiple task federated learning user scheduling strategy based on a multi-arm bandit and matching is proposed.This strategy does not require tasks to know a priori information such as wireless channel state information and computing power in advance.User scheduling is carried out through continuous online learning.The server constructs the preference list of the task to the user according to the reward value of each user feedback and uses the iterative matching process to solve the problem of multiple task federated learning user scheduling conflict.Finally,the effectiveness of the proposed user scheduling strategy is verified by simulation analysis.To solve the problem that malicious and unreliable users participate in the training process of federated learning tasks in wireless networks,a user scheduling and resource allocation strategy based on reputation and matching is proposed.First of all,the reputation is introduced as the measure of the user’s behavior in the process of federated learning task training.The reputation model represents the user’s behavior as positive and negative behavior,and the server evaluates the user’s credibility through the reputation value.Secondly,the server sorts the users according to the reputation,evaluates the user’s behavior in the task training process through the reputation,and schedules the users with high reputation to participate in the federated learning task training process,thus eliminating the local model of underperforming users.Finally,the method of a bipartite graph is used to allocate the corresponding communication resources to the scheduled users.Simulation results verify the effectiveness of the proposed user scheduling and resource allocation strategy.This thesis has 23 figures,10 tables,and 82 references.