Research On Distributed And Centralized Learning In Edge-Intelligent Networks

With the development of 5G technology and the popularization of Internet of Things(Io T),the number of mobile terminals and Io T devices has experienced explosive growth.These geographically distributed mobile terminals and Io T devices generate massive data at the edge of the network.In order to improve the immediacy and intelligence of data processing,edge intelligent network has received a lot of attention and research.Edge intelligent network is a new network paradigm that combines edge computing and artificial intelligence(AI).The key of edge intelligent network is to design effective algorithms to train machine learning models.In the presence of the channel state information(CSI)error,this dissertation aims to study robust optimization algorithms in distributed learning and centralized learning to improve the performance of edge intelligent network according to its characteristics.Firstly,a robust algorithm for edge intelligent network based on federated learning is proposed for the scenarios where data privacy needs to be protected.The user devices send the model parameters to the edge server after model training locally,and the edge server utilizes over-the-air computation to achieve model aggregation.According to the CSI error,the mean square error(MSE)between the signal received by the edge server and the ideal signal is established.Combining the MSE constraint with the power constraint of user devices,and in the presence of CSI error,an optimization problem is formulated to maximize the number of devices participating federated learning.By solving the problem,a joint optimization algorithm for device selection,transmitting power and receive beamforming is obtained.The simulation results confirm the robustness of the proposed algorithm in federated learning.Secondly,a robust algorithm for edge intelligent network based on centralized learning is proposed for the scenarios where the computing power of user devices is limited.The user devices send data samples to the edge server,and the edge server uses the received data for model training.Combining the relationship between the model generalization error and the number of data samples,and in the presence of CSI error,an optimization problem is formulated to minimize the generalization error of the global model.By solving the problem,a joint optimization algorithm for the transmission power of devices and the receive beamforming of the edge server is obtained.The simulation results confirm the effectiveness of the proposed algorithm in centralized learning.Finally,in order to fully exploit the potential of user devices and the edge server,a joint distributed learning and centralized learning robust algorithm is proposed.User devices conduct local model training while transmitting data.After the user training is completed,the model parameters are sent to the edge server.The edge server obtains the aggregation model and user device data to calculate the global machine learning model.In the presence of CSI error,the optimization problem of minimizing the loss function error of the global model is established.By solving this optimization problem,a joint optimization algorithm for user device selection,transmit power and receive beamforming is obtained.The simulation results confirm that the proposed algorithm is feasible and robust.