Research On Wireless Resource Scheduling Algorithm Based On Deep Learning In User-Centric Networks

In order to address the EB-level mobile traffic requirement caused by services such as augmented reality and ultra-high-definition video streaming and the serious cumulative interference caused by ultra-dense deployment of base stations(BS),the industry has proposed a user-centric network(UCN),which also brings huge scheduling complexity and introduces challenges to real-time applications.In recent years,some studies have used deep learning(DL)technology to solve the real-time problem of complex wireless resource scheduling.The basic idea is to leverage deep neural network(DNN)to offline learn the relationship between input and output of traditional scheduling algorithm,and finally predict the results online through simple calculations.However,the huge complexity from network size,base station(BS)clustering and interference coordination in UCN makes the approximation performance of learning-based solutions degrade greatly.Therefore,this paper proposes a residual learning-based convolutional neural network(CNN)to approximate the iterative algorithm in user-centric networks,and implements real-time wireless network resource scheduling and interference management through online deployment.The main work of this article is divided into the following three parts:Aiming at the problem of serious cumulative interference in densely deployed cells,this paper first models clustering and beamforming in user-centric network and proposes a non-convex optimization problem to maximize capacity,and then introduces a weight matrix to convert the original problem to a minimum mean square error(MMSE)problem;then the closed-form solutions of clustering and beamforming are given by using the Lagrangian multiplier method.Simulations verify that the user-centric architecture improves the performance of edge users compared to traditional base station-centric algorithms.Aiming at the real-time problem of traditional power allocation algorithms in user-centric networks,this paper first proposes power allocation algorithms based on weighted MMSE in user-centric networks.Then,a deep convolution network based on residual learning is proposed to effectively avoid the problems of poor approximation ability for complex scenarios and the gradient vanishing in very deep networks.Simultaneously,the feasibility of the universal approximation theorem of neural networks in wireless resource scheduling is provedBased on the power allocation problem of real number domains in user-centric networks,this article extends it to more complex multi-antenna beam scheduling in plural domains.First,a large amount of learning data is generated based on the weighted MMSE algorithm for model training.For plural input the output and user-centric mechanism,we propose a deep residual neural network based on three sub-networks.By increasing the feature dimensions of the input,it effectively reduces information loss and improves the model's awareness of the user-centric mechanism.Finally,the performance of the proposed method compared with the traditional complex processing method is analyzed and various factors affecting the performance of approximation are analyzed from different perspectives.