Research On Precoding And User Pairing Scheme Of MIMO Based On Deep Learning

MIMO transmission has become a technology that 5G and 6G need to further expand the scale of use because of its extremely high spectrum utilization.Since the traditional MIMO precoding transmission and detection schemes are designed based on the statistical model of the system parameters,the actual operation of the system parameters is not completely consistent with the statistical model,which makes the transmission and detection scheme based on the model unable to achieve The best results.In order to overcome this difficulty,the academic community proposes to use data-based machine learning methods to obtain sending and receiving schemes that are highly consistent with the actual scene of the system.The thesis focuses on the problem of MIMO precoding and user pairing,and studies its machine learning implementation methods.The main work of the thesis is as follows:1.Aiming at the single-cell multi-user MIMO downlink precoding scenario,taking user weight and rate as the objective function,the CNN implementation method of the precoding scheme design is given.The main work is as follows: 1)Convert the mathematical optimization problem corresponding to the precoding design into a matrix weighting and mean square error minimization problem,use the WMMSE iterative algorithm to solve,and use the input and output of the algorithm as the training data of the CNN network;2)Constructed a CNN structure with strong feature extraction capabilities and convolution kernel parameter sharing,as a machine learning architecture designed by the precoding scheme,which can perform feature learning and network calculations with a small amount of calculation;3)A training method combining supervised and unsupervised is proposed.The supervised learning method is used for pre-training,and then unsupervised learning is used for retraining through the constructed regularization term loss function.The simulation results show that the proposed precoding machine learning implementation method can achieve a rate performance equivalent to that of the WMMSE algorithm,and can greatly reduce the computational complexity of the precoding design.2.Aiming at the single-cell multi-user MISO precoding scenario,taking the user weight and rate as the objective function,a dedicated RNN implementation method for precoding scheme design is given.The main work is as follows: 1)For the mathematical optimization problem corresponding to the precoding design,first use the Lagrangian dual reconstruction and quadratic transformation method to reconstruct the objective function,and then use the fractional programming algorithm to optimize the reconstructed problem Perform iterative solution,and use the input and output of the algorithm as the training data of the special RNN network;2)In order to overcome the shortcomings of the general neural network being too large or too small,according to the iterative characteristics of the fractional programming algorithm,a prediction method is designed Encoding-specific RNN structure;3)A training method combining supervised and unsupervised is proposed.The supervised learning method is used for pre-training,and then unsupervised learning is used for retraining through the constructed regularization term loss function.The simulation results show that the proposed algorithm can use a smaller neural network scale to obtain comparable speed performance.3.Aiming at the precoding problem scenario of multi-cell cooperative multi-user MIMO downlink transmission,taking user weight and rate as the objective function,a distributed MARL implementation method for precoding scheme design is given.The main work is as follows: 1)Convert the mathematical optimization problem corresponding to the precoding design into a matrix weighting and mean square error minimization problem,and use the WMMSE iterative algorithm to solve it;2)According to the characteristics of the dynamic change of the wireless environment channel,through the design of MARL Structure and reinforcement learning distributed training and execution method,a dynamic precoding design algorithm based on distributed multi-agent reinforcement learning is proposed;the training process and execution process of the algorithm are carried out in a distributed manner,and each agent follows Neighboring agents exchange information to collect state information,and each agent designs the precoding of users in the cell by observing its state.The simulation results show that the proposed algorithm has a lower delay,and the system and rate are very close to the WMMSE algorithm using global channel information,which is feasible in a dynamic wireless environment.4.Aiming at the problem of single-cell multi-user virtual MIMO user pairing,taking the energy efficiency of the system as the objective function,a reinforcement learning implementation method for the user pairing scheme design is given.The main work is as follows: 1)In order to solve this non-convex integer programming problem,the optimal algorithm and greedy algorithm for the user pairing problem are first given.2)Based on reinforcement learning and deep neural networks,an intelligent user matching algorithm is proposed,and in order to reduce the action space of the algorithm,a pre-processing mechanism for pre-selecting some users is proposed;the proposed algorithm can interact with the environment and can be based on the environment state Intelligent online adjustment and continuous learning to approach the target.In addition,historical channel information,especially channel information of the previous time slot,can be used to infer current channel information to make user pairing decisions.The simulation results show that the energy efficiency of the proposed algorithm is relatively close to the optimal algorithm,and can slightly exceed the performance of the greedy algorithm when sufficient training time slots are guaranteed.