Research On Machine Learning Based Multi-Cell Coordinated Beamforming Algorithm

To meet the growing demand for wireless data and high-speed connections,wireless communication systems continuously explore higher frequency bands such as millime-ter wave and terahertz bands,while also using various advanced technologies to improve the spectral efficiency of existing frequency bands.Among them,massive antenna ar-rays are considered the most effective technology for improving spectral efficiency^[1],and beamforming is an important way to achieve the gain of massive antenna arrays.In multi-cell communication systems,each cell base station can also use the same frequency band to provide services to users within the cell,further improving spectral efficiency.In order to reduce the performance loss caused by inter-cell interference in the system,co-operative beamforming technology can improve the overall performance of the system by jointly designing beamforming vectors for each base station.However,traditional coop-erative beamforming methods require obtaining global channel state information,which suffers from computational complexity and high complexity,making it impractical for actual communication systems.On the other hand,with the breakthrough development of machine learning and neural networks in recent years,big data,machine learning and other technologies have been applied to various scenarios.Therefore,this paper considers using deep learning and deep reinforcement learning technologies to achieve high sys-tem capacity with relatively low system overhead,to meet the requirements of practical systems.The specific research content is as follows:Firstly,in order to solve the problem that traditional cooperative beamforming al-gorithms cannot be applied to practical communication systems due to their high com-putational complexity and large transmission overhead,this paper proposes a multi-cell cooperative beamforming method based on supervised learning.Specifically,this paper designs a channel information compression network,beam direction vector design net-work,and loss function based on supervised learning.Simulation results show that the performance of the multi-cell cooperative beamforming based on supervised learning can approach the Weighted Minimum Mean Square Error（WMMSE）algorithm,while signif-icantly reducing computational complexity and transmission overhead.Furthermore,although the multi-cell cooperative beamforming based on supervised learning can meet the requirements of practical communication systems,its power allo-cation scheme is still based on equal power allocation,leaving room for optimization.Therefore,this paper proposes a dynamic power allocation scheme based on multi-agent reinforcement learning.Specifically,based on the design of beam direction vectors us-ing supervised learning techniques,this paper proposes to use multi-agent reinforcement learning techniques to achieve dynamic power allocation for multi-cell systems.Sim-ulation results show that the multi-cell cooperative beamforming algorithm integrating supervised learning and deep reinforcement learning can further improve the sum rate of multi-cell systems while achieving low computational complexity and transmission over-head.In addition,to address the issue that the performance of the beam direction vector design scheme based on supervised learning is difficult to be superior to the label algo-rithm,this paper proposes a beam direction vector design scheme based on unsupervised learning.Specifically,this paper designs a channel information compression network,beam direction vector design network,and loss function based on unsupervised learning.Simulation results show that the multi-cell cooperative beamforming algorithm combin-ing unsupervised learning and deep reinforcement learning can further improve the sum rate of multi-cell systems while achieving low computational complexity and transmission overhead.