Optimization Research On Massive Antenna Systems Based On Machine Learning

Massive multiple-input multiple-output(M-MIMO)network is one of the key technologies in the fifth generation and beyond(5G/B5G)mobile communication systems,which can dramatically promote the spectral efficiency and energy efficiency of wireless cellular networks by space division multiplexing.On top of traditional horizontal two-dimensional(2D)precoding based on 2D M-MIMO,adjusting antenna array tilt angles(AATAs)can acquire an extra spatial degree of freedom on the vertical dimension and further increase the system capacity,which forms the three-dimensional(3D)MIMO.By adjusting the AATAs,the 3D-MIMO system can achieve traffic/user distribution aware array configuration to achieve interference management and performance enhancement of the wireless cellular network.However,the adjustment of antenna parameters in multi-cell environment is particu-larly important and complex.Most of the relevant existing works all adopt relatively complicated conventional AATA optimization methods,which may not be able to meet the real-time requirement of communication systems.In this paper,we research the optimization methods of AATAs under typical 3D-MIMO system scenarios based on machine learning(ML),which significantly reduce the system overhead for antenna adjustment while ensuring excellent performance,and are able to meet the real-time demand of parameter adjustment in 3D-MIMO systems.The main works of this paper are as follows:1)In this thesis,we propose a novel proactive AATA adjustment scheme based on ML,aiming at the multicells 3D-MIMO scenario.By exploiting ML,the base stations(BSs)are able to predict the AATA parameter groups of the next time slot that closely maximizes the spectral efficiency(SE)of the area within the jurisdiction.In this way,BSs can directly adjust AATAs without frequently acquiring users' actual locations of the current time slot to shorten the adjustment time and save system overheads.Moreover,the prediction process is carried out by well-structured and trained neural networks,which further saves the online optimization time.Simulation results show that the achievable SE of the proposed ML-empowered proactive AATA adjustment can occupies over 98.7%on average of the performance with known real-time user positions and exhaustive optimal AATA searching,while the proposed scheme significantly saves the system overheads and online computational complexity.2)In order to further improve the scalability of the ML-empowered proactive AATA optimization scheme,this paper further propose a gridization-based and ML-empowered AATA optimization scheme,where we conduct distance-based gridization,enabling the base stations to adjust the antenna tilt angles without knowing the accurate angular information from users.In particular,for cases with relatively fewer grids,we introduce the semi-supervised clustering(SSC)scheme,which can map the gridized user states to antenna array tilt angles(AATAs)in groups quickly.Numerical simulations validate the excellent online performance of our proposed scheme with much lower time overhead,indicating its potential to meet the real-time requirement for 5G/B5G communication systems.3)Finally,this paper conducts optimization design and comprehend-sive simulation verification evaluation for the robustness and scalability of the proposed scheme.We tighten the condition from perfect channel state information(CSI)to the imperfect CSI estimated by pilots,and consider a more realistic multi-cell deployment scenario.Simulation results show that the proposed proactive AATA adjustment scheme has good robustness for the change of conditions and scenarios.For the problem that the amount of data and computing time increase dramatically led by the increase of the grid number in multi-cells,we propose an optimization mechanism adjusting the SSC method to use trained neural network(NN)to map online user states directly to AATA groups,which improves the algorithm efficiency and scalability.Numerical simulations reveal that the NN-fitting method achieves good SE performance for the case with relatively more grids,which also indicates that the proposed AATA adjustment scheme based on grid and ML method has strong scalability for different conditions and scenarios.