| Large-scale multiple-input multiple-output(MIMO)systems with a large number of antennas to serve a relatively small number of users can significantly improve the capacity and reliability of wireless systems.Ac-cording to the coverage requirements of different 5G/B5G communica-tions,the thesis investigates the joint optimization method of intelligent multi-array antenna parameters for sub-6GHz large-scale MIMO heteroge-neous networking and millimeter wave(mm Wave)large-scale MIMO with intelligent reflecting surface(IRS)assisted coverage enhancement.There-fore,in the thesis,we obtain key information such as user location from the dynamic network environment,and make preemptive decisions by im-proved deep reinforcement learning(DRL)and transfer learning algo-rithms to improve the effectiveness,reliability and scalability of the sys-tem,and develop an intelligent joint tuning method for multi-array antenna parameters in the following two aspects:1)For the optimal configuration of multiple base-station antenna ar-rays in parameters in multi-cell 3D large-scale MIMO heterogeneous net-works where macro-cell and micro-cell overlay,this thesis proposes an in-telligent multi-array antenna parameters joint tuning method based on im-proved DRL,which automatically adjusts three key antenna parameters of multi-array,i.e.,down-tilt angle,vertical and horizontal half-power beam-width,in the dynamic environment with user mobility,to achieve user weighted sum-rate improvement.Specifically,this thesis adopts an im-proved hybrid Q-learning algorithm that utilizes gridded and relatively coarse-grained user location information,integrates a two-layer neural sub-net hyperparameter parallel update mechanism and a priority replay buffer pool technique,and the designed neural network can effectively improve the convergence performance of user weighted sum-rate in dynamic opti-mization problems by efficiently learning historical experience.In addi-tion,the gridded processing helps the autonomous framework of base sta-tion antenna parameter configuration to better adapt to heterogeneous cel-lular networks with different distribution densities of users.The simulation results show that the weighted sum-rate as well as the network coverage are significantly improved compared to the reference scheme without grid-ding.Besides,the performance of the proposed algorithm in this thesis is closer to that of the exhaustive search algorithm,which is significantly bet-ter than the empirical-based approach,with greatly reduced computational complexity.2)Considering the scenario of IRS-assisted millimeter-wave commu-nication in B5G networks,this thesis proposes a joint multi-array antenna parameter tuning method for millimeter-wave base station antenna array and multi-IRS reflector antenna array.The method employs machine learn-ing to empower intelligent transmission beam optimization,and effectively solves the challenging dynamic nonconvex problem of joint optimization of intelligent multi-array parameters by fine-grained multi-array antenna parameter tuning,i.e.,joint tuning of hybrid beam assignment and IRS phase shift.Specifically,since the probability of users in the cell edge or the area not directly reached by the base station signal is increased due to mobility,the adaptive beam modulation mechanism for robust mobility is particularly important,so in this thesis,by improving the deep determinis-tic policy gradient(DDPG)algorithm,and the simulations show that the convergence speed and stability of knowledge empowerment and dual ex-perience pool of DRL and transfer learning are generally higher than those of the general DRL algorithms.In addition,the performance of the pro-posed algorithm in this thesis is closer to that of the exhaustive search al-gorithm under the premise of greatly reducing the computational complex-ity.Finally,a summary of the research content and an outlook on scenar-ios such as intelligent multi-array antenna systems for 5G/B5G is given. |
PDF Full Text Request