Research On Intelligent Beamforming Optimization Method Based On Fractional Programming

In recent years,the proliferation of cellular base station antennas and smart devices has made beamforming optimization a key focus of antenna array technology.Fractional programming(FP)technology can be used to reconstruct the beamforming problem into a series of convex problems.However,the iterative process of FP may include many complex operations,such as binary search,eigen-decomposition,matrix inversion,etc.,making it unsuitable for delay-sensitive services.Although deep learning(DL)methods can transfer online computing to offline training,providing real-time response capabilities,training complexity will rapidly increase as the scale of communication systems increases.How to design a beamforming algorithm that takes into account the system performance and complexity is one of the urgent problems to be solved.In view of this,this research proposes an optimization framework of Deep Unfolding Fractional Programming(DUFP)based on the combination of expert knowledge and deep neural network.The iterative process is unfolding into the layer of neural network,and the optimal beamforming solution is found by training the neural network with a few parameters.Specifically,this research mainly includes the following two aspects of innovation:1.Aiming at the optimization of downlink beamforming in MISO(Multiple-Input Single-Output)system,the DUFP optimization framework is proposed which maps the fixed number of iterations of FP algorithm to the trainable neural network with good interpretability,greatly reducing the trainable parameters and preserving the structure of FP algorithm.This framework is applied to the multi-rate FP problem of MISO downlink beamforming.Simple matrix multiplication and addition operations are used to replace the complex operations such as binary search,eigen-decomposition,matrix inversion,etc.in the original algorithm to solve the trade-off between the complexity and performance of base station(BS).Compared with traditional DL technology,it also reduces the training time and the number of training parameters.The results show that the proposed DUFP method can naturally integrate expert knowledge into the learning process,so as to achieve the system weighted sum rate(WSR)performance approaching the WMMSE(Weighted Minimum MeanSquare Error)algorithm,and reduce the algorithm running time by more than 60%.2.Aiming at the downlink joint beamforming and phase optimization assisted by the Reconfigurable Intelligent Surface(RIS)in the MISO system,a low complexity iterative algorithm is designed using the DUFP framework for joint optimization.Firstly,the non-convex problem is decoupled by Lagrange dual transformation,and the objective function is reconstructed by FP.The original problem is decomposed into the beamforming optimization sub-problem of BS and the phase optimization sub-problem of RIS.The DUFP framework is used to optimize the two sub-problems alternately.By jointly optimizing the beamforming at BS and the phase vector at RIS,the WSR of all users in the system is maximized.The results show that the proposed joint beamforming and phase optimization solution has achieved significant performance improvement in WSR and operation time.